1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "X86ISelLowering.h"
16 #include "Utils/X86ShuffleDecode.h"
17 #include "X86CallingConv.h"
18 #include "X86FrameLowering.h"
19 #include "X86InstrBuilder.h"
20 #include "X86MachineFunctionInfo.h"
21 #include "X86TargetMachine.h"
22 #include "X86TargetObjectFile.h"
23 #include "llvm/ADT/SmallBitVector.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/ADT/StringSwitch.h"
28 #include "llvm/CodeGen/IntrinsicLowering.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/WinEHFuncInfo.h"
36 #include "llvm/IR/CallSite.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/GlobalAlias.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/Intrinsics.h"
45 #include "llvm/MC/MCAsmInfo.h"
46 #include "llvm/MC/MCContext.h"
47 #include "llvm/MC/MCExpr.h"
48 #include "llvm/MC/MCSymbol.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Target/TargetOptions.h"
54 #include "X86IntrinsicsInfo.h"
60 #define DEBUG_TYPE "x86-isel"
62 STATISTIC(NumTailCalls, "Number of tail calls");
64 static cl::opt<bool> ExperimentalVectorWideningLegalization(
65 "x86-experimental-vector-widening-legalization", cl::init(false),
66 cl::desc("Enable an experimental vector type legalization through widening "
67 "rather than promotion."),
70 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
71 const X86Subtarget &STI)
72 : TargetLowering(TM), Subtarget(&STI) {
73 X86ScalarSSEf64 = Subtarget->hasSSE2();
74 X86ScalarSSEf32 = Subtarget->hasSSE1();
75 MVT PtrVT = MVT::getIntegerVT(8 * TM.getPointerSize());
77 // Set up the TargetLowering object.
78 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
80 // X86 is weird. It always uses i8 for shift amounts and setcc results.
81 setBooleanContents(ZeroOrOneBooleanContent);
82 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
83 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
85 // For 64-bit, since we have so many registers, use the ILP scheduler.
86 // For 32-bit, use the register pressure specific scheduling.
87 // For Atom, always use ILP scheduling.
88 if (Subtarget->isAtom())
89 setSchedulingPreference(Sched::ILP);
90 else if (Subtarget->is64Bit())
91 setSchedulingPreference(Sched::ILP);
93 setSchedulingPreference(Sched::RegPressure);
94 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
95 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
97 // Bypass expensive divides on Atom when compiling with O2.
98 if (TM.getOptLevel() >= CodeGenOpt::Default) {
99 if (Subtarget->hasSlowDivide32())
100 addBypassSlowDiv(32, 8);
101 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
102 addBypassSlowDiv(64, 16);
105 if (Subtarget->isTargetKnownWindowsMSVC()) {
106 // Setup Windows compiler runtime calls.
107 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
108 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
109 setLibcallName(RTLIB::SREM_I64, "_allrem");
110 setLibcallName(RTLIB::UREM_I64, "_aullrem");
111 setLibcallName(RTLIB::MUL_I64, "_allmul");
112 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
113 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
114 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
115 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
116 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
119 if (Subtarget->isTargetDarwin()) {
120 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
121 setUseUnderscoreSetJmp(false);
122 setUseUnderscoreLongJmp(false);
123 } else if (Subtarget->isTargetWindowsGNU()) {
124 // MS runtime is weird: it exports _setjmp, but longjmp!
125 setUseUnderscoreSetJmp(true);
126 setUseUnderscoreLongJmp(false);
128 setUseUnderscoreSetJmp(true);
129 setUseUnderscoreLongJmp(true);
132 // Set up the register classes.
133 addRegisterClass(MVT::i8, &X86::GR8RegClass);
134 addRegisterClass(MVT::i16, &X86::GR16RegClass);
135 addRegisterClass(MVT::i32, &X86::GR32RegClass);
136 if (Subtarget->is64Bit())
137 addRegisterClass(MVT::i64, &X86::GR64RegClass);
139 for (MVT VT : MVT::integer_valuetypes())
140 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
142 // We don't accept any truncstore of integer registers.
143 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
144 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
145 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
146 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
147 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
148 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
150 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
152 // SETOEQ and SETUNE require checking two conditions.
153 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
154 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
155 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
156 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
157 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
158 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
160 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
162 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
163 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
164 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
166 if (Subtarget->is64Bit()) {
167 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
168 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
169 } else if (!Subtarget->useSoftFloat()) {
170 // We have an algorithm for SSE2->double, and we turn this into a
171 // 64-bit FILD followed by conditional FADD for other targets.
172 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
173 // We have an algorithm for SSE2, and we turn this into a 64-bit
174 // FILD for other targets.
175 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
178 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
180 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
181 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
183 if (!Subtarget->useSoftFloat()) {
184 // SSE has no i16 to fp conversion, only i32
185 if (X86ScalarSSEf32) {
186 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
187 // f32 and f64 cases are Legal, f80 case is not
188 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
190 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
191 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
194 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
195 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
198 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
199 // are Legal, f80 is custom lowered.
200 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
201 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
203 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
205 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
206 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
208 if (X86ScalarSSEf32) {
209 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
210 // f32 and f64 cases are Legal, f80 case is not
211 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
213 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
214 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
217 // Handle FP_TO_UINT by promoting the destination to a larger signed
219 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
220 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
221 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
223 if (Subtarget->is64Bit()) {
224 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
225 // FP_TO_UINT-i32/i64 is legal for f32/f64, but custom for f80.
226 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
227 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
229 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
230 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
232 } else if (!Subtarget->useSoftFloat()) {
233 // Since AVX is a superset of SSE3, only check for SSE here.
234 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
235 // Expand FP_TO_UINT into a select.
236 // FIXME: We would like to use a Custom expander here eventually to do
237 // the optimal thing for SSE vs. the default expansion in the legalizer.
238 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
240 // With AVX512 we can use vcvts[ds]2usi for f32/f64->i32, f80 is custom.
241 // With SSE3 we can use fisttpll to convert to a signed i64; without
242 // SSE, we're stuck with a fistpll.
243 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
245 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
248 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
249 if (!X86ScalarSSEf64) {
250 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
251 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
252 if (Subtarget->is64Bit()) {
253 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
254 // Without SSE, i64->f64 goes through memory.
255 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
259 // Scalar integer divide and remainder are lowered to use operations that
260 // produce two results, to match the available instructions. This exposes
261 // the two-result form to trivial CSE, which is able to combine x/y and x%y
262 // into a single instruction.
264 // Scalar integer multiply-high is also lowered to use two-result
265 // operations, to match the available instructions. However, plain multiply
266 // (low) operations are left as Legal, as there are single-result
267 // instructions for this in x86. Using the two-result multiply instructions
268 // when both high and low results are needed must be arranged by dagcombine.
269 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
271 setOperationAction(ISD::MULHS, VT, Expand);
272 setOperationAction(ISD::MULHU, VT, Expand);
273 setOperationAction(ISD::SDIV, VT, Expand);
274 setOperationAction(ISD::UDIV, VT, Expand);
275 setOperationAction(ISD::SREM, VT, Expand);
276 setOperationAction(ISD::UREM, VT, Expand);
278 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
279 setOperationAction(ISD::ADDC, VT, Custom);
280 setOperationAction(ISD::ADDE, VT, Custom);
281 setOperationAction(ISD::SUBC, VT, Custom);
282 setOperationAction(ISD::SUBE, VT, Custom);
285 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
286 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
287 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
288 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
289 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
290 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
291 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
292 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
293 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
294 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
295 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
296 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
297 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
298 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
299 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
300 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
301 if (Subtarget->is64Bit())
302 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
303 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
304 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
305 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
306 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
308 if (Subtarget->is32Bit() && Subtarget->isTargetKnownWindowsMSVC()) {
309 // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)`
310 // is. We should promote the value to 64-bits to solve this.
311 // This is what the CRT headers do - `fmodf` is an inline header
312 // function casting to f64 and calling `fmod`.
313 setOperationAction(ISD::FREM , MVT::f32 , Promote);
315 setOperationAction(ISD::FREM , MVT::f32 , Expand);
318 setOperationAction(ISD::FREM , MVT::f64 , Expand);
319 setOperationAction(ISD::FREM , MVT::f80 , Expand);
320 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
322 // Promote the i8 variants and force them on up to i32 which has a shorter
324 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
325 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
326 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
327 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
328 if (Subtarget->hasBMI()) {
329 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
330 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
331 if (Subtarget->is64Bit())
332 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
334 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
335 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
336 if (Subtarget->is64Bit())
337 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
340 if (Subtarget->hasLZCNT()) {
341 // When promoting the i8 variants, force them to i32 for a shorter
343 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
344 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
345 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
346 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
347 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
348 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
349 if (Subtarget->is64Bit())
350 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
352 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
353 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
354 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
355 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
356 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
357 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
358 if (Subtarget->is64Bit()) {
359 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
360 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
364 // Special handling for half-precision floating point conversions.
365 // If we don't have F16C support, then lower half float conversions
366 // into library calls.
367 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
368 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
369 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
372 // There's never any support for operations beyond MVT::f32.
373 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
374 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
375 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
376 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
378 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
379 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
380 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
381 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
382 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
383 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
385 if (Subtarget->hasPOPCNT()) {
386 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
388 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
389 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
390 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
391 if (Subtarget->is64Bit())
392 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
395 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
397 if (!Subtarget->hasMOVBE())
398 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
400 // These should be promoted to a larger select which is supported.
401 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
402 // X86 wants to expand cmov itself.
403 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
404 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
405 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
406 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
407 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
408 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
409 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
410 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
411 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
412 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
413 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
414 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
415 if (Subtarget->is64Bit()) {
416 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
417 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
419 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
420 setOperationAction(ISD::CATCHRET , MVT::Other, Custom);
421 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
422 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
423 // support continuation, user-level threading, and etc.. As a result, no
424 // other SjLj exception interfaces are implemented and please don't build
425 // your own exception handling based on them.
426 // LLVM/Clang supports zero-cost DWARF exception handling.
427 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
428 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
431 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
432 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
433 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
434 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
435 if (Subtarget->is64Bit())
436 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
437 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
438 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
439 if (Subtarget->is64Bit()) {
440 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
441 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
442 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
443 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
444 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
446 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
447 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
448 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
449 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
450 if (Subtarget->is64Bit()) {
451 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
452 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
453 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
456 if (Subtarget->hasSSE1())
457 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
459 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
461 // Expand certain atomics
462 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
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 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
852 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
853 MVT VT = (MVT::SimpleValueType)i;
854 // Do not attempt to custom lower non-power-of-2 vectors
855 if (!isPowerOf2_32(VT.getVectorNumElements()))
857 // Do not attempt to custom lower non-128-bit vectors
858 if (!VT.is128BitVector())
860 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
861 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
862 setOperationAction(ISD::VSELECT, VT, Custom);
863 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
866 // We support custom legalizing of sext and anyext loads for specific
867 // memory vector types which we can load as a scalar (or sequence of
868 // scalars) and extend in-register to a legal 128-bit vector type. For sext
869 // loads these must work with a single scalar load.
870 for (MVT VT : MVT::integer_vector_valuetypes()) {
871 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
872 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
873 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
874 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
875 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
876 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
877 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
878 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
879 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
882 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
883 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
884 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
885 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
886 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
887 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
888 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
889 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
891 if (Subtarget->is64Bit()) {
892 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
893 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
896 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
897 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
898 MVT VT = (MVT::SimpleValueType)i;
900 // Do not attempt to promote non-128-bit vectors
901 if (!VT.is128BitVector())
904 setOperationAction(ISD::AND, VT, Promote);
905 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
906 setOperationAction(ISD::OR, VT, Promote);
907 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
908 setOperationAction(ISD::XOR, VT, Promote);
909 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
910 setOperationAction(ISD::LOAD, VT, Promote);
911 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
912 setOperationAction(ISD::SELECT, VT, Promote);
913 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
916 // Custom lower v2i64 and v2f64 selects.
917 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
918 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
919 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
920 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
922 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
923 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
925 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
927 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
928 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
929 // As there is no 64-bit GPR available, we need build a special custom
930 // sequence to convert from v2i32 to v2f32.
931 if (!Subtarget->is64Bit())
932 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
934 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
935 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
937 for (MVT VT : MVT::fp_vector_valuetypes())
938 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
940 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
941 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
942 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
945 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
946 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
947 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
948 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
949 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
950 setOperationAction(ISD::FRINT, RoundedTy, Legal);
951 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
954 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
955 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
956 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
957 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
958 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
959 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
960 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
961 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
963 // FIXME: Do we need to handle scalar-to-vector here?
964 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
966 // We directly match byte blends in the backend as they match the VSELECT
968 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
970 // SSE41 brings specific instructions for doing vector sign extend even in
971 // cases where we don't have SRA.
972 for (MVT VT : MVT::integer_vector_valuetypes()) {
973 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
974 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
975 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
978 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
979 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
980 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
981 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
982 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
983 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
984 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
986 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
987 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
988 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
989 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
990 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
991 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
993 // i8 and i16 vectors are custom because the source register and source
994 // source memory operand types are not the same width. f32 vectors are
995 // custom since the immediate controlling the insert encodes additional
997 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
998 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
999 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1000 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1002 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1003 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1004 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1005 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1007 // FIXME: these should be Legal, but that's only for the case where
1008 // the index is constant. For now custom expand to deal with that.
1009 if (Subtarget->is64Bit()) {
1010 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1011 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1015 if (Subtarget->hasSSE2()) {
1016 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1017 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1018 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1020 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1021 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1023 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1024 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1026 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1027 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1029 // In the customized shift lowering, the legal cases in AVX2 will be
1031 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1032 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1034 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1035 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1037 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1038 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1041 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1042 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1043 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1044 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1045 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1046 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1047 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1049 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1050 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1051 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1053 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1054 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1055 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1056 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1057 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1058 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1059 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1060 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1061 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1062 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1063 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1064 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1066 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1067 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1068 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1069 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1070 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1071 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1072 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1073 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1074 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1075 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1076 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1077 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1079 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1080 // even though v8i16 is a legal type.
1081 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1082 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1083 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1085 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1086 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1087 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1089 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1090 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1092 for (MVT VT : MVT::fp_vector_valuetypes())
1093 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1095 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1096 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1098 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1099 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1101 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1102 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1104 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1105 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1106 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1107 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1109 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1110 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1111 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1113 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1114 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1115 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1116 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1117 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1118 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1119 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1120 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1121 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1122 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1123 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1124 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1126 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1127 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1128 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1129 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1131 if (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()) {
1132 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1133 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1134 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1135 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1136 setOperationAction(ISD::FMA, MVT::f32, Legal);
1137 setOperationAction(ISD::FMA, MVT::f64, Legal);
1140 if (Subtarget->hasInt256()) {
1141 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1142 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1143 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1144 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1146 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1147 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1148 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1149 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1151 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1152 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1153 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1154 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1156 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1157 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1158 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1159 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1161 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1162 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1163 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1164 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1165 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1166 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1167 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1168 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1169 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1170 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1171 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1172 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1174 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1175 // when we have a 256bit-wide blend with immediate.
1176 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1178 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1179 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1180 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1181 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1182 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1183 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1184 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1186 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1187 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1188 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1189 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1190 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1191 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1193 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1194 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1195 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1196 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1198 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1199 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1200 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1201 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1203 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1204 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1205 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1206 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1208 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1209 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1210 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1211 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1212 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1213 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1214 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1215 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1216 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1217 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1218 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1219 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1222 // In the customized shift lowering, the legal cases in AVX2 will be
1224 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1225 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1227 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1228 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1230 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1231 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1233 // Custom lower several nodes for 256-bit types.
1234 for (MVT VT : MVT::vector_valuetypes()) {
1235 if (VT.getScalarSizeInBits() >= 32) {
1236 setOperationAction(ISD::MLOAD, VT, Legal);
1237 setOperationAction(ISD::MSTORE, VT, Legal);
1239 // Extract subvector is special because the value type
1240 // (result) is 128-bit but the source is 256-bit wide.
1241 if (VT.is128BitVector()) {
1242 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1244 // Do not attempt to custom lower other non-256-bit vectors
1245 if (!VT.is256BitVector())
1248 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1249 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1250 setOperationAction(ISD::VSELECT, VT, Custom);
1251 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1252 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1253 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1254 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1255 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1258 if (Subtarget->hasInt256())
1259 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1262 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1263 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1264 MVT VT = (MVT::SimpleValueType)i;
1266 // Do not attempt to promote non-256-bit vectors
1267 if (!VT.is256BitVector())
1270 setOperationAction(ISD::AND, VT, Promote);
1271 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1272 setOperationAction(ISD::OR, VT, Promote);
1273 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1274 setOperationAction(ISD::XOR, VT, Promote);
1275 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1276 setOperationAction(ISD::LOAD, VT, Promote);
1277 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1278 setOperationAction(ISD::SELECT, VT, Promote);
1279 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1283 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1284 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1285 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1286 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1287 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1289 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1290 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1291 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1293 for (MVT VT : MVT::fp_vector_valuetypes())
1294 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1296 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1297 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1298 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1299 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1300 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1301 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1302 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1303 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1304 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1305 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1306 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1307 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1309 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1310 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1311 setOperationAction(ISD::XOR, MVT::i1, Legal);
1312 setOperationAction(ISD::OR, MVT::i1, Legal);
1313 setOperationAction(ISD::AND, MVT::i1, Legal);
1314 setOperationAction(ISD::SUB, MVT::i1, Custom);
1315 setOperationAction(ISD::ADD, MVT::i1, Custom);
1316 setOperationAction(ISD::MUL, MVT::i1, Custom);
1317 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1318 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1319 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1320 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1321 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1323 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1324 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1325 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1326 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1327 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1328 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1330 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1331 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1332 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1333 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1334 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1335 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1336 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1337 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1339 // FIXME: [US]INT_TO_FP are not legal for f80.
1340 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1341 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1342 if (Subtarget->is64Bit()) {
1343 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1344 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1346 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1347 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1348 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1349 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1350 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1351 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1352 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1353 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1354 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1355 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1356 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1357 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1358 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1359 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1360 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1361 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1363 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1364 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1365 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1366 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1367 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1368 if (Subtarget->hasVLX()){
1369 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1370 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1371 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1372 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1373 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1375 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1376 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1377 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1378 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1379 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1381 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1382 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1383 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1384 if (Subtarget->hasDQI()) {
1385 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1386 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1388 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1389 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1390 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1391 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1392 if (Subtarget->hasVLX()) {
1393 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1394 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1395 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1396 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1397 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1398 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1399 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1400 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1403 if (Subtarget->hasVLX()) {
1404 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1405 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1406 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1407 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1408 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1409 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1410 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1411 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1413 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1414 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1415 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1416 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1417 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1418 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1419 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1420 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1421 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1422 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1423 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1424 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1425 if (Subtarget->hasDQI()) {
1426 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1427 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1429 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1430 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1431 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1432 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1433 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1434 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1435 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1436 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1437 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1438 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1440 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1441 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1442 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1443 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1444 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1446 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1447 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1449 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1451 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1452 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1453 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1454 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1455 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1456 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1457 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1458 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1459 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1460 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1461 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1463 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1464 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1465 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1466 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1467 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1468 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1469 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1470 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1472 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1473 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1475 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1476 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1478 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1480 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1481 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1483 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1484 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1486 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1487 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1489 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1490 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1491 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1492 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1493 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1494 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1496 if (Subtarget->hasCDI()) {
1497 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1498 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1500 if (Subtarget->hasDQI()) {
1501 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1502 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1503 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1505 // Custom lower several nodes.
1506 for (MVT VT : MVT::vector_valuetypes()) {
1507 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1509 setOperationAction(ISD::AND, VT, Legal);
1510 setOperationAction(ISD::OR, VT, Legal);
1511 setOperationAction(ISD::XOR, VT, Legal);
1513 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1514 setOperationAction(ISD::MGATHER, VT, Custom);
1515 setOperationAction(ISD::MSCATTER, VT, Custom);
1517 // Extract subvector is special because the value type
1518 // (result) is 256/128-bit but the source is 512-bit wide.
1519 if (VT.is128BitVector() || VT.is256BitVector()) {
1520 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1522 if (VT.getVectorElementType() == MVT::i1)
1523 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1525 // Do not attempt to custom lower other non-512-bit vectors
1526 if (!VT.is512BitVector())
1529 if (EltSize >= 32) {
1530 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1531 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1532 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1533 setOperationAction(ISD::VSELECT, VT, Legal);
1534 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1535 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1536 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1537 setOperationAction(ISD::MLOAD, VT, Legal);
1538 setOperationAction(ISD::MSTORE, VT, Legal);
1541 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1542 MVT VT = (MVT::SimpleValueType)i;
1544 // Do not attempt to promote non-512-bit vectors.
1545 if (!VT.is512BitVector())
1548 setOperationAction(ISD::SELECT, VT, Promote);
1549 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1553 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1554 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1555 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1557 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1558 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1560 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1561 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1562 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1563 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1564 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1565 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1566 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1567 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1568 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1569 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1570 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1571 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1572 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1573 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1574 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1575 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1576 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1577 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1578 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1579 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1580 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1581 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1582 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1583 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1584 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1585 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1586 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1587 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1588 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1589 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1591 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1592 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1593 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1594 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1595 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1596 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1597 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1598 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1600 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1601 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1602 if (Subtarget->hasVLX())
1603 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1605 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1606 const MVT VT = (MVT::SimpleValueType)i;
1608 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1610 // Do not attempt to promote non-512-bit vectors.
1611 if (!VT.is512BitVector())
1615 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1616 setOperationAction(ISD::VSELECT, VT, Legal);
1621 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1622 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1623 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1625 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1626 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1627 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1628 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1629 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1630 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1631 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1632 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1633 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1634 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1636 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1637 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1638 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1639 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1640 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1641 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1642 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1643 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1645 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1646 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1647 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1648 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1649 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1650 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1651 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1652 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1655 // We want to custom lower some of our intrinsics.
1656 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1657 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1658 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1659 if (!Subtarget->is64Bit())
1660 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1662 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1663 // handle type legalization for these operations here.
1665 // FIXME: We really should do custom legalization for addition and
1666 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1667 // than generic legalization for 64-bit multiplication-with-overflow, though.
1668 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1669 // Add/Sub/Mul with overflow operations are custom lowered.
1671 setOperationAction(ISD::SADDO, VT, Custom);
1672 setOperationAction(ISD::UADDO, VT, Custom);
1673 setOperationAction(ISD::SSUBO, VT, Custom);
1674 setOperationAction(ISD::USUBO, VT, Custom);
1675 setOperationAction(ISD::SMULO, VT, Custom);
1676 setOperationAction(ISD::UMULO, VT, Custom);
1680 if (!Subtarget->is64Bit()) {
1681 // These libcalls are not available in 32-bit.
1682 setLibcallName(RTLIB::SHL_I128, nullptr);
1683 setLibcallName(RTLIB::SRL_I128, nullptr);
1684 setLibcallName(RTLIB::SRA_I128, nullptr);
1687 // Combine sin / cos into one node or libcall if possible.
1688 if (Subtarget->hasSinCos()) {
1689 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1690 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1691 if (Subtarget->isTargetDarwin()) {
1692 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1693 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1694 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1695 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1699 if (Subtarget->isTargetWin64()) {
1700 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1701 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1702 setOperationAction(ISD::SREM, MVT::i128, Custom);
1703 setOperationAction(ISD::UREM, MVT::i128, Custom);
1704 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1705 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1708 // We have target-specific dag combine patterns for the following nodes:
1709 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1710 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1711 setTargetDAGCombine(ISD::BITCAST);
1712 setTargetDAGCombine(ISD::VSELECT);
1713 setTargetDAGCombine(ISD::SELECT);
1714 setTargetDAGCombine(ISD::SHL);
1715 setTargetDAGCombine(ISD::SRA);
1716 setTargetDAGCombine(ISD::SRL);
1717 setTargetDAGCombine(ISD::OR);
1718 setTargetDAGCombine(ISD::AND);
1719 setTargetDAGCombine(ISD::ADD);
1720 setTargetDAGCombine(ISD::FADD);
1721 setTargetDAGCombine(ISD::FSUB);
1722 setTargetDAGCombine(ISD::FMA);
1723 setTargetDAGCombine(ISD::SUB);
1724 setTargetDAGCombine(ISD::LOAD);
1725 setTargetDAGCombine(ISD::MLOAD);
1726 setTargetDAGCombine(ISD::STORE);
1727 setTargetDAGCombine(ISD::MSTORE);
1728 setTargetDAGCombine(ISD::ZERO_EXTEND);
1729 setTargetDAGCombine(ISD::ANY_EXTEND);
1730 setTargetDAGCombine(ISD::SIGN_EXTEND);
1731 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1732 setTargetDAGCombine(ISD::SINT_TO_FP);
1733 setTargetDAGCombine(ISD::UINT_TO_FP);
1734 setTargetDAGCombine(ISD::SETCC);
1735 setTargetDAGCombine(ISD::BUILD_VECTOR);
1736 setTargetDAGCombine(ISD::MUL);
1737 setTargetDAGCombine(ISD::XOR);
1739 computeRegisterProperties(Subtarget->getRegisterInfo());
1741 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1742 MaxStoresPerMemsetOptSize = 8;
1743 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1744 MaxStoresPerMemcpyOptSize = 4;
1745 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1746 MaxStoresPerMemmoveOptSize = 4;
1747 setPrefLoopAlignment(4); // 2^4 bytes.
1749 // Predictable cmov don't hurt on atom because it's in-order.
1750 PredictableSelectIsExpensive = !Subtarget->isAtom();
1751 EnableExtLdPromotion = true;
1752 setPrefFunctionAlignment(4); // 2^4 bytes.
1754 verifyIntrinsicTables();
1757 // This has so far only been implemented for 64-bit MachO.
1758 bool X86TargetLowering::useLoadStackGuardNode() const {
1759 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1762 TargetLoweringBase::LegalizeTypeAction
1763 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1764 if (ExperimentalVectorWideningLegalization &&
1765 VT.getVectorNumElements() != 1 &&
1766 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1767 return TypeWidenVector;
1769 return TargetLoweringBase::getPreferredVectorAction(VT);
1772 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1775 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1777 const unsigned NumElts = VT.getVectorNumElements();
1778 const EVT EltVT = VT.getVectorElementType();
1779 if (VT.is512BitVector()) {
1780 if (Subtarget->hasAVX512())
1781 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1782 EltVT == MVT::f32 || EltVT == MVT::f64)
1784 case 8: return MVT::v8i1;
1785 case 16: return MVT::v16i1;
1787 if (Subtarget->hasBWI())
1788 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1790 case 32: return MVT::v32i1;
1791 case 64: return MVT::v64i1;
1795 if (VT.is256BitVector() || VT.is128BitVector()) {
1796 if (Subtarget->hasVLX())
1797 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1798 EltVT == MVT::f32 || EltVT == MVT::f64)
1800 case 2: return MVT::v2i1;
1801 case 4: return MVT::v4i1;
1802 case 8: return MVT::v8i1;
1804 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1805 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1807 case 8: return MVT::v8i1;
1808 case 16: return MVT::v16i1;
1809 case 32: return MVT::v32i1;
1813 return VT.changeVectorElementTypeToInteger();
1816 /// Helper for getByValTypeAlignment to determine
1817 /// the desired ByVal argument alignment.
1818 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1821 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1822 if (VTy->getBitWidth() == 128)
1824 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1825 unsigned EltAlign = 0;
1826 getMaxByValAlign(ATy->getElementType(), EltAlign);
1827 if (EltAlign > MaxAlign)
1828 MaxAlign = EltAlign;
1829 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1830 for (auto *EltTy : STy->elements()) {
1831 unsigned EltAlign = 0;
1832 getMaxByValAlign(EltTy, EltAlign);
1833 if (EltAlign > MaxAlign)
1834 MaxAlign = EltAlign;
1841 /// Return the desired alignment for ByVal aggregate
1842 /// function arguments in the caller parameter area. For X86, aggregates
1843 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1844 /// are at 4-byte boundaries.
1845 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1846 const DataLayout &DL) const {
1847 if (Subtarget->is64Bit()) {
1848 // Max of 8 and alignment of type.
1849 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1856 if (Subtarget->hasSSE1())
1857 getMaxByValAlign(Ty, Align);
1861 /// Returns the target specific optimal type for load
1862 /// and store operations as a result of memset, memcpy, and memmove
1863 /// lowering. If DstAlign is zero that means it's safe to destination
1864 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1865 /// means there isn't a need to check it against alignment requirement,
1866 /// probably because the source does not need to be loaded. If 'IsMemset' is
1867 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1868 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1869 /// source is constant so it does not need to be loaded.
1870 /// It returns EVT::Other if the type should be determined using generic
1871 /// target-independent logic.
1873 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1874 unsigned DstAlign, unsigned SrcAlign,
1875 bool IsMemset, bool ZeroMemset,
1877 MachineFunction &MF) const {
1878 const Function *F = MF.getFunction();
1879 if ((!IsMemset || ZeroMemset) &&
1880 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1882 (!Subtarget->isUnalignedMem16Slow() ||
1883 ((DstAlign == 0 || DstAlign >= 16) &&
1884 (SrcAlign == 0 || SrcAlign >= 16)))) {
1886 // FIXME: Check if unaligned 32-byte accesses are slow.
1887 if (Subtarget->hasInt256())
1889 if (Subtarget->hasFp256())
1892 if (Subtarget->hasSSE2())
1894 if (Subtarget->hasSSE1())
1896 } else if (!MemcpyStrSrc && Size >= 8 &&
1897 !Subtarget->is64Bit() &&
1898 Subtarget->hasSSE2()) {
1899 // Do not use f64 to lower memcpy if source is string constant. It's
1900 // better to use i32 to avoid the loads.
1904 // This is a compromise. If we reach here, unaligned accesses may be slow on
1905 // this target. However, creating smaller, aligned accesses could be even
1906 // slower and would certainly be a lot more code.
1907 if (Subtarget->is64Bit() && Size >= 8)
1912 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1914 return X86ScalarSSEf32;
1915 else if (VT == MVT::f64)
1916 return X86ScalarSSEf64;
1921 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1926 switch (VT.getSizeInBits()) {
1928 // 8-byte and under are always assumed to be fast.
1932 *Fast = !Subtarget->isUnalignedMem16Slow();
1935 *Fast = !Subtarget->isUnalignedMem32Slow();
1937 // TODO: What about AVX-512 (512-bit) accesses?
1940 // Misaligned accesses of any size are always allowed.
1944 /// Return the entry encoding for a jump table in the
1945 /// current function. The returned value is a member of the
1946 /// MachineJumpTableInfo::JTEntryKind enum.
1947 unsigned X86TargetLowering::getJumpTableEncoding() const {
1948 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1950 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1951 Subtarget->isPICStyleGOT())
1952 return MachineJumpTableInfo::EK_Custom32;
1954 // Otherwise, use the normal jump table encoding heuristics.
1955 return TargetLowering::getJumpTableEncoding();
1958 bool X86TargetLowering::useSoftFloat() const {
1959 return Subtarget->useSoftFloat();
1963 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1964 const MachineBasicBlock *MBB,
1965 unsigned uid,MCContext &Ctx) const{
1966 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1967 Subtarget->isPICStyleGOT());
1968 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1970 return MCSymbolRefExpr::create(MBB->getSymbol(),
1971 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1974 /// Returns relocation base for the given PIC jumptable.
1975 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1976 SelectionDAG &DAG) const {
1977 if (!Subtarget->is64Bit())
1978 // This doesn't have SDLoc associated with it, but is not really the
1979 // same as a Register.
1980 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
1981 getPointerTy(DAG.getDataLayout()));
1985 /// This returns the relocation base for the given PIC jumptable,
1986 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1987 const MCExpr *X86TargetLowering::
1988 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1989 MCContext &Ctx) const {
1990 // X86-64 uses RIP relative addressing based on the jump table label.
1991 if (Subtarget->isPICStyleRIPRel())
1992 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1994 // Otherwise, the reference is relative to the PIC base.
1995 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
1998 std::pair<const TargetRegisterClass *, uint8_t>
1999 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2001 const TargetRegisterClass *RRC = nullptr;
2003 switch (VT.SimpleTy) {
2005 return TargetLowering::findRepresentativeClass(TRI, VT);
2006 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
2007 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
2010 RRC = &X86::VR64RegClass;
2012 case MVT::f32: case MVT::f64:
2013 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2014 case MVT::v4f32: case MVT::v2f64:
2015 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2017 RRC = &X86::VR128RegClass;
2020 return std::make_pair(RRC, Cost);
2023 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2024 unsigned &Offset) const {
2025 if (!Subtarget->isTargetLinux())
2028 if (Subtarget->is64Bit()) {
2029 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2031 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2043 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2044 unsigned DestAS) const {
2045 assert(SrcAS != DestAS && "Expected different address spaces!");
2047 return SrcAS < 256 && DestAS < 256;
2050 //===----------------------------------------------------------------------===//
2051 // Return Value Calling Convention Implementation
2052 //===----------------------------------------------------------------------===//
2054 #include "X86GenCallingConv.inc"
2057 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2058 MachineFunction &MF, bool isVarArg,
2059 const SmallVectorImpl<ISD::OutputArg> &Outs,
2060 LLVMContext &Context) const {
2061 SmallVector<CCValAssign, 16> RVLocs;
2062 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2063 return CCInfo.CheckReturn(Outs, RetCC_X86);
2066 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2067 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2072 X86TargetLowering::LowerReturn(SDValue Chain,
2073 CallingConv::ID CallConv, bool isVarArg,
2074 const SmallVectorImpl<ISD::OutputArg> &Outs,
2075 const SmallVectorImpl<SDValue> &OutVals,
2076 SDLoc dl, SelectionDAG &DAG) const {
2077 MachineFunction &MF = DAG.getMachineFunction();
2078 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2080 SmallVector<CCValAssign, 16> RVLocs;
2081 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2082 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2085 SmallVector<SDValue, 6> RetOps;
2086 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2087 // Operand #1 = Bytes To Pop
2088 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2091 // Copy the result values into the output registers.
2092 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2093 CCValAssign &VA = RVLocs[i];
2094 assert(VA.isRegLoc() && "Can only return in registers!");
2095 SDValue ValToCopy = OutVals[i];
2096 EVT ValVT = ValToCopy.getValueType();
2098 // Promote values to the appropriate types.
2099 if (VA.getLocInfo() == CCValAssign::SExt)
2100 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2101 else if (VA.getLocInfo() == CCValAssign::ZExt)
2102 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2103 else if (VA.getLocInfo() == CCValAssign::AExt) {
2104 if (ValVT.isVector() && ValVT.getScalarType() == MVT::i1)
2105 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2107 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2109 else if (VA.getLocInfo() == CCValAssign::BCvt)
2110 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2112 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2113 "Unexpected FP-extend for return value.");
2115 // If this is x86-64, and we disabled SSE, we can't return FP values,
2116 // or SSE or MMX vectors.
2117 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2118 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2119 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2120 report_fatal_error("SSE register return with SSE disabled");
2122 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2123 // llvm-gcc has never done it right and no one has noticed, so this
2124 // should be OK for now.
2125 if (ValVT == MVT::f64 &&
2126 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2127 report_fatal_error("SSE2 register return with SSE2 disabled");
2129 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2130 // the RET instruction and handled by the FP Stackifier.
2131 if (VA.getLocReg() == X86::FP0 ||
2132 VA.getLocReg() == X86::FP1) {
2133 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2134 // change the value to the FP stack register class.
2135 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2136 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2137 RetOps.push_back(ValToCopy);
2138 // Don't emit a copytoreg.
2142 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2143 // which is returned in RAX / RDX.
2144 if (Subtarget->is64Bit()) {
2145 if (ValVT == MVT::x86mmx) {
2146 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2147 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2148 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2150 // If we don't have SSE2 available, convert to v4f32 so the generated
2151 // register is legal.
2152 if (!Subtarget->hasSSE2())
2153 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2158 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2159 Flag = Chain.getValue(1);
2160 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2163 // All x86 ABIs require that for returning structs by value we copy
2164 // the sret argument into %rax/%eax (depending on ABI) for the return.
2165 // We saved the argument into a virtual register in the entry block,
2166 // so now we copy the value out and into %rax/%eax.
2168 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2169 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2170 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2171 // either case FuncInfo->setSRetReturnReg() will have been called.
2172 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2173 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2174 getPointerTy(MF.getDataLayout()));
2177 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2178 X86::RAX : X86::EAX;
2179 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2180 Flag = Chain.getValue(1);
2182 // RAX/EAX now acts like a return value.
2184 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2187 RetOps[0] = Chain; // Update chain.
2189 // Add the flag if we have it.
2191 RetOps.push_back(Flag);
2193 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2196 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2197 if (N->getNumValues() != 1)
2199 if (!N->hasNUsesOfValue(1, 0))
2202 SDValue TCChain = Chain;
2203 SDNode *Copy = *N->use_begin();
2204 if (Copy->getOpcode() == ISD::CopyToReg) {
2205 // If the copy has a glue operand, we conservatively assume it isn't safe to
2206 // perform a tail call.
2207 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2209 TCChain = Copy->getOperand(0);
2210 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2213 bool HasRet = false;
2214 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2216 if (UI->getOpcode() != X86ISD::RET_FLAG)
2218 // If we are returning more than one value, we can definitely
2219 // not make a tail call see PR19530
2220 if (UI->getNumOperands() > 4)
2222 if (UI->getNumOperands() == 4 &&
2223 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2236 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2237 ISD::NodeType ExtendKind) const {
2239 // TODO: Is this also valid on 32-bit?
2240 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2241 ReturnMVT = MVT::i8;
2243 ReturnMVT = MVT::i32;
2245 EVT MinVT = getRegisterType(Context, ReturnMVT);
2246 return VT.bitsLT(MinVT) ? MinVT : VT;
2249 /// Lower the result values of a call into the
2250 /// appropriate copies out of appropriate physical registers.
2253 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2254 CallingConv::ID CallConv, bool isVarArg,
2255 const SmallVectorImpl<ISD::InputArg> &Ins,
2256 SDLoc dl, SelectionDAG &DAG,
2257 SmallVectorImpl<SDValue> &InVals) const {
2259 // Assign locations to each value returned by this call.
2260 SmallVector<CCValAssign, 16> RVLocs;
2261 bool Is64Bit = Subtarget->is64Bit();
2262 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2264 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2266 // Copy all of the result registers out of their specified physreg.
2267 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2268 CCValAssign &VA = RVLocs[i];
2269 EVT CopyVT = VA.getLocVT();
2271 // If this is x86-64, and we disabled SSE, we can't return FP values
2272 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2273 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2274 report_fatal_error("SSE register return with SSE disabled");
2277 // If we prefer to use the value in xmm registers, copy it out as f80 and
2278 // use a truncate to move it from fp stack reg to xmm reg.
2279 bool RoundAfterCopy = false;
2280 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2281 isScalarFPTypeInSSEReg(VA.getValVT())) {
2283 RoundAfterCopy = (CopyVT != VA.getLocVT());
2286 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2287 CopyVT, InFlag).getValue(1);
2288 SDValue Val = Chain.getValue(0);
2291 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2292 // This truncation won't change the value.
2293 DAG.getIntPtrConstant(1, dl));
2295 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2296 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2298 InFlag = Chain.getValue(2);
2299 InVals.push_back(Val);
2305 //===----------------------------------------------------------------------===//
2306 // C & StdCall & Fast Calling Convention implementation
2307 //===----------------------------------------------------------------------===//
2308 // StdCall calling convention seems to be standard for many Windows' API
2309 // routines and around. It differs from C calling convention just a little:
2310 // callee should clean up the stack, not caller. Symbols should be also
2311 // decorated in some fancy way :) It doesn't support any vector arguments.
2312 // For info on fast calling convention see Fast Calling Convention (tail call)
2313 // implementation LowerX86_32FastCCCallTo.
2315 /// CallIsStructReturn - Determines whether a call uses struct return
2317 enum StructReturnType {
2322 static StructReturnType
2323 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2325 return NotStructReturn;
2327 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2328 if (!Flags.isSRet())
2329 return NotStructReturn;
2330 if (Flags.isInReg())
2331 return RegStructReturn;
2332 return StackStructReturn;
2335 /// Determines whether a function uses struct return semantics.
2336 static StructReturnType
2337 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2339 return NotStructReturn;
2341 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2342 if (!Flags.isSRet())
2343 return NotStructReturn;
2344 if (Flags.isInReg())
2345 return RegStructReturn;
2346 return StackStructReturn;
2349 /// Make a copy of an aggregate at address specified by "Src" to address
2350 /// "Dst" with size and alignment information specified by the specific
2351 /// parameter attribute. The copy will be passed as a byval function parameter.
2353 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2354 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2356 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2358 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2359 /*isVolatile*/false, /*AlwaysInline=*/true,
2360 /*isTailCall*/false,
2361 MachinePointerInfo(), MachinePointerInfo());
2364 /// Return true if the calling convention is one that
2365 /// supports tail call optimization.
2366 static bool IsTailCallConvention(CallingConv::ID CC) {
2367 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2368 CC == CallingConv::HiPE);
2371 /// \brief Return true if the calling convention is a C calling convention.
2372 static bool IsCCallConvention(CallingConv::ID CC) {
2373 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2374 CC == CallingConv::X86_64_SysV);
2377 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2379 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2380 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2384 CallingConv::ID CalleeCC = CS.getCallingConv();
2385 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2391 /// Return true if the function is being made into
2392 /// a tailcall target by changing its ABI.
2393 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2394 bool GuaranteedTailCallOpt) {
2395 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2399 X86TargetLowering::LowerMemArgument(SDValue Chain,
2400 CallingConv::ID CallConv,
2401 const SmallVectorImpl<ISD::InputArg> &Ins,
2402 SDLoc dl, SelectionDAG &DAG,
2403 const CCValAssign &VA,
2404 MachineFrameInfo *MFI,
2406 // Create the nodes corresponding to a load from this parameter slot.
2407 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2408 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2409 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2410 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2413 // If value is passed by pointer we have address passed instead of the value
2415 bool ExtendedInMem = VA.isExtInLoc() &&
2416 VA.getValVT().getScalarType() == MVT::i1;
2418 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2419 ValVT = VA.getLocVT();
2421 ValVT = VA.getValVT();
2423 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2424 // changed with more analysis.
2425 // In case of tail call optimization mark all arguments mutable. Since they
2426 // could be overwritten by lowering of arguments in case of a tail call.
2427 if (Flags.isByVal()) {
2428 unsigned Bytes = Flags.getByValSize();
2429 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2430 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2431 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2433 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2434 VA.getLocMemOffset(), isImmutable);
2435 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2436 SDValue Val = DAG.getLoad(
2437 ValVT, dl, Chain, FIN,
2438 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2440 return ExtendedInMem ?
2441 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2445 // FIXME: Get this from tablegen.
2446 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2447 const X86Subtarget *Subtarget) {
2448 assert(Subtarget->is64Bit());
2450 if (Subtarget->isCallingConvWin64(CallConv)) {
2451 static const MCPhysReg GPR64ArgRegsWin64[] = {
2452 X86::RCX, X86::RDX, X86::R8, X86::R9
2454 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2457 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2458 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2460 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2463 // FIXME: Get this from tablegen.
2464 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2465 CallingConv::ID CallConv,
2466 const X86Subtarget *Subtarget) {
2467 assert(Subtarget->is64Bit());
2468 if (Subtarget->isCallingConvWin64(CallConv)) {
2469 // The XMM registers which might contain var arg parameters are shadowed
2470 // in their paired GPR. So we only need to save the GPR to their home
2472 // TODO: __vectorcall will change this.
2476 const Function *Fn = MF.getFunction();
2477 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2478 bool isSoftFloat = Subtarget->useSoftFloat();
2479 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2480 "SSE register cannot be used when SSE is disabled!");
2481 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2482 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2486 static const MCPhysReg XMMArgRegs64Bit[] = {
2487 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2488 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2490 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2494 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2495 CallingConv::ID CallConv,
2497 const SmallVectorImpl<ISD::InputArg> &Ins,
2500 SmallVectorImpl<SDValue> &InVals)
2502 MachineFunction &MF = DAG.getMachineFunction();
2503 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2504 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2506 const Function* Fn = MF.getFunction();
2507 if (Fn->hasExternalLinkage() &&
2508 Subtarget->isTargetCygMing() &&
2509 Fn->getName() == "main")
2510 FuncInfo->setForceFramePointer(true);
2512 MachineFrameInfo *MFI = MF.getFrameInfo();
2513 bool Is64Bit = Subtarget->is64Bit();
2514 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2516 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2517 "Var args not supported with calling convention fastcc, ghc or hipe");
2519 // Assign locations to all of the incoming arguments.
2520 SmallVector<CCValAssign, 16> ArgLocs;
2521 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2523 // Allocate shadow area for Win64
2525 CCInfo.AllocateStack(32, 8);
2527 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2529 unsigned LastVal = ~0U;
2531 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2532 CCValAssign &VA = ArgLocs[i];
2533 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2535 assert(VA.getValNo() != LastVal &&
2536 "Don't support value assigned to multiple locs yet");
2538 LastVal = VA.getValNo();
2540 if (VA.isRegLoc()) {
2541 EVT RegVT = VA.getLocVT();
2542 const TargetRegisterClass *RC;
2543 if (RegVT == MVT::i32)
2544 RC = &X86::GR32RegClass;
2545 else if (Is64Bit && RegVT == MVT::i64)
2546 RC = &X86::GR64RegClass;
2547 else if (RegVT == MVT::f32)
2548 RC = &X86::FR32RegClass;
2549 else if (RegVT == MVT::f64)
2550 RC = &X86::FR64RegClass;
2551 else if (RegVT.is512BitVector())
2552 RC = &X86::VR512RegClass;
2553 else if (RegVT.is256BitVector())
2554 RC = &X86::VR256RegClass;
2555 else if (RegVT.is128BitVector())
2556 RC = &X86::VR128RegClass;
2557 else if (RegVT == MVT::x86mmx)
2558 RC = &X86::VR64RegClass;
2559 else if (RegVT == MVT::i1)
2560 RC = &X86::VK1RegClass;
2561 else if (RegVT == MVT::v8i1)
2562 RC = &X86::VK8RegClass;
2563 else if (RegVT == MVT::v16i1)
2564 RC = &X86::VK16RegClass;
2565 else if (RegVT == MVT::v32i1)
2566 RC = &X86::VK32RegClass;
2567 else if (RegVT == MVT::v64i1)
2568 RC = &X86::VK64RegClass;
2570 llvm_unreachable("Unknown argument type!");
2572 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2573 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2575 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2576 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2578 if (VA.getLocInfo() == CCValAssign::SExt)
2579 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2580 DAG.getValueType(VA.getValVT()));
2581 else if (VA.getLocInfo() == CCValAssign::ZExt)
2582 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2583 DAG.getValueType(VA.getValVT()));
2584 else if (VA.getLocInfo() == CCValAssign::BCvt)
2585 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2587 if (VA.isExtInLoc()) {
2588 // Handle MMX values passed in XMM regs.
2589 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2590 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2592 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2595 assert(VA.isMemLoc());
2596 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2599 // If value is passed via pointer - do a load.
2600 if (VA.getLocInfo() == CCValAssign::Indirect)
2601 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2602 MachinePointerInfo(), false, false, false, 0);
2604 InVals.push_back(ArgValue);
2607 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2608 // All x86 ABIs require that for returning structs by value we copy the
2609 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2610 // the argument into a virtual register so that we can access it from the
2612 if (Ins[i].Flags.isSRet()) {
2613 unsigned Reg = FuncInfo->getSRetReturnReg();
2615 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2616 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2617 FuncInfo->setSRetReturnReg(Reg);
2619 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2620 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2625 unsigned StackSize = CCInfo.getNextStackOffset();
2626 // Align stack specially for tail calls.
2627 if (FuncIsMadeTailCallSafe(CallConv,
2628 MF.getTarget().Options.GuaranteedTailCallOpt))
2629 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2631 // If the function takes variable number of arguments, make a frame index for
2632 // the start of the first vararg value... for expansion of llvm.va_start. We
2633 // can skip this if there are no va_start calls.
2634 if (MFI->hasVAStart() &&
2635 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2636 CallConv != CallingConv::X86_ThisCall))) {
2637 FuncInfo->setVarArgsFrameIndex(
2638 MFI->CreateFixedObject(1, StackSize, true));
2641 MachineModuleInfo &MMI = MF.getMMI();
2642 const Function *WinEHParent = nullptr;
2643 if (MMI.hasWinEHFuncInfo(Fn))
2644 WinEHParent = MMI.getWinEHParent(Fn);
2645 bool IsWinEHOutlined = WinEHParent && WinEHParent != Fn;
2646 bool IsWinEHParent = WinEHParent && WinEHParent == Fn;
2648 // Figure out if XMM registers are in use.
2649 assert(!(Subtarget->useSoftFloat() &&
2650 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2651 "SSE register cannot be used when SSE is disabled!");
2653 // 64-bit calling conventions support varargs and register parameters, so we
2654 // have to do extra work to spill them in the prologue.
2655 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2656 // Find the first unallocated argument registers.
2657 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2658 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2659 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2660 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2661 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2662 "SSE register cannot be used when SSE is disabled!");
2664 // Gather all the live in physical registers.
2665 SmallVector<SDValue, 6> LiveGPRs;
2666 SmallVector<SDValue, 8> LiveXMMRegs;
2668 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2669 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2671 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2673 if (!ArgXMMs.empty()) {
2674 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2675 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2676 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2677 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2678 LiveXMMRegs.push_back(
2679 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2684 // Get to the caller-allocated home save location. Add 8 to account
2685 // for the return address.
2686 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2687 FuncInfo->setRegSaveFrameIndex(
2688 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2689 // Fixup to set vararg frame on shadow area (4 x i64).
2691 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2693 // For X86-64, if there are vararg parameters that are passed via
2694 // registers, then we must store them to their spots on the stack so
2695 // they may be loaded by deferencing the result of va_next.
2696 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2697 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2698 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2699 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2702 // Store the integer parameter registers.
2703 SmallVector<SDValue, 8> MemOps;
2704 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2705 getPointerTy(DAG.getDataLayout()));
2706 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2707 for (SDValue Val : LiveGPRs) {
2708 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2709 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2711 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2712 MachinePointerInfo::getFixedStack(
2713 DAG.getMachineFunction(),
2714 FuncInfo->getRegSaveFrameIndex(), Offset),
2716 MemOps.push_back(Store);
2720 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2721 // Now store the XMM (fp + vector) parameter registers.
2722 SmallVector<SDValue, 12> SaveXMMOps;
2723 SaveXMMOps.push_back(Chain);
2724 SaveXMMOps.push_back(ALVal);
2725 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2726 FuncInfo->getRegSaveFrameIndex(), dl));
2727 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2728 FuncInfo->getVarArgsFPOffset(), dl));
2729 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2731 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2732 MVT::Other, SaveXMMOps));
2735 if (!MemOps.empty())
2736 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2737 } else if (IsWin64 && IsWinEHOutlined) {
2738 // Get to the caller-allocated home save location. Add 8 to account
2739 // for the return address.
2740 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2741 FuncInfo->setRegSaveFrameIndex(MFI->CreateFixedObject(
2742 /*Size=*/1, /*SPOffset=*/HomeOffset + 8, /*Immutable=*/false));
2744 MMI.getWinEHFuncInfo(Fn)
2745 .CatchHandlerParentFrameObjIdx[const_cast<Function *>(Fn)] =
2746 FuncInfo->getRegSaveFrameIndex();
2748 // Store the second integer parameter (rdx) into rsp+16 relative to the
2749 // stack pointer at the entry of the function.
2750 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2751 getPointerTy(DAG.getDataLayout()));
2752 unsigned GPR = MF.addLiveIn(X86::RDX, &X86::GR64RegClass);
2753 SDValue Val = DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64);
2754 Chain = DAG.getStore(
2755 Val.getValue(1), dl, Val, RSFIN,
2756 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(),
2757 FuncInfo->getRegSaveFrameIndex()),
2758 /*isVolatile=*/true, /*isNonTemporal=*/false, /*Alignment=*/0);
2761 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2762 // Find the largest legal vector type.
2763 MVT VecVT = MVT::Other;
2764 // FIXME: Only some x86_32 calling conventions support AVX512.
2765 if (Subtarget->hasAVX512() &&
2766 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2767 CallConv == CallingConv::Intel_OCL_BI)))
2768 VecVT = MVT::v16f32;
2769 else if (Subtarget->hasAVX())
2771 else if (Subtarget->hasSSE2())
2774 // We forward some GPRs and some vector types.
2775 SmallVector<MVT, 2> RegParmTypes;
2776 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2777 RegParmTypes.push_back(IntVT);
2778 if (VecVT != MVT::Other)
2779 RegParmTypes.push_back(VecVT);
2781 // Compute the set of forwarded registers. The rest are scratch.
2782 SmallVectorImpl<ForwardedRegister> &Forwards =
2783 FuncInfo->getForwardedMustTailRegParms();
2784 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2786 // Conservatively forward AL on x86_64, since it might be used for varargs.
2787 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2788 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2789 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2792 // Copy all forwards from physical to virtual registers.
2793 for (ForwardedRegister &F : Forwards) {
2794 // FIXME: Can we use a less constrained schedule?
2795 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2796 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2797 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2801 // Some CCs need callee pop.
2802 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2803 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2804 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2806 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2807 // If this is an sret function, the return should pop the hidden pointer.
2808 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2809 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2810 argsAreStructReturn(Ins) == StackStructReturn)
2811 FuncInfo->setBytesToPopOnReturn(4);
2815 // RegSaveFrameIndex is X86-64 only.
2816 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2817 if (CallConv == CallingConv::X86_FastCall ||
2818 CallConv == CallingConv::X86_ThisCall)
2819 // fastcc functions can't have varargs.
2820 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2823 FuncInfo->setArgumentStackSize(StackSize);
2825 if (IsWinEHParent) {
2827 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2828 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2829 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2830 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2831 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2832 MachinePointerInfo::getFixedStack(
2833 DAG.getMachineFunction(), UnwindHelpFI),
2834 /*isVolatile=*/true,
2835 /*isNonTemporal=*/false, /*Alignment=*/0);
2837 // Functions using Win32 EH are considered to have opaque SP adjustments
2838 // to force local variables to be addressed from the frame or base
2840 MFI->setHasOpaqueSPAdjustment(true);
2848 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2849 SDValue StackPtr, SDValue Arg,
2850 SDLoc dl, SelectionDAG &DAG,
2851 const CCValAssign &VA,
2852 ISD::ArgFlagsTy Flags) const {
2853 unsigned LocMemOffset = VA.getLocMemOffset();
2854 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2855 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2857 if (Flags.isByVal())
2858 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2860 return DAG.getStore(
2861 Chain, dl, Arg, PtrOff,
2862 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
2866 /// Emit a load of return address if tail call
2867 /// optimization is performed and it is required.
2869 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2870 SDValue &OutRetAddr, SDValue Chain,
2871 bool IsTailCall, bool Is64Bit,
2872 int FPDiff, SDLoc dl) const {
2873 // Adjust the Return address stack slot.
2874 EVT VT = getPointerTy(DAG.getDataLayout());
2875 OutRetAddr = getReturnAddressFrameIndex(DAG);
2877 // Load the "old" Return address.
2878 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2879 false, false, false, 0);
2880 return SDValue(OutRetAddr.getNode(), 1);
2883 /// Emit a store of the return address if tail call
2884 /// optimization is performed and it is required (FPDiff!=0).
2885 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2886 SDValue Chain, SDValue RetAddrFrIdx,
2887 EVT PtrVT, unsigned SlotSize,
2888 int FPDiff, SDLoc dl) {
2889 // Store the return address to the appropriate stack slot.
2890 if (!FPDiff) return Chain;
2891 // Calculate the new stack slot for the return address.
2892 int NewReturnAddrFI =
2893 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2895 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2896 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2897 MachinePointerInfo::getFixedStack(
2898 DAG.getMachineFunction(), NewReturnAddrFI),
2903 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2904 /// operation of specified width.
2905 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
2907 unsigned NumElems = VT.getVectorNumElements();
2908 SmallVector<int, 8> Mask;
2909 Mask.push_back(NumElems);
2910 for (unsigned i = 1; i != NumElems; ++i)
2912 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2916 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2917 SmallVectorImpl<SDValue> &InVals) const {
2918 SelectionDAG &DAG = CLI.DAG;
2920 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2921 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2922 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2923 SDValue Chain = CLI.Chain;
2924 SDValue Callee = CLI.Callee;
2925 CallingConv::ID CallConv = CLI.CallConv;
2926 bool &isTailCall = CLI.IsTailCall;
2927 bool isVarArg = CLI.IsVarArg;
2929 MachineFunction &MF = DAG.getMachineFunction();
2930 bool Is64Bit = Subtarget->is64Bit();
2931 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2932 StructReturnType SR = callIsStructReturn(Outs);
2933 bool IsSibcall = false;
2934 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2935 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
2937 if (Attr.getValueAsString() == "true")
2940 if (Subtarget->isPICStyleGOT() &&
2941 !MF.getTarget().Options.GuaranteedTailCallOpt) {
2942 // If we are using a GOT, disable tail calls to external symbols with
2943 // default visibility. Tail calling such a symbol requires using a GOT
2944 // relocation, which forces early binding of the symbol. This breaks code
2945 // that require lazy function symbol resolution. Using musttail or
2946 // GuaranteedTailCallOpt will override this.
2947 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2948 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
2949 G->getGlobal()->hasDefaultVisibility()))
2953 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2955 // Force this to be a tail call. The verifier rules are enough to ensure
2956 // that we can lower this successfully without moving the return address
2959 } else if (isTailCall) {
2960 // Check if it's really possible to do a tail call.
2961 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2962 isVarArg, SR != NotStructReturn,
2963 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2964 Outs, OutVals, Ins, DAG);
2966 // Sibcalls are automatically detected tailcalls which do not require
2968 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2975 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2976 "Var args not supported with calling convention fastcc, ghc or hipe");
2978 // Analyze operands of the call, assigning locations to each operand.
2979 SmallVector<CCValAssign, 16> ArgLocs;
2980 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2982 // Allocate shadow area for Win64
2984 CCInfo.AllocateStack(32, 8);
2986 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2988 // Get a count of how many bytes are to be pushed on the stack.
2989 unsigned NumBytes = CCInfo.getNextStackOffset();
2991 // This is a sibcall. The memory operands are available in caller's
2992 // own caller's stack.
2994 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2995 IsTailCallConvention(CallConv))
2996 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2999 if (isTailCall && !IsSibcall && !IsMustTail) {
3000 // Lower arguments at fp - stackoffset + fpdiff.
3001 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3003 FPDiff = NumBytesCallerPushed - NumBytes;
3005 // Set the delta of movement of the returnaddr stackslot.
3006 // But only set if delta is greater than previous delta.
3007 if (FPDiff < X86Info->getTCReturnAddrDelta())
3008 X86Info->setTCReturnAddrDelta(FPDiff);
3011 unsigned NumBytesToPush = NumBytes;
3012 unsigned NumBytesToPop = NumBytes;
3014 // If we have an inalloca argument, all stack space has already been allocated
3015 // for us and be right at the top of the stack. We don't support multiple
3016 // arguments passed in memory when using inalloca.
3017 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3019 if (!ArgLocs.back().isMemLoc())
3020 report_fatal_error("cannot use inalloca attribute on a register "
3022 if (ArgLocs.back().getLocMemOffset() != 0)
3023 report_fatal_error("any parameter with the inalloca attribute must be "
3024 "the only memory argument");
3028 Chain = DAG.getCALLSEQ_START(
3029 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3031 SDValue RetAddrFrIdx;
3032 // Load return address for tail calls.
3033 if (isTailCall && FPDiff)
3034 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3035 Is64Bit, FPDiff, dl);
3037 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3038 SmallVector<SDValue, 8> MemOpChains;
3041 // Walk the register/memloc assignments, inserting copies/loads. In the case
3042 // of tail call optimization arguments are handle later.
3043 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3044 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3045 // Skip inalloca arguments, they have already been written.
3046 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3047 if (Flags.isInAlloca())
3050 CCValAssign &VA = ArgLocs[i];
3051 EVT RegVT = VA.getLocVT();
3052 SDValue Arg = OutVals[i];
3053 bool isByVal = Flags.isByVal();
3055 // Promote the value if needed.
3056 switch (VA.getLocInfo()) {
3057 default: llvm_unreachable("Unknown loc info!");
3058 case CCValAssign::Full: break;
3059 case CCValAssign::SExt:
3060 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3062 case CCValAssign::ZExt:
3063 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3065 case CCValAssign::AExt:
3066 if (Arg.getValueType().isVector() &&
3067 Arg.getValueType().getScalarType() == MVT::i1)
3068 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3069 else if (RegVT.is128BitVector()) {
3070 // Special case: passing MMX values in XMM registers.
3071 Arg = DAG.getBitcast(MVT::i64, Arg);
3072 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3073 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3075 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3077 case CCValAssign::BCvt:
3078 Arg = DAG.getBitcast(RegVT, Arg);
3080 case CCValAssign::Indirect: {
3081 // Store the argument.
3082 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3083 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3084 Chain = DAG.getStore(
3085 Chain, dl, Arg, SpillSlot,
3086 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3093 if (VA.isRegLoc()) {
3094 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3095 if (isVarArg && IsWin64) {
3096 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3097 // shadow reg if callee is a varargs function.
3098 unsigned ShadowReg = 0;
3099 switch (VA.getLocReg()) {
3100 case X86::XMM0: ShadowReg = X86::RCX; break;
3101 case X86::XMM1: ShadowReg = X86::RDX; break;
3102 case X86::XMM2: ShadowReg = X86::R8; break;
3103 case X86::XMM3: ShadowReg = X86::R9; break;
3106 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3108 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3109 assert(VA.isMemLoc());
3110 if (!StackPtr.getNode())
3111 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3112 getPointerTy(DAG.getDataLayout()));
3113 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3114 dl, DAG, VA, Flags));
3118 if (!MemOpChains.empty())
3119 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3121 if (Subtarget->isPICStyleGOT()) {
3122 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3125 RegsToPass.push_back(std::make_pair(
3126 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3127 getPointerTy(DAG.getDataLayout()))));
3129 // If we are tail calling and generating PIC/GOT style code load the
3130 // address of the callee into ECX. The value in ecx is used as target of
3131 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3132 // for tail calls on PIC/GOT architectures. Normally we would just put the
3133 // address of GOT into ebx and then call target@PLT. But for tail calls
3134 // ebx would be restored (since ebx is callee saved) before jumping to the
3137 // Note: The actual moving to ECX is done further down.
3138 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3139 if (G && !G->getGlobal()->hasLocalLinkage() &&
3140 G->getGlobal()->hasDefaultVisibility())
3141 Callee = LowerGlobalAddress(Callee, DAG);
3142 else if (isa<ExternalSymbolSDNode>(Callee))
3143 Callee = LowerExternalSymbol(Callee, DAG);
3147 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3148 // From AMD64 ABI document:
3149 // For calls that may call functions that use varargs or stdargs
3150 // (prototype-less calls or calls to functions containing ellipsis (...) in
3151 // the declaration) %al is used as hidden argument to specify the number
3152 // of SSE registers used. The contents of %al do not need to match exactly
3153 // the number of registers, but must be an ubound on the number of SSE
3154 // registers used and is in the range 0 - 8 inclusive.
3156 // Count the number of XMM registers allocated.
3157 static const MCPhysReg XMMArgRegs[] = {
3158 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3159 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3161 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3162 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3163 && "SSE registers cannot be used when SSE is disabled");
3165 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3166 DAG.getConstant(NumXMMRegs, dl,
3170 if (isVarArg && IsMustTail) {
3171 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3172 for (const auto &F : Forwards) {
3173 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3174 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3178 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3179 // don't need this because the eligibility check rejects calls that require
3180 // shuffling arguments passed in memory.
3181 if (!IsSibcall && isTailCall) {
3182 // Force all the incoming stack arguments to be loaded from the stack
3183 // before any new outgoing arguments are stored to the stack, because the
3184 // outgoing stack slots may alias the incoming argument stack slots, and
3185 // the alias isn't otherwise explicit. This is slightly more conservative
3186 // than necessary, because it means that each store effectively depends
3187 // on every argument instead of just those arguments it would clobber.
3188 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3190 SmallVector<SDValue, 8> MemOpChains2;
3193 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3194 CCValAssign &VA = ArgLocs[i];
3197 assert(VA.isMemLoc());
3198 SDValue Arg = OutVals[i];
3199 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3200 // Skip inalloca arguments. They don't require any work.
3201 if (Flags.isInAlloca())
3203 // Create frame index.
3204 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3205 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3206 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3207 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3209 if (Flags.isByVal()) {
3210 // Copy relative to framepointer.
3211 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3212 if (!StackPtr.getNode())
3213 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3214 getPointerTy(DAG.getDataLayout()));
3215 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3218 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3222 // Store relative to framepointer.
3223 MemOpChains2.push_back(DAG.getStore(
3224 ArgChain, dl, Arg, FIN,
3225 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3230 if (!MemOpChains2.empty())
3231 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3233 // Store the return address to the appropriate stack slot.
3234 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3235 getPointerTy(DAG.getDataLayout()),
3236 RegInfo->getSlotSize(), FPDiff, dl);
3239 // Build a sequence of copy-to-reg nodes chained together with token chain
3240 // and flag operands which copy the outgoing args into registers.
3242 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3243 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3244 RegsToPass[i].second, InFlag);
3245 InFlag = Chain.getValue(1);
3248 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3249 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3250 // In the 64-bit large code model, we have to make all calls
3251 // through a register, since the call instruction's 32-bit
3252 // pc-relative offset may not be large enough to hold the whole
3254 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3255 // If the callee is a GlobalAddress node (quite common, every direct call
3256 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3258 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3260 // We should use extra load for direct calls to dllimported functions in
3262 const GlobalValue *GV = G->getGlobal();
3263 if (!GV->hasDLLImportStorageClass()) {
3264 unsigned char OpFlags = 0;
3265 bool ExtraLoad = false;
3266 unsigned WrapperKind = ISD::DELETED_NODE;
3268 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3269 // external symbols most go through the PLT in PIC mode. If the symbol
3270 // has hidden or protected visibility, or if it is static or local, then
3271 // we don't need to use the PLT - we can directly call it.
3272 if (Subtarget->isTargetELF() &&
3273 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3274 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3275 OpFlags = X86II::MO_PLT;
3276 } else if (Subtarget->isPICStyleStubAny() &&
3277 !GV->isStrongDefinitionForLinker() &&
3278 (!Subtarget->getTargetTriple().isMacOSX() ||
3279 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3280 // PC-relative references to external symbols should go through $stub,
3281 // unless we're building with the leopard linker or later, which
3282 // automatically synthesizes these stubs.
3283 OpFlags = X86II::MO_DARWIN_STUB;
3284 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3285 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3286 // If the function is marked as non-lazy, generate an indirect call
3287 // which loads from the GOT directly. This avoids runtime overhead
3288 // at the cost of eager binding (and one extra byte of encoding).
3289 OpFlags = X86II::MO_GOTPCREL;
3290 WrapperKind = X86ISD::WrapperRIP;
3294 Callee = DAG.getTargetGlobalAddress(
3295 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3297 // Add a wrapper if needed.
3298 if (WrapperKind != ISD::DELETED_NODE)
3299 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3300 getPointerTy(DAG.getDataLayout()), Callee);
3301 // Add extra indirection if needed.
3303 Callee = DAG.getLoad(
3304 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3305 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3308 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3309 unsigned char OpFlags = 0;
3311 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3312 // external symbols should go through the PLT.
3313 if (Subtarget->isTargetELF() &&
3314 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3315 OpFlags = X86II::MO_PLT;
3316 } else if (Subtarget->isPICStyleStubAny() &&
3317 (!Subtarget->getTargetTriple().isMacOSX() ||
3318 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3319 // PC-relative references to external symbols should go through $stub,
3320 // unless we're building with the leopard linker or later, which
3321 // automatically synthesizes these stubs.
3322 OpFlags = X86II::MO_DARWIN_STUB;
3325 Callee = DAG.getTargetExternalSymbol(
3326 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3327 } else if (Subtarget->isTarget64BitILP32() &&
3328 Callee->getValueType(0) == MVT::i32) {
3329 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3330 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3333 // Returns a chain & a flag for retval copy to use.
3334 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3335 SmallVector<SDValue, 8> Ops;
3337 if (!IsSibcall && isTailCall) {
3338 Chain = DAG.getCALLSEQ_END(Chain,
3339 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3340 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3341 InFlag = Chain.getValue(1);
3344 Ops.push_back(Chain);
3345 Ops.push_back(Callee);
3348 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3350 // Add argument registers to the end of the list so that they are known live
3352 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3353 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3354 RegsToPass[i].second.getValueType()));
3356 // Add a register mask operand representing the call-preserved registers.
3357 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3358 assert(Mask && "Missing call preserved mask for calling convention");
3360 // If this is an invoke in a 32-bit function using an MSVC personality, assume
3361 // the function clobbers all registers. If an exception is thrown, the runtime
3362 // will not restore CSRs.
3363 // FIXME: Model this more precisely so that we can register allocate across
3364 // the normal edge and spill and fill across the exceptional edge.
3365 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3366 const Function *CallerFn = MF.getFunction();
3367 EHPersonality Pers =
3368 CallerFn->hasPersonalityFn()
3369 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3370 : EHPersonality::Unknown;
3371 if (isMSVCEHPersonality(Pers))
3372 Mask = RegInfo->getNoPreservedMask();
3375 Ops.push_back(DAG.getRegisterMask(Mask));
3377 if (InFlag.getNode())
3378 Ops.push_back(InFlag);
3382 //// If this is the first return lowered for this function, add the regs
3383 //// to the liveout set for the function.
3384 // This isn't right, although it's probably harmless on x86; liveouts
3385 // should be computed from returns not tail calls. Consider a void
3386 // function making a tail call to a function returning int.
3387 MF.getFrameInfo()->setHasTailCall();
3388 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3391 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3392 InFlag = Chain.getValue(1);
3394 // Create the CALLSEQ_END node.
3395 unsigned NumBytesForCalleeToPop;
3396 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3397 DAG.getTarget().Options.GuaranteedTailCallOpt))
3398 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3399 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3400 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3401 SR == StackStructReturn)
3402 // If this is a call to a struct-return function, the callee
3403 // pops the hidden struct pointer, so we have to push it back.
3404 // This is common for Darwin/X86, Linux & Mingw32 targets.
3405 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3406 NumBytesForCalleeToPop = 4;
3408 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3410 // Returns a flag for retval copy to use.
3412 Chain = DAG.getCALLSEQ_END(Chain,
3413 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3414 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3417 InFlag = Chain.getValue(1);
3420 // Handle result values, copying them out of physregs into vregs that we
3422 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3423 Ins, dl, DAG, InVals);
3426 //===----------------------------------------------------------------------===//
3427 // Fast Calling Convention (tail call) implementation
3428 //===----------------------------------------------------------------------===//
3430 // Like std call, callee cleans arguments, convention except that ECX is
3431 // reserved for storing the tail called function address. Only 2 registers are
3432 // free for argument passing (inreg). Tail call optimization is performed
3434 // * tailcallopt is enabled
3435 // * caller/callee are fastcc
3436 // On X86_64 architecture with GOT-style position independent code only local
3437 // (within module) calls are supported at the moment.
3438 // To keep the stack aligned according to platform abi the function
3439 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3440 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3441 // If a tail called function callee has more arguments than the caller the
3442 // caller needs to make sure that there is room to move the RETADDR to. This is
3443 // achieved by reserving an area the size of the argument delta right after the
3444 // original RETADDR, but before the saved framepointer or the spilled registers
3445 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3457 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3460 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3461 SelectionDAG& DAG) const {
3462 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3463 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3464 unsigned StackAlignment = TFI.getStackAlignment();
3465 uint64_t AlignMask = StackAlignment - 1;
3466 int64_t Offset = StackSize;
3467 unsigned SlotSize = RegInfo->getSlotSize();
3468 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3469 // Number smaller than 12 so just add the difference.
3470 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3472 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3473 Offset = ((~AlignMask) & Offset) + StackAlignment +
3474 (StackAlignment-SlotSize);
3479 /// Return true if the given stack call argument is already available in the
3480 /// same position (relatively) of the caller's incoming argument stack.
3482 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3483 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3484 const X86InstrInfo *TII) {
3485 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3487 if (Arg.getOpcode() == ISD::CopyFromReg) {
3488 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3489 if (!TargetRegisterInfo::isVirtualRegister(VR))
3491 MachineInstr *Def = MRI->getVRegDef(VR);
3494 if (!Flags.isByVal()) {
3495 if (!TII->isLoadFromStackSlot(Def, FI))
3498 unsigned Opcode = Def->getOpcode();
3499 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3500 Opcode == X86::LEA64_32r) &&
3501 Def->getOperand(1).isFI()) {
3502 FI = Def->getOperand(1).getIndex();
3503 Bytes = Flags.getByValSize();
3507 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3508 if (Flags.isByVal())
3509 // ByVal argument is passed in as a pointer but it's now being
3510 // dereferenced. e.g.
3511 // define @foo(%struct.X* %A) {
3512 // tail call @bar(%struct.X* byval %A)
3515 SDValue Ptr = Ld->getBasePtr();
3516 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3519 FI = FINode->getIndex();
3520 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3521 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3522 FI = FINode->getIndex();
3523 Bytes = Flags.getByValSize();
3527 assert(FI != INT_MAX);
3528 if (!MFI->isFixedObjectIndex(FI))
3530 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3533 /// Check whether the call is eligible for tail call optimization. Targets
3534 /// that want to do tail call optimization should implement this function.
3536 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3537 CallingConv::ID CalleeCC,
3539 bool isCalleeStructRet,
3540 bool isCallerStructRet,
3542 const SmallVectorImpl<ISD::OutputArg> &Outs,
3543 const SmallVectorImpl<SDValue> &OutVals,
3544 const SmallVectorImpl<ISD::InputArg> &Ins,
3545 SelectionDAG &DAG) const {
3546 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3549 // If -tailcallopt is specified, make fastcc functions tail-callable.
3550 const MachineFunction &MF = DAG.getMachineFunction();
3551 const Function *CallerF = MF.getFunction();
3553 // If the function return type is x86_fp80 and the callee return type is not,
3554 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3555 // perform a tailcall optimization here.
3556 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3559 CallingConv::ID CallerCC = CallerF->getCallingConv();
3560 bool CCMatch = CallerCC == CalleeCC;
3561 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3562 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3564 // Win64 functions have extra shadow space for argument homing. Don't do the
3565 // sibcall if the caller and callee have mismatched expectations for this
3567 if (IsCalleeWin64 != IsCallerWin64)
3570 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3571 if (IsTailCallConvention(CalleeCC) && CCMatch)
3576 // Look for obvious safe cases to perform tail call optimization that do not
3577 // require ABI changes. This is what gcc calls sibcall.
3579 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3580 // emit a special epilogue.
3581 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3582 if (RegInfo->needsStackRealignment(MF))
3585 // Also avoid sibcall optimization if either caller or callee uses struct
3586 // return semantics.
3587 if (isCalleeStructRet || isCallerStructRet)
3590 // An stdcall/thiscall caller is expected to clean up its arguments; the
3591 // callee isn't going to do that.
3592 // FIXME: this is more restrictive than needed. We could produce a tailcall
3593 // when the stack adjustment matches. For example, with a thiscall that takes
3594 // only one argument.
3595 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3596 CallerCC == CallingConv::X86_ThisCall))
3599 // Do not sibcall optimize vararg calls unless all arguments are passed via
3601 if (isVarArg && !Outs.empty()) {
3603 // Optimizing for varargs on Win64 is unlikely to be safe without
3604 // additional testing.
3605 if (IsCalleeWin64 || IsCallerWin64)
3608 SmallVector<CCValAssign, 16> ArgLocs;
3609 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3612 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3613 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3614 if (!ArgLocs[i].isRegLoc())
3618 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3619 // stack. Therefore, if it's not used by the call it is not safe to optimize
3620 // this into a sibcall.
3621 bool Unused = false;
3622 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3629 SmallVector<CCValAssign, 16> RVLocs;
3630 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3632 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3633 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3634 CCValAssign &VA = RVLocs[i];
3635 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3640 // If the calling conventions do not match, then we'd better make sure the
3641 // results are returned in the same way as what the caller expects.
3643 SmallVector<CCValAssign, 16> RVLocs1;
3644 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3646 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3648 SmallVector<CCValAssign, 16> RVLocs2;
3649 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3651 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3653 if (RVLocs1.size() != RVLocs2.size())
3655 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3656 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3658 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3660 if (RVLocs1[i].isRegLoc()) {
3661 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3664 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3670 // If the callee takes no arguments then go on to check the results of the
3672 if (!Outs.empty()) {
3673 // Check if stack adjustment is needed. For now, do not do this if any
3674 // argument is passed on the stack.
3675 SmallVector<CCValAssign, 16> ArgLocs;
3676 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3679 // Allocate shadow area for Win64
3681 CCInfo.AllocateStack(32, 8);
3683 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3684 if (CCInfo.getNextStackOffset()) {
3685 MachineFunction &MF = DAG.getMachineFunction();
3686 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3689 // Check if the arguments are already laid out in the right way as
3690 // the caller's fixed stack objects.
3691 MachineFrameInfo *MFI = MF.getFrameInfo();
3692 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3693 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3694 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3695 CCValAssign &VA = ArgLocs[i];
3696 SDValue Arg = OutVals[i];
3697 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3698 if (VA.getLocInfo() == CCValAssign::Indirect)
3700 if (!VA.isRegLoc()) {
3701 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3708 // If the tailcall address may be in a register, then make sure it's
3709 // possible to register allocate for it. In 32-bit, the call address can
3710 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3711 // callee-saved registers are restored. These happen to be the same
3712 // registers used to pass 'inreg' arguments so watch out for those.
3713 if (!Subtarget->is64Bit() &&
3714 ((!isa<GlobalAddressSDNode>(Callee) &&
3715 !isa<ExternalSymbolSDNode>(Callee)) ||
3716 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3717 unsigned NumInRegs = 0;
3718 // In PIC we need an extra register to formulate the address computation
3720 unsigned MaxInRegs =
3721 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3723 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3724 CCValAssign &VA = ArgLocs[i];
3727 unsigned Reg = VA.getLocReg();
3730 case X86::EAX: case X86::EDX: case X86::ECX:
3731 if (++NumInRegs == MaxInRegs)
3743 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3744 const TargetLibraryInfo *libInfo) const {
3745 return X86::createFastISel(funcInfo, libInfo);
3748 //===----------------------------------------------------------------------===//
3749 // Other Lowering Hooks
3750 //===----------------------------------------------------------------------===//
3752 static bool MayFoldLoad(SDValue Op) {
3753 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3756 static bool MayFoldIntoStore(SDValue Op) {
3757 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3760 static bool isTargetShuffle(unsigned Opcode) {
3762 default: return false;
3763 case X86ISD::BLENDI:
3764 case X86ISD::PSHUFB:
3765 case X86ISD::PSHUFD:
3766 case X86ISD::PSHUFHW:
3767 case X86ISD::PSHUFLW:
3769 case X86ISD::PALIGNR:
3770 case X86ISD::MOVLHPS:
3771 case X86ISD::MOVLHPD:
3772 case X86ISD::MOVHLPS:
3773 case X86ISD::MOVLPS:
3774 case X86ISD::MOVLPD:
3775 case X86ISD::MOVSHDUP:
3776 case X86ISD::MOVSLDUP:
3777 case X86ISD::MOVDDUP:
3780 case X86ISD::UNPCKL:
3781 case X86ISD::UNPCKH:
3782 case X86ISD::VPERMILPI:
3783 case X86ISD::VPERM2X128:
3784 case X86ISD::VPERMI:
3789 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3790 SDValue V1, unsigned TargetMask,
3791 SelectionDAG &DAG) {
3793 default: llvm_unreachable("Unknown x86 shuffle node");
3794 case X86ISD::PSHUFD:
3795 case X86ISD::PSHUFHW:
3796 case X86ISD::PSHUFLW:
3797 case X86ISD::VPERMILPI:
3798 case X86ISD::VPERMI:
3799 return DAG.getNode(Opc, dl, VT, V1,
3800 DAG.getConstant(TargetMask, dl, MVT::i8));
3804 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3805 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3807 default: llvm_unreachable("Unknown x86 shuffle node");
3808 case X86ISD::MOVLHPS:
3809 case X86ISD::MOVLHPD:
3810 case X86ISD::MOVHLPS:
3811 case X86ISD::MOVLPS:
3812 case X86ISD::MOVLPD:
3815 case X86ISD::UNPCKL:
3816 case X86ISD::UNPCKH:
3817 return DAG.getNode(Opc, dl, VT, V1, V2);
3821 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3822 MachineFunction &MF = DAG.getMachineFunction();
3823 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3824 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3825 int ReturnAddrIndex = FuncInfo->getRAIndex();
3827 if (ReturnAddrIndex == 0) {
3828 // Set up a frame object for the return address.
3829 unsigned SlotSize = RegInfo->getSlotSize();
3830 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3833 FuncInfo->setRAIndex(ReturnAddrIndex);
3836 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3839 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3840 bool hasSymbolicDisplacement) {
3841 // Offset should fit into 32 bit immediate field.
3842 if (!isInt<32>(Offset))
3845 // If we don't have a symbolic displacement - we don't have any extra
3847 if (!hasSymbolicDisplacement)
3850 // FIXME: Some tweaks might be needed for medium code model.
3851 if (M != CodeModel::Small && M != CodeModel::Kernel)
3854 // For small code model we assume that latest object is 16MB before end of 31
3855 // bits boundary. We may also accept pretty large negative constants knowing
3856 // that all objects are in the positive half of address space.
3857 if (M == CodeModel::Small && Offset < 16*1024*1024)
3860 // For kernel code model we know that all object resist in the negative half
3861 // of 32bits address space. We may not accept negative offsets, since they may
3862 // be just off and we may accept pretty large positive ones.
3863 if (M == CodeModel::Kernel && Offset >= 0)
3869 /// Determines whether the callee is required to pop its own arguments.
3870 /// Callee pop is necessary to support tail calls.
3871 bool X86::isCalleePop(CallingConv::ID CallingConv,
3872 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3873 switch (CallingConv) {
3876 case CallingConv::X86_StdCall:
3877 case CallingConv::X86_FastCall:
3878 case CallingConv::X86_ThisCall:
3880 case CallingConv::Fast:
3881 case CallingConv::GHC:
3882 case CallingConv::HiPE:
3889 /// \brief Return true if the condition is an unsigned comparison operation.
3890 static bool isX86CCUnsigned(unsigned X86CC) {
3892 default: llvm_unreachable("Invalid integer condition!");
3893 case X86::COND_E: return true;
3894 case X86::COND_G: return false;
3895 case X86::COND_GE: return false;
3896 case X86::COND_L: return false;
3897 case X86::COND_LE: return false;
3898 case X86::COND_NE: return true;
3899 case X86::COND_B: return true;
3900 case X86::COND_A: return true;
3901 case X86::COND_BE: return true;
3902 case X86::COND_AE: return true;
3904 llvm_unreachable("covered switch fell through?!");
3907 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
3908 /// condition code, returning the condition code and the LHS/RHS of the
3909 /// comparison to make.
3910 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3911 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3913 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3914 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3915 // X > -1 -> X == 0, jump !sign.
3916 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3917 return X86::COND_NS;
3919 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3920 // X < 0 -> X == 0, jump on sign.
3923 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3925 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3926 return X86::COND_LE;
3930 switch (SetCCOpcode) {
3931 default: llvm_unreachable("Invalid integer condition!");
3932 case ISD::SETEQ: return X86::COND_E;
3933 case ISD::SETGT: return X86::COND_G;
3934 case ISD::SETGE: return X86::COND_GE;
3935 case ISD::SETLT: return X86::COND_L;
3936 case ISD::SETLE: return X86::COND_LE;
3937 case ISD::SETNE: return X86::COND_NE;
3938 case ISD::SETULT: return X86::COND_B;
3939 case ISD::SETUGT: return X86::COND_A;
3940 case ISD::SETULE: return X86::COND_BE;
3941 case ISD::SETUGE: return X86::COND_AE;
3945 // First determine if it is required or is profitable to flip the operands.
3947 // If LHS is a foldable load, but RHS is not, flip the condition.
3948 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3949 !ISD::isNON_EXTLoad(RHS.getNode())) {
3950 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3951 std::swap(LHS, RHS);
3954 switch (SetCCOpcode) {
3960 std::swap(LHS, RHS);
3964 // On a floating point condition, the flags are set as follows:
3966 // 0 | 0 | 0 | X > Y
3967 // 0 | 0 | 1 | X < Y
3968 // 1 | 0 | 0 | X == Y
3969 // 1 | 1 | 1 | unordered
3970 switch (SetCCOpcode) {
3971 default: llvm_unreachable("Condcode should be pre-legalized away");
3973 case ISD::SETEQ: return X86::COND_E;
3974 case ISD::SETOLT: // flipped
3976 case ISD::SETGT: return X86::COND_A;
3977 case ISD::SETOLE: // flipped
3979 case ISD::SETGE: return X86::COND_AE;
3980 case ISD::SETUGT: // flipped
3982 case ISD::SETLT: return X86::COND_B;
3983 case ISD::SETUGE: // flipped
3985 case ISD::SETLE: return X86::COND_BE;
3987 case ISD::SETNE: return X86::COND_NE;
3988 case ISD::SETUO: return X86::COND_P;
3989 case ISD::SETO: return X86::COND_NP;
3991 case ISD::SETUNE: return X86::COND_INVALID;
3995 /// Is there a floating point cmov for the specific X86 condition code?
3996 /// Current x86 isa includes the following FP cmov instructions:
3997 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3998 static bool hasFPCMov(unsigned X86CC) {
4014 /// Returns true if the target can instruction select the
4015 /// specified FP immediate natively. If false, the legalizer will
4016 /// materialize the FP immediate as a load from a constant pool.
4017 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4018 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4019 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4025 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4026 ISD::LoadExtType ExtTy,
4028 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4029 // relocation target a movq or addq instruction: don't let the load shrink.
4030 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4031 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4032 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4033 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4037 /// \brief Returns true if it is beneficial to convert a load of a constant
4038 /// to just the constant itself.
4039 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4041 assert(Ty->isIntegerTy());
4043 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4044 if (BitSize == 0 || BitSize > 64)
4049 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4050 unsigned Index) const {
4051 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4054 return (Index == 0 || Index == ResVT.getVectorNumElements());
4057 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4058 // Speculate cttz only if we can directly use TZCNT.
4059 return Subtarget->hasBMI();
4062 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4063 // Speculate ctlz only if we can directly use LZCNT.
4064 return Subtarget->hasLZCNT();
4067 /// Return true if every element in Mask, beginning
4068 /// from position Pos and ending in Pos+Size is undef.
4069 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4070 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4076 /// Return true if Val is undef or if its value falls within the
4077 /// specified range (L, H].
4078 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4079 return (Val < 0) || (Val >= Low && Val < Hi);
4082 /// Val is either less than zero (undef) or equal to the specified value.
4083 static bool isUndefOrEqual(int Val, int CmpVal) {
4084 return (Val < 0 || Val == CmpVal);
4087 /// Return true if every element in Mask, beginning
4088 /// from position Pos and ending in Pos+Size, falls within the specified
4089 /// sequential range (Low, Low+Size]. or is undef.
4090 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4091 unsigned Pos, unsigned Size, int Low) {
4092 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4093 if (!isUndefOrEqual(Mask[i], Low))
4098 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4099 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4100 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4101 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4102 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4105 // The index should be aligned on a vecWidth-bit boundary.
4107 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4109 MVT VT = N->getSimpleValueType(0);
4110 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4111 bool Result = (Index * ElSize) % vecWidth == 0;
4116 /// Return true if the specified INSERT_SUBVECTOR
4117 /// operand specifies a subvector insert that is suitable for input to
4118 /// insertion of 128 or 256-bit subvectors
4119 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4120 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4121 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4123 // The index should be aligned on a vecWidth-bit boundary.
4125 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4127 MVT VT = N->getSimpleValueType(0);
4128 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4129 bool Result = (Index * ElSize) % vecWidth == 0;
4134 bool X86::isVINSERT128Index(SDNode *N) {
4135 return isVINSERTIndex(N, 128);
4138 bool X86::isVINSERT256Index(SDNode *N) {
4139 return isVINSERTIndex(N, 256);
4142 bool X86::isVEXTRACT128Index(SDNode *N) {
4143 return isVEXTRACTIndex(N, 128);
4146 bool X86::isVEXTRACT256Index(SDNode *N) {
4147 return isVEXTRACTIndex(N, 256);
4150 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4151 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4152 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4153 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4156 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4158 MVT VecVT = N->getOperand(0).getSimpleValueType();
4159 MVT ElVT = VecVT.getVectorElementType();
4161 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4162 return Index / NumElemsPerChunk;
4165 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4166 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4167 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4168 llvm_unreachable("Illegal insert subvector for VINSERT");
4171 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4173 MVT VecVT = N->getSimpleValueType(0);
4174 MVT ElVT = VecVT.getVectorElementType();
4176 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4177 return Index / NumElemsPerChunk;
4180 /// Return the appropriate immediate to extract the specified
4181 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4182 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4183 return getExtractVEXTRACTImmediate(N, 128);
4186 /// Return the appropriate immediate to extract the specified
4187 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4188 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4189 return getExtractVEXTRACTImmediate(N, 256);
4192 /// Return the appropriate immediate to insert at the specified
4193 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4194 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4195 return getInsertVINSERTImmediate(N, 128);
4198 /// Return the appropriate immediate to insert at the specified
4199 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4200 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4201 return getInsertVINSERTImmediate(N, 256);
4204 /// Returns true if Elt is a constant integer zero
4205 static bool isZero(SDValue V) {
4206 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4207 return C && C->isNullValue();
4210 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4211 bool X86::isZeroNode(SDValue Elt) {
4214 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4215 return CFP->getValueAPF().isPosZero();
4219 /// Returns a vector of specified type with all zero elements.
4220 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4221 SelectionDAG &DAG, SDLoc dl) {
4222 assert(VT.isVector() && "Expected a vector type");
4224 // Always build SSE zero vectors as <4 x i32> bitcasted
4225 // to their dest type. This ensures they get CSE'd.
4227 if (VT.is128BitVector()) { // SSE
4228 if (Subtarget->hasSSE2()) { // SSE2
4229 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4230 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4232 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4233 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4235 } else if (VT.is256BitVector()) { // AVX
4236 if (Subtarget->hasInt256()) { // AVX2
4237 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4238 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4239 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4241 // 256-bit logic and arithmetic instructions in AVX are all
4242 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4243 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4244 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4245 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4247 } else if (VT.is512BitVector()) { // AVX-512
4248 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4249 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4250 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4251 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4252 } else if (VT.getScalarType() == MVT::i1) {
4254 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4255 && "Unexpected vector type");
4256 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4257 && "Unexpected vector type");
4258 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4259 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4260 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4262 llvm_unreachable("Unexpected vector type");
4264 return DAG.getBitcast(VT, Vec);
4267 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4268 SelectionDAG &DAG, SDLoc dl,
4269 unsigned vectorWidth) {
4270 assert((vectorWidth == 128 || vectorWidth == 256) &&
4271 "Unsupported vector width");
4272 EVT VT = Vec.getValueType();
4273 EVT ElVT = VT.getVectorElementType();
4274 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4275 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4276 VT.getVectorNumElements()/Factor);
4278 // Extract from UNDEF is UNDEF.
4279 if (Vec.getOpcode() == ISD::UNDEF)
4280 return DAG.getUNDEF(ResultVT);
4282 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4283 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4285 // This is the index of the first element of the vectorWidth-bit chunk
4287 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4290 // If the input is a buildvector just emit a smaller one.
4291 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4292 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4293 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4296 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4297 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4300 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4301 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4302 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4303 /// instructions or a simple subregister reference. Idx is an index in the
4304 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4305 /// lowering EXTRACT_VECTOR_ELT operations easier.
4306 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4307 SelectionDAG &DAG, SDLoc dl) {
4308 assert((Vec.getValueType().is256BitVector() ||
4309 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4310 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4313 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4314 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4315 SelectionDAG &DAG, SDLoc dl) {
4316 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4317 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4320 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4321 unsigned IdxVal, SelectionDAG &DAG,
4322 SDLoc dl, unsigned vectorWidth) {
4323 assert((vectorWidth == 128 || vectorWidth == 256) &&
4324 "Unsupported vector width");
4325 // Inserting UNDEF is Result
4326 if (Vec.getOpcode() == ISD::UNDEF)
4328 EVT VT = Vec.getValueType();
4329 EVT ElVT = VT.getVectorElementType();
4330 EVT ResultVT = Result.getValueType();
4332 // Insert the relevant vectorWidth bits.
4333 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4335 // This is the index of the first element of the vectorWidth-bit chunk
4337 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4340 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4341 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4344 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4345 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4346 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4347 /// simple superregister reference. Idx is an index in the 128 bits
4348 /// we want. It need not be aligned to a 128-bit boundary. That makes
4349 /// lowering INSERT_VECTOR_ELT operations easier.
4350 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4351 SelectionDAG &DAG, SDLoc dl) {
4352 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4354 // For insertion into the zero index (low half) of a 256-bit vector, it is
4355 // more efficient to generate a blend with immediate instead of an insert*128.
4356 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4357 // extend the subvector to the size of the result vector. Make sure that
4358 // we are not recursing on that node by checking for undef here.
4359 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4360 Result.getOpcode() != ISD::UNDEF) {
4361 EVT ResultVT = Result.getValueType();
4362 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4363 SDValue Undef = DAG.getUNDEF(ResultVT);
4364 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4367 // The blend instruction, and therefore its mask, depend on the data type.
4368 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4369 if (ScalarType.isFloatingPoint()) {
4370 // Choose either vblendps (float) or vblendpd (double).
4371 unsigned ScalarSize = ScalarType.getSizeInBits();
4372 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4373 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4374 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4375 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4378 const X86Subtarget &Subtarget =
4379 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4381 // AVX2 is needed for 256-bit integer blend support.
4382 // Integers must be cast to 32-bit because there is only vpblendd;
4383 // vpblendw can't be used for this because it has a handicapped mask.
4385 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4386 // is still more efficient than using the wrong domain vinsertf128 that
4387 // will be created by InsertSubVector().
4388 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4390 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4391 Vec256 = DAG.getBitcast(CastVT, Vec256);
4392 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4393 return DAG.getBitcast(ResultVT, Vec256);
4396 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4399 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4400 SelectionDAG &DAG, SDLoc dl) {
4401 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4402 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4405 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4406 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4407 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4408 /// large BUILD_VECTORS.
4409 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4410 unsigned NumElems, SelectionDAG &DAG,
4412 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4413 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4416 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4417 unsigned NumElems, SelectionDAG &DAG,
4419 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4420 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4423 /// Returns a vector of specified type with all bits set.
4424 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4425 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4426 /// Then bitcast to their original type, ensuring they get CSE'd.
4427 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4429 assert(VT.isVector() && "Expected a vector type");
4431 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4433 if (VT.is256BitVector()) {
4434 if (HasInt256) { // AVX2
4435 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4436 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4438 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4439 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4441 } else if (VT.is128BitVector()) {
4442 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4444 llvm_unreachable("Unexpected vector type");
4446 return DAG.getBitcast(VT, Vec);
4449 /// Returns a vector_shuffle node for an unpackl operation.
4450 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4452 unsigned NumElems = VT.getVectorNumElements();
4453 SmallVector<int, 8> Mask;
4454 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4456 Mask.push_back(i + NumElems);
4458 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4461 /// Returns a vector_shuffle node for an unpackh operation.
4462 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4464 unsigned NumElems = VT.getVectorNumElements();
4465 SmallVector<int, 8> Mask;
4466 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4467 Mask.push_back(i + Half);
4468 Mask.push_back(i + NumElems + Half);
4470 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4473 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4474 /// This produces a shuffle where the low element of V2 is swizzled into the
4475 /// zero/undef vector, landing at element Idx.
4476 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4477 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4479 const X86Subtarget *Subtarget,
4480 SelectionDAG &DAG) {
4481 MVT VT = V2.getSimpleValueType();
4483 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4484 unsigned NumElems = VT.getVectorNumElements();
4485 SmallVector<int, 16> MaskVec;
4486 for (unsigned i = 0; i != NumElems; ++i)
4487 // If this is the insertion idx, put the low elt of V2 here.
4488 MaskVec.push_back(i == Idx ? NumElems : i);
4489 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4492 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4493 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4494 /// uses one source. Note that this will set IsUnary for shuffles which use a
4495 /// single input multiple times, and in those cases it will
4496 /// adjust the mask to only have indices within that single input.
4497 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4498 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4499 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4500 unsigned NumElems = VT.getVectorNumElements();
4504 bool IsFakeUnary = false;
4505 switch(N->getOpcode()) {
4506 case X86ISD::BLENDI:
4507 ImmN = N->getOperand(N->getNumOperands()-1);
4508 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4511 ImmN = N->getOperand(N->getNumOperands()-1);
4512 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4513 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4515 case X86ISD::UNPCKH:
4516 DecodeUNPCKHMask(VT, Mask);
4517 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4519 case X86ISD::UNPCKL:
4520 DecodeUNPCKLMask(VT, Mask);
4521 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4523 case X86ISD::MOVHLPS:
4524 DecodeMOVHLPSMask(NumElems, Mask);
4525 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4527 case X86ISD::MOVLHPS:
4528 DecodeMOVLHPSMask(NumElems, Mask);
4529 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4531 case X86ISD::PALIGNR:
4532 ImmN = N->getOperand(N->getNumOperands()-1);
4533 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4535 case X86ISD::PSHUFD:
4536 case X86ISD::VPERMILPI:
4537 ImmN = N->getOperand(N->getNumOperands()-1);
4538 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4541 case X86ISD::PSHUFHW:
4542 ImmN = N->getOperand(N->getNumOperands()-1);
4543 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4546 case X86ISD::PSHUFLW:
4547 ImmN = N->getOperand(N->getNumOperands()-1);
4548 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4551 case X86ISD::PSHUFB: {
4553 SDValue MaskNode = N->getOperand(1);
4554 while (MaskNode->getOpcode() == ISD::BITCAST)
4555 MaskNode = MaskNode->getOperand(0);
4557 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4558 // If we have a build-vector, then things are easy.
4559 EVT VT = MaskNode.getValueType();
4560 assert(VT.isVector() &&
4561 "Can't produce a non-vector with a build_vector!");
4562 if (!VT.isInteger())
4565 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4567 SmallVector<uint64_t, 32> RawMask;
4568 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4569 SDValue Op = MaskNode->getOperand(i);
4570 if (Op->getOpcode() == ISD::UNDEF) {
4571 RawMask.push_back((uint64_t)SM_SentinelUndef);
4574 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4577 APInt MaskElement = CN->getAPIntValue();
4579 // We now have to decode the element which could be any integer size and
4580 // extract each byte of it.
4581 for (int j = 0; j < NumBytesPerElement; ++j) {
4582 // Note that this is x86 and so always little endian: the low byte is
4583 // the first byte of the mask.
4584 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4585 MaskElement = MaskElement.lshr(8);
4588 DecodePSHUFBMask(RawMask, Mask);
4592 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4596 SDValue Ptr = MaskLoad->getBasePtr();
4597 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4598 Ptr->getOpcode() == X86ISD::WrapperRIP)
4599 Ptr = Ptr->getOperand(0);
4601 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4602 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4605 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4606 DecodePSHUFBMask(C, Mask);
4614 case X86ISD::VPERMI:
4615 ImmN = N->getOperand(N->getNumOperands()-1);
4616 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4621 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4623 case X86ISD::VPERM2X128:
4624 ImmN = N->getOperand(N->getNumOperands()-1);
4625 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4626 if (Mask.empty()) return false;
4627 // Mask only contains negative index if an element is zero.
4628 if (std::any_of(Mask.begin(), Mask.end(),
4629 [](int M){ return M == SM_SentinelZero; }))
4632 case X86ISD::MOVSLDUP:
4633 DecodeMOVSLDUPMask(VT, Mask);
4636 case X86ISD::MOVSHDUP:
4637 DecodeMOVSHDUPMask(VT, Mask);
4640 case X86ISD::MOVDDUP:
4641 DecodeMOVDDUPMask(VT, Mask);
4644 case X86ISD::MOVLHPD:
4645 case X86ISD::MOVLPD:
4646 case X86ISD::MOVLPS:
4647 // Not yet implemented
4649 default: llvm_unreachable("unknown target shuffle node");
4652 // If we have a fake unary shuffle, the shuffle mask is spread across two
4653 // inputs that are actually the same node. Re-map the mask to always point
4654 // into the first input.
4657 if (M >= (int)Mask.size())
4663 /// Returns the scalar element that will make up the ith
4664 /// element of the result of the vector shuffle.
4665 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4668 return SDValue(); // Limit search depth.
4670 SDValue V = SDValue(N, 0);
4671 EVT VT = V.getValueType();
4672 unsigned Opcode = V.getOpcode();
4674 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4675 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4676 int Elt = SV->getMaskElt(Index);
4679 return DAG.getUNDEF(VT.getVectorElementType());
4681 unsigned NumElems = VT.getVectorNumElements();
4682 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4683 : SV->getOperand(1);
4684 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4687 // Recurse into target specific vector shuffles to find scalars.
4688 if (isTargetShuffle(Opcode)) {
4689 MVT ShufVT = V.getSimpleValueType();
4690 unsigned NumElems = ShufVT.getVectorNumElements();
4691 SmallVector<int, 16> ShuffleMask;
4694 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4697 int Elt = ShuffleMask[Index];
4699 return DAG.getUNDEF(ShufVT.getVectorElementType());
4701 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4703 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4707 // Actual nodes that may contain scalar elements
4708 if (Opcode == ISD::BITCAST) {
4709 V = V.getOperand(0);
4710 EVT SrcVT = V.getValueType();
4711 unsigned NumElems = VT.getVectorNumElements();
4713 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4717 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4718 return (Index == 0) ? V.getOperand(0)
4719 : DAG.getUNDEF(VT.getVectorElementType());
4721 if (V.getOpcode() == ISD::BUILD_VECTOR)
4722 return V.getOperand(Index);
4727 /// Custom lower build_vector of v16i8.
4728 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4729 unsigned NumNonZero, unsigned NumZero,
4731 const X86Subtarget* Subtarget,
4732 const TargetLowering &TLI) {
4740 // SSE4.1 - use PINSRB to insert each byte directly.
4741 if (Subtarget->hasSSE41()) {
4742 for (unsigned i = 0; i < 16; ++i) {
4743 bool isNonZero = (NonZeros & (1 << i)) != 0;
4747 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4749 V = DAG.getUNDEF(MVT::v16i8);
4752 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4753 MVT::v16i8, V, Op.getOperand(i),
4754 DAG.getIntPtrConstant(i, dl));
4761 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4762 for (unsigned i = 0; i < 16; ++i) {
4763 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4764 if (ThisIsNonZero && First) {
4766 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4768 V = DAG.getUNDEF(MVT::v8i16);
4773 SDValue ThisElt, LastElt;
4774 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4775 if (LastIsNonZero) {
4776 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4777 MVT::i16, Op.getOperand(i-1));
4779 if (ThisIsNonZero) {
4780 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4781 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4782 ThisElt, DAG.getConstant(8, dl, MVT::i8));
4784 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4788 if (ThisElt.getNode())
4789 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4790 DAG.getIntPtrConstant(i/2, dl));
4794 return DAG.getBitcast(MVT::v16i8, V);
4797 /// Custom lower build_vector of v8i16.
4798 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4799 unsigned NumNonZero, unsigned NumZero,
4801 const X86Subtarget* Subtarget,
4802 const TargetLowering &TLI) {
4809 for (unsigned i = 0; i < 8; ++i) {
4810 bool isNonZero = (NonZeros & (1 << i)) != 0;
4814 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4816 V = DAG.getUNDEF(MVT::v8i16);
4819 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4820 MVT::v8i16, V, Op.getOperand(i),
4821 DAG.getIntPtrConstant(i, dl));
4828 /// Custom lower build_vector of v4i32 or v4f32.
4829 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4830 const X86Subtarget *Subtarget,
4831 const TargetLowering &TLI) {
4832 // Find all zeroable elements.
4833 std::bitset<4> Zeroable;
4834 for (int i=0; i < 4; ++i) {
4835 SDValue Elt = Op->getOperand(i);
4836 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4838 assert(Zeroable.size() - Zeroable.count() > 1 &&
4839 "We expect at least two non-zero elements!");
4841 // We only know how to deal with build_vector nodes where elements are either
4842 // zeroable or extract_vector_elt with constant index.
4843 SDValue FirstNonZero;
4844 unsigned FirstNonZeroIdx;
4845 for (unsigned i=0; i < 4; ++i) {
4848 SDValue Elt = Op->getOperand(i);
4849 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4850 !isa<ConstantSDNode>(Elt.getOperand(1)))
4852 // Make sure that this node is extracting from a 128-bit vector.
4853 MVT VT = Elt.getOperand(0).getSimpleValueType();
4854 if (!VT.is128BitVector())
4856 if (!FirstNonZero.getNode()) {
4858 FirstNonZeroIdx = i;
4862 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
4863 SDValue V1 = FirstNonZero.getOperand(0);
4864 MVT VT = V1.getSimpleValueType();
4866 // See if this build_vector can be lowered as a blend with zero.
4868 unsigned EltMaskIdx, EltIdx;
4870 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
4871 if (Zeroable[EltIdx]) {
4872 // The zero vector will be on the right hand side.
4873 Mask[EltIdx] = EltIdx+4;
4877 Elt = Op->getOperand(EltIdx);
4878 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
4879 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
4880 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
4882 Mask[EltIdx] = EltIdx;
4886 // Let the shuffle legalizer deal with blend operations.
4887 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
4888 if (V1.getSimpleValueType() != VT)
4889 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
4890 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
4893 // See if we can lower this build_vector to a INSERTPS.
4894 if (!Subtarget->hasSSE41())
4897 SDValue V2 = Elt.getOperand(0);
4898 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
4901 bool CanFold = true;
4902 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
4906 SDValue Current = Op->getOperand(i);
4907 SDValue SrcVector = Current->getOperand(0);
4910 CanFold = SrcVector == V1 &&
4911 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
4917 assert(V1.getNode() && "Expected at least two non-zero elements!");
4918 if (V1.getSimpleValueType() != MVT::v4f32)
4919 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
4920 if (V2.getSimpleValueType() != MVT::v4f32)
4921 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
4923 // Ok, we can emit an INSERTPS instruction.
4924 unsigned ZMask = Zeroable.to_ulong();
4926 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
4927 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
4929 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
4930 DAG.getIntPtrConstant(InsertPSMask, DL));
4931 return DAG.getBitcast(VT, Result);
4934 /// Return a vector logical shift node.
4935 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4936 unsigned NumBits, SelectionDAG &DAG,
4937 const TargetLowering &TLI, SDLoc dl) {
4938 assert(VT.is128BitVector() && "Unknown type for VShift");
4939 MVT ShVT = MVT::v2i64;
4940 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4941 SrcOp = DAG.getBitcast(ShVT, SrcOp);
4942 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
4943 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
4944 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
4945 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
4949 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
4951 // Check if the scalar load can be widened into a vector load. And if
4952 // the address is "base + cst" see if the cst can be "absorbed" into
4953 // the shuffle mask.
4954 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4955 SDValue Ptr = LD->getBasePtr();
4956 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4958 EVT PVT = LD->getValueType(0);
4959 if (PVT != MVT::i32 && PVT != MVT::f32)
4964 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4965 FI = FINode->getIndex();
4967 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4968 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4969 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4970 Offset = Ptr.getConstantOperandVal(1);
4971 Ptr = Ptr.getOperand(0);
4976 // FIXME: 256-bit vector instructions don't require a strict alignment,
4977 // improve this code to support it better.
4978 unsigned RequiredAlign = VT.getSizeInBits()/8;
4979 SDValue Chain = LD->getChain();
4980 // Make sure the stack object alignment is at least 16 or 32.
4981 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4982 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4983 if (MFI->isFixedObjectIndex(FI)) {
4984 // Can't change the alignment. FIXME: It's possible to compute
4985 // the exact stack offset and reference FI + adjust offset instead.
4986 // If someone *really* cares about this. That's the way to implement it.
4989 MFI->setObjectAlignment(FI, RequiredAlign);
4993 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4994 // Ptr + (Offset & ~15).
4997 if ((Offset % RequiredAlign) & 3)
4999 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5002 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5003 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5006 int EltNo = (Offset - StartOffset) >> 2;
5007 unsigned NumElems = VT.getVectorNumElements();
5009 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5010 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5011 LD->getPointerInfo().getWithOffset(StartOffset),
5012 false, false, false, 0);
5014 SmallVector<int, 8> Mask(NumElems, EltNo);
5016 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5022 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5023 /// elements can be replaced by a single large load which has the same value as
5024 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5026 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5028 /// FIXME: we'd also like to handle the case where the last elements are zero
5029 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5030 /// There's even a handy isZeroNode for that purpose.
5031 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5032 SDLoc &DL, SelectionDAG &DAG,
5033 bool isAfterLegalize) {
5034 unsigned NumElems = Elts.size();
5036 LoadSDNode *LDBase = nullptr;
5037 unsigned LastLoadedElt = -1U;
5039 // For each element in the initializer, see if we've found a load or an undef.
5040 // If we don't find an initial load element, or later load elements are
5041 // non-consecutive, bail out.
5042 for (unsigned i = 0; i < NumElems; ++i) {
5043 SDValue Elt = Elts[i];
5044 // Look through a bitcast.
5045 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5046 Elt = Elt.getOperand(0);
5047 if (!Elt.getNode() ||
5048 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5051 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5053 LDBase = cast<LoadSDNode>(Elt.getNode());
5057 if (Elt.getOpcode() == ISD::UNDEF)
5060 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5061 EVT LdVT = Elt.getValueType();
5062 // Each loaded element must be the correct fractional portion of the
5063 // requested vector load.
5064 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5066 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5071 // If we have found an entire vector of loads and undefs, then return a large
5072 // load of the entire vector width starting at the base pointer. If we found
5073 // consecutive loads for the low half, generate a vzext_load node.
5074 if (LastLoadedElt == NumElems - 1) {
5075 assert(LDBase && "Did not find base load for merging consecutive loads");
5076 EVT EltVT = LDBase->getValueType(0);
5077 // Ensure that the input vector size for the merged loads matches the
5078 // cumulative size of the input elements.
5079 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5082 if (isAfterLegalize &&
5083 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5086 SDValue NewLd = SDValue();
5088 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5089 LDBase->getPointerInfo(), LDBase->isVolatile(),
5090 LDBase->isNonTemporal(), LDBase->isInvariant(),
5091 LDBase->getAlignment());
5093 if (LDBase->hasAnyUseOfValue(1)) {
5094 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5096 SDValue(NewLd.getNode(), 1));
5097 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5098 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5099 SDValue(NewLd.getNode(), 1));
5105 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5106 //of a v4i32 / v4f32. It's probably worth generalizing.
5107 EVT EltVT = VT.getVectorElementType();
5108 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5109 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5110 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5111 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5113 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5114 LDBase->getPointerInfo(),
5115 LDBase->getAlignment(),
5116 false/*isVolatile*/, true/*ReadMem*/,
5119 // Make sure the newly-created LOAD is in the same position as LDBase in
5120 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5121 // update uses of LDBase's output chain to use the TokenFactor.
5122 if (LDBase->hasAnyUseOfValue(1)) {
5123 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5124 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5125 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5126 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5127 SDValue(ResNode.getNode(), 1));
5130 return DAG.getBitcast(VT, ResNode);
5135 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5136 /// to generate a splat value for the following cases:
5137 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5138 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5139 /// a scalar load, or a constant.
5140 /// The VBROADCAST node is returned when a pattern is found,
5141 /// or SDValue() otherwise.
5142 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5143 SelectionDAG &DAG) {
5144 // VBROADCAST requires AVX.
5145 // TODO: Splats could be generated for non-AVX CPUs using SSE
5146 // instructions, but there's less potential gain for only 128-bit vectors.
5147 if (!Subtarget->hasAVX())
5150 MVT VT = Op.getSimpleValueType();
5153 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5154 "Unsupported vector type for broadcast.");
5159 switch (Op.getOpcode()) {
5161 // Unknown pattern found.
5164 case ISD::BUILD_VECTOR: {
5165 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5166 BitVector UndefElements;
5167 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5169 // We need a splat of a single value to use broadcast, and it doesn't
5170 // make any sense if the value is only in one element of the vector.
5171 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5175 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5176 Ld.getOpcode() == ISD::ConstantFP);
5178 // Make sure that all of the users of a non-constant load are from the
5179 // BUILD_VECTOR node.
5180 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5185 case ISD::VECTOR_SHUFFLE: {
5186 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5188 // Shuffles must have a splat mask where the first element is
5190 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5193 SDValue Sc = Op.getOperand(0);
5194 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5195 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5197 if (!Subtarget->hasInt256())
5200 // Use the register form of the broadcast instruction available on AVX2.
5201 if (VT.getSizeInBits() >= 256)
5202 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5203 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5206 Ld = Sc.getOperand(0);
5207 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5208 Ld.getOpcode() == ISD::ConstantFP);
5210 // The scalar_to_vector node and the suspected
5211 // load node must have exactly one user.
5212 // Constants may have multiple users.
5214 // AVX-512 has register version of the broadcast
5215 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5216 Ld.getValueType().getSizeInBits() >= 32;
5217 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5224 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5225 bool IsGE256 = (VT.getSizeInBits() >= 256);
5227 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5228 // instruction to save 8 or more bytes of constant pool data.
5229 // TODO: If multiple splats are generated to load the same constant,
5230 // it may be detrimental to overall size. There needs to be a way to detect
5231 // that condition to know if this is truly a size win.
5232 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5234 // Handle broadcasting a single constant scalar from the constant pool
5236 // On Sandybridge (no AVX2), it is still better to load a constant vector
5237 // from the constant pool and not to broadcast it from a scalar.
5238 // But override that restriction when optimizing for size.
5239 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5240 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5241 EVT CVT = Ld.getValueType();
5242 assert(!CVT.isVector() && "Must not broadcast a vector type");
5244 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5245 // For size optimization, also splat v2f64 and v2i64, and for size opt
5246 // with AVX2, also splat i8 and i16.
5247 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5248 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5249 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5250 const Constant *C = nullptr;
5251 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5252 C = CI->getConstantIntValue();
5253 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5254 C = CF->getConstantFPValue();
5256 assert(C && "Invalid constant type");
5258 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5260 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5261 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5263 CVT, dl, DAG.getEntryNode(), CP,
5264 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5265 false, false, Alignment);
5267 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5271 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5273 // Handle AVX2 in-register broadcasts.
5274 if (!IsLoad && Subtarget->hasInt256() &&
5275 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5276 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5278 // The scalar source must be a normal load.
5282 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5283 (Subtarget->hasVLX() && ScalarSize == 64))
5284 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5286 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5287 // double since there is no vbroadcastsd xmm
5288 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5289 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5290 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5293 // Unsupported broadcast.
5297 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5298 /// underlying vector and index.
5300 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5302 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5304 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5305 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5308 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5310 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5312 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5313 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5316 // In this case the vector is the extract_subvector expression and the index
5317 // is 2, as specified by the shuffle.
5318 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5319 SDValue ShuffleVec = SVOp->getOperand(0);
5320 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5321 assert(ShuffleVecVT.getVectorElementType() ==
5322 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5324 int ShuffleIdx = SVOp->getMaskElt(Idx);
5325 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5326 ExtractedFromVec = ShuffleVec;
5332 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5333 MVT VT = Op.getSimpleValueType();
5335 // Skip if insert_vec_elt is not supported.
5336 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5337 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5341 unsigned NumElems = Op.getNumOperands();
5345 SmallVector<unsigned, 4> InsertIndices;
5346 SmallVector<int, 8> Mask(NumElems, -1);
5348 for (unsigned i = 0; i != NumElems; ++i) {
5349 unsigned Opc = Op.getOperand(i).getOpcode();
5351 if (Opc == ISD::UNDEF)
5354 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5355 // Quit if more than 1 elements need inserting.
5356 if (InsertIndices.size() > 1)
5359 InsertIndices.push_back(i);
5363 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5364 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5365 // Quit if non-constant index.
5366 if (!isa<ConstantSDNode>(ExtIdx))
5368 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5370 // Quit if extracted from vector of different type.
5371 if (ExtractedFromVec.getValueType() != VT)
5374 if (!VecIn1.getNode())
5375 VecIn1 = ExtractedFromVec;
5376 else if (VecIn1 != ExtractedFromVec) {
5377 if (!VecIn2.getNode())
5378 VecIn2 = ExtractedFromVec;
5379 else if (VecIn2 != ExtractedFromVec)
5380 // Quit if more than 2 vectors to shuffle
5384 if (ExtractedFromVec == VecIn1)
5386 else if (ExtractedFromVec == VecIn2)
5387 Mask[i] = Idx + NumElems;
5390 if (!VecIn1.getNode())
5393 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5394 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5395 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5396 unsigned Idx = InsertIndices[i];
5397 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5398 DAG.getIntPtrConstant(Idx, DL));
5404 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5405 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5406 Op.getScalarValueSizeInBits() == 1 &&
5407 "Can not convert non-constant vector");
5408 uint64_t Immediate = 0;
5409 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5410 SDValue In = Op.getOperand(idx);
5411 if (In.getOpcode() != ISD::UNDEF)
5412 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5416 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5417 return DAG.getConstant(Immediate, dl, VT);
5419 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5421 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5423 MVT VT = Op.getSimpleValueType();
5424 assert((VT.getVectorElementType() == MVT::i1) &&
5425 "Unexpected type in LowerBUILD_VECTORvXi1!");
5428 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5429 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5430 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5431 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5434 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5435 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5436 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5437 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5440 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5441 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5442 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5443 return DAG.getBitcast(VT, Imm);
5444 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5445 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5446 DAG.getIntPtrConstant(0, dl));
5449 // Vector has one or more non-const elements
5450 uint64_t Immediate = 0;
5451 SmallVector<unsigned, 16> NonConstIdx;
5452 bool IsSplat = true;
5453 bool HasConstElts = false;
5455 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5456 SDValue In = Op.getOperand(idx);
5457 if (In.getOpcode() == ISD::UNDEF)
5459 if (!isa<ConstantSDNode>(In))
5460 NonConstIdx.push_back(idx);
5462 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5463 HasConstElts = true;
5467 else if (In != Op.getOperand(SplatIdx))
5471 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5473 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5474 DAG.getConstant(1, dl, VT),
5475 DAG.getConstant(0, dl, VT));
5477 // insert elements one by one
5481 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5482 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5484 else if (HasConstElts)
5485 Imm = DAG.getConstant(0, dl, VT);
5487 Imm = DAG.getUNDEF(VT);
5488 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5489 DstVec = DAG.getBitcast(VT, Imm);
5491 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5492 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5493 DAG.getIntPtrConstant(0, dl));
5496 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5497 unsigned InsertIdx = NonConstIdx[i];
5498 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5499 Op.getOperand(InsertIdx),
5500 DAG.getIntPtrConstant(InsertIdx, dl));
5505 /// \brief Return true if \p N implements a horizontal binop and return the
5506 /// operands for the horizontal binop into V0 and V1.
5508 /// This is a helper function of LowerToHorizontalOp().
5509 /// This function checks that the build_vector \p N in input implements a
5510 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5511 /// operation to match.
5512 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5513 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5514 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5517 /// This function only analyzes elements of \p N whose indices are
5518 /// in range [BaseIdx, LastIdx).
5519 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5521 unsigned BaseIdx, unsigned LastIdx,
5522 SDValue &V0, SDValue &V1) {
5523 EVT VT = N->getValueType(0);
5525 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5526 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5527 "Invalid Vector in input!");
5529 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5530 bool CanFold = true;
5531 unsigned ExpectedVExtractIdx = BaseIdx;
5532 unsigned NumElts = LastIdx - BaseIdx;
5533 V0 = DAG.getUNDEF(VT);
5534 V1 = DAG.getUNDEF(VT);
5536 // Check if N implements a horizontal binop.
5537 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5538 SDValue Op = N->getOperand(i + BaseIdx);
5541 if (Op->getOpcode() == ISD::UNDEF) {
5542 // Update the expected vector extract index.
5543 if (i * 2 == NumElts)
5544 ExpectedVExtractIdx = BaseIdx;
5545 ExpectedVExtractIdx += 2;
5549 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5554 SDValue Op0 = Op.getOperand(0);
5555 SDValue Op1 = Op.getOperand(1);
5557 // Try to match the following pattern:
5558 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5559 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5560 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5561 Op0.getOperand(0) == Op1.getOperand(0) &&
5562 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5563 isa<ConstantSDNode>(Op1.getOperand(1)));
5567 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5568 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5570 if (i * 2 < NumElts) {
5571 if (V0.getOpcode() == ISD::UNDEF) {
5572 V0 = Op0.getOperand(0);
5573 if (V0.getValueType() != VT)
5577 if (V1.getOpcode() == ISD::UNDEF) {
5578 V1 = Op0.getOperand(0);
5579 if (V1.getValueType() != VT)
5582 if (i * 2 == NumElts)
5583 ExpectedVExtractIdx = BaseIdx;
5586 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5587 if (I0 == ExpectedVExtractIdx)
5588 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5589 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5590 // Try to match the following dag sequence:
5591 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5592 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5596 ExpectedVExtractIdx += 2;
5602 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5603 /// a concat_vector.
5605 /// This is a helper function of LowerToHorizontalOp().
5606 /// This function expects two 256-bit vectors called V0 and V1.
5607 /// At first, each vector is split into two separate 128-bit vectors.
5608 /// Then, the resulting 128-bit vectors are used to implement two
5609 /// horizontal binary operations.
5611 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5613 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5614 /// the two new horizontal binop.
5615 /// When Mode is set, the first horizontal binop dag node would take as input
5616 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5617 /// horizontal binop dag node would take as input the lower 128-bit of V1
5618 /// and the upper 128-bit of V1.
5620 /// HADD V0_LO, V0_HI
5621 /// HADD V1_LO, V1_HI
5623 /// Otherwise, the first horizontal binop dag node takes as input the lower
5624 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5625 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5627 /// HADD V0_LO, V1_LO
5628 /// HADD V0_HI, V1_HI
5630 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5631 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5632 /// the upper 128-bits of the result.
5633 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5634 SDLoc DL, SelectionDAG &DAG,
5635 unsigned X86Opcode, bool Mode,
5636 bool isUndefLO, bool isUndefHI) {
5637 EVT VT = V0.getValueType();
5638 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5639 "Invalid nodes in input!");
5641 unsigned NumElts = VT.getVectorNumElements();
5642 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5643 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5644 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5645 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5646 EVT NewVT = V0_LO.getValueType();
5648 SDValue LO = DAG.getUNDEF(NewVT);
5649 SDValue HI = DAG.getUNDEF(NewVT);
5652 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5653 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5654 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5655 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5656 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5658 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5659 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5660 V1_LO->getOpcode() != ISD::UNDEF))
5661 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5663 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5664 V1_HI->getOpcode() != ISD::UNDEF))
5665 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5668 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5671 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5673 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5674 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5675 EVT VT = BV->getValueType(0);
5676 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5677 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5681 unsigned NumElts = VT.getVectorNumElements();
5682 SDValue InVec0 = DAG.getUNDEF(VT);
5683 SDValue InVec1 = DAG.getUNDEF(VT);
5685 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5686 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5688 // Odd-numbered elements in the input build vector are obtained from
5689 // adding two integer/float elements.
5690 // Even-numbered elements in the input build vector are obtained from
5691 // subtracting two integer/float elements.
5692 unsigned ExpectedOpcode = ISD::FSUB;
5693 unsigned NextExpectedOpcode = ISD::FADD;
5694 bool AddFound = false;
5695 bool SubFound = false;
5697 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5698 SDValue Op = BV->getOperand(i);
5700 // Skip 'undef' values.
5701 unsigned Opcode = Op.getOpcode();
5702 if (Opcode == ISD::UNDEF) {
5703 std::swap(ExpectedOpcode, NextExpectedOpcode);
5707 // Early exit if we found an unexpected opcode.
5708 if (Opcode != ExpectedOpcode)
5711 SDValue Op0 = Op.getOperand(0);
5712 SDValue Op1 = Op.getOperand(1);
5714 // Try to match the following pattern:
5715 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5716 // Early exit if we cannot match that sequence.
5717 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5718 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5719 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5720 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5721 Op0.getOperand(1) != Op1.getOperand(1))
5724 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5728 // We found a valid add/sub node. Update the information accordingly.
5734 // Update InVec0 and InVec1.
5735 if (InVec0.getOpcode() == ISD::UNDEF) {
5736 InVec0 = Op0.getOperand(0);
5737 if (InVec0.getValueType() != VT)
5740 if (InVec1.getOpcode() == ISD::UNDEF) {
5741 InVec1 = Op1.getOperand(0);
5742 if (InVec1.getValueType() != VT)
5746 // Make sure that operands in input to each add/sub node always
5747 // come from a same pair of vectors.
5748 if (InVec0 != Op0.getOperand(0)) {
5749 if (ExpectedOpcode == ISD::FSUB)
5752 // FADD is commutable. Try to commute the operands
5753 // and then test again.
5754 std::swap(Op0, Op1);
5755 if (InVec0 != Op0.getOperand(0))
5759 if (InVec1 != Op1.getOperand(0))
5762 // Update the pair of expected opcodes.
5763 std::swap(ExpectedOpcode, NextExpectedOpcode);
5766 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5767 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5768 InVec1.getOpcode() != ISD::UNDEF)
5769 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5774 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5775 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5776 const X86Subtarget *Subtarget,
5777 SelectionDAG &DAG) {
5778 EVT VT = BV->getValueType(0);
5779 unsigned NumElts = VT.getVectorNumElements();
5780 unsigned NumUndefsLO = 0;
5781 unsigned NumUndefsHI = 0;
5782 unsigned Half = NumElts/2;
5784 // Count the number of UNDEF operands in the build_vector in input.
5785 for (unsigned i = 0, e = Half; i != e; ++i)
5786 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5789 for (unsigned i = Half, e = NumElts; i != e; ++i)
5790 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5793 // Early exit if this is either a build_vector of all UNDEFs or all the
5794 // operands but one are UNDEF.
5795 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5799 SDValue InVec0, InVec1;
5800 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5801 // Try to match an SSE3 float HADD/HSUB.
5802 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5803 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5805 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5806 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5807 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5808 // Try to match an SSSE3 integer HADD/HSUB.
5809 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5810 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5812 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5813 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5816 if (!Subtarget->hasAVX())
5819 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5820 // Try to match an AVX horizontal add/sub of packed single/double
5821 // precision floating point values from 256-bit vectors.
5822 SDValue InVec2, InVec3;
5823 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5824 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5825 ((InVec0.getOpcode() == ISD::UNDEF ||
5826 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5827 ((InVec1.getOpcode() == ISD::UNDEF ||
5828 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5829 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5831 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5832 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5833 ((InVec0.getOpcode() == ISD::UNDEF ||
5834 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5835 ((InVec1.getOpcode() == ISD::UNDEF ||
5836 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5837 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5838 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5839 // Try to match an AVX2 horizontal add/sub of signed integers.
5840 SDValue InVec2, InVec3;
5842 bool CanFold = true;
5844 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5845 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5846 ((InVec0.getOpcode() == ISD::UNDEF ||
5847 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5848 ((InVec1.getOpcode() == ISD::UNDEF ||
5849 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5850 X86Opcode = X86ISD::HADD;
5851 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5852 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5853 ((InVec0.getOpcode() == ISD::UNDEF ||
5854 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5855 ((InVec1.getOpcode() == ISD::UNDEF ||
5856 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5857 X86Opcode = X86ISD::HSUB;
5862 // Fold this build_vector into a single horizontal add/sub.
5863 // Do this only if the target has AVX2.
5864 if (Subtarget->hasAVX2())
5865 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
5867 // Do not try to expand this build_vector into a pair of horizontal
5868 // add/sub if we can emit a pair of scalar add/sub.
5869 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5872 // Convert this build_vector into a pair of horizontal binop followed by
5874 bool isUndefLO = NumUndefsLO == Half;
5875 bool isUndefHI = NumUndefsHI == Half;
5876 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
5877 isUndefLO, isUndefHI);
5881 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
5882 VT == MVT::v16i16) && Subtarget->hasAVX()) {
5884 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5885 X86Opcode = X86ISD::HADD;
5886 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5887 X86Opcode = X86ISD::HSUB;
5888 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5889 X86Opcode = X86ISD::FHADD;
5890 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5891 X86Opcode = X86ISD::FHSUB;
5895 // Don't try to expand this build_vector into a pair of horizontal add/sub
5896 // if we can simply emit a pair of scalar add/sub.
5897 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5900 // Convert this build_vector into two horizontal add/sub followed by
5902 bool isUndefLO = NumUndefsLO == Half;
5903 bool isUndefHI = NumUndefsHI == Half;
5904 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
5905 isUndefLO, isUndefHI);
5912 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5915 MVT VT = Op.getSimpleValueType();
5916 MVT ExtVT = VT.getVectorElementType();
5917 unsigned NumElems = Op.getNumOperands();
5919 // Generate vectors for predicate vectors.
5920 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5921 return LowerBUILD_VECTORvXi1(Op, DAG);
5923 // Vectors containing all zeros can be matched by pxor and xorps later
5924 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5925 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5926 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5927 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
5930 return getZeroVector(VT, Subtarget, DAG, dl);
5933 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5934 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5935 // vpcmpeqd on 256-bit vectors.
5936 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5937 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5940 if (!VT.is512BitVector())
5941 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5944 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
5945 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
5947 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
5948 return HorizontalOp;
5949 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
5952 unsigned EVTBits = ExtVT.getSizeInBits();
5954 unsigned NumZero = 0;
5955 unsigned NumNonZero = 0;
5956 unsigned NonZeros = 0;
5957 bool IsAllConstants = true;
5958 SmallSet<SDValue, 8> Values;
5959 for (unsigned i = 0; i < NumElems; ++i) {
5960 SDValue Elt = Op.getOperand(i);
5961 if (Elt.getOpcode() == ISD::UNDEF)
5964 if (Elt.getOpcode() != ISD::Constant &&
5965 Elt.getOpcode() != ISD::ConstantFP)
5966 IsAllConstants = false;
5967 if (X86::isZeroNode(Elt))
5970 NonZeros |= (1 << i);
5975 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5976 if (NumNonZero == 0)
5977 return DAG.getUNDEF(VT);
5979 // Special case for single non-zero, non-undef, element.
5980 if (NumNonZero == 1) {
5981 unsigned Idx = countTrailingZeros(NonZeros);
5982 SDValue Item = Op.getOperand(Idx);
5984 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5985 // the value are obviously zero, truncate the value to i32 and do the
5986 // insertion that way. Only do this if the value is non-constant or if the
5987 // value is a constant being inserted into element 0. It is cheaper to do
5988 // a constant pool load than it is to do a movd + shuffle.
5989 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5990 (!IsAllConstants || Idx == 0)) {
5991 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5993 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5994 EVT VecVT = MVT::v4i32;
5996 // Truncate the value (which may itself be a constant) to i32, and
5997 // convert it to a vector with movd (S2V+shuffle to zero extend).
5998 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5999 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6000 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6001 Item, Idx * 2, true, Subtarget, DAG));
6005 // If we have a constant or non-constant insertion into the low element of
6006 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6007 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6008 // depending on what the source datatype is.
6011 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6013 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6014 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6015 if (VT.is512BitVector()) {
6016 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6017 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6018 Item, DAG.getIntPtrConstant(0, dl));
6020 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6021 "Expected an SSE value type!");
6022 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6023 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6024 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6027 // We can't directly insert an i8 or i16 into a vector, so zero extend
6029 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6030 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6031 if (VT.is256BitVector()) {
6032 if (Subtarget->hasAVX()) {
6033 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6034 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6036 // Without AVX, we need to extend to a 128-bit vector and then
6037 // insert into the 256-bit vector.
6038 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6039 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6040 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6043 assert(VT.is128BitVector() && "Expected an SSE value type!");
6044 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6045 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6047 return DAG.getBitcast(VT, Item);
6051 // Is it a vector logical left shift?
6052 if (NumElems == 2 && Idx == 1 &&
6053 X86::isZeroNode(Op.getOperand(0)) &&
6054 !X86::isZeroNode(Op.getOperand(1))) {
6055 unsigned NumBits = VT.getSizeInBits();
6056 return getVShift(true, VT,
6057 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6058 VT, Op.getOperand(1)),
6059 NumBits/2, DAG, *this, dl);
6062 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6065 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6066 // is a non-constant being inserted into an element other than the low one,
6067 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6068 // movd/movss) to move this into the low element, then shuffle it into
6070 if (EVTBits == 32) {
6071 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6072 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6076 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6077 if (Values.size() == 1) {
6078 if (EVTBits == 32) {
6079 // Instead of a shuffle like this:
6080 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6081 // Check if it's possible to issue this instead.
6082 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6083 unsigned Idx = countTrailingZeros(NonZeros);
6084 SDValue Item = Op.getOperand(Idx);
6085 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6086 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6091 // A vector full of immediates; various special cases are already
6092 // handled, so this is best done with a single constant-pool load.
6096 // For AVX-length vectors, see if we can use a vector load to get all of the
6097 // elements, otherwise build the individual 128-bit pieces and use
6098 // shuffles to put them in place.
6099 if (VT.is256BitVector() || VT.is512BitVector()) {
6100 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6102 // Check for a build vector of consecutive loads.
6103 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6106 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6108 // Build both the lower and upper subvector.
6109 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6110 makeArrayRef(&V[0], NumElems/2));
6111 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6112 makeArrayRef(&V[NumElems / 2], NumElems/2));
6114 // Recreate the wider vector with the lower and upper part.
6115 if (VT.is256BitVector())
6116 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6117 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6120 // Let legalizer expand 2-wide build_vectors.
6121 if (EVTBits == 64) {
6122 if (NumNonZero == 1) {
6123 // One half is zero or undef.
6124 unsigned Idx = countTrailingZeros(NonZeros);
6125 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6126 Op.getOperand(Idx));
6127 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6132 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6133 if (EVTBits == 8 && NumElems == 16)
6134 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6138 if (EVTBits == 16 && NumElems == 8)
6139 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6143 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6144 if (EVTBits == 32 && NumElems == 4)
6145 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6148 // If element VT is == 32 bits, turn it into a number of shuffles.
6149 SmallVector<SDValue, 8> V(NumElems);
6150 if (NumElems == 4 && NumZero > 0) {
6151 for (unsigned i = 0; i < 4; ++i) {
6152 bool isZero = !(NonZeros & (1 << i));
6154 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6156 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6159 for (unsigned i = 0; i < 2; ++i) {
6160 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6163 V[i] = V[i*2]; // Must be a zero vector.
6166 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6169 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6172 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6177 bool Reverse1 = (NonZeros & 0x3) == 2;
6178 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6182 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6183 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6185 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6188 if (Values.size() > 1 && VT.is128BitVector()) {
6189 // Check for a build vector of consecutive loads.
6190 for (unsigned i = 0; i < NumElems; ++i)
6191 V[i] = Op.getOperand(i);
6193 // Check for elements which are consecutive loads.
6194 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6197 // Check for a build vector from mostly shuffle plus few inserting.
6198 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6201 // For SSE 4.1, use insertps to put the high elements into the low element.
6202 if (Subtarget->hasSSE41()) {
6204 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6205 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6207 Result = DAG.getUNDEF(VT);
6209 for (unsigned i = 1; i < NumElems; ++i) {
6210 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6211 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6212 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6217 // Otherwise, expand into a number of unpckl*, start by extending each of
6218 // our (non-undef) elements to the full vector width with the element in the
6219 // bottom slot of the vector (which generates no code for SSE).
6220 for (unsigned i = 0; i < NumElems; ++i) {
6221 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6222 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6224 V[i] = DAG.getUNDEF(VT);
6227 // Next, we iteratively mix elements, e.g. for v4f32:
6228 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6229 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6230 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6231 unsigned EltStride = NumElems >> 1;
6232 while (EltStride != 0) {
6233 for (unsigned i = 0; i < EltStride; ++i) {
6234 // If V[i+EltStride] is undef and this is the first round of mixing,
6235 // then it is safe to just drop this shuffle: V[i] is already in the
6236 // right place, the one element (since it's the first round) being
6237 // inserted as undef can be dropped. This isn't safe for successive
6238 // rounds because they will permute elements within both vectors.
6239 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6240 EltStride == NumElems/2)
6243 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6252 // 256-bit AVX can use the vinsertf128 instruction
6253 // to create 256-bit vectors from two other 128-bit ones.
6254 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6256 MVT ResVT = Op.getSimpleValueType();
6258 assert((ResVT.is256BitVector() ||
6259 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6261 SDValue V1 = Op.getOperand(0);
6262 SDValue V2 = Op.getOperand(1);
6263 unsigned NumElems = ResVT.getVectorNumElements();
6264 if (ResVT.is256BitVector())
6265 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6267 if (Op.getNumOperands() == 4) {
6268 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6269 ResVT.getVectorNumElements()/2);
6270 SDValue V3 = Op.getOperand(2);
6271 SDValue V4 = Op.getOperand(3);
6272 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6273 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6275 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6278 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6279 const X86Subtarget *Subtarget,
6280 SelectionDAG & DAG) {
6282 MVT ResVT = Op.getSimpleValueType();
6283 unsigned NumOfOperands = Op.getNumOperands();
6285 assert(isPowerOf2_32(NumOfOperands) &&
6286 "Unexpected number of operands in CONCAT_VECTORS");
6288 if (NumOfOperands > 2) {
6289 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6290 ResVT.getVectorNumElements()/2);
6291 SmallVector<SDValue, 2> Ops;
6292 for (unsigned i = 0; i < NumOfOperands/2; i++)
6293 Ops.push_back(Op.getOperand(i));
6294 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6296 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6297 Ops.push_back(Op.getOperand(i));
6298 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6299 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6302 SDValue V1 = Op.getOperand(0);
6303 SDValue V2 = Op.getOperand(1);
6304 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6305 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6307 if (IsZeroV1 && IsZeroV2)
6308 return getZeroVector(ResVT, Subtarget, DAG, dl);
6310 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6311 SDValue Undef = DAG.getUNDEF(ResVT);
6312 unsigned NumElems = ResVT.getVectorNumElements();
6313 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6315 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6316 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6320 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6321 // Zero the upper bits of V1
6322 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6323 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6326 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6329 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6330 const X86Subtarget *Subtarget,
6331 SelectionDAG &DAG) {
6332 MVT VT = Op.getSimpleValueType();
6333 if (VT.getVectorElementType() == MVT::i1)
6334 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6336 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6337 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6338 Op.getNumOperands() == 4)));
6340 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6341 // from two other 128-bit ones.
6343 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6344 return LowerAVXCONCAT_VECTORS(Op, DAG);
6348 //===----------------------------------------------------------------------===//
6349 // Vector shuffle lowering
6351 // This is an experimental code path for lowering vector shuffles on x86. It is
6352 // designed to handle arbitrary vector shuffles and blends, gracefully
6353 // degrading performance as necessary. It works hard to recognize idiomatic
6354 // shuffles and lower them to optimal instruction patterns without leaving
6355 // a framework that allows reasonably efficient handling of all vector shuffle
6357 //===----------------------------------------------------------------------===//
6359 /// \brief Tiny helper function to identify a no-op mask.
6361 /// This is a somewhat boring predicate function. It checks whether the mask
6362 /// array input, which is assumed to be a single-input shuffle mask of the kind
6363 /// used by the X86 shuffle instructions (not a fully general
6364 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6365 /// in-place shuffle are 'no-op's.
6366 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6367 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6368 if (Mask[i] != -1 && Mask[i] != i)
6373 /// \brief Helper function to classify a mask as a single-input mask.
6375 /// This isn't a generic single-input test because in the vector shuffle
6376 /// lowering we canonicalize single inputs to be the first input operand. This
6377 /// means we can more quickly test for a single input by only checking whether
6378 /// an input from the second operand exists. We also assume that the size of
6379 /// mask corresponds to the size of the input vectors which isn't true in the
6380 /// fully general case.
6381 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6383 if (M >= (int)Mask.size())
6388 /// \brief Test whether there are elements crossing 128-bit lanes in this
6391 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6392 /// and we routinely test for these.
6393 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6394 int LaneSize = 128 / VT.getScalarSizeInBits();
6395 int Size = Mask.size();
6396 for (int i = 0; i < Size; ++i)
6397 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6402 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6404 /// This checks a shuffle mask to see if it is performing the same
6405 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6406 /// that it is also not lane-crossing. It may however involve a blend from the
6407 /// same lane of a second vector.
6409 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6410 /// non-trivial to compute in the face of undef lanes. The representation is
6411 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6412 /// entries from both V1 and V2 inputs to the wider mask.
6414 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6415 SmallVectorImpl<int> &RepeatedMask) {
6416 int LaneSize = 128 / VT.getScalarSizeInBits();
6417 RepeatedMask.resize(LaneSize, -1);
6418 int Size = Mask.size();
6419 for (int i = 0; i < Size; ++i) {
6422 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6423 // This entry crosses lanes, so there is no way to model this shuffle.
6426 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6427 if (RepeatedMask[i % LaneSize] == -1)
6428 // This is the first non-undef entry in this slot of a 128-bit lane.
6429 RepeatedMask[i % LaneSize] =
6430 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6431 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6432 // Found a mismatch with the repeated mask.
6438 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6441 /// This is a fast way to test a shuffle mask against a fixed pattern:
6443 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6445 /// It returns true if the mask is exactly as wide as the argument list, and
6446 /// each element of the mask is either -1 (signifying undef) or the value given
6447 /// in the argument.
6448 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6449 ArrayRef<int> ExpectedMask) {
6450 if (Mask.size() != ExpectedMask.size())
6453 int Size = Mask.size();
6455 // If the values are build vectors, we can look through them to find
6456 // equivalent inputs that make the shuffles equivalent.
6457 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6458 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6460 for (int i = 0; i < Size; ++i)
6461 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6462 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6463 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6464 if (!MaskBV || !ExpectedBV ||
6465 MaskBV->getOperand(Mask[i] % Size) !=
6466 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6473 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6475 /// This helper function produces an 8-bit shuffle immediate corresponding to
6476 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6477 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6480 /// NB: We rely heavily on "undef" masks preserving the input lane.
6481 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6482 SelectionDAG &DAG) {
6483 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6484 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6485 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6486 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6487 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6490 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6491 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6492 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6493 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6494 return DAG.getConstant(Imm, DL, MVT::i8);
6497 /// \brief Compute whether each element of a shuffle is zeroable.
6499 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6500 /// Either it is an undef element in the shuffle mask, the element of the input
6501 /// referenced is undef, or the element of the input referenced is known to be
6502 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6503 /// as many lanes with this technique as possible to simplify the remaining
6505 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6506 SDValue V1, SDValue V2) {
6507 SmallBitVector Zeroable(Mask.size(), false);
6509 while (V1.getOpcode() == ISD::BITCAST)
6510 V1 = V1->getOperand(0);
6511 while (V2.getOpcode() == ISD::BITCAST)
6512 V2 = V2->getOperand(0);
6514 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6515 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6517 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6519 // Handle the easy cases.
6520 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6525 // If this is an index into a build_vector node (which has the same number
6526 // of elements), dig out the input value and use it.
6527 SDValue V = M < Size ? V1 : V2;
6528 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6531 SDValue Input = V.getOperand(M % Size);
6532 // The UNDEF opcode check really should be dead code here, but not quite
6533 // worth asserting on (it isn't invalid, just unexpected).
6534 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6541 /// \brief Try to emit a bitmask instruction for a shuffle.
6543 /// This handles cases where we can model a blend exactly as a bitmask due to
6544 /// one of the inputs being zeroable.
6545 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6546 SDValue V2, ArrayRef<int> Mask,
6547 SelectionDAG &DAG) {
6548 MVT EltVT = VT.getScalarType();
6549 int NumEltBits = EltVT.getSizeInBits();
6550 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6551 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6552 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6554 if (EltVT.isFloatingPoint()) {
6555 Zero = DAG.getBitcast(EltVT, Zero);
6556 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6558 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6559 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6561 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6564 if (Mask[i] % Size != i)
6565 return SDValue(); // Not a blend.
6567 V = Mask[i] < Size ? V1 : V2;
6568 else if (V != (Mask[i] < Size ? V1 : V2))
6569 return SDValue(); // Can only let one input through the mask.
6571 VMaskOps[i] = AllOnes;
6574 return SDValue(); // No non-zeroable elements!
6576 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6577 V = DAG.getNode(VT.isFloatingPoint()
6578 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6583 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6585 /// This is used as a fallback approach when first class blend instructions are
6586 /// unavailable. Currently it is only suitable for integer vectors, but could
6587 /// be generalized for floating point vectors if desirable.
6588 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6589 SDValue V2, ArrayRef<int> Mask,
6590 SelectionDAG &DAG) {
6591 assert(VT.isInteger() && "Only supports integer vector types!");
6592 MVT EltVT = VT.getScalarType();
6593 int NumEltBits = EltVT.getSizeInBits();
6594 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6595 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6597 SmallVector<SDValue, 16> MaskOps;
6598 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6599 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6600 return SDValue(); // Shuffled input!
6601 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6604 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6605 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6606 // We have to cast V2 around.
6607 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6608 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6609 DAG.getBitcast(MaskVT, V1Mask),
6610 DAG.getBitcast(MaskVT, V2)));
6611 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6614 /// \brief Try to emit a blend instruction for a shuffle.
6616 /// This doesn't do any checks for the availability of instructions for blending
6617 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6618 /// be matched in the backend with the type given. What it does check for is
6619 /// that the shuffle mask is in fact a blend.
6620 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6621 SDValue V2, ArrayRef<int> Mask,
6622 const X86Subtarget *Subtarget,
6623 SelectionDAG &DAG) {
6624 unsigned BlendMask = 0;
6625 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6626 if (Mask[i] >= Size) {
6627 if (Mask[i] != i + Size)
6628 return SDValue(); // Shuffled V2 input!
6629 BlendMask |= 1u << i;
6632 if (Mask[i] >= 0 && Mask[i] != i)
6633 return SDValue(); // Shuffled V1 input!
6635 switch (VT.SimpleTy) {
6640 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6641 DAG.getConstant(BlendMask, DL, MVT::i8));
6645 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6649 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6650 // that instruction.
6651 if (Subtarget->hasAVX2()) {
6652 // Scale the blend by the number of 32-bit dwords per element.
6653 int Scale = VT.getScalarSizeInBits() / 32;
6655 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6656 if (Mask[i] >= Size)
6657 for (int j = 0; j < Scale; ++j)
6658 BlendMask |= 1u << (i * Scale + j);
6660 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6661 V1 = DAG.getBitcast(BlendVT, V1);
6662 V2 = DAG.getBitcast(BlendVT, V2);
6663 return DAG.getBitcast(
6664 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6665 DAG.getConstant(BlendMask, DL, MVT::i8)));
6669 // For integer shuffles we need to expand the mask and cast the inputs to
6670 // v8i16s prior to blending.
6671 int Scale = 8 / VT.getVectorNumElements();
6673 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6674 if (Mask[i] >= Size)
6675 for (int j = 0; j < Scale; ++j)
6676 BlendMask |= 1u << (i * Scale + j);
6678 V1 = DAG.getBitcast(MVT::v8i16, V1);
6679 V2 = DAG.getBitcast(MVT::v8i16, V2);
6680 return DAG.getBitcast(VT,
6681 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6682 DAG.getConstant(BlendMask, DL, MVT::i8)));
6686 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6687 SmallVector<int, 8> RepeatedMask;
6688 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6689 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6690 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6692 for (int i = 0; i < 8; ++i)
6693 if (RepeatedMask[i] >= 16)
6694 BlendMask |= 1u << i;
6695 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6696 DAG.getConstant(BlendMask, DL, MVT::i8));
6702 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6703 "256-bit byte-blends require AVX2 support!");
6705 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
6706 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
6709 // Scale the blend by the number of bytes per element.
6710 int Scale = VT.getScalarSizeInBits() / 8;
6712 // This form of blend is always done on bytes. Compute the byte vector
6714 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6716 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6717 // mix of LLVM's code generator and the x86 backend. We tell the code
6718 // generator that boolean values in the elements of an x86 vector register
6719 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6720 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6721 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6722 // of the element (the remaining are ignored) and 0 in that high bit would
6723 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6724 // the LLVM model for boolean values in vector elements gets the relevant
6725 // bit set, it is set backwards and over constrained relative to x86's
6727 SmallVector<SDValue, 32> VSELECTMask;
6728 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6729 for (int j = 0; j < Scale; ++j)
6730 VSELECTMask.push_back(
6731 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6732 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
6735 V1 = DAG.getBitcast(BlendVT, V1);
6736 V2 = DAG.getBitcast(BlendVT, V2);
6737 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
6738 DAG.getNode(ISD::BUILD_VECTOR, DL,
6739 BlendVT, VSELECTMask),
6744 llvm_unreachable("Not a supported integer vector type!");
6748 /// \brief Try to lower as a blend of elements from two inputs followed by
6749 /// a single-input permutation.
6751 /// This matches the pattern where we can blend elements from two inputs and
6752 /// then reduce the shuffle to a single-input permutation.
6753 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6756 SelectionDAG &DAG) {
6757 // We build up the blend mask while checking whether a blend is a viable way
6758 // to reduce the shuffle.
6759 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6760 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6762 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6766 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6768 if (BlendMask[Mask[i] % Size] == -1)
6769 BlendMask[Mask[i] % Size] = Mask[i];
6770 else if (BlendMask[Mask[i] % Size] != Mask[i])
6771 return SDValue(); // Can't blend in the needed input!
6773 PermuteMask[i] = Mask[i] % Size;
6776 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6777 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6780 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6781 /// blends and permutes.
6783 /// This matches the extremely common pattern for handling combined
6784 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6785 /// operations. It will try to pick the best arrangement of shuffles and
6787 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6791 SelectionDAG &DAG) {
6792 // Shuffle the input elements into the desired positions in V1 and V2 and
6793 // blend them together.
6794 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6795 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6796 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6797 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6798 if (Mask[i] >= 0 && Mask[i] < Size) {
6799 V1Mask[i] = Mask[i];
6801 } else if (Mask[i] >= Size) {
6802 V2Mask[i] = Mask[i] - Size;
6803 BlendMask[i] = i + Size;
6806 // Try to lower with the simpler initial blend strategy unless one of the
6807 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6808 // shuffle may be able to fold with a load or other benefit. However, when
6809 // we'll have to do 2x as many shuffles in order to achieve this, blending
6810 // first is a better strategy.
6811 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6812 if (SDValue BlendPerm =
6813 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6816 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6817 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6818 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6821 /// \brief Try to lower a vector shuffle as a byte rotation.
6823 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6824 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6825 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6826 /// try to generically lower a vector shuffle through such an pattern. It
6827 /// does not check for the profitability of lowering either as PALIGNR or
6828 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6829 /// This matches shuffle vectors that look like:
6831 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6833 /// Essentially it concatenates V1 and V2, shifts right by some number of
6834 /// elements, and takes the low elements as the result. Note that while this is
6835 /// specified as a *right shift* because x86 is little-endian, it is a *left
6836 /// rotate* of the vector lanes.
6837 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
6840 const X86Subtarget *Subtarget,
6841 SelectionDAG &DAG) {
6842 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
6844 int NumElts = Mask.size();
6845 int NumLanes = VT.getSizeInBits() / 128;
6846 int NumLaneElts = NumElts / NumLanes;
6848 // We need to detect various ways of spelling a rotation:
6849 // [11, 12, 13, 14, 15, 0, 1, 2]
6850 // [-1, 12, 13, 14, -1, -1, 1, -1]
6851 // [-1, -1, -1, -1, -1, -1, 1, 2]
6852 // [ 3, 4, 5, 6, 7, 8, 9, 10]
6853 // [-1, 4, 5, 6, -1, -1, 9, -1]
6854 // [-1, 4, 5, 6, -1, -1, -1, -1]
6857 for (int l = 0; l < NumElts; l += NumLaneElts) {
6858 for (int i = 0; i < NumLaneElts; ++i) {
6859 if (Mask[l + i] == -1)
6861 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
6863 // Get the mod-Size index and lane correct it.
6864 int LaneIdx = (Mask[l + i] % NumElts) - l;
6865 // Make sure it was in this lane.
6866 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
6869 // Determine where a rotated vector would have started.
6870 int StartIdx = i - LaneIdx;
6872 // The identity rotation isn't interesting, stop.
6875 // If we found the tail of a vector the rotation must be the missing
6876 // front. If we found the head of a vector, it must be how much of the
6878 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
6881 Rotation = CandidateRotation;
6882 else if (Rotation != CandidateRotation)
6883 // The rotations don't match, so we can't match this mask.
6886 // Compute which value this mask is pointing at.
6887 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
6889 // Compute which of the two target values this index should be assigned
6890 // to. This reflects whether the high elements are remaining or the low
6891 // elements are remaining.
6892 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
6894 // Either set up this value if we've not encountered it before, or check
6895 // that it remains consistent.
6898 else if (TargetV != MaskV)
6899 // This may be a rotation, but it pulls from the inputs in some
6900 // unsupported interleaving.
6905 // Check that we successfully analyzed the mask, and normalize the results.
6906 assert(Rotation != 0 && "Failed to locate a viable rotation!");
6907 assert((Lo || Hi) && "Failed to find a rotated input vector!");
6913 // The actual rotate instruction rotates bytes, so we need to scale the
6914 // rotation based on how many bytes are in the vector lane.
6915 int Scale = 16 / NumLaneElts;
6917 // SSSE3 targets can use the palignr instruction.
6918 if (Subtarget->hasSSSE3()) {
6919 // Cast the inputs to i8 vector of correct length to match PALIGNR.
6920 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
6921 Lo = DAG.getBitcast(AlignVT, Lo);
6922 Hi = DAG.getBitcast(AlignVT, Hi);
6924 return DAG.getBitcast(
6925 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
6926 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
6929 assert(VT.getSizeInBits() == 128 &&
6930 "Rotate-based lowering only supports 128-bit lowering!");
6931 assert(Mask.size() <= 16 &&
6932 "Can shuffle at most 16 bytes in a 128-bit vector!");
6934 // Default SSE2 implementation
6935 int LoByteShift = 16 - Rotation * Scale;
6936 int HiByteShift = Rotation * Scale;
6938 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
6939 Lo = DAG.getBitcast(MVT::v2i64, Lo);
6940 Hi = DAG.getBitcast(MVT::v2i64, Hi);
6942 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
6943 DAG.getConstant(LoByteShift, DL, MVT::i8));
6944 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
6945 DAG.getConstant(HiByteShift, DL, MVT::i8));
6946 return DAG.getBitcast(VT,
6947 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
6950 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
6952 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
6953 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
6954 /// matches elements from one of the input vectors shuffled to the left or
6955 /// right with zeroable elements 'shifted in'. It handles both the strictly
6956 /// bit-wise element shifts and the byte shift across an entire 128-bit double
6959 /// PSHL : (little-endian) left bit shift.
6960 /// [ zz, 0, zz, 2 ]
6961 /// [ -1, 4, zz, -1 ]
6962 /// PSRL : (little-endian) right bit shift.
6964 /// [ -1, -1, 7, zz]
6965 /// PSLLDQ : (little-endian) left byte shift
6966 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
6967 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
6968 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
6969 /// PSRLDQ : (little-endian) right byte shift
6970 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
6971 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
6972 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
6973 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
6974 SDValue V2, ArrayRef<int> Mask,
6975 SelectionDAG &DAG) {
6976 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6978 int Size = Mask.size();
6979 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
6981 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
6982 for (int i = 0; i < Size; i += Scale)
6983 for (int j = 0; j < Shift; ++j)
6984 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
6990 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
6991 for (int i = 0; i != Size; i += Scale) {
6992 unsigned Pos = Left ? i + Shift : i;
6993 unsigned Low = Left ? i : i + Shift;
6994 unsigned Len = Scale - Shift;
6995 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
6996 Low + (V == V1 ? 0 : Size)))
7000 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7001 bool ByteShift = ShiftEltBits > 64;
7002 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7003 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7004 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7006 // Normalize the scale for byte shifts to still produce an i64 element
7008 Scale = ByteShift ? Scale / 2 : Scale;
7010 // We need to round trip through the appropriate type for the shift.
7011 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7012 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7013 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7014 "Illegal integer vector type");
7015 V = DAG.getBitcast(ShiftVT, V);
7017 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7018 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7019 return DAG.getBitcast(VT, V);
7022 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7023 // keep doubling the size of the integer elements up to that. We can
7024 // then shift the elements of the integer vector by whole multiples of
7025 // their width within the elements of the larger integer vector. Test each
7026 // multiple to see if we can find a match with the moved element indices
7027 // and that the shifted in elements are all zeroable.
7028 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7029 for (int Shift = 1; Shift != Scale; ++Shift)
7030 for (bool Left : {true, false})
7031 if (CheckZeros(Shift, Scale, Left))
7032 for (SDValue V : {V1, V2})
7033 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7040 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7041 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7042 SDValue V2, ArrayRef<int> Mask,
7043 SelectionDAG &DAG) {
7044 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7045 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7047 int Size = Mask.size();
7048 int HalfSize = Size / 2;
7049 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7051 // Upper half must be undefined.
7052 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7055 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7056 // Remainder of lower half result is zero and upper half is all undef.
7057 auto LowerAsEXTRQ = [&]() {
7058 // Determine the extraction length from the part of the
7059 // lower half that isn't zeroable.
7061 for (; Len >= 0; --Len)
7062 if (!Zeroable[Len - 1])
7064 assert(Len > 0 && "Zeroable shuffle mask");
7066 // Attempt to match first Len sequential elements from the lower half.
7069 for (int i = 0; i != Len; ++i) {
7073 SDValue &V = (M < Size ? V1 : V2);
7076 // All mask elements must be in the lower half.
7080 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7091 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7092 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7093 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7094 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7095 DAG.getConstant(BitLen, DL, MVT::i8),
7096 DAG.getConstant(BitIdx, DL, MVT::i8));
7099 if (SDValue ExtrQ = LowerAsEXTRQ())
7102 // INSERTQ: Extract lowest Len elements from lower half of second source and
7103 // insert over first source, starting at Idx.
7104 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7105 auto LowerAsInsertQ = [&]() {
7106 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7109 // Attempt to match first source from mask before insertion point.
7110 if (isUndefInRange(Mask, 0, Idx)) {
7112 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7114 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7120 // Extend the extraction length looking to match both the insertion of
7121 // the second source and the remaining elements of the first.
7122 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7127 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7129 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7135 // Match the remaining elements of the lower half.
7136 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7138 } else if ((!Base || (Base == V1)) &&
7139 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7141 } else if ((!Base || (Base == V2)) &&
7142 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7149 // We may not have a base (first source) - this can safely be undefined.
7151 Base = DAG.getUNDEF(VT);
7153 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7154 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7155 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7156 DAG.getConstant(BitLen, DL, MVT::i8),
7157 DAG.getConstant(BitIdx, DL, MVT::i8));
7164 if (SDValue InsertQ = LowerAsInsertQ())
7170 /// \brief Lower a vector shuffle as a zero or any extension.
7172 /// Given a specific number of elements, element bit width, and extension
7173 /// stride, produce either a zero or any extension based on the available
7174 /// features of the subtarget.
7175 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7176 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
7177 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7178 assert(Scale > 1 && "Need a scale to extend.");
7179 int NumElements = VT.getVectorNumElements();
7180 int EltBits = VT.getScalarSizeInBits();
7181 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7182 "Only 8, 16, and 32 bit elements can be extended.");
7183 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7185 // Found a valid zext mask! Try various lowering strategies based on the
7186 // input type and available ISA extensions.
7187 if (Subtarget->hasSSE41()) {
7188 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7189 NumElements / Scale);
7190 return DAG.getBitcast(VT, DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
7193 // For any extends we can cheat for larger element sizes and use shuffle
7194 // instructions that can fold with a load and/or copy.
7195 if (AnyExt && EltBits == 32) {
7196 int PSHUFDMask[4] = {0, -1, 1, -1};
7197 return DAG.getBitcast(
7198 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7199 DAG.getBitcast(MVT::v4i32, InputV),
7200 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7202 if (AnyExt && EltBits == 16 && Scale > 2) {
7203 int PSHUFDMask[4] = {0, -1, 0, -1};
7204 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7205 DAG.getBitcast(MVT::v4i32, InputV),
7206 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7207 int PSHUFHWMask[4] = {1, -1, -1, -1};
7208 return DAG.getBitcast(
7209 VT, DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
7210 DAG.getBitcast(MVT::v8i16, InputV),
7211 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DL, DAG)));
7214 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7216 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7217 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7218 assert(VT.getSizeInBits() == 128 && "Unexpected vector width!");
7220 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7221 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7222 DAG.getConstant(EltBits, DL, MVT::i8),
7223 DAG.getConstant(0, DL, MVT::i8)));
7224 if (isUndefInRange(Mask, NumElements/2, NumElements/2))
7225 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7228 DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7229 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7230 DAG.getConstant(EltBits, DL, MVT::i8),
7231 DAG.getConstant(EltBits, DL, MVT::i8)));
7232 return DAG.getNode(ISD::BITCAST, DL, VT,
7233 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7236 // If this would require more than 2 unpack instructions to expand, use
7237 // pshufb when available. We can only use more than 2 unpack instructions
7238 // when zero extending i8 elements which also makes it easier to use pshufb.
7239 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7240 assert(NumElements == 16 && "Unexpected byte vector width!");
7241 SDValue PSHUFBMask[16];
7242 for (int i = 0; i < 16; ++i)
7244 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, DL, MVT::i8);
7245 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7246 return DAG.getBitcast(VT,
7247 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7248 DAG.getNode(ISD::BUILD_VECTOR, DL,
7249 MVT::v16i8, PSHUFBMask)));
7252 // Otherwise emit a sequence of unpacks.
7254 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7255 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7256 : getZeroVector(InputVT, Subtarget, DAG, DL);
7257 InputV = DAG.getBitcast(InputVT, InputV);
7258 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
7262 } while (Scale > 1);
7263 return DAG.getBitcast(VT, InputV);
7266 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7268 /// This routine will try to do everything in its power to cleverly lower
7269 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7270 /// check for the profitability of this lowering, it tries to aggressively
7271 /// match this pattern. It will use all of the micro-architectural details it
7272 /// can to emit an efficient lowering. It handles both blends with all-zero
7273 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7274 /// masking out later).
7276 /// The reason we have dedicated lowering for zext-style shuffles is that they
7277 /// are both incredibly common and often quite performance sensitive.
7278 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7279 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7280 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7281 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7283 int Bits = VT.getSizeInBits();
7284 int NumElements = VT.getVectorNumElements();
7285 assert(VT.getScalarSizeInBits() <= 32 &&
7286 "Exceeds 32-bit integer zero extension limit");
7287 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7289 // Define a helper function to check a particular ext-scale and lower to it if
7291 auto Lower = [&](int Scale) -> SDValue {
7294 for (int i = 0; i < NumElements; ++i) {
7296 continue; // Valid anywhere but doesn't tell us anything.
7297 if (i % Scale != 0) {
7298 // Each of the extended elements need to be zeroable.
7302 // We no longer are in the anyext case.
7307 // Each of the base elements needs to be consecutive indices into the
7308 // same input vector.
7309 SDValue V = Mask[i] < NumElements ? V1 : V2;
7312 else if (InputV != V)
7313 return SDValue(); // Flip-flopping inputs.
7315 if (Mask[i] % NumElements != i / Scale)
7316 return SDValue(); // Non-consecutive strided elements.
7319 // If we fail to find an input, we have a zero-shuffle which should always
7320 // have already been handled.
7321 // FIXME: Maybe handle this here in case during blending we end up with one?
7325 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7326 DL, VT, Scale, AnyExt, InputV, Mask, Subtarget, DAG);
7329 // The widest scale possible for extending is to a 64-bit integer.
7330 assert(Bits % 64 == 0 &&
7331 "The number of bits in a vector must be divisible by 64 on x86!");
7332 int NumExtElements = Bits / 64;
7334 // Each iteration, try extending the elements half as much, but into twice as
7336 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7337 assert(NumElements % NumExtElements == 0 &&
7338 "The input vector size must be divisible by the extended size.");
7339 if (SDValue V = Lower(NumElements / NumExtElements))
7343 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7347 // Returns one of the source operands if the shuffle can be reduced to a
7348 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7349 auto CanZExtLowHalf = [&]() {
7350 for (int i = NumElements / 2; i != NumElements; ++i)
7353 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7355 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7360 if (SDValue V = CanZExtLowHalf()) {
7361 V = DAG.getBitcast(MVT::v2i64, V);
7362 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7363 return DAG.getBitcast(VT, V);
7366 // No viable ext lowering found.
7370 /// \brief Try to get a scalar value for a specific element of a vector.
7372 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7373 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7374 SelectionDAG &DAG) {
7375 MVT VT = V.getSimpleValueType();
7376 MVT EltVT = VT.getVectorElementType();
7377 while (V.getOpcode() == ISD::BITCAST)
7378 V = V.getOperand(0);
7379 // If the bitcasts shift the element size, we can't extract an equivalent
7381 MVT NewVT = V.getSimpleValueType();
7382 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7385 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7386 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7387 // Ensure the scalar operand is the same size as the destination.
7388 // FIXME: Add support for scalar truncation where possible.
7389 SDValue S = V.getOperand(Idx);
7390 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7391 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7397 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7399 /// This is particularly important because the set of instructions varies
7400 /// significantly based on whether the operand is a load or not.
7401 static bool isShuffleFoldableLoad(SDValue V) {
7402 while (V.getOpcode() == ISD::BITCAST)
7403 V = V.getOperand(0);
7405 return ISD::isNON_EXTLoad(V.getNode());
7408 /// \brief Try to lower insertion of a single element into a zero vector.
7410 /// This is a common pattern that we have especially efficient patterns to lower
7411 /// across all subtarget feature sets.
7412 static SDValue lowerVectorShuffleAsElementInsertion(
7413 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7414 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7415 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7417 MVT EltVT = VT.getVectorElementType();
7419 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7420 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7422 bool IsV1Zeroable = true;
7423 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7424 if (i != V2Index && !Zeroable[i]) {
7425 IsV1Zeroable = false;
7429 // Check for a single input from a SCALAR_TO_VECTOR node.
7430 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7431 // all the smarts here sunk into that routine. However, the current
7432 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7433 // vector shuffle lowering is dead.
7434 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7436 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7437 // We need to zext the scalar if it is smaller than an i32.
7438 V2S = DAG.getBitcast(EltVT, V2S);
7439 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7440 // Using zext to expand a narrow element won't work for non-zero
7445 // Zero-extend directly to i32.
7447 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7449 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7450 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7451 EltVT == MVT::i16) {
7452 // Either not inserting from the low element of the input or the input
7453 // element size is too small to use VZEXT_MOVL to clear the high bits.
7457 if (!IsV1Zeroable) {
7458 // If V1 can't be treated as a zero vector we have fewer options to lower
7459 // this. We can't support integer vectors or non-zero targets cheaply, and
7460 // the V1 elements can't be permuted in any way.
7461 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7462 if (!VT.isFloatingPoint() || V2Index != 0)
7464 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7465 V1Mask[V2Index] = -1;
7466 if (!isNoopShuffleMask(V1Mask))
7468 // This is essentially a special case blend operation, but if we have
7469 // general purpose blend operations, they are always faster. Bail and let
7470 // the rest of the lowering handle these as blends.
7471 if (Subtarget->hasSSE41())
7474 // Otherwise, use MOVSD or MOVSS.
7475 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7476 "Only two types of floating point element types to handle!");
7477 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7481 // This lowering only works for the low element with floating point vectors.
7482 if (VT.isFloatingPoint() && V2Index != 0)
7485 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7487 V2 = DAG.getBitcast(VT, V2);
7490 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7491 // the desired position. Otherwise it is more efficient to do a vector
7492 // shift left. We know that we can do a vector shift left because all
7493 // the inputs are zero.
7494 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7495 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7496 V2Shuffle[V2Index] = 0;
7497 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7499 V2 = DAG.getBitcast(MVT::v2i64, V2);
7501 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7502 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7503 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7504 DAG.getDataLayout(), VT)));
7505 V2 = DAG.getBitcast(VT, V2);
7511 /// \brief Try to lower broadcast of a single element.
7513 /// For convenience, this code also bundles all of the subtarget feature set
7514 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7515 /// a convenient way to factor it out.
7516 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7518 const X86Subtarget *Subtarget,
7519 SelectionDAG &DAG) {
7520 if (!Subtarget->hasAVX())
7522 if (VT.isInteger() && !Subtarget->hasAVX2())
7525 // Check that the mask is a broadcast.
7526 int BroadcastIdx = -1;
7528 if (M >= 0 && BroadcastIdx == -1)
7530 else if (M >= 0 && M != BroadcastIdx)
7533 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7534 "a sorted mask where the broadcast "
7537 // Go up the chain of (vector) values to find a scalar load that we can
7538 // combine with the broadcast.
7540 switch (V.getOpcode()) {
7541 case ISD::CONCAT_VECTORS: {
7542 int OperandSize = Mask.size() / V.getNumOperands();
7543 V = V.getOperand(BroadcastIdx / OperandSize);
7544 BroadcastIdx %= OperandSize;
7548 case ISD::INSERT_SUBVECTOR: {
7549 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7550 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7554 int BeginIdx = (int)ConstantIdx->getZExtValue();
7556 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7557 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7558 BroadcastIdx -= BeginIdx;
7569 // Check if this is a broadcast of a scalar. We special case lowering
7570 // for scalars so that we can more effectively fold with loads.
7571 // First, look through bitcast: if the original value has a larger element
7572 // type than the shuffle, the broadcast element is in essence truncated.
7573 // Make that explicit to ease folding.
7574 if (V.getOpcode() == ISD::BITCAST && VT.isInteger()) {
7575 EVT EltVT = VT.getVectorElementType();
7576 SDValue V0 = V.getOperand(0);
7577 EVT V0VT = V0.getValueType();
7579 if (V0VT.isInteger() && V0VT.getVectorElementType().bitsGT(EltVT) &&
7580 ((V0.getOpcode() == ISD::BUILD_VECTOR ||
7581 (V0.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)))) {
7582 V = DAG.getNode(ISD::TRUNCATE, DL, EltVT, V0.getOperand(BroadcastIdx));
7587 // Also check the simpler case, where we can directly reuse the scalar.
7588 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7589 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7590 V = V.getOperand(BroadcastIdx);
7592 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7593 // Only AVX2 has register broadcasts.
7594 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7596 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7597 // We can't broadcast from a vector register without AVX2, and we can only
7598 // broadcast from the zero-element of a vector register.
7602 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7605 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7606 // INSERTPS when the V1 elements are already in the correct locations
7607 // because otherwise we can just always use two SHUFPS instructions which
7608 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7609 // perform INSERTPS if a single V1 element is out of place and all V2
7610 // elements are zeroable.
7611 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7613 SelectionDAG &DAG) {
7614 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7615 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7616 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7617 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7619 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7622 int V1DstIndex = -1;
7623 int V2DstIndex = -1;
7624 bool V1UsedInPlace = false;
7626 for (int i = 0; i < 4; ++i) {
7627 // Synthesize a zero mask from the zeroable elements (includes undefs).
7633 // Flag if we use any V1 inputs in place.
7635 V1UsedInPlace = true;
7639 // We can only insert a single non-zeroable element.
7640 if (V1DstIndex != -1 || V2DstIndex != -1)
7644 // V1 input out of place for insertion.
7647 // V2 input for insertion.
7652 // Don't bother if we have no (non-zeroable) element for insertion.
7653 if (V1DstIndex == -1 && V2DstIndex == -1)
7656 // Determine element insertion src/dst indices. The src index is from the
7657 // start of the inserted vector, not the start of the concatenated vector.
7658 unsigned V2SrcIndex = 0;
7659 if (V1DstIndex != -1) {
7660 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7661 // and don't use the original V2 at all.
7662 V2SrcIndex = Mask[V1DstIndex];
7663 V2DstIndex = V1DstIndex;
7666 V2SrcIndex = Mask[V2DstIndex] - 4;
7669 // If no V1 inputs are used in place, then the result is created only from
7670 // the zero mask and the V2 insertion - so remove V1 dependency.
7672 V1 = DAG.getUNDEF(MVT::v4f32);
7674 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7675 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7677 // Insert the V2 element into the desired position.
7679 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7680 DAG.getConstant(InsertPSMask, DL, MVT::i8));
7683 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7684 /// UNPCK instruction.
7686 /// This specifically targets cases where we end up with alternating between
7687 /// the two inputs, and so can permute them into something that feeds a single
7688 /// UNPCK instruction. Note that this routine only targets integer vectors
7689 /// because for floating point vectors we have a generalized SHUFPS lowering
7690 /// strategy that handles everything that doesn't *exactly* match an unpack,
7691 /// making this clever lowering unnecessary.
7692 static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
7693 SDValue V2, ArrayRef<int> Mask,
7694 SelectionDAG &DAG) {
7695 assert(!VT.isFloatingPoint() &&
7696 "This routine only supports integer vectors.");
7697 assert(!isSingleInputShuffleMask(Mask) &&
7698 "This routine should only be used when blending two inputs.");
7699 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7701 int Size = Mask.size();
7703 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7704 return M >= 0 && M % Size < Size / 2;
7706 int NumHiInputs = std::count_if(
7707 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7709 bool UnpackLo = NumLoInputs >= NumHiInputs;
7711 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7712 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7713 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7715 for (int i = 0; i < Size; ++i) {
7719 // Each element of the unpack contains Scale elements from this mask.
7720 int UnpackIdx = i / Scale;
7722 // We only handle the case where V1 feeds the first slots of the unpack.
7723 // We rely on canonicalization to ensure this is the case.
7724 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7727 // Setup the mask for this input. The indexing is tricky as we have to
7728 // handle the unpack stride.
7729 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7730 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7734 // If we will have to shuffle both inputs to use the unpack, check whether
7735 // we can just unpack first and shuffle the result. If so, skip this unpack.
7736 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7737 !isNoopShuffleMask(V2Mask))
7740 // Shuffle the inputs into place.
7741 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7742 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7744 // Cast the inputs to the type we will use to unpack them.
7745 V1 = DAG.getBitcast(UnpackVT, V1);
7746 V2 = DAG.getBitcast(UnpackVT, V2);
7748 // Unpack the inputs and cast the result back to the desired type.
7749 return DAG.getBitcast(
7750 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
7754 // We try each unpack from the largest to the smallest to try and find one
7755 // that fits this mask.
7756 int OrigNumElements = VT.getVectorNumElements();
7757 int OrigScalarSize = VT.getScalarSizeInBits();
7758 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7759 int Scale = ScalarSize / OrigScalarSize;
7760 int NumElements = OrigNumElements / Scale;
7761 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7762 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7766 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7768 if (NumLoInputs == 0 || NumHiInputs == 0) {
7769 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7770 "We have to have *some* inputs!");
7771 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7773 // FIXME: We could consider the total complexity of the permute of each
7774 // possible unpacking. Or at the least we should consider how many
7775 // half-crossings are created.
7776 // FIXME: We could consider commuting the unpacks.
7778 SmallVector<int, 32> PermMask;
7779 PermMask.assign(Size, -1);
7780 for (int i = 0; i < Size; ++i) {
7784 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7787 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7789 return DAG.getVectorShuffle(
7790 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7792 DAG.getUNDEF(VT), PermMask);
7798 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7800 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7801 /// support for floating point shuffles but not integer shuffles. These
7802 /// instructions will incur a domain crossing penalty on some chips though so
7803 /// it is better to avoid lowering through this for integer vectors where
7805 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7806 const X86Subtarget *Subtarget,
7807 SelectionDAG &DAG) {
7809 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7810 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7811 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7812 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7813 ArrayRef<int> Mask = SVOp->getMask();
7814 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7816 if (isSingleInputShuffleMask(Mask)) {
7817 // Use low duplicate instructions for masks that match their pattern.
7818 if (Subtarget->hasSSE3())
7819 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7820 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7822 // Straight shuffle of a single input vector. Simulate this by using the
7823 // single input as both of the "inputs" to this instruction..
7824 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7826 if (Subtarget->hasAVX()) {
7827 // If we have AVX, we can use VPERMILPS which will allow folding a load
7828 // into the shuffle.
7829 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7830 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7833 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
7834 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7836 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7837 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7839 // If we have a single input, insert that into V1 if we can do so cheaply.
7840 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
7841 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7842 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
7844 // Try inverting the insertion since for v2 masks it is easy to do and we
7845 // can't reliably sort the mask one way or the other.
7846 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7847 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7848 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7849 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
7853 // Try to use one of the special instruction patterns to handle two common
7854 // blend patterns if a zero-blend above didn't work.
7855 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
7856 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7857 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
7858 // We can either use a special instruction to load over the low double or
7859 // to move just the low double.
7861 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
7863 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
7865 if (Subtarget->hasSSE41())
7866 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7870 // Use dedicated unpack instructions for masks that match their pattern.
7871 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7872 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7873 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7874 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7876 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7877 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
7878 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7881 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7883 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7884 /// the integer unit to minimize domain crossing penalties. However, for blends
7885 /// it falls back to the floating point shuffle operation with appropriate bit
7887 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7888 const X86Subtarget *Subtarget,
7889 SelectionDAG &DAG) {
7891 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7892 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7893 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7894 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7895 ArrayRef<int> Mask = SVOp->getMask();
7896 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7898 if (isSingleInputShuffleMask(Mask)) {
7899 // Check for being able to broadcast a single element.
7900 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
7901 Mask, Subtarget, DAG))
7904 // Straight shuffle of a single input vector. For everything from SSE2
7905 // onward this has a single fast instruction with no scary immediates.
7906 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7907 V1 = DAG.getBitcast(MVT::v4i32, V1);
7908 int WidenedMask[4] = {
7909 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7910 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7911 return DAG.getBitcast(
7913 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7914 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
7916 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
7917 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
7918 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
7919 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
7921 // If we have a blend of two PACKUS operations an the blend aligns with the
7922 // low and half halves, we can just merge the PACKUS operations. This is
7923 // particularly important as it lets us merge shuffles that this routine itself
7925 auto GetPackNode = [](SDValue V) {
7926 while (V.getOpcode() == ISD::BITCAST)
7927 V = V.getOperand(0);
7929 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
7931 if (SDValue V1Pack = GetPackNode(V1))
7932 if (SDValue V2Pack = GetPackNode(V2))
7933 return DAG.getBitcast(MVT::v2i64,
7934 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
7935 Mask[0] == 0 ? V1Pack.getOperand(0)
7936 : V1Pack.getOperand(1),
7937 Mask[1] == 2 ? V2Pack.getOperand(0)
7938 : V2Pack.getOperand(1)));
7940 // Try to use shift instructions.
7942 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
7945 // When loading a scalar and then shuffling it into a vector we can often do
7946 // the insertion cheaply.
7947 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7948 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7950 // Try inverting the insertion since for v2 masks it is easy to do and we
7951 // can't reliably sort the mask one way or the other.
7952 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
7953 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7954 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
7957 // We have different paths for blend lowering, but they all must use the
7958 // *exact* same predicate.
7959 bool IsBlendSupported = Subtarget->hasSSE41();
7960 if (IsBlendSupported)
7961 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7965 // Use dedicated unpack instructions for masks that match their pattern.
7966 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7967 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7968 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7969 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7971 // Try to use byte rotation instructions.
7972 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7973 if (Subtarget->hasSSSE3())
7974 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7975 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7978 // If we have direct support for blends, we should lower by decomposing into
7979 // a permute. That will be faster than the domain cross.
7980 if (IsBlendSupported)
7981 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
7984 // We implement this with SHUFPD which is pretty lame because it will likely
7985 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7986 // However, all the alternatives are still more cycles and newer chips don't
7987 // have this problem. It would be really nice if x86 had better shuffles here.
7988 V1 = DAG.getBitcast(MVT::v2f64, V1);
7989 V2 = DAG.getBitcast(MVT::v2f64, V2);
7990 return DAG.getBitcast(MVT::v2i64,
7991 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7994 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
7996 /// This is used to disable more specialized lowerings when the shufps lowering
7997 /// will happen to be efficient.
7998 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
7999 // This routine only handles 128-bit shufps.
8000 assert(Mask.size() == 4 && "Unsupported mask size!");
8002 // To lower with a single SHUFPS we need to have the low half and high half
8003 // each requiring a single input.
8004 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8006 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8012 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8014 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8015 /// It makes no assumptions about whether this is the *best* lowering, it simply
8017 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8018 ArrayRef<int> Mask, SDValue V1,
8019 SDValue V2, SelectionDAG &DAG) {
8020 SDValue LowV = V1, HighV = V2;
8021 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8024 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8026 if (NumV2Elements == 1) {
8028 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8031 // Compute the index adjacent to V2Index and in the same half by toggling
8033 int V2AdjIndex = V2Index ^ 1;
8035 if (Mask[V2AdjIndex] == -1) {
8036 // Handles all the cases where we have a single V2 element and an undef.
8037 // This will only ever happen in the high lanes because we commute the
8038 // vector otherwise.
8040 std::swap(LowV, HighV);
8041 NewMask[V2Index] -= 4;
8043 // Handle the case where the V2 element ends up adjacent to a V1 element.
8044 // To make this work, blend them together as the first step.
8045 int V1Index = V2AdjIndex;
8046 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8047 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8048 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8050 // Now proceed to reconstruct the final blend as we have the necessary
8051 // high or low half formed.
8058 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8059 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8061 } else if (NumV2Elements == 2) {
8062 if (Mask[0] < 4 && Mask[1] < 4) {
8063 // Handle the easy case where we have V1 in the low lanes and V2 in the
8067 } else if (Mask[2] < 4 && Mask[3] < 4) {
8068 // We also handle the reversed case because this utility may get called
8069 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8070 // arrange things in the right direction.
8076 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8077 // trying to place elements directly, just blend them and set up the final
8078 // shuffle to place them.
8080 // The first two blend mask elements are for V1, the second two are for
8082 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8083 Mask[2] < 4 ? Mask[2] : Mask[3],
8084 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8085 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8086 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8087 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8089 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8092 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8093 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8094 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8095 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8098 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8099 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8102 /// \brief Lower 4-lane 32-bit floating point shuffles.
8104 /// Uses instructions exclusively from the floating point unit to minimize
8105 /// domain crossing penalties, as these are sufficient to implement all v4f32
8107 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8108 const X86Subtarget *Subtarget,
8109 SelectionDAG &DAG) {
8111 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8112 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8113 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8114 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8115 ArrayRef<int> Mask = SVOp->getMask();
8116 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8119 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8121 if (NumV2Elements == 0) {
8122 // Check for being able to broadcast a single element.
8123 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8124 Mask, Subtarget, DAG))
8127 // Use even/odd duplicate instructions for masks that match their pattern.
8128 if (Subtarget->hasSSE3()) {
8129 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8130 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8131 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8132 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8135 if (Subtarget->hasAVX()) {
8136 // If we have AVX, we can use VPERMILPS which will allow folding a load
8137 // into the shuffle.
8138 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8139 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8142 // Otherwise, use a straight shuffle of a single input vector. We pass the
8143 // input vector to both operands to simulate this with a SHUFPS.
8144 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8145 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8148 // There are special ways we can lower some single-element blends. However, we
8149 // have custom ways we can lower more complex single-element blends below that
8150 // we defer to if both this and BLENDPS fail to match, so restrict this to
8151 // when the V2 input is targeting element 0 of the mask -- that is the fast
8153 if (NumV2Elements == 1 && Mask[0] >= 4)
8154 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8155 Mask, Subtarget, DAG))
8158 if (Subtarget->hasSSE41()) {
8159 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8163 // Use INSERTPS if we can complete the shuffle efficiently.
8164 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8167 if (!isSingleSHUFPSMask(Mask))
8168 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8169 DL, MVT::v4f32, V1, V2, Mask, DAG))
8173 // Use dedicated unpack instructions for masks that match their pattern.
8174 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8175 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
8176 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8177 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
8178 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8179 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
8180 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8181 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
8183 // Otherwise fall back to a SHUFPS lowering strategy.
8184 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8187 /// \brief Lower 4-lane i32 vector shuffles.
8189 /// We try to handle these with integer-domain shuffles where we can, but for
8190 /// blends we use the floating point domain blend instructions.
8191 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8192 const X86Subtarget *Subtarget,
8193 SelectionDAG &DAG) {
8195 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8196 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8197 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8198 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8199 ArrayRef<int> Mask = SVOp->getMask();
8200 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8202 // Whenever we can lower this as a zext, that instruction is strictly faster
8203 // than any alternative. It also allows us to fold memory operands into the
8204 // shuffle in many cases.
8205 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8206 Mask, Subtarget, DAG))
8210 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8212 if (NumV2Elements == 0) {
8213 // Check for being able to broadcast a single element.
8214 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8215 Mask, Subtarget, DAG))
8218 // Straight shuffle of a single input vector. For everything from SSE2
8219 // onward this has a single fast instruction with no scary immediates.
8220 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8221 // but we aren't actually going to use the UNPCK instruction because doing
8222 // so prevents folding a load into this instruction or making a copy.
8223 const int UnpackLoMask[] = {0, 0, 1, 1};
8224 const int UnpackHiMask[] = {2, 2, 3, 3};
8225 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8226 Mask = UnpackLoMask;
8227 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8228 Mask = UnpackHiMask;
8230 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8231 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8234 // Try to use shift instructions.
8236 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8239 // There are special ways we can lower some single-element blends.
8240 if (NumV2Elements == 1)
8241 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8242 Mask, Subtarget, DAG))
8245 // We have different paths for blend lowering, but they all must use the
8246 // *exact* same predicate.
8247 bool IsBlendSupported = Subtarget->hasSSE41();
8248 if (IsBlendSupported)
8249 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8253 if (SDValue Masked =
8254 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8257 // Use dedicated unpack instructions for masks that match their pattern.
8258 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8259 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
8260 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8261 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
8262 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8263 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
8264 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8265 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
8267 // Try to use byte rotation instructions.
8268 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8269 if (Subtarget->hasSSSE3())
8270 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8271 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8274 // If we have direct support for blends, we should lower by decomposing into
8275 // a permute. That will be faster than the domain cross.
8276 if (IsBlendSupported)
8277 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8280 // Try to lower by permuting the inputs into an unpack instruction.
8281 if (SDValue Unpack =
8282 lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
8285 // We implement this with SHUFPS because it can blend from two vectors.
8286 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8287 // up the inputs, bypassing domain shift penalties that we would encur if we
8288 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8290 return DAG.getBitcast(
8292 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8293 DAG.getBitcast(MVT::v4f32, V2), Mask));
8296 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8297 /// shuffle lowering, and the most complex part.
8299 /// The lowering strategy is to try to form pairs of input lanes which are
8300 /// targeted at the same half of the final vector, and then use a dword shuffle
8301 /// to place them onto the right half, and finally unpack the paired lanes into
8302 /// their final position.
8304 /// The exact breakdown of how to form these dword pairs and align them on the
8305 /// correct sides is really tricky. See the comments within the function for
8306 /// more of the details.
8308 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8309 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8310 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8311 /// vector, form the analogous 128-bit 8-element Mask.
8312 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8313 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8314 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8315 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
8316 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8318 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8319 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8320 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8322 SmallVector<int, 4> LoInputs;
8323 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8324 [](int M) { return M >= 0; });
8325 std::sort(LoInputs.begin(), LoInputs.end());
8326 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8327 SmallVector<int, 4> HiInputs;
8328 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8329 [](int M) { return M >= 0; });
8330 std::sort(HiInputs.begin(), HiInputs.end());
8331 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8333 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8334 int NumHToL = LoInputs.size() - NumLToL;
8336 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8337 int NumHToH = HiInputs.size() - NumLToH;
8338 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8339 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8340 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8341 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8343 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8344 // such inputs we can swap two of the dwords across the half mark and end up
8345 // with <=2 inputs to each half in each half. Once there, we can fall through
8346 // to the generic code below. For example:
8348 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8349 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8351 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8352 // and an existing 2-into-2 on the other half. In this case we may have to
8353 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8354 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8355 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8356 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8357 // half than the one we target for fixing) will be fixed when we re-enter this
8358 // path. We will also combine away any sequence of PSHUFD instructions that
8359 // result into a single instruction. Here is an example of the tricky case:
8361 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8362 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8364 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8366 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8367 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8369 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8370 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8372 // The result is fine to be handled by the generic logic.
8373 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8374 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8375 int AOffset, int BOffset) {
8376 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8377 "Must call this with A having 3 or 1 inputs from the A half.");
8378 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8379 "Must call this with B having 1 or 3 inputs from the B half.");
8380 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8381 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8383 bool ThreeAInputs = AToAInputs.size() == 3;
8385 // Compute the index of dword with only one word among the three inputs in
8386 // a half by taking the sum of the half with three inputs and subtracting
8387 // the sum of the actual three inputs. The difference is the remaining
8390 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8391 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8392 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8393 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8394 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8395 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8396 int TripleNonInputIdx =
8397 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8398 TripleDWord = TripleNonInputIdx / 2;
8400 // We use xor with one to compute the adjacent DWord to whichever one the
8402 OneInputDWord = (OneInput / 2) ^ 1;
8404 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8405 // and BToA inputs. If there is also such a problem with the BToB and AToB
8406 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8407 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8408 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8409 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8410 // Compute how many inputs will be flipped by swapping these DWords. We
8412 // to balance this to ensure we don't form a 3-1 shuffle in the other
8414 int NumFlippedAToBInputs =
8415 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8416 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8417 int NumFlippedBToBInputs =
8418 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8419 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8420 if ((NumFlippedAToBInputs == 1 &&
8421 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8422 (NumFlippedBToBInputs == 1 &&
8423 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8424 // We choose whether to fix the A half or B half based on whether that
8425 // half has zero flipped inputs. At zero, we may not be able to fix it
8426 // with that half. We also bias towards fixing the B half because that
8427 // will more commonly be the high half, and we have to bias one way.
8428 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8429 ArrayRef<int> Inputs) {
8430 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8431 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8432 PinnedIdx ^ 1) != Inputs.end();
8433 // Determine whether the free index is in the flipped dword or the
8434 // unflipped dword based on where the pinned index is. We use this bit
8435 // in an xor to conditionally select the adjacent dword.
8436 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8437 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8438 FixFreeIdx) != Inputs.end();
8439 if (IsFixIdxInput == IsFixFreeIdxInput)
8441 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8442 FixFreeIdx) != Inputs.end();
8443 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8444 "We need to be changing the number of flipped inputs!");
8445 int PSHUFHalfMask[] = {0, 1, 2, 3};
8446 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8447 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8449 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8452 if (M != -1 && M == FixIdx)
8454 else if (M != -1 && M == FixFreeIdx)
8457 if (NumFlippedBToBInputs != 0) {
8459 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8460 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8462 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8463 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8464 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8469 int PSHUFDMask[] = {0, 1, 2, 3};
8470 PSHUFDMask[ADWord] = BDWord;
8471 PSHUFDMask[BDWord] = ADWord;
8474 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8475 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8477 // Adjust the mask to match the new locations of A and B.
8479 if (M != -1 && M/2 == ADWord)
8480 M = 2 * BDWord + M % 2;
8481 else if (M != -1 && M/2 == BDWord)
8482 M = 2 * ADWord + M % 2;
8484 // Recurse back into this routine to re-compute state now that this isn't
8485 // a 3 and 1 problem.
8486 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8489 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8490 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8491 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8492 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8494 // At this point there are at most two inputs to the low and high halves from
8495 // each half. That means the inputs can always be grouped into dwords and
8496 // those dwords can then be moved to the correct half with a dword shuffle.
8497 // We use at most one low and one high word shuffle to collect these paired
8498 // inputs into dwords, and finally a dword shuffle to place them.
8499 int PSHUFLMask[4] = {-1, -1, -1, -1};
8500 int PSHUFHMask[4] = {-1, -1, -1, -1};
8501 int PSHUFDMask[4] = {-1, -1, -1, -1};
8503 // First fix the masks for all the inputs that are staying in their
8504 // original halves. This will then dictate the targets of the cross-half
8506 auto fixInPlaceInputs =
8507 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8508 MutableArrayRef<int> SourceHalfMask,
8509 MutableArrayRef<int> HalfMask, int HalfOffset) {
8510 if (InPlaceInputs.empty())
8512 if (InPlaceInputs.size() == 1) {
8513 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8514 InPlaceInputs[0] - HalfOffset;
8515 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8518 if (IncomingInputs.empty()) {
8519 // Just fix all of the in place inputs.
8520 for (int Input : InPlaceInputs) {
8521 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8522 PSHUFDMask[Input / 2] = Input / 2;
8527 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8528 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8529 InPlaceInputs[0] - HalfOffset;
8530 // Put the second input next to the first so that they are packed into
8531 // a dword. We find the adjacent index by toggling the low bit.
8532 int AdjIndex = InPlaceInputs[0] ^ 1;
8533 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8534 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8535 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8537 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8538 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8540 // Now gather the cross-half inputs and place them into a free dword of
8541 // their target half.
8542 // FIXME: This operation could almost certainly be simplified dramatically to
8543 // look more like the 3-1 fixing operation.
8544 auto moveInputsToRightHalf = [&PSHUFDMask](
8545 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8546 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8547 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8549 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8550 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8552 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8554 int LowWord = Word & ~1;
8555 int HighWord = Word | 1;
8556 return isWordClobbered(SourceHalfMask, LowWord) ||
8557 isWordClobbered(SourceHalfMask, HighWord);
8560 if (IncomingInputs.empty())
8563 if (ExistingInputs.empty()) {
8564 // Map any dwords with inputs from them into the right half.
8565 for (int Input : IncomingInputs) {
8566 // If the source half mask maps over the inputs, turn those into
8567 // swaps and use the swapped lane.
8568 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8569 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8570 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8571 Input - SourceOffset;
8572 // We have to swap the uses in our half mask in one sweep.
8573 for (int &M : HalfMask)
8574 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8576 else if (M == Input)
8577 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8579 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8580 Input - SourceOffset &&
8581 "Previous placement doesn't match!");
8583 // Note that this correctly re-maps both when we do a swap and when
8584 // we observe the other side of the swap above. We rely on that to
8585 // avoid swapping the members of the input list directly.
8586 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8589 // Map the input's dword into the correct half.
8590 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8591 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8593 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8595 "Previous placement doesn't match!");
8598 // And just directly shift any other-half mask elements to be same-half
8599 // as we will have mirrored the dword containing the element into the
8600 // same position within that half.
8601 for (int &M : HalfMask)
8602 if (M >= SourceOffset && M < SourceOffset + 4) {
8603 M = M - SourceOffset + DestOffset;
8604 assert(M >= 0 && "This should never wrap below zero!");
8609 // Ensure we have the input in a viable dword of its current half. This
8610 // is particularly tricky because the original position may be clobbered
8611 // by inputs being moved and *staying* in that half.
8612 if (IncomingInputs.size() == 1) {
8613 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8614 int InputFixed = std::find(std::begin(SourceHalfMask),
8615 std::end(SourceHalfMask), -1) -
8616 std::begin(SourceHalfMask) + SourceOffset;
8617 SourceHalfMask[InputFixed - SourceOffset] =
8618 IncomingInputs[0] - SourceOffset;
8619 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8621 IncomingInputs[0] = InputFixed;
8623 } else if (IncomingInputs.size() == 2) {
8624 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8625 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8626 // We have two non-adjacent or clobbered inputs we need to extract from
8627 // the source half. To do this, we need to map them into some adjacent
8628 // dword slot in the source mask.
8629 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8630 IncomingInputs[1] - SourceOffset};
8632 // If there is a free slot in the source half mask adjacent to one of
8633 // the inputs, place the other input in it. We use (Index XOR 1) to
8634 // compute an adjacent index.
8635 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8636 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8637 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8638 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8639 InputsFixed[1] = InputsFixed[0] ^ 1;
8640 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8641 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8642 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8643 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8644 InputsFixed[0] = InputsFixed[1] ^ 1;
8645 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8646 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8647 // The two inputs are in the same DWord but it is clobbered and the
8648 // adjacent DWord isn't used at all. Move both inputs to the free
8650 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8651 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8652 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8653 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8655 // The only way we hit this point is if there is no clobbering
8656 // (because there are no off-half inputs to this half) and there is no
8657 // free slot adjacent to one of the inputs. In this case, we have to
8658 // swap an input with a non-input.
8659 for (int i = 0; i < 4; ++i)
8660 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8661 "We can't handle any clobbers here!");
8662 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8663 "Cannot have adjacent inputs here!");
8665 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8666 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8668 // We also have to update the final source mask in this case because
8669 // it may need to undo the above swap.
8670 for (int &M : FinalSourceHalfMask)
8671 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8672 M = InputsFixed[1] + SourceOffset;
8673 else if (M == InputsFixed[1] + SourceOffset)
8674 M = (InputsFixed[0] ^ 1) + SourceOffset;
8676 InputsFixed[1] = InputsFixed[0] ^ 1;
8679 // Point everything at the fixed inputs.
8680 for (int &M : HalfMask)
8681 if (M == IncomingInputs[0])
8682 M = InputsFixed[0] + SourceOffset;
8683 else if (M == IncomingInputs[1])
8684 M = InputsFixed[1] + SourceOffset;
8686 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8687 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8690 llvm_unreachable("Unhandled input size!");
8693 // Now hoist the DWord down to the right half.
8694 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8695 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8696 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8697 for (int &M : HalfMask)
8698 for (int Input : IncomingInputs)
8700 M = FreeDWord * 2 + Input % 2;
8702 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8703 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8704 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8705 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8707 // Now enact all the shuffles we've computed to move the inputs into their
8709 if (!isNoopShuffleMask(PSHUFLMask))
8710 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8711 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
8712 if (!isNoopShuffleMask(PSHUFHMask))
8713 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8714 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
8715 if (!isNoopShuffleMask(PSHUFDMask))
8718 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8719 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8721 // At this point, each half should contain all its inputs, and we can then
8722 // just shuffle them into their final position.
8723 assert(std::count_if(LoMask.begin(), LoMask.end(),
8724 [](int M) { return M >= 4; }) == 0 &&
8725 "Failed to lift all the high half inputs to the low mask!");
8726 assert(std::count_if(HiMask.begin(), HiMask.end(),
8727 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8728 "Failed to lift all the low half inputs to the high mask!");
8730 // Do a half shuffle for the low mask.
8731 if (!isNoopShuffleMask(LoMask))
8732 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8733 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
8735 // Do a half shuffle with the high mask after shifting its values down.
8736 for (int &M : HiMask)
8739 if (!isNoopShuffleMask(HiMask))
8740 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8741 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
8746 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8747 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8748 SDValue V2, ArrayRef<int> Mask,
8749 SelectionDAG &DAG, bool &V1InUse,
8751 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8757 int Size = Mask.size();
8758 int Scale = 16 / Size;
8759 for (int i = 0; i < 16; ++i) {
8760 if (Mask[i / Scale] == -1) {
8761 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8763 const int ZeroMask = 0x80;
8764 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8766 int V2Idx = Mask[i / Scale] < Size
8768 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8769 if (Zeroable[i / Scale])
8770 V1Idx = V2Idx = ZeroMask;
8771 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
8772 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
8773 V1InUse |= (ZeroMask != V1Idx);
8774 V2InUse |= (ZeroMask != V2Idx);
8779 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8780 DAG.getBitcast(MVT::v16i8, V1),
8781 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8783 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8784 DAG.getBitcast(MVT::v16i8, V2),
8785 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8787 // If we need shuffled inputs from both, blend the two.
8789 if (V1InUse && V2InUse)
8790 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8792 V = V1InUse ? V1 : V2;
8794 // Cast the result back to the correct type.
8795 return DAG.getBitcast(VT, V);
8798 /// \brief Generic lowering of 8-lane i16 shuffles.
8800 /// This handles both single-input shuffles and combined shuffle/blends with
8801 /// two inputs. The single input shuffles are immediately delegated to
8802 /// a dedicated lowering routine.
8804 /// The blends are lowered in one of three fundamental ways. If there are few
8805 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8806 /// of the input is significantly cheaper when lowered as an interleaving of
8807 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8808 /// halves of the inputs separately (making them have relatively few inputs)
8809 /// and then concatenate them.
8810 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8811 const X86Subtarget *Subtarget,
8812 SelectionDAG &DAG) {
8814 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8815 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8816 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8817 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8818 ArrayRef<int> OrigMask = SVOp->getMask();
8819 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8820 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8821 MutableArrayRef<int> Mask(MaskStorage);
8823 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8825 // Whenever we can lower this as a zext, that instruction is strictly faster
8826 // than any alternative.
8827 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8828 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8831 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8833 auto isV2 = [](int M) { return M >= 8; };
8835 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8837 if (NumV2Inputs == 0) {
8838 // Check for being able to broadcast a single element.
8839 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
8840 Mask, Subtarget, DAG))
8843 // Try to use shift instructions.
8845 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
8848 // Use dedicated unpack instructions for masks that match their pattern.
8849 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
8850 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
8851 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
8852 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
8854 // Try to use byte rotation instructions.
8855 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
8856 Mask, Subtarget, DAG))
8859 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
8863 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
8864 "All single-input shuffles should be canonicalized to be V1-input "
8867 // Try to use shift instructions.
8869 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
8872 // See if we can use SSE4A Extraction / Insertion.
8873 if (Subtarget->hasSSE4A())
8874 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
8877 // There are special ways we can lower some single-element blends.
8878 if (NumV2Inputs == 1)
8879 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
8880 Mask, Subtarget, DAG))
8883 // We have different paths for blend lowering, but they all must use the
8884 // *exact* same predicate.
8885 bool IsBlendSupported = Subtarget->hasSSE41();
8886 if (IsBlendSupported)
8887 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8891 if (SDValue Masked =
8892 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
8895 // Use dedicated unpack instructions for masks that match their pattern.
8896 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
8897 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
8898 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
8899 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
8901 // Try to use byte rotation instructions.
8902 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8903 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
8906 if (SDValue BitBlend =
8907 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
8910 if (SDValue Unpack =
8911 lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
8914 // If we can't directly blend but can use PSHUFB, that will be better as it
8915 // can both shuffle and set up the inefficient blend.
8916 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
8917 bool V1InUse, V2InUse;
8918 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
8922 // We can always bit-blend if we have to so the fallback strategy is to
8923 // decompose into single-input permutes and blends.
8924 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
8928 /// \brief Check whether a compaction lowering can be done by dropping even
8929 /// elements and compute how many times even elements must be dropped.
8931 /// This handles shuffles which take every Nth element where N is a power of
8932 /// two. Example shuffle masks:
8934 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8935 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8936 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8937 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8938 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8939 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8941 /// Any of these lanes can of course be undef.
8943 /// This routine only supports N <= 3.
8944 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8947 /// \returns N above, or the number of times even elements must be dropped if
8948 /// there is such a number. Otherwise returns zero.
8949 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8950 // Figure out whether we're looping over two inputs or just one.
8951 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8953 // The modulus for the shuffle vector entries is based on whether this is
8954 // a single input or not.
8955 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8956 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8957 "We should only be called with masks with a power-of-2 size!");
8959 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8961 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8962 // and 2^3 simultaneously. This is because we may have ambiguity with
8963 // partially undef inputs.
8964 bool ViableForN[3] = {true, true, true};
8966 for (int i = 0, e = Mask.size(); i < e; ++i) {
8967 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8972 bool IsAnyViable = false;
8973 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8974 if (ViableForN[j]) {
8977 // The shuffle mask must be equal to (i * 2^N) % M.
8978 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8981 ViableForN[j] = false;
8983 // Early exit if we exhaust the possible powers of two.
8988 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8992 // Return 0 as there is no viable power of two.
8996 /// \brief Generic lowering of v16i8 shuffles.
8998 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8999 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9000 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9001 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9003 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9004 const X86Subtarget *Subtarget,
9005 SelectionDAG &DAG) {
9007 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9008 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9009 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9010 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9011 ArrayRef<int> Mask = SVOp->getMask();
9012 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9014 // Try to use shift instructions.
9016 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9019 // Try to use byte rotation instructions.
9020 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9021 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9024 // Try to use a zext lowering.
9025 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9026 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9029 // See if we can use SSE4A Extraction / Insertion.
9030 if (Subtarget->hasSSE4A())
9031 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9035 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9037 // For single-input shuffles, there are some nicer lowering tricks we can use.
9038 if (NumV2Elements == 0) {
9039 // Check for being able to broadcast a single element.
9040 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9041 Mask, Subtarget, DAG))
9044 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9045 // Notably, this handles splat and partial-splat shuffles more efficiently.
9046 // However, it only makes sense if the pre-duplication shuffle simplifies
9047 // things significantly. Currently, this means we need to be able to
9048 // express the pre-duplication shuffle as an i16 shuffle.
9050 // FIXME: We should check for other patterns which can be widened into an
9051 // i16 shuffle as well.
9052 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9053 for (int i = 0; i < 16; i += 2)
9054 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9059 auto tryToWidenViaDuplication = [&]() -> SDValue {
9060 if (!canWidenViaDuplication(Mask))
9062 SmallVector<int, 4> LoInputs;
9063 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9064 [](int M) { return M >= 0 && M < 8; });
9065 std::sort(LoInputs.begin(), LoInputs.end());
9066 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9068 SmallVector<int, 4> HiInputs;
9069 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9070 [](int M) { return M >= 8; });
9071 std::sort(HiInputs.begin(), HiInputs.end());
9072 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9075 bool TargetLo = LoInputs.size() >= HiInputs.size();
9076 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9077 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9079 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9080 SmallDenseMap<int, int, 8> LaneMap;
9081 for (int I : InPlaceInputs) {
9082 PreDupI16Shuffle[I/2] = I/2;
9085 int j = TargetLo ? 0 : 4, je = j + 4;
9086 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9087 // Check if j is already a shuffle of this input. This happens when
9088 // there are two adjacent bytes after we move the low one.
9089 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9090 // If we haven't yet mapped the input, search for a slot into which
9092 while (j < je && PreDupI16Shuffle[j] != -1)
9096 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9099 // Map this input with the i16 shuffle.
9100 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9103 // Update the lane map based on the mapping we ended up with.
9104 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9106 V1 = DAG.getBitcast(
9108 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9109 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9111 // Unpack the bytes to form the i16s that will be shuffled into place.
9112 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9113 MVT::v16i8, V1, V1);
9115 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9116 for (int i = 0; i < 16; ++i)
9117 if (Mask[i] != -1) {
9118 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9119 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9120 if (PostDupI16Shuffle[i / 2] == -1)
9121 PostDupI16Shuffle[i / 2] = MappedMask;
9123 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9124 "Conflicting entrties in the original shuffle!");
9126 return DAG.getBitcast(
9128 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9129 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9131 if (SDValue V = tryToWidenViaDuplication())
9135 if (SDValue Masked =
9136 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9139 // Use dedicated unpack instructions for masks that match their pattern.
9140 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9141 0, 16, 1, 17, 2, 18, 3, 19,
9143 4, 20, 5, 21, 6, 22, 7, 23}))
9144 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
9145 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9146 8, 24, 9, 25, 10, 26, 11, 27,
9148 12, 28, 13, 29, 14, 30, 15, 31}))
9149 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
9151 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9152 // with PSHUFB. It is important to do this before we attempt to generate any
9153 // blends but after all of the single-input lowerings. If the single input
9154 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9155 // want to preserve that and we can DAG combine any longer sequences into
9156 // a PSHUFB in the end. But once we start blending from multiple inputs,
9157 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9158 // and there are *very* few patterns that would actually be faster than the
9159 // PSHUFB approach because of its ability to zero lanes.
9161 // FIXME: The only exceptions to the above are blends which are exact
9162 // interleavings with direct instructions supporting them. We currently don't
9163 // handle those well here.
9164 if (Subtarget->hasSSSE3()) {
9165 bool V1InUse = false;
9166 bool V2InUse = false;
9168 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9169 DAG, V1InUse, V2InUse);
9171 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9172 // do so. This avoids using them to handle blends-with-zero which is
9173 // important as a single pshufb is significantly faster for that.
9174 if (V1InUse && V2InUse) {
9175 if (Subtarget->hasSSE41())
9176 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9177 Mask, Subtarget, DAG))
9180 // We can use an unpack to do the blending rather than an or in some
9181 // cases. Even though the or may be (very minorly) more efficient, we
9182 // preference this lowering because there are common cases where part of
9183 // the complexity of the shuffles goes away when we do the final blend as
9185 // FIXME: It might be worth trying to detect if the unpack-feeding
9186 // shuffles will both be pshufb, in which case we shouldn't bother with
9188 if (SDValue Unpack =
9189 lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
9196 // There are special ways we can lower some single-element blends.
9197 if (NumV2Elements == 1)
9198 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9199 Mask, Subtarget, DAG))
9202 if (SDValue BitBlend =
9203 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9206 // Check whether a compaction lowering can be done. This handles shuffles
9207 // which take every Nth element for some even N. See the helper function for
9210 // We special case these as they can be particularly efficiently handled with
9211 // the PACKUSB instruction on x86 and they show up in common patterns of
9212 // rearranging bytes to truncate wide elements.
9213 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9214 // NumEvenDrops is the power of two stride of the elements. Another way of
9215 // thinking about it is that we need to drop the even elements this many
9216 // times to get the original input.
9217 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9219 // First we need to zero all the dropped bytes.
9220 assert(NumEvenDrops <= 3 &&
9221 "No support for dropping even elements more than 3 times.");
9222 // We use the mask type to pick which bytes are preserved based on how many
9223 // elements are dropped.
9224 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9225 SDValue ByteClearMask = DAG.getBitcast(
9226 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9227 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9229 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9231 // Now pack things back together.
9232 V1 = DAG.getBitcast(MVT::v8i16, V1);
9233 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9234 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9235 for (int i = 1; i < NumEvenDrops; ++i) {
9236 Result = DAG.getBitcast(MVT::v8i16, Result);
9237 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9243 // Handle multi-input cases by blending single-input shuffles.
9244 if (NumV2Elements > 0)
9245 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9248 // The fallback path for single-input shuffles widens this into two v8i16
9249 // vectors with unpacks, shuffles those, and then pulls them back together
9253 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9254 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9255 for (int i = 0; i < 16; ++i)
9257 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9259 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9261 SDValue VLoHalf, VHiHalf;
9262 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9263 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9265 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9266 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9267 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9268 [](int M) { return M >= 0 && M % 2 == 1; })) {
9269 // Use a mask to drop the high bytes.
9270 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9271 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9272 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9274 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9275 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9277 // Squash the masks to point directly into VLoHalf.
9278 for (int &M : LoBlendMask)
9281 for (int &M : HiBlendMask)
9285 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9286 // VHiHalf so that we can blend them as i16s.
9287 VLoHalf = DAG.getBitcast(
9288 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9289 VHiHalf = DAG.getBitcast(
9290 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9293 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9294 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9296 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9299 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9301 /// This routine breaks down the specific type of 128-bit shuffle and
9302 /// dispatches to the lowering routines accordingly.
9303 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9304 MVT VT, const X86Subtarget *Subtarget,
9305 SelectionDAG &DAG) {
9306 switch (VT.SimpleTy) {
9308 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9310 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9312 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9314 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9316 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9318 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9321 llvm_unreachable("Unimplemented!");
9325 /// \brief Helper function to test whether a shuffle mask could be
9326 /// simplified by widening the elements being shuffled.
9328 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9329 /// leaves it in an unspecified state.
9331 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9332 /// shuffle masks. The latter have the special property of a '-2' representing
9333 /// a zero-ed lane of a vector.
9334 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9335 SmallVectorImpl<int> &WidenedMask) {
9336 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9337 // If both elements are undef, its trivial.
9338 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9339 WidenedMask.push_back(SM_SentinelUndef);
9343 // Check for an undef mask and a mask value properly aligned to fit with
9344 // a pair of values. If we find such a case, use the non-undef mask's value.
9345 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9346 WidenedMask.push_back(Mask[i + 1] / 2);
9349 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9350 WidenedMask.push_back(Mask[i] / 2);
9354 // When zeroing, we need to spread the zeroing across both lanes to widen.
9355 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9356 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9357 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9358 WidenedMask.push_back(SM_SentinelZero);
9364 // Finally check if the two mask values are adjacent and aligned with
9366 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9367 WidenedMask.push_back(Mask[i] / 2);
9371 // Otherwise we can't safely widen the elements used in this shuffle.
9374 assert(WidenedMask.size() == Mask.size() / 2 &&
9375 "Incorrect size of mask after widening the elements!");
9380 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9382 /// This routine just extracts two subvectors, shuffles them independently, and
9383 /// then concatenates them back together. This should work effectively with all
9384 /// AVX vector shuffle types.
9385 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9386 SDValue V2, ArrayRef<int> Mask,
9387 SelectionDAG &DAG) {
9388 assert(VT.getSizeInBits() >= 256 &&
9389 "Only for 256-bit or wider vector shuffles!");
9390 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9391 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9393 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9394 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9396 int NumElements = VT.getVectorNumElements();
9397 int SplitNumElements = NumElements / 2;
9398 MVT ScalarVT = VT.getScalarType();
9399 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9401 // Rather than splitting build-vectors, just build two narrower build
9402 // vectors. This helps shuffling with splats and zeros.
9403 auto SplitVector = [&](SDValue V) {
9404 while (V.getOpcode() == ISD::BITCAST)
9405 V = V->getOperand(0);
9407 MVT OrigVT = V.getSimpleValueType();
9408 int OrigNumElements = OrigVT.getVectorNumElements();
9409 int OrigSplitNumElements = OrigNumElements / 2;
9410 MVT OrigScalarVT = OrigVT.getScalarType();
9411 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9415 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9417 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9418 DAG.getIntPtrConstant(0, DL));
9419 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9420 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9423 SmallVector<SDValue, 16> LoOps, HiOps;
9424 for (int i = 0; i < OrigSplitNumElements; ++i) {
9425 LoOps.push_back(BV->getOperand(i));
9426 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9428 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9429 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9431 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9432 DAG.getBitcast(SplitVT, HiV));
9435 SDValue LoV1, HiV1, LoV2, HiV2;
9436 std::tie(LoV1, HiV1) = SplitVector(V1);
9437 std::tie(LoV2, HiV2) = SplitVector(V2);
9439 // Now create two 4-way blends of these half-width vectors.
9440 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9441 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9442 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9443 for (int i = 0; i < SplitNumElements; ++i) {
9444 int M = HalfMask[i];
9445 if (M >= NumElements) {
9446 if (M >= NumElements + SplitNumElements)
9450 V2BlendMask.push_back(M - NumElements);
9451 V1BlendMask.push_back(-1);
9452 BlendMask.push_back(SplitNumElements + i);
9453 } else if (M >= 0) {
9454 if (M >= SplitNumElements)
9458 V2BlendMask.push_back(-1);
9459 V1BlendMask.push_back(M);
9460 BlendMask.push_back(i);
9462 V2BlendMask.push_back(-1);
9463 V1BlendMask.push_back(-1);
9464 BlendMask.push_back(-1);
9468 // Because the lowering happens after all combining takes place, we need to
9469 // manually combine these blend masks as much as possible so that we create
9470 // a minimal number of high-level vector shuffle nodes.
9472 // First try just blending the halves of V1 or V2.
9473 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9474 return DAG.getUNDEF(SplitVT);
9475 if (!UseLoV2 && !UseHiV2)
9476 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9477 if (!UseLoV1 && !UseHiV1)
9478 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9480 SDValue V1Blend, V2Blend;
9481 if (UseLoV1 && UseHiV1) {
9483 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9485 // We only use half of V1 so map the usage down into the final blend mask.
9486 V1Blend = UseLoV1 ? LoV1 : HiV1;
9487 for (int i = 0; i < SplitNumElements; ++i)
9488 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9489 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9491 if (UseLoV2 && UseHiV2) {
9493 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9495 // We only use half of V2 so map the usage down into the final blend mask.
9496 V2Blend = UseLoV2 ? LoV2 : HiV2;
9497 for (int i = 0; i < SplitNumElements; ++i)
9498 if (BlendMask[i] >= SplitNumElements)
9499 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9501 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9503 SDValue Lo = HalfBlend(LoMask);
9504 SDValue Hi = HalfBlend(HiMask);
9505 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9508 /// \brief Either split a vector in halves or decompose the shuffles and the
9511 /// This is provided as a good fallback for many lowerings of non-single-input
9512 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9513 /// between splitting the shuffle into 128-bit components and stitching those
9514 /// back together vs. extracting the single-input shuffles and blending those
9516 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9517 SDValue V2, ArrayRef<int> Mask,
9518 SelectionDAG &DAG) {
9519 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9520 "lower single-input shuffles as it "
9521 "could then recurse on itself.");
9522 int Size = Mask.size();
9524 // If this can be modeled as a broadcast of two elements followed by a blend,
9525 // prefer that lowering. This is especially important because broadcasts can
9526 // often fold with memory operands.
9527 auto DoBothBroadcast = [&] {
9528 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9531 if (V2BroadcastIdx == -1)
9532 V2BroadcastIdx = M - Size;
9533 else if (M - Size != V2BroadcastIdx)
9535 } else if (M >= 0) {
9536 if (V1BroadcastIdx == -1)
9538 else if (M != V1BroadcastIdx)
9543 if (DoBothBroadcast())
9544 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9547 // If the inputs all stem from a single 128-bit lane of each input, then we
9548 // split them rather than blending because the split will decompose to
9549 // unusually few instructions.
9550 int LaneCount = VT.getSizeInBits() / 128;
9551 int LaneSize = Size / LaneCount;
9552 SmallBitVector LaneInputs[2];
9553 LaneInputs[0].resize(LaneCount, false);
9554 LaneInputs[1].resize(LaneCount, false);
9555 for (int i = 0; i < Size; ++i)
9557 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9558 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9559 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9561 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9562 // that the decomposed single-input shuffles don't end up here.
9563 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9566 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9567 /// a permutation and blend of those lanes.
9569 /// This essentially blends the out-of-lane inputs to each lane into the lane
9570 /// from a permuted copy of the vector. This lowering strategy results in four
9571 /// instructions in the worst case for a single-input cross lane shuffle which
9572 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9573 /// of. Special cases for each particular shuffle pattern should be handled
9574 /// prior to trying this lowering.
9575 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9576 SDValue V1, SDValue V2,
9578 SelectionDAG &DAG) {
9579 // FIXME: This should probably be generalized for 512-bit vectors as well.
9580 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9581 int LaneSize = Mask.size() / 2;
9583 // If there are only inputs from one 128-bit lane, splitting will in fact be
9584 // less expensive. The flags track whether the given lane contains an element
9585 // that crosses to another lane.
9586 bool LaneCrossing[2] = {false, false};
9587 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9588 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9589 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9590 if (!LaneCrossing[0] || !LaneCrossing[1])
9591 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9593 if (isSingleInputShuffleMask(Mask)) {
9594 SmallVector<int, 32> FlippedBlendMask;
9595 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9596 FlippedBlendMask.push_back(
9597 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9599 : Mask[i] % LaneSize +
9600 (i / LaneSize) * LaneSize + Size));
9602 // Flip the vector, and blend the results which should now be in-lane. The
9603 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9604 // 5 for the high source. The value 3 selects the high half of source 2 and
9605 // the value 2 selects the low half of source 2. We only use source 2 to
9606 // allow folding it into a memory operand.
9607 unsigned PERMMask = 3 | 2 << 4;
9608 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9609 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9610 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9613 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9614 // will be handled by the above logic and a blend of the results, much like
9615 // other patterns in AVX.
9616 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9619 /// \brief Handle lowering 2-lane 128-bit shuffles.
9620 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9621 SDValue V2, ArrayRef<int> Mask,
9622 const X86Subtarget *Subtarget,
9623 SelectionDAG &DAG) {
9624 // TODO: If minimizing size and one of the inputs is a zero vector and the
9625 // the zero vector has only one use, we could use a VPERM2X128 to save the
9626 // instruction bytes needed to explicitly generate the zero vector.
9628 // Blends are faster and handle all the non-lane-crossing cases.
9629 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9633 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9634 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9636 // If either input operand is a zero vector, use VPERM2X128 because its mask
9637 // allows us to replace the zero input with an implicit zero.
9638 if (!IsV1Zero && !IsV2Zero) {
9639 // Check for patterns which can be matched with a single insert of a 128-bit
9641 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9642 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9643 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9644 VT.getVectorNumElements() / 2);
9645 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9646 DAG.getIntPtrConstant(0, DL));
9647 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9648 OnlyUsesV1 ? V1 : V2,
9649 DAG.getIntPtrConstant(0, DL));
9650 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9654 // Otherwise form a 128-bit permutation. After accounting for undefs,
9655 // convert the 64-bit shuffle mask selection values into 128-bit
9656 // selection bits by dividing the indexes by 2 and shifting into positions
9657 // defined by a vperm2*128 instruction's immediate control byte.
9659 // The immediate permute control byte looks like this:
9660 // [1:0] - select 128 bits from sources for low half of destination
9662 // [3] - zero low half of destination
9663 // [5:4] - select 128 bits from sources for high half of destination
9665 // [7] - zero high half of destination
9667 int MaskLO = Mask[0];
9668 if (MaskLO == SM_SentinelUndef)
9669 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
9671 int MaskHI = Mask[2];
9672 if (MaskHI == SM_SentinelUndef)
9673 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
9675 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
9677 // If either input is a zero vector, replace it with an undef input.
9678 // Shuffle mask values < 4 are selecting elements of V1.
9679 // Shuffle mask values >= 4 are selecting elements of V2.
9680 // Adjust each half of the permute mask by clearing the half that was
9681 // selecting the zero vector and setting the zero mask bit.
9683 V1 = DAG.getUNDEF(VT);
9685 PermMask = (PermMask & 0xf0) | 0x08;
9687 PermMask = (PermMask & 0x0f) | 0x80;
9690 V2 = DAG.getUNDEF(VT);
9692 PermMask = (PermMask & 0xf0) | 0x08;
9694 PermMask = (PermMask & 0x0f) | 0x80;
9697 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9698 DAG.getConstant(PermMask, DL, MVT::i8));
9701 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9702 /// shuffling each lane.
9704 /// This will only succeed when the result of fixing the 128-bit lanes results
9705 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9706 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9707 /// the lane crosses early and then use simpler shuffles within each lane.
9709 /// FIXME: It might be worthwhile at some point to support this without
9710 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9711 /// in x86 only floating point has interesting non-repeating shuffles, and even
9712 /// those are still *marginally* more expensive.
9713 static SDValue lowerVectorShuffleByMerging128BitLanes(
9714 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9715 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9716 assert(!isSingleInputShuffleMask(Mask) &&
9717 "This is only useful with multiple inputs.");
9719 int Size = Mask.size();
9720 int LaneSize = 128 / VT.getScalarSizeInBits();
9721 int NumLanes = Size / LaneSize;
9722 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9724 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9725 // check whether the in-128-bit lane shuffles share a repeating pattern.
9726 SmallVector<int, 4> Lanes;
9727 Lanes.resize(NumLanes, -1);
9728 SmallVector<int, 4> InLaneMask;
9729 InLaneMask.resize(LaneSize, -1);
9730 for (int i = 0; i < Size; ++i) {
9734 int j = i / LaneSize;
9737 // First entry we've seen for this lane.
9738 Lanes[j] = Mask[i] / LaneSize;
9739 } else if (Lanes[j] != Mask[i] / LaneSize) {
9740 // This doesn't match the lane selected previously!
9744 // Check that within each lane we have a consistent shuffle mask.
9745 int k = i % LaneSize;
9746 if (InLaneMask[k] < 0) {
9747 InLaneMask[k] = Mask[i] % LaneSize;
9748 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9749 // This doesn't fit a repeating in-lane mask.
9754 // First shuffle the lanes into place.
9755 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9756 VT.getSizeInBits() / 64);
9757 SmallVector<int, 8> LaneMask;
9758 LaneMask.resize(NumLanes * 2, -1);
9759 for (int i = 0; i < NumLanes; ++i)
9760 if (Lanes[i] >= 0) {
9761 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9762 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9765 V1 = DAG.getBitcast(LaneVT, V1);
9766 V2 = DAG.getBitcast(LaneVT, V2);
9767 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9769 // Cast it back to the type we actually want.
9770 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
9772 // Now do a simple shuffle that isn't lane crossing.
9773 SmallVector<int, 8> NewMask;
9774 NewMask.resize(Size, -1);
9775 for (int i = 0; i < Size; ++i)
9777 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9778 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9779 "Must not introduce lane crosses at this point!");
9781 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9784 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9787 /// This returns true if the elements from a particular input are already in the
9788 /// slot required by the given mask and require no permutation.
9789 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9790 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9791 int Size = Mask.size();
9792 for (int i = 0; i < Size; ++i)
9793 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9799 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
9800 ArrayRef<int> Mask, SDValue V1,
9801 SDValue V2, SelectionDAG &DAG) {
9803 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
9804 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
9805 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
9806 int NumElts = VT.getVectorNumElements();
9807 bool ShufpdMask = true;
9808 bool CommutableMask = true;
9809 unsigned Immediate = 0;
9810 for (int i = 0; i < NumElts; ++i) {
9813 int Val = (i & 6) + NumElts * (i & 1);
9814 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
9815 if (Mask[i] < Val || Mask[i] > Val + 1)
9817 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
9818 CommutableMask = false;
9819 Immediate |= (Mask[i] % 2) << i;
9822 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
9823 DAG.getConstant(Immediate, DL, MVT::i8));
9825 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
9826 DAG.getConstant(Immediate, DL, MVT::i8));
9830 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9832 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9833 /// isn't available.
9834 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9835 const X86Subtarget *Subtarget,
9836 SelectionDAG &DAG) {
9838 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9839 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9840 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9841 ArrayRef<int> Mask = SVOp->getMask();
9842 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9844 SmallVector<int, 4> WidenedMask;
9845 if (canWidenShuffleElements(Mask, WidenedMask))
9846 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
9849 if (isSingleInputShuffleMask(Mask)) {
9850 // Check for being able to broadcast a single element.
9851 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
9852 Mask, Subtarget, DAG))
9855 // Use low duplicate instructions for masks that match their pattern.
9856 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
9857 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
9859 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9860 // Non-half-crossing single input shuffles can be lowerid with an
9861 // interleaved permutation.
9862 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9863 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9864 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9865 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
9868 // With AVX2 we have direct support for this permutation.
9869 if (Subtarget->hasAVX2())
9870 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9871 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9873 // Otherwise, fall back.
9874 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9878 // X86 has dedicated unpack instructions that can handle specific blend
9879 // operations: UNPCKH and UNPCKL.
9880 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9881 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9882 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9883 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9884 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9885 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
9886 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9887 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
9889 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9893 // Check if the blend happens to exactly fit that of SHUFPD.
9895 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
9898 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9899 // shuffle. However, if we have AVX2 and either inputs are already in place,
9900 // we will be able to shuffle even across lanes the other input in a single
9901 // instruction so skip this pattern.
9902 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9903 isShuffleMaskInputInPlace(1, Mask))))
9904 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9905 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9908 // If we have AVX2 then we always want to lower with a blend because an v4 we
9909 // can fully permute the elements.
9910 if (Subtarget->hasAVX2())
9911 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9914 // Otherwise fall back on generic lowering.
9915 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
9918 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9920 /// This routine is only called when we have AVX2 and thus a reasonable
9921 /// instruction set for v4i64 shuffling..
9922 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9923 const X86Subtarget *Subtarget,
9924 SelectionDAG &DAG) {
9926 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9927 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9928 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9929 ArrayRef<int> Mask = SVOp->getMask();
9930 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9931 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9933 SmallVector<int, 4> WidenedMask;
9934 if (canWidenShuffleElements(Mask, WidenedMask))
9935 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
9938 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9942 // Check for being able to broadcast a single element.
9943 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
9944 Mask, Subtarget, DAG))
9947 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9948 // use lower latency instructions that will operate on both 128-bit lanes.
9949 SmallVector<int, 2> RepeatedMask;
9950 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9951 if (isSingleInputShuffleMask(Mask)) {
9952 int PSHUFDMask[] = {-1, -1, -1, -1};
9953 for (int i = 0; i < 2; ++i)
9954 if (RepeatedMask[i] >= 0) {
9955 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9956 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9958 return DAG.getBitcast(
9960 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9961 DAG.getBitcast(MVT::v8i32, V1),
9962 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9966 // AVX2 provides a direct instruction for permuting a single input across
9968 if (isSingleInputShuffleMask(Mask))
9969 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9970 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9972 // Try to use shift instructions.
9974 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
9977 // Use dedicated unpack instructions for masks that match their pattern.
9978 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9979 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9980 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9981 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9982 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9983 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
9984 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9985 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
9987 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9988 // shuffle. However, if we have AVX2 and either inputs are already in place,
9989 // we will be able to shuffle even across lanes the other input in a single
9990 // instruction so skip this pattern.
9991 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9992 isShuffleMaskInputInPlace(1, Mask))))
9993 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9994 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
9997 // Otherwise fall back on generic blend lowering.
9998 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10002 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10004 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10005 /// isn't available.
10006 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10007 const X86Subtarget *Subtarget,
10008 SelectionDAG &DAG) {
10010 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10011 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10012 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10013 ArrayRef<int> Mask = SVOp->getMask();
10014 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10016 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10020 // Check for being able to broadcast a single element.
10021 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10022 Mask, Subtarget, DAG))
10025 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10026 // options to efficiently lower the shuffle.
10027 SmallVector<int, 4> RepeatedMask;
10028 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10029 assert(RepeatedMask.size() == 4 &&
10030 "Repeated masks must be half the mask width!");
10032 // Use even/odd duplicate instructions for masks that match their pattern.
10033 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10034 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10035 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10036 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10038 if (isSingleInputShuffleMask(Mask))
10039 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10040 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10042 // Use dedicated unpack instructions for masks that match their pattern.
10043 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10044 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
10045 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10046 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
10047 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10048 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
10049 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10050 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
10052 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10053 // have already handled any direct blends. We also need to squash the
10054 // repeated mask into a simulated v4f32 mask.
10055 for (int i = 0; i < 4; ++i)
10056 if (RepeatedMask[i] >= 8)
10057 RepeatedMask[i] -= 4;
10058 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10061 // If we have a single input shuffle with different shuffle patterns in the
10062 // two 128-bit lanes use the variable mask to VPERMILPS.
10063 if (isSingleInputShuffleMask(Mask)) {
10064 SDValue VPermMask[8];
10065 for (int i = 0; i < 8; ++i)
10066 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10067 : DAG.getConstant(Mask[i], DL, MVT::i32);
10068 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10069 return DAG.getNode(
10070 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10071 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10073 if (Subtarget->hasAVX2())
10074 return DAG.getNode(
10075 X86ISD::VPERMV, DL, MVT::v8f32,
10076 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10077 MVT::v8i32, VPermMask)),
10080 // Otherwise, fall back.
10081 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10085 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10087 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10088 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10091 // If we have AVX2 then we always want to lower with a blend because at v8 we
10092 // can fully permute the elements.
10093 if (Subtarget->hasAVX2())
10094 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10097 // Otherwise fall back on generic lowering.
10098 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10101 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10103 /// This routine is only called when we have AVX2 and thus a reasonable
10104 /// instruction set for v8i32 shuffling..
10105 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10106 const X86Subtarget *Subtarget,
10107 SelectionDAG &DAG) {
10109 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10110 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10111 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10112 ArrayRef<int> Mask = SVOp->getMask();
10113 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10114 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10116 // Whenever we can lower this as a zext, that instruction is strictly faster
10117 // than any alternative. It also allows us to fold memory operands into the
10118 // shuffle in many cases.
10119 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10120 Mask, Subtarget, DAG))
10123 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10127 // Check for being able to broadcast a single element.
10128 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10129 Mask, Subtarget, DAG))
10132 // If the shuffle mask is repeated in each 128-bit lane we can use more
10133 // efficient instructions that mirror the shuffles across the two 128-bit
10135 SmallVector<int, 4> RepeatedMask;
10136 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10137 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10138 if (isSingleInputShuffleMask(Mask))
10139 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10140 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10142 // Use dedicated unpack instructions for masks that match their pattern.
10143 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10144 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
10145 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10146 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
10147 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10148 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
10149 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10150 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
10153 // Try to use shift instructions.
10154 if (SDValue Shift =
10155 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10158 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10159 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10162 // If the shuffle patterns aren't repeated but it is a single input, directly
10163 // generate a cross-lane VPERMD instruction.
10164 if (isSingleInputShuffleMask(Mask)) {
10165 SDValue VPermMask[8];
10166 for (int i = 0; i < 8; ++i)
10167 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10168 : DAG.getConstant(Mask[i], DL, MVT::i32);
10169 return DAG.getNode(
10170 X86ISD::VPERMV, DL, MVT::v8i32,
10171 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10174 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10176 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10177 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10180 // Otherwise fall back on generic blend lowering.
10181 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10185 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10187 /// This routine is only called when we have AVX2 and thus a reasonable
10188 /// instruction set for v16i16 shuffling..
10189 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10190 const X86Subtarget *Subtarget,
10191 SelectionDAG &DAG) {
10193 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10194 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10195 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10196 ArrayRef<int> Mask = SVOp->getMask();
10197 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10198 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10200 // Whenever we can lower this as a zext, that instruction is strictly faster
10201 // than any alternative. It also allows us to fold memory operands into the
10202 // shuffle in many cases.
10203 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10204 Mask, Subtarget, DAG))
10207 // Check for being able to broadcast a single element.
10208 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10209 Mask, Subtarget, DAG))
10212 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10216 // Use dedicated unpack instructions for masks that match their pattern.
10217 if (isShuffleEquivalent(V1, V2, Mask,
10218 {// First 128-bit lane:
10219 0, 16, 1, 17, 2, 18, 3, 19,
10220 // Second 128-bit lane:
10221 8, 24, 9, 25, 10, 26, 11, 27}))
10222 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
10223 if (isShuffleEquivalent(V1, V2, Mask,
10224 {// First 128-bit lane:
10225 4, 20, 5, 21, 6, 22, 7, 23,
10226 // Second 128-bit lane:
10227 12, 28, 13, 29, 14, 30, 15, 31}))
10228 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
10230 // Try to use shift instructions.
10231 if (SDValue Shift =
10232 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10235 // Try to use byte rotation instructions.
10236 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10237 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10240 if (isSingleInputShuffleMask(Mask)) {
10241 // There are no generalized cross-lane shuffle operations available on i16
10243 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10244 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10247 SmallVector<int, 8> RepeatedMask;
10248 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10249 // As this is a single-input shuffle, the repeated mask should be
10250 // a strictly valid v8i16 mask that we can pass through to the v8i16
10251 // lowering to handle even the v16 case.
10252 return lowerV8I16GeneralSingleInputVectorShuffle(
10253 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10256 SDValue PSHUFBMask[32];
10257 for (int i = 0; i < 16; ++i) {
10258 if (Mask[i] == -1) {
10259 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10263 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10264 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10265 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10266 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10268 return DAG.getBitcast(MVT::v16i16,
10269 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10270 DAG.getBitcast(MVT::v32i8, V1),
10271 DAG.getNode(ISD::BUILD_VECTOR, DL,
10272 MVT::v32i8, PSHUFBMask)));
10275 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10277 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10278 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10281 // Otherwise fall back on generic lowering.
10282 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10285 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10287 /// This routine is only called when we have AVX2 and thus a reasonable
10288 /// instruction set for v32i8 shuffling..
10289 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10290 const X86Subtarget *Subtarget,
10291 SelectionDAG &DAG) {
10293 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10294 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10295 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10296 ArrayRef<int> Mask = SVOp->getMask();
10297 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10298 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10300 // Whenever we can lower this as a zext, that instruction is strictly faster
10301 // than any alternative. It also allows us to fold memory operands into the
10302 // shuffle in many cases.
10303 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10304 Mask, Subtarget, DAG))
10307 // Check for being able to broadcast a single element.
10308 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10309 Mask, Subtarget, DAG))
10312 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10316 // Use dedicated unpack instructions for masks that match their pattern.
10317 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
10319 if (isShuffleEquivalent(
10321 {// First 128-bit lane:
10322 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
10323 // Second 128-bit lane:
10324 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
10325 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
10326 if (isShuffleEquivalent(
10328 {// First 128-bit lane:
10329 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
10330 // Second 128-bit lane:
10331 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
10332 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
10334 // Try to use shift instructions.
10335 if (SDValue Shift =
10336 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10339 // Try to use byte rotation instructions.
10340 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10341 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10344 if (isSingleInputShuffleMask(Mask)) {
10345 // There are no generalized cross-lane shuffle operations available on i8
10347 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10348 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10351 SDValue PSHUFBMask[32];
10352 for (int i = 0; i < 32; ++i)
10355 ? DAG.getUNDEF(MVT::i8)
10356 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10359 return DAG.getNode(
10360 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10361 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10364 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10366 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10367 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10370 // Otherwise fall back on generic lowering.
10371 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10374 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10376 /// This routine either breaks down the specific type of a 256-bit x86 vector
10377 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10378 /// together based on the available instructions.
10379 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10380 MVT VT, const X86Subtarget *Subtarget,
10381 SelectionDAG &DAG) {
10383 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10384 ArrayRef<int> Mask = SVOp->getMask();
10386 // If we have a single input to the zero element, insert that into V1 if we
10387 // can do so cheaply.
10388 int NumElts = VT.getVectorNumElements();
10389 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10390 return M >= NumElts;
10393 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10394 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10395 DL, VT, V1, V2, Mask, Subtarget, DAG))
10398 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
10399 // check for those subtargets here and avoid much of the subtarget querying in
10400 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
10401 // ability to manipulate a 256-bit vector with integer types. Since we'll use
10402 // floating point types there eventually, just immediately cast everything to
10403 // a float and operate entirely in that domain.
10404 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10405 int ElementBits = VT.getScalarSizeInBits();
10406 if (ElementBits < 32)
10407 // No floating point type available, decompose into 128-bit vectors.
10408 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10410 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10411 VT.getVectorNumElements());
10412 V1 = DAG.getBitcast(FpVT, V1);
10413 V2 = DAG.getBitcast(FpVT, V2);
10414 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10417 switch (VT.SimpleTy) {
10419 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10421 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10423 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10425 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10427 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10429 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10432 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10436 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10437 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10438 const X86Subtarget *Subtarget,
10439 SelectionDAG &DAG) {
10441 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10442 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10443 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10444 ArrayRef<int> Mask = SVOp->getMask();
10445 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10447 // X86 has dedicated unpack instructions that can handle specific blend
10448 // operations: UNPCKH and UNPCKL.
10449 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10450 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
10451 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10452 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
10454 // FIXME: Implement direct support for this type!
10455 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
10458 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10459 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10460 const X86Subtarget *Subtarget,
10461 SelectionDAG &DAG) {
10463 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10464 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10465 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10466 ArrayRef<int> Mask = SVOp->getMask();
10467 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10469 // Use dedicated unpack instructions for masks that match their pattern.
10470 if (isShuffleEquivalent(V1, V2, Mask,
10471 {// First 128-bit lane.
10472 0, 16, 1, 17, 4, 20, 5, 21,
10473 // Second 128-bit lane.
10474 8, 24, 9, 25, 12, 28, 13, 29}))
10475 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
10476 if (isShuffleEquivalent(V1, V2, Mask,
10477 {// First 128-bit lane.
10478 2, 18, 3, 19, 6, 22, 7, 23,
10479 // Second 128-bit lane.
10480 10, 26, 11, 27, 14, 30, 15, 31}))
10481 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
10483 // FIXME: Implement direct support for this type!
10484 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
10487 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10488 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10489 const X86Subtarget *Subtarget,
10490 SelectionDAG &DAG) {
10492 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10493 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10494 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10495 ArrayRef<int> Mask = SVOp->getMask();
10496 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10498 // X86 has dedicated unpack instructions that can handle specific blend
10499 // operations: UNPCKH and UNPCKL.
10500 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10501 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
10502 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10503 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
10505 // FIXME: Implement direct support for this type!
10506 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
10509 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10510 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10511 const X86Subtarget *Subtarget,
10512 SelectionDAG &DAG) {
10514 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10515 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10516 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10517 ArrayRef<int> Mask = SVOp->getMask();
10518 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10520 // Use dedicated unpack instructions for masks that match their pattern.
10521 if (isShuffleEquivalent(V1, V2, Mask,
10522 {// First 128-bit lane.
10523 0, 16, 1, 17, 4, 20, 5, 21,
10524 // Second 128-bit lane.
10525 8, 24, 9, 25, 12, 28, 13, 29}))
10526 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
10527 if (isShuffleEquivalent(V1, V2, Mask,
10528 {// First 128-bit lane.
10529 2, 18, 3, 19, 6, 22, 7, 23,
10530 // Second 128-bit lane.
10531 10, 26, 11, 27, 14, 30, 15, 31}))
10532 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
10534 // FIXME: Implement direct support for this type!
10535 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
10538 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10539 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10540 const X86Subtarget *Subtarget,
10541 SelectionDAG &DAG) {
10543 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10544 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10545 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10546 ArrayRef<int> Mask = SVOp->getMask();
10547 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10548 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10550 // FIXME: Implement direct support for this type!
10551 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
10554 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10555 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10556 const X86Subtarget *Subtarget,
10557 SelectionDAG &DAG) {
10559 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10560 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10561 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10562 ArrayRef<int> Mask = SVOp->getMask();
10563 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10564 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10566 // FIXME: Implement direct support for this type!
10567 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10570 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10572 /// This routine either breaks down the specific type of a 512-bit x86 vector
10573 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10574 /// together based on the available instructions.
10575 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10576 MVT VT, const X86Subtarget *Subtarget,
10577 SelectionDAG &DAG) {
10579 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10580 ArrayRef<int> Mask = SVOp->getMask();
10581 assert(Subtarget->hasAVX512() &&
10582 "Cannot lower 512-bit vectors w/ basic ISA!");
10584 // Check for being able to broadcast a single element.
10585 if (SDValue Broadcast =
10586 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10589 // Dispatch to each element type for lowering. If we don't have supprot for
10590 // specific element type shuffles at 512 bits, immediately split them and
10591 // lower them. Each lowering routine of a given type is allowed to assume that
10592 // the requisite ISA extensions for that element type are available.
10593 switch (VT.SimpleTy) {
10595 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10597 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10599 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10601 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10603 if (Subtarget->hasBWI())
10604 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10607 if (Subtarget->hasBWI())
10608 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10612 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10615 // Otherwise fall back on splitting.
10616 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10619 /// \brief Top-level lowering for x86 vector shuffles.
10621 /// This handles decomposition, canonicalization, and lowering of all x86
10622 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10623 /// above in helper routines. The canonicalization attempts to widen shuffles
10624 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10625 /// s.t. only one of the two inputs needs to be tested, etc.
10626 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10627 SelectionDAG &DAG) {
10628 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10629 ArrayRef<int> Mask = SVOp->getMask();
10630 SDValue V1 = Op.getOperand(0);
10631 SDValue V2 = Op.getOperand(1);
10632 MVT VT = Op.getSimpleValueType();
10633 int NumElements = VT.getVectorNumElements();
10636 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
10638 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10639 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10640 if (V1IsUndef && V2IsUndef)
10641 return DAG.getUNDEF(VT);
10643 // When we create a shuffle node we put the UNDEF node to second operand,
10644 // but in some cases the first operand may be transformed to UNDEF.
10645 // In this case we should just commute the node.
10647 return DAG.getCommutedVectorShuffle(*SVOp);
10649 // Check for non-undef masks pointing at an undef vector and make the masks
10650 // undef as well. This makes it easier to match the shuffle based solely on
10654 if (M >= NumElements) {
10655 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
10656 for (int &M : NewMask)
10657 if (M >= NumElements)
10659 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
10662 // We actually see shuffles that are entirely re-arrangements of a set of
10663 // zero inputs. This mostly happens while decomposing complex shuffles into
10664 // simple ones. Directly lower these as a buildvector of zeros.
10665 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
10666 if (Zeroable.all())
10667 return getZeroVector(VT, Subtarget, DAG, dl);
10669 // Try to collapse shuffles into using a vector type with fewer elements but
10670 // wider element types. We cap this to not form integers or floating point
10671 // elements wider than 64 bits, but it might be interesting to form i128
10672 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
10673 SmallVector<int, 16> WidenedMask;
10674 if (VT.getScalarSizeInBits() < 64 &&
10675 canWidenShuffleElements(Mask, WidenedMask)) {
10676 MVT NewEltVT = VT.isFloatingPoint()
10677 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
10678 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
10679 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
10680 // Make sure that the new vector type is legal. For example, v2f64 isn't
10682 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
10683 V1 = DAG.getBitcast(NewVT, V1);
10684 V2 = DAG.getBitcast(NewVT, V2);
10685 return DAG.getBitcast(
10686 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
10690 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
10691 for (int M : SVOp->getMask())
10693 ++NumUndefElements;
10694 else if (M < NumElements)
10699 // Commute the shuffle as needed such that more elements come from V1 than
10700 // V2. This allows us to match the shuffle pattern strictly on how many
10701 // elements come from V1 without handling the symmetric cases.
10702 if (NumV2Elements > NumV1Elements)
10703 return DAG.getCommutedVectorShuffle(*SVOp);
10705 // When the number of V1 and V2 elements are the same, try to minimize the
10706 // number of uses of V2 in the low half of the vector. When that is tied,
10707 // ensure that the sum of indices for V1 is equal to or lower than the sum
10708 // indices for V2. When those are equal, try to ensure that the number of odd
10709 // indices for V1 is lower than the number of odd indices for V2.
10710 if (NumV1Elements == NumV2Elements) {
10711 int LowV1Elements = 0, LowV2Elements = 0;
10712 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10713 if (M >= NumElements)
10717 if (LowV2Elements > LowV1Elements) {
10718 return DAG.getCommutedVectorShuffle(*SVOp);
10719 } else if (LowV2Elements == LowV1Elements) {
10720 int SumV1Indices = 0, SumV2Indices = 0;
10721 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10722 if (SVOp->getMask()[i] >= NumElements)
10724 else if (SVOp->getMask()[i] >= 0)
10726 if (SumV2Indices < SumV1Indices) {
10727 return DAG.getCommutedVectorShuffle(*SVOp);
10728 } else if (SumV2Indices == SumV1Indices) {
10729 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10730 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10731 if (SVOp->getMask()[i] >= NumElements)
10732 NumV2OddIndices += i % 2;
10733 else if (SVOp->getMask()[i] >= 0)
10734 NumV1OddIndices += i % 2;
10735 if (NumV2OddIndices < NumV1OddIndices)
10736 return DAG.getCommutedVectorShuffle(*SVOp);
10741 // For each vector width, delegate to a specialized lowering routine.
10742 if (VT.getSizeInBits() == 128)
10743 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10745 if (VT.getSizeInBits() == 256)
10746 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10748 // Force AVX-512 vectors to be scalarized for now.
10749 // FIXME: Implement AVX-512 support!
10750 if (VT.getSizeInBits() == 512)
10751 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10753 llvm_unreachable("Unimplemented!");
10756 // This function assumes its argument is a BUILD_VECTOR of constants or
10757 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10759 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10760 unsigned &MaskValue) {
10762 unsigned NumElems = BuildVector->getNumOperands();
10763 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10764 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10765 unsigned NumElemsInLane = NumElems / NumLanes;
10767 // Blend for v16i16 should be symmetric for the both lanes.
10768 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10769 SDValue EltCond = BuildVector->getOperand(i);
10770 SDValue SndLaneEltCond =
10771 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10773 int Lane1Cond = -1, Lane2Cond = -1;
10774 if (isa<ConstantSDNode>(EltCond))
10775 Lane1Cond = !isZero(EltCond);
10776 if (isa<ConstantSDNode>(SndLaneEltCond))
10777 Lane2Cond = !isZero(SndLaneEltCond);
10779 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10780 // Lane1Cond != 0, means we want the first argument.
10781 // Lane1Cond == 0, means we want the second argument.
10782 // The encoding of this argument is 0 for the first argument, 1
10783 // for the second. Therefore, invert the condition.
10784 MaskValue |= !Lane1Cond << i;
10785 else if (Lane1Cond < 0)
10786 MaskValue |= !Lane2Cond << i;
10793 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
10794 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
10795 const X86Subtarget *Subtarget,
10796 SelectionDAG &DAG) {
10797 SDValue Cond = Op.getOperand(0);
10798 SDValue LHS = Op.getOperand(1);
10799 SDValue RHS = Op.getOperand(2);
10801 MVT VT = Op.getSimpleValueType();
10803 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
10805 auto *CondBV = cast<BuildVectorSDNode>(Cond);
10807 // Only non-legal VSELECTs reach this lowering, convert those into generic
10808 // shuffles and re-use the shuffle lowering path for blends.
10809 SmallVector<int, 32> Mask;
10810 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
10811 SDValue CondElt = CondBV->getOperand(i);
10813 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
10815 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
10818 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
10819 // A vselect where all conditions and data are constants can be optimized into
10820 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
10821 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
10822 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
10823 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
10826 // Try to lower this to a blend-style vector shuffle. This can handle all
10827 // constant condition cases.
10828 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
10831 // Variable blends are only legal from SSE4.1 onward.
10832 if (!Subtarget->hasSSE41())
10835 // Only some types will be legal on some subtargets. If we can emit a legal
10836 // VSELECT-matching blend, return Op, and but if we need to expand, return
10838 switch (Op.getSimpleValueType().SimpleTy) {
10840 // Most of the vector types have blends past SSE4.1.
10844 // The byte blends for AVX vectors were introduced only in AVX2.
10845 if (Subtarget->hasAVX2())
10852 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
10853 if (Subtarget->hasBWI() && Subtarget->hasVLX())
10856 // FIXME: We should custom lower this by fixing the condition and using i8
10862 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
10863 MVT VT = Op.getSimpleValueType();
10866 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
10869 if (VT.getSizeInBits() == 8) {
10870 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
10871 Op.getOperand(0), Op.getOperand(1));
10872 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10873 DAG.getValueType(VT));
10874 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10877 if (VT.getSizeInBits() == 16) {
10878 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10879 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
10881 return DAG.getNode(
10882 ISD::TRUNCATE, dl, MVT::i16,
10883 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10884 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
10885 Op.getOperand(1)));
10886 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
10887 Op.getOperand(0), Op.getOperand(1));
10888 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10889 DAG.getValueType(VT));
10890 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10893 if (VT == MVT::f32) {
10894 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
10895 // the result back to FR32 register. It's only worth matching if the
10896 // result has a single use which is a store or a bitcast to i32. And in
10897 // the case of a store, it's not worth it if the index is a constant 0,
10898 // because a MOVSSmr can be used instead, which is smaller and faster.
10899 if (!Op.hasOneUse())
10901 SDNode *User = *Op.getNode()->use_begin();
10902 if ((User->getOpcode() != ISD::STORE ||
10903 (isa<ConstantSDNode>(Op.getOperand(1)) &&
10904 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
10905 (User->getOpcode() != ISD::BITCAST ||
10906 User->getValueType(0) != MVT::i32))
10908 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10909 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
10911 return DAG.getBitcast(MVT::f32, Extract);
10914 if (VT == MVT::i32 || VT == MVT::i64) {
10915 // ExtractPS/pextrq works with constant index.
10916 if (isa<ConstantSDNode>(Op.getOperand(1)))
10922 /// Extract one bit from mask vector, like v16i1 or v8i1.
10923 /// AVX-512 feature.
10925 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
10926 SDValue Vec = Op.getOperand(0);
10928 MVT VecVT = Vec.getSimpleValueType();
10929 SDValue Idx = Op.getOperand(1);
10930 MVT EltVT = Op.getSimpleValueType();
10932 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
10933 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
10934 "Unexpected vector type in ExtractBitFromMaskVector");
10936 // variable index can't be handled in mask registers,
10937 // extend vector to VR512
10938 if (!isa<ConstantSDNode>(Idx)) {
10939 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10940 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
10941 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
10942 ExtVT.getVectorElementType(), Ext, Idx);
10943 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
10946 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10947 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10948 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
10949 rc = getRegClassFor(MVT::v16i1);
10950 unsigned MaxSift = rc->getSize()*8 - 1;
10951 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
10952 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
10953 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
10954 DAG.getConstant(MaxSift, dl, MVT::i8));
10955 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
10956 DAG.getIntPtrConstant(0, dl));
10960 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
10961 SelectionDAG &DAG) const {
10963 SDValue Vec = Op.getOperand(0);
10964 MVT VecVT = Vec.getSimpleValueType();
10965 SDValue Idx = Op.getOperand(1);
10967 if (Op.getSimpleValueType() == MVT::i1)
10968 return ExtractBitFromMaskVector(Op, DAG);
10970 if (!isa<ConstantSDNode>(Idx)) {
10971 if (VecVT.is512BitVector() ||
10972 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
10973 VecVT.getVectorElementType().getSizeInBits() == 32)) {
10976 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
10977 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
10978 MaskEltVT.getSizeInBits());
10980 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
10981 auto PtrVT = getPointerTy(DAG.getDataLayout());
10982 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
10983 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
10984 DAG.getConstant(0, dl, PtrVT));
10985 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
10986 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
10987 DAG.getConstant(0, dl, PtrVT));
10992 // If this is a 256-bit vector result, first extract the 128-bit vector and
10993 // then extract the element from the 128-bit vector.
10994 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
10996 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10997 // Get the 128-bit vector.
10998 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
10999 MVT EltVT = VecVT.getVectorElementType();
11001 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11003 //if (IdxVal >= NumElems/2)
11004 // IdxVal -= NumElems/2;
11005 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
11006 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11007 DAG.getConstant(IdxVal, dl, MVT::i32));
11010 assert(VecVT.is128BitVector() && "Unexpected vector length");
11012 if (Subtarget->hasSSE41())
11013 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11016 MVT VT = Op.getSimpleValueType();
11017 // TODO: handle v16i8.
11018 if (VT.getSizeInBits() == 16) {
11019 SDValue Vec = Op.getOperand(0);
11020 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11022 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11023 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11024 DAG.getBitcast(MVT::v4i32, Vec),
11025 Op.getOperand(1)));
11026 // Transform it so it match pextrw which produces a 32-bit result.
11027 MVT EltVT = MVT::i32;
11028 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11029 Op.getOperand(0), Op.getOperand(1));
11030 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11031 DAG.getValueType(VT));
11032 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11035 if (VT.getSizeInBits() == 32) {
11036 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11040 // SHUFPS the element to the lowest double word, then movss.
11041 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11042 MVT VVT = Op.getOperand(0).getSimpleValueType();
11043 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11044 DAG.getUNDEF(VVT), Mask);
11045 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11046 DAG.getIntPtrConstant(0, dl));
11049 if (VT.getSizeInBits() == 64) {
11050 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11051 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11052 // to match extract_elt for f64.
11053 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11057 // UNPCKHPD the element to the lowest double word, then movsd.
11058 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11059 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11060 int Mask[2] = { 1, -1 };
11061 MVT VVT = Op.getOperand(0).getSimpleValueType();
11062 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11063 DAG.getUNDEF(VVT), Mask);
11064 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11065 DAG.getIntPtrConstant(0, dl));
11071 /// Insert one bit to mask vector, like v16i1 or v8i1.
11072 /// AVX-512 feature.
11074 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11076 SDValue Vec = Op.getOperand(0);
11077 SDValue Elt = Op.getOperand(1);
11078 SDValue Idx = Op.getOperand(2);
11079 MVT VecVT = Vec.getSimpleValueType();
11081 if (!isa<ConstantSDNode>(Idx)) {
11082 // Non constant index. Extend source and destination,
11083 // insert element and then truncate the result.
11084 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11085 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11086 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11087 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11088 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11089 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11092 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11093 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11095 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11096 DAG.getConstant(IdxVal, dl, MVT::i8));
11097 if (Vec.getOpcode() == ISD::UNDEF)
11099 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11102 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11103 SelectionDAG &DAG) const {
11104 MVT VT = Op.getSimpleValueType();
11105 MVT EltVT = VT.getVectorElementType();
11107 if (EltVT == MVT::i1)
11108 return InsertBitToMaskVector(Op, DAG);
11111 SDValue N0 = Op.getOperand(0);
11112 SDValue N1 = Op.getOperand(1);
11113 SDValue N2 = Op.getOperand(2);
11114 if (!isa<ConstantSDNode>(N2))
11116 auto *N2C = cast<ConstantSDNode>(N2);
11117 unsigned IdxVal = N2C->getZExtValue();
11119 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11120 // into that, and then insert the subvector back into the result.
11121 if (VT.is256BitVector() || VT.is512BitVector()) {
11122 // With a 256-bit vector, we can insert into the zero element efficiently
11123 // using a blend if we have AVX or AVX2 and the right data type.
11124 if (VT.is256BitVector() && IdxVal == 0) {
11125 // TODO: It is worthwhile to cast integer to floating point and back
11126 // and incur a domain crossing penalty if that's what we'll end up
11127 // doing anyway after extracting to a 128-bit vector.
11128 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11129 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11130 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11131 N2 = DAG.getIntPtrConstant(1, dl);
11132 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11136 // Get the desired 128-bit vector chunk.
11137 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11139 // Insert the element into the desired chunk.
11140 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11141 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
11143 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11144 DAG.getConstant(IdxIn128, dl, MVT::i32));
11146 // Insert the changed part back into the bigger vector
11147 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11149 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11151 if (Subtarget->hasSSE41()) {
11152 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11154 if (VT == MVT::v8i16) {
11155 Opc = X86ISD::PINSRW;
11157 assert(VT == MVT::v16i8);
11158 Opc = X86ISD::PINSRB;
11161 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11163 if (N1.getValueType() != MVT::i32)
11164 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11165 if (N2.getValueType() != MVT::i32)
11166 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11167 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11170 if (EltVT == MVT::f32) {
11171 // Bits [7:6] of the constant are the source select. This will always be
11172 // zero here. The DAG Combiner may combine an extract_elt index into
11173 // these bits. For example (insert (extract, 3), 2) could be matched by
11174 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11175 // Bits [5:4] of the constant are the destination select. This is the
11176 // value of the incoming immediate.
11177 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11178 // combine either bitwise AND or insert of float 0.0 to set these bits.
11180 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11181 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11182 // If this is an insertion of 32-bits into the low 32-bits of
11183 // a vector, we prefer to generate a blend with immediate rather
11184 // than an insertps. Blends are simpler operations in hardware and so
11185 // will always have equal or better performance than insertps.
11186 // But if optimizing for size and there's a load folding opportunity,
11187 // generate insertps because blendps does not have a 32-bit memory
11189 N2 = DAG.getIntPtrConstant(1, dl);
11190 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11191 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11193 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11194 // Create this as a scalar to vector..
11195 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11196 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11199 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11200 // PINSR* works with constant index.
11205 if (EltVT == MVT::i8)
11208 if (EltVT.getSizeInBits() == 16) {
11209 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11210 // as its second argument.
11211 if (N1.getValueType() != MVT::i32)
11212 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11213 if (N2.getValueType() != MVT::i32)
11214 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11215 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11220 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11222 MVT OpVT = Op.getSimpleValueType();
11224 // If this is a 256-bit vector result, first insert into a 128-bit
11225 // vector and then insert into the 256-bit vector.
11226 if (!OpVT.is128BitVector()) {
11227 // Insert into a 128-bit vector.
11228 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11229 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11230 OpVT.getVectorNumElements() / SizeFactor);
11232 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11234 // Insert the 128-bit vector.
11235 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11238 if (OpVT == MVT::v1i64 &&
11239 Op.getOperand(0).getValueType() == MVT::i64)
11240 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11242 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11243 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11244 return DAG.getBitcast(
11245 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11248 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11249 // a simple subregister reference or explicit instructions to grab
11250 // upper bits of a vector.
11251 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11252 SelectionDAG &DAG) {
11254 SDValue In = Op.getOperand(0);
11255 SDValue Idx = Op.getOperand(1);
11256 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11257 MVT ResVT = Op.getSimpleValueType();
11258 MVT InVT = In.getSimpleValueType();
11260 if (Subtarget->hasFp256()) {
11261 if (ResVT.is128BitVector() &&
11262 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11263 isa<ConstantSDNode>(Idx)) {
11264 return Extract128BitVector(In, IdxVal, DAG, dl);
11266 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11267 isa<ConstantSDNode>(Idx)) {
11268 return Extract256BitVector(In, IdxVal, DAG, dl);
11274 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11275 // simple superregister reference or explicit instructions to insert
11276 // the upper bits of a vector.
11277 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11278 SelectionDAG &DAG) {
11279 if (!Subtarget->hasAVX())
11283 SDValue Vec = Op.getOperand(0);
11284 SDValue SubVec = Op.getOperand(1);
11285 SDValue Idx = Op.getOperand(2);
11287 if (!isa<ConstantSDNode>(Idx))
11290 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11291 MVT OpVT = Op.getSimpleValueType();
11292 MVT SubVecVT = SubVec.getSimpleValueType();
11294 // Fold two 16-byte subvector loads into one 32-byte load:
11295 // (insert_subvector (insert_subvector undef, (load addr), 0),
11296 // (load addr + 16), Elts/2)
11298 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11299 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11300 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11301 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11302 if (Idx2 && Idx2->getZExtValue() == 0) {
11303 SDValue SubVec2 = Vec.getOperand(1);
11304 // If needed, look through a bitcast to get to the load.
11305 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11306 SubVec2 = SubVec2.getOperand(0);
11308 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11310 unsigned Alignment = FirstLd->getAlignment();
11311 unsigned AS = FirstLd->getAddressSpace();
11312 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11313 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11314 OpVT, AS, Alignment, &Fast) && Fast) {
11315 SDValue Ops[] = { SubVec2, SubVec };
11316 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11323 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11324 SubVecVT.is128BitVector())
11325 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11327 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11328 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11330 if (OpVT.getVectorElementType() == MVT::i1) {
11331 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11333 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11334 SDValue Undef = DAG.getUNDEF(OpVT);
11335 unsigned NumElems = OpVT.getVectorNumElements();
11336 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11338 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11339 // Zero upper bits of the Vec
11340 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11341 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11343 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11345 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11346 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11349 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11351 // Zero upper bits of the Vec2
11352 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11353 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11354 // Zero lower bits of the Vec
11355 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11356 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11357 // Merge them together
11358 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11364 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11365 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11366 // one of the above mentioned nodes. It has to be wrapped because otherwise
11367 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11368 // be used to form addressing mode. These wrapped nodes will be selected
11371 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11372 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11374 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11375 // global base reg.
11376 unsigned char OpFlag = 0;
11377 unsigned WrapperKind = X86ISD::Wrapper;
11378 CodeModel::Model M = DAG.getTarget().getCodeModel();
11380 if (Subtarget->isPICStyleRIPRel() &&
11381 (M == CodeModel::Small || M == CodeModel::Kernel))
11382 WrapperKind = X86ISD::WrapperRIP;
11383 else if (Subtarget->isPICStyleGOT())
11384 OpFlag = X86II::MO_GOTOFF;
11385 else if (Subtarget->isPICStyleStubPIC())
11386 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11388 auto PtrVT = getPointerTy(DAG.getDataLayout());
11389 SDValue Result = DAG.getTargetConstantPool(
11390 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11392 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11393 // With PIC, the address is actually $g + Offset.
11396 DAG.getNode(ISD::ADD, DL, PtrVT,
11397 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11403 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11404 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11406 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11407 // global base reg.
11408 unsigned char OpFlag = 0;
11409 unsigned WrapperKind = X86ISD::Wrapper;
11410 CodeModel::Model M = DAG.getTarget().getCodeModel();
11412 if (Subtarget->isPICStyleRIPRel() &&
11413 (M == CodeModel::Small || M == CodeModel::Kernel))
11414 WrapperKind = X86ISD::WrapperRIP;
11415 else if (Subtarget->isPICStyleGOT())
11416 OpFlag = X86II::MO_GOTOFF;
11417 else if (Subtarget->isPICStyleStubPIC())
11418 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11420 auto PtrVT = getPointerTy(DAG.getDataLayout());
11421 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11423 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11425 // With PIC, the address is actually $g + Offset.
11428 DAG.getNode(ISD::ADD, DL, PtrVT,
11429 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11435 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11436 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11438 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11439 // global base reg.
11440 unsigned char OpFlag = 0;
11441 unsigned WrapperKind = X86ISD::Wrapper;
11442 CodeModel::Model M = DAG.getTarget().getCodeModel();
11444 if (Subtarget->isPICStyleRIPRel() &&
11445 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11446 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11447 OpFlag = X86II::MO_GOTPCREL;
11448 WrapperKind = X86ISD::WrapperRIP;
11449 } else if (Subtarget->isPICStyleGOT()) {
11450 OpFlag = X86II::MO_GOT;
11451 } else if (Subtarget->isPICStyleStubPIC()) {
11452 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11453 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11454 OpFlag = X86II::MO_DARWIN_NONLAZY;
11457 auto PtrVT = getPointerTy(DAG.getDataLayout());
11458 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11461 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11463 // With PIC, the address is actually $g + Offset.
11464 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11465 !Subtarget->is64Bit()) {
11467 DAG.getNode(ISD::ADD, DL, PtrVT,
11468 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11471 // For symbols that require a load from a stub to get the address, emit the
11473 if (isGlobalStubReference(OpFlag))
11474 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11475 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11476 false, false, false, 0);
11482 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11483 // Create the TargetBlockAddressAddress node.
11484 unsigned char OpFlags =
11485 Subtarget->ClassifyBlockAddressReference();
11486 CodeModel::Model M = DAG.getTarget().getCodeModel();
11487 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11488 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11490 auto PtrVT = getPointerTy(DAG.getDataLayout());
11491 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11493 if (Subtarget->isPICStyleRIPRel() &&
11494 (M == CodeModel::Small || M == CodeModel::Kernel))
11495 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11497 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11499 // With PIC, the address is actually $g + Offset.
11500 if (isGlobalRelativeToPICBase(OpFlags)) {
11501 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11502 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11509 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11510 int64_t Offset, SelectionDAG &DAG) const {
11511 // Create the TargetGlobalAddress node, folding in the constant
11512 // offset if it is legal.
11513 unsigned char OpFlags =
11514 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11515 CodeModel::Model M = DAG.getTarget().getCodeModel();
11516 auto PtrVT = getPointerTy(DAG.getDataLayout());
11518 if (OpFlags == X86II::MO_NO_FLAG &&
11519 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11520 // A direct static reference to a global.
11521 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11524 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11527 if (Subtarget->isPICStyleRIPRel() &&
11528 (M == CodeModel::Small || M == CodeModel::Kernel))
11529 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11531 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11533 // With PIC, the address is actually $g + Offset.
11534 if (isGlobalRelativeToPICBase(OpFlags)) {
11535 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11536 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11539 // For globals that require a load from a stub to get the address, emit the
11541 if (isGlobalStubReference(OpFlags))
11542 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11543 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11544 false, false, false, 0);
11546 // If there was a non-zero offset that we didn't fold, create an explicit
11547 // addition for it.
11549 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11550 DAG.getConstant(Offset, dl, PtrVT));
11556 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11557 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11558 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11559 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11563 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11564 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11565 unsigned char OperandFlags, bool LocalDynamic = false) {
11566 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11567 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11569 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11570 GA->getValueType(0),
11574 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11578 SDValue Ops[] = { Chain, TGA, *InFlag };
11579 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11581 SDValue Ops[] = { Chain, TGA };
11582 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11585 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11586 MFI->setAdjustsStack(true);
11587 MFI->setHasCalls(true);
11589 SDValue Flag = Chain.getValue(1);
11590 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11593 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11595 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11598 SDLoc dl(GA); // ? function entry point might be better
11599 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11600 DAG.getNode(X86ISD::GlobalBaseReg,
11601 SDLoc(), PtrVT), InFlag);
11602 InFlag = Chain.getValue(1);
11604 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11607 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11609 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11611 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
11612 X86::RAX, X86II::MO_TLSGD);
11615 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
11621 // Get the start address of the TLS block for this module.
11622 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
11623 .getInfo<X86MachineFunctionInfo>();
11624 MFI->incNumLocalDynamicTLSAccesses();
11628 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
11629 X86II::MO_TLSLD, /*LocalDynamic=*/true);
11632 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11633 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
11634 InFlag = Chain.getValue(1);
11635 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
11636 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
11639 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
11643 unsigned char OperandFlags = X86II::MO_DTPOFF;
11644 unsigned WrapperKind = X86ISD::Wrapper;
11645 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11646 GA->getValueType(0),
11647 GA->getOffset(), OperandFlags);
11648 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11650 // Add x@dtpoff with the base.
11651 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
11654 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
11655 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11656 const EVT PtrVT, TLSModel::Model model,
11657 bool is64Bit, bool isPIC) {
11660 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
11661 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
11662 is64Bit ? 257 : 256));
11664 SDValue ThreadPointer =
11665 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
11666 MachinePointerInfo(Ptr), false, false, false, 0);
11668 unsigned char OperandFlags = 0;
11669 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
11671 unsigned WrapperKind = X86ISD::Wrapper;
11672 if (model == TLSModel::LocalExec) {
11673 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
11674 } else if (model == TLSModel::InitialExec) {
11676 OperandFlags = X86II::MO_GOTTPOFF;
11677 WrapperKind = X86ISD::WrapperRIP;
11679 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
11682 llvm_unreachable("Unexpected model");
11685 // emit "addl x@ntpoff,%eax" (local exec)
11686 // or "addl x@indntpoff,%eax" (initial exec)
11687 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
11689 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
11690 GA->getOffset(), OperandFlags);
11691 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11693 if (model == TLSModel::InitialExec) {
11694 if (isPIC && !is64Bit) {
11695 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
11696 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11700 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
11701 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11702 false, false, false, 0);
11705 // The address of the thread local variable is the add of the thread
11706 // pointer with the offset of the variable.
11707 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
11711 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
11713 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
11714 const GlobalValue *GV = GA->getGlobal();
11715 auto PtrVT = getPointerTy(DAG.getDataLayout());
11717 if (Subtarget->isTargetELF()) {
11718 if (DAG.getTarget().Options.EmulatedTLS)
11719 return LowerToTLSEmulatedModel(GA, DAG);
11720 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
11722 case TLSModel::GeneralDynamic:
11723 if (Subtarget->is64Bit())
11724 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
11725 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
11726 case TLSModel::LocalDynamic:
11727 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
11728 Subtarget->is64Bit());
11729 case TLSModel::InitialExec:
11730 case TLSModel::LocalExec:
11731 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
11732 DAG.getTarget().getRelocationModel() ==
11735 llvm_unreachable("Unknown TLS model.");
11738 if (Subtarget->isTargetDarwin()) {
11739 // Darwin only has one model of TLS. Lower to that.
11740 unsigned char OpFlag = 0;
11741 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
11742 X86ISD::WrapperRIP : X86ISD::Wrapper;
11744 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11745 // global base reg.
11746 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
11747 !Subtarget->is64Bit();
11749 OpFlag = X86II::MO_TLVP_PIC_BASE;
11751 OpFlag = X86II::MO_TLVP;
11753 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
11754 GA->getValueType(0),
11755 GA->getOffset(), OpFlag);
11756 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11758 // With PIC32, the address is actually $g + Offset.
11760 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
11761 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11764 // Lowering the machine isd will make sure everything is in the right
11766 SDValue Chain = DAG.getEntryNode();
11767 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11768 SDValue Args[] = { Chain, Offset };
11769 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
11771 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
11772 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11773 MFI->setAdjustsStack(true);
11775 // And our return value (tls address) is in the standard call return value
11777 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11778 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
11781 if (Subtarget->isTargetKnownWindowsMSVC() ||
11782 Subtarget->isTargetWindowsGNU()) {
11783 // Just use the implicit TLS architecture
11784 // Need to generate someting similar to:
11785 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
11787 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
11788 // mov rcx, qword [rdx+rcx*8]
11789 // mov eax, .tls$:tlsvar
11790 // [rax+rcx] contains the address
11791 // Windows 64bit: gs:0x58
11792 // Windows 32bit: fs:__tls_array
11795 SDValue Chain = DAG.getEntryNode();
11797 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
11798 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
11799 // use its literal value of 0x2C.
11800 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
11801 ? Type::getInt8PtrTy(*DAG.getContext(),
11803 : Type::getInt32PtrTy(*DAG.getContext(),
11806 SDValue TlsArray = Subtarget->is64Bit()
11807 ? DAG.getIntPtrConstant(0x58, dl)
11808 : (Subtarget->isTargetWindowsGNU()
11809 ? DAG.getIntPtrConstant(0x2C, dl)
11810 : DAG.getExternalSymbol("_tls_array", PtrVT));
11812 SDValue ThreadPointer =
11813 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
11817 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
11818 res = ThreadPointer;
11820 // Load the _tls_index variable
11821 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
11822 if (Subtarget->is64Bit())
11823 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
11824 MachinePointerInfo(), MVT::i32, false, false,
11827 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
11830 auto &DL = DAG.getDataLayout();
11832 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
11833 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
11835 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
11838 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
11841 // Get the offset of start of .tls section
11842 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11843 GA->getValueType(0),
11844 GA->getOffset(), X86II::MO_SECREL);
11845 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
11847 // The address of the thread local variable is the add of the thread
11848 // pointer with the offset of the variable.
11849 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
11852 llvm_unreachable("TLS not implemented for this target.");
11855 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
11856 /// and take a 2 x i32 value to shift plus a shift amount.
11857 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
11858 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
11859 MVT VT = Op.getSimpleValueType();
11860 unsigned VTBits = VT.getSizeInBits();
11862 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
11863 SDValue ShOpLo = Op.getOperand(0);
11864 SDValue ShOpHi = Op.getOperand(1);
11865 SDValue ShAmt = Op.getOperand(2);
11866 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
11867 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
11869 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11870 DAG.getConstant(VTBits - 1, dl, MVT::i8));
11871 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
11872 DAG.getConstant(VTBits - 1, dl, MVT::i8))
11873 : DAG.getConstant(0, dl, VT);
11875 SDValue Tmp2, Tmp3;
11876 if (Op.getOpcode() == ISD::SHL_PARTS) {
11877 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
11878 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
11880 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
11881 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
11884 // If the shift amount is larger or equal than the width of a part we can't
11885 // rely on the results of shld/shrd. Insert a test and select the appropriate
11886 // values for large shift amounts.
11887 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11888 DAG.getConstant(VTBits, dl, MVT::i8));
11889 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
11890 AndNode, DAG.getConstant(0, dl, MVT::i8));
11893 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
11894 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
11895 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
11897 if (Op.getOpcode() == ISD::SHL_PARTS) {
11898 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11899 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11901 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11902 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11905 SDValue Ops[2] = { Lo, Hi };
11906 return DAG.getMergeValues(Ops, dl);
11909 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
11910 SelectionDAG &DAG) const {
11911 SDValue Src = Op.getOperand(0);
11912 MVT SrcVT = Src.getSimpleValueType();
11913 MVT VT = Op.getSimpleValueType();
11916 if (SrcVT.isVector()) {
11917 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
11918 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
11919 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
11920 DAG.getUNDEF(SrcVT)));
11922 if (SrcVT.getVectorElementType() == MVT::i1) {
11923 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
11924 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11925 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
11930 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
11931 "Unknown SINT_TO_FP to lower!");
11933 // These are really Legal; return the operand so the caller accepts it as
11935 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
11937 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
11938 Subtarget->is64Bit()) {
11942 unsigned Size = SrcVT.getSizeInBits()/8;
11943 MachineFunction &MF = DAG.getMachineFunction();
11944 auto PtrVT = getPointerTy(MF.getDataLayout());
11945 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
11946 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
11947 SDValue Chain = DAG.getStore(
11948 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
11949 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
11951 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
11954 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
11956 SelectionDAG &DAG) const {
11960 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
11962 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
11964 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
11966 unsigned ByteSize = SrcVT.getSizeInBits()/8;
11968 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
11969 MachineMemOperand *MMO;
11971 int SSFI = FI->getIndex();
11972 MMO = DAG.getMachineFunction().getMachineMemOperand(
11973 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
11974 MachineMemOperand::MOLoad, ByteSize, ByteSize);
11976 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
11977 StackSlot = StackSlot.getOperand(1);
11979 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
11980 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
11982 Tys, Ops, SrcVT, MMO);
11985 Chain = Result.getValue(1);
11986 SDValue InFlag = Result.getValue(2);
11988 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
11989 // shouldn't be necessary except that RFP cannot be live across
11990 // multiple blocks. When stackifier is fixed, they can be uncoupled.
11991 MachineFunction &MF = DAG.getMachineFunction();
11992 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
11993 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
11994 auto PtrVT = getPointerTy(MF.getDataLayout());
11995 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
11996 Tys = DAG.getVTList(MVT::Other);
11998 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12000 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12001 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12002 MachineMemOperand::MOStore, SSFISize, SSFISize);
12004 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12005 Ops, Op.getValueType(), MMO);
12006 Result = DAG.getLoad(
12007 Op.getValueType(), DL, Chain, StackSlot,
12008 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12009 false, false, false, 0);
12015 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12016 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12017 SelectionDAG &DAG) const {
12018 // This algorithm is not obvious. Here it is what we're trying to output:
12021 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12022 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12024 haddpd %xmm0, %xmm0
12026 pshufd $0x4e, %xmm0, %xmm1
12032 LLVMContext *Context = DAG.getContext();
12034 // Build some magic constants.
12035 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12036 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12037 auto PtrVT = getPointerTy(DAG.getDataLayout());
12038 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12040 SmallVector<Constant*,2> CV1;
12042 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12043 APInt(64, 0x4330000000000000ULL))));
12045 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12046 APInt(64, 0x4530000000000000ULL))));
12047 Constant *C1 = ConstantVector::get(CV1);
12048 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12050 // Load the 64-bit value into an XMM register.
12051 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12054 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12055 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12056 false, false, false, 16);
12058 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12061 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12062 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12063 false, false, false, 16);
12064 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12065 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12068 if (Subtarget->hasSSE3()) {
12069 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12070 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12072 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12073 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12075 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12076 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12079 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12080 DAG.getIntPtrConstant(0, dl));
12083 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12084 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12085 SelectionDAG &DAG) const {
12087 // FP constant to bias correct the final result.
12088 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12091 // Load the 32-bit value into an XMM register.
12092 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12095 // Zero out the upper parts of the register.
12096 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12098 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12099 DAG.getBitcast(MVT::v2f64, Load),
12100 DAG.getIntPtrConstant(0, dl));
12102 // Or the load with the bias.
12103 SDValue Or = DAG.getNode(
12104 ISD::OR, dl, MVT::v2i64,
12105 DAG.getBitcast(MVT::v2i64,
12106 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12107 DAG.getBitcast(MVT::v2i64,
12108 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12110 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12111 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12113 // Subtract the bias.
12114 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12116 // Handle final rounding.
12117 EVT DestVT = Op.getValueType();
12119 if (DestVT.bitsLT(MVT::f64))
12120 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12121 DAG.getIntPtrConstant(0, dl));
12122 if (DestVT.bitsGT(MVT::f64))
12123 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12125 // Handle final rounding.
12129 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12130 const X86Subtarget &Subtarget) {
12131 // The algorithm is the following:
12132 // #ifdef __SSE4_1__
12133 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12134 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12135 // (uint4) 0x53000000, 0xaa);
12137 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12138 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12140 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12141 // return (float4) lo + fhi;
12144 SDValue V = Op->getOperand(0);
12145 EVT VecIntVT = V.getValueType();
12146 bool Is128 = VecIntVT == MVT::v4i32;
12147 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12148 // If we convert to something else than the supported type, e.g., to v4f64,
12150 if (VecFloatVT != Op->getValueType(0))
12153 unsigned NumElts = VecIntVT.getVectorNumElements();
12154 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12155 "Unsupported custom type");
12156 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12158 // In the #idef/#else code, we have in common:
12159 // - The vector of constants:
12165 // Create the splat vector for 0x4b000000.
12166 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12167 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12168 CstLow, CstLow, CstLow, CstLow};
12169 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12170 makeArrayRef(&CstLowArray[0], NumElts));
12171 // Create the splat vector for 0x53000000.
12172 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12173 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12174 CstHigh, CstHigh, CstHigh, CstHigh};
12175 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12176 makeArrayRef(&CstHighArray[0], NumElts));
12178 // Create the right shift.
12179 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12180 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12181 CstShift, CstShift, CstShift, CstShift};
12182 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12183 makeArrayRef(&CstShiftArray[0], NumElts));
12184 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12187 if (Subtarget.hasSSE41()) {
12188 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12189 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12190 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12191 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12192 // Low will be bitcasted right away, so do not bother bitcasting back to its
12194 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12195 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12196 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12197 // (uint4) 0x53000000, 0xaa);
12198 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12199 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12200 // High will be bitcasted right away, so do not bother bitcasting back to
12201 // its original type.
12202 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12203 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12205 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12206 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12207 CstMask, CstMask, CstMask);
12208 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12209 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12210 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12212 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12213 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12216 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12217 SDValue CstFAdd = DAG.getConstantFP(
12218 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12219 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12220 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12221 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12222 makeArrayRef(&CstFAddArray[0], NumElts));
12224 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12225 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12227 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12228 // return (float4) lo + fhi;
12229 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12230 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12233 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12234 SelectionDAG &DAG) const {
12235 SDValue N0 = Op.getOperand(0);
12236 MVT SVT = N0.getSimpleValueType();
12239 switch (SVT.SimpleTy) {
12241 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12246 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12247 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12248 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12252 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12255 if (Subtarget->hasAVX512())
12256 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12257 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12259 llvm_unreachable(nullptr);
12262 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12263 SelectionDAG &DAG) const {
12264 SDValue N0 = Op.getOperand(0);
12266 auto PtrVT = getPointerTy(DAG.getDataLayout());
12268 if (Op.getValueType().isVector())
12269 return lowerUINT_TO_FP_vec(Op, DAG);
12271 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12272 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12273 // the optimization here.
12274 if (DAG.SignBitIsZero(N0))
12275 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12277 MVT SrcVT = N0.getSimpleValueType();
12278 MVT DstVT = Op.getSimpleValueType();
12279 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12280 return LowerUINT_TO_FP_i64(Op, DAG);
12281 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12282 return LowerUINT_TO_FP_i32(Op, DAG);
12283 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12286 // Make a 64-bit buffer, and use it to build an FILD.
12287 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12288 if (SrcVT == MVT::i32) {
12289 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12290 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12291 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12292 StackSlot, MachinePointerInfo(),
12294 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12295 OffsetSlot, MachinePointerInfo(),
12297 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12301 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12302 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12303 StackSlot, MachinePointerInfo(),
12305 // For i64 source, we need to add the appropriate power of 2 if the input
12306 // was negative. This is the same as the optimization in
12307 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12308 // we must be careful to do the computation in x87 extended precision, not
12309 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12310 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12311 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12312 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12313 MachineMemOperand::MOLoad, 8, 8);
12315 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12316 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12317 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12320 APInt FF(32, 0x5F800000ULL);
12322 // Check whether the sign bit is set.
12323 SDValue SignSet = DAG.getSetCC(
12324 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12325 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12327 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12328 SDValue FudgePtr = DAG.getConstantPool(
12329 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12331 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12332 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12333 SDValue Four = DAG.getIntPtrConstant(4, dl);
12334 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12336 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12338 // Load the value out, extending it from f32 to f80.
12339 // FIXME: Avoid the extend by constructing the right constant pool?
12340 SDValue Fudge = DAG.getExtLoad(
12341 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12342 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12343 false, false, false, 4);
12344 // Extend everything to 80 bits to force it to be done on x87.
12345 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12346 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12347 DAG.getIntPtrConstant(0, dl));
12350 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
12351 // is legal, or has an f16 source (which needs to be promoted to f32),
12352 // just return an <SDValue(), SDValue()> pair.
12353 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
12354 // to i16, i32 or i64, and we lower it to a legal sequence.
12355 // If lowered to the final integer result we return a <result, SDValue()> pair.
12356 // Otherwise we lower it to a sequence ending with a FIST, return a
12357 // <FIST, StackSlot> pair, and the caller is responsible for loading
12358 // the final integer result from StackSlot.
12359 std::pair<SDValue,SDValue>
12360 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12361 bool IsSigned, bool IsReplace) const {
12364 EVT DstTy = Op.getValueType();
12365 EVT TheVT = Op.getOperand(0).getValueType();
12366 auto PtrVT = getPointerTy(DAG.getDataLayout());
12368 if (TheVT == MVT::f16)
12369 // We need to promote the f16 to f32 before using the lowering
12370 // in this routine.
12371 return std::make_pair(SDValue(), SDValue());
12373 assert((TheVT == MVT::f32 ||
12374 TheVT == MVT::f64 ||
12375 TheVT == MVT::f80) &&
12376 "Unexpected FP operand type in FP_TO_INTHelper");
12378 // If using FIST to compute an unsigned i64, we'll need some fixup
12379 // to handle values above the maximum signed i64. A FIST is always
12380 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
12381 bool UnsignedFixup = !IsSigned &&
12382 DstTy == MVT::i64 &&
12383 (!Subtarget->is64Bit() ||
12384 !isScalarFPTypeInSSEReg(TheVT));
12386 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
12387 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
12388 // The low 32 bits of the fist result will have the correct uint32 result.
12389 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12393 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12394 DstTy.getSimpleVT() >= MVT::i16 &&
12395 "Unknown FP_TO_INT to lower!");
12397 // These are really Legal.
12398 if (DstTy == MVT::i32 &&
12399 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12400 return std::make_pair(SDValue(), SDValue());
12401 if (Subtarget->is64Bit() &&
12402 DstTy == MVT::i64 &&
12403 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12404 return std::make_pair(SDValue(), SDValue());
12406 // We lower FP->int64 into FISTP64 followed by a load from a temporary
12408 MachineFunction &MF = DAG.getMachineFunction();
12409 unsigned MemSize = DstTy.getSizeInBits()/8;
12410 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12411 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12414 switch (DstTy.getSimpleVT().SimpleTy) {
12415 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12416 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12417 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12418 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12421 SDValue Chain = DAG.getEntryNode();
12422 SDValue Value = Op.getOperand(0);
12423 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
12425 if (UnsignedFixup) {
12427 // Conversion to unsigned i64 is implemented with a select,
12428 // depending on whether the source value fits in the range
12429 // of a signed i64. Let Thresh be the FP equivalent of
12430 // 0x8000000000000000ULL.
12432 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
12433 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
12434 // Fist-to-mem64 FistSrc
12435 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
12436 // to XOR'ing the high 32 bits with Adjust.
12438 // Being a power of 2, Thresh is exactly representable in all FP formats.
12439 // For X87 we'd like to use the smallest FP type for this constant, but
12440 // for DAG type consistency we have to match the FP operand type.
12442 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
12443 APFloat::opStatus Status = APFloat::opOK;
12444 bool LosesInfo = false;
12445 if (TheVT == MVT::f64)
12446 // The rounding mode is irrelevant as the conversion should be exact.
12447 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
12449 else if (TheVT == MVT::f80)
12450 Status = Thresh.convert(APFloat::x87DoubleExtended,
12451 APFloat::rmNearestTiesToEven, &LosesInfo);
12453 assert(Status == APFloat::opOK && !LosesInfo &&
12454 "FP conversion should have been exact");
12456 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
12458 SDValue Cmp = DAG.getSetCC(DL,
12459 getSetCCResultType(DAG.getDataLayout(),
12460 *DAG.getContext(), TheVT),
12461 Value, ThreshVal, ISD::SETLT);
12462 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
12463 DAG.getConstant(0, DL, MVT::i32),
12464 DAG.getConstant(0x80000000, DL, MVT::i32));
12465 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
12466 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
12467 *DAG.getContext(), TheVT),
12468 Value, ThreshVal, ISD::SETLT);
12469 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
12472 // FIXME This causes a redundant load/store if the SSE-class value is already
12473 // in memory, such as if it is on the callstack.
12474 if (isScalarFPTypeInSSEReg(TheVT)) {
12475 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12476 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12477 MachinePointerInfo::getFixedStack(MF, SSFI), false,
12479 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12481 Chain, StackSlot, DAG.getValueType(TheVT)
12484 MachineMemOperand *MMO =
12485 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12486 MachineMemOperand::MOLoad, MemSize, MemSize);
12487 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12488 Chain = Value.getValue(1);
12489 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12490 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12493 MachineMemOperand *MMO =
12494 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12495 MachineMemOperand::MOStore, MemSize, MemSize);
12497 if (UnsignedFixup) {
12499 // Insert the FIST, load its result as two i32's,
12500 // and XOR the high i32 with Adjust.
12502 SDValue FistOps[] = { Chain, Value, StackSlot };
12503 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12504 FistOps, DstTy, MMO);
12506 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
12507 MachinePointerInfo(),
12508 false, false, false, 0);
12509 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
12510 DAG.getConstant(4, DL, PtrVT));
12512 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
12513 MachinePointerInfo(),
12514 false, false, false, 0);
12515 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
12517 if (Subtarget->is64Bit()) {
12518 // Join High32 and Low32 into a 64-bit result.
12519 // (High32 << 32) | Low32
12520 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
12521 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
12522 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
12523 DAG.getConstant(32, DL, MVT::i8));
12524 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
12525 return std::make_pair(Result, SDValue());
12528 SDValue ResultOps[] = { Low32, High32 };
12530 SDValue pair = IsReplace
12531 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
12532 : DAG.getMergeValues(ResultOps, DL);
12533 return std::make_pair(pair, SDValue());
12535 // Build the FP_TO_INT*_IN_MEM
12536 SDValue Ops[] = { Chain, Value, StackSlot };
12537 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12539 return std::make_pair(FIST, StackSlot);
12543 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12544 const X86Subtarget *Subtarget) {
12545 MVT VT = Op->getSimpleValueType(0);
12546 SDValue In = Op->getOperand(0);
12547 MVT InVT = In.getSimpleValueType();
12550 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12551 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12553 // Optimize vectors in AVX mode:
12556 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12557 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12558 // Concat upper and lower parts.
12561 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12562 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12563 // Concat upper and lower parts.
12566 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12567 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12568 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12571 if (Subtarget->hasInt256())
12572 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12574 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12575 SDValue Undef = DAG.getUNDEF(InVT);
12576 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12577 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12578 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12580 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12581 VT.getVectorNumElements()/2);
12583 OpLo = DAG.getBitcast(HVT, OpLo);
12584 OpHi = DAG.getBitcast(HVT, OpHi);
12586 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12589 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
12590 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
12591 MVT VT = Op->getSimpleValueType(0);
12592 SDValue In = Op->getOperand(0);
12593 MVT InVT = In.getSimpleValueType();
12595 unsigned int NumElts = VT.getVectorNumElements();
12596 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
12599 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12600 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
12602 assert(InVT.getVectorElementType() == MVT::i1);
12603 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
12605 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
12607 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
12609 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
12610 if (VT.is512BitVector())
12612 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
12615 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12616 SelectionDAG &DAG) {
12617 if (Subtarget->hasFp256())
12618 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12624 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12625 SelectionDAG &DAG) {
12627 MVT VT = Op.getSimpleValueType();
12628 SDValue In = Op.getOperand(0);
12629 MVT SVT = In.getSimpleValueType();
12631 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
12632 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
12634 if (Subtarget->hasFp256())
12635 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12638 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
12639 VT.getVectorNumElements() != SVT.getVectorNumElements());
12643 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
12645 MVT VT = Op.getSimpleValueType();
12646 SDValue In = Op.getOperand(0);
12647 MVT InVT = In.getSimpleValueType();
12649 if (VT == MVT::i1) {
12650 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
12651 "Invalid scalar TRUNCATE operation");
12652 if (InVT.getSizeInBits() >= 32)
12654 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
12655 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
12657 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
12658 "Invalid TRUNCATE operation");
12660 // move vector to mask - truncate solution for SKX
12661 if (VT.getVectorElementType() == MVT::i1) {
12662 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
12663 Subtarget->hasBWI())
12664 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12665 if ((InVT.is256BitVector() || InVT.is128BitVector())
12666 && InVT.getScalarSizeInBits() <= 16 &&
12667 Subtarget->hasBWI() && Subtarget->hasVLX())
12668 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12669 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
12670 Subtarget->hasDQI())
12671 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
12672 if ((InVT.is256BitVector() || InVT.is128BitVector())
12673 && InVT.getScalarSizeInBits() >= 32 &&
12674 Subtarget->hasDQI() && Subtarget->hasVLX())
12675 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
12678 if (VT.getVectorElementType() == MVT::i1) {
12679 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
12680 unsigned NumElts = InVT.getVectorNumElements();
12681 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
12682 if (InVT.getSizeInBits() < 512) {
12683 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
12684 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
12689 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
12690 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
12691 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
12694 // vpmovqb/w/d, vpmovdb/w, vpmovwb
12695 if (((!InVT.is512BitVector() && Subtarget->hasVLX()) || InVT.is512BitVector()) &&
12696 (InVT.getVectorElementType() != MVT::i16 || Subtarget->hasBWI()))
12697 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
12699 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
12700 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
12701 if (Subtarget->hasInt256()) {
12702 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
12703 In = DAG.getBitcast(MVT::v8i32, In);
12704 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
12706 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
12707 DAG.getIntPtrConstant(0, DL));
12710 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12711 DAG.getIntPtrConstant(0, DL));
12712 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12713 DAG.getIntPtrConstant(2, DL));
12714 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12715 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12716 static const int ShufMask[] = {0, 2, 4, 6};
12717 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
12720 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
12721 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
12722 if (Subtarget->hasInt256()) {
12723 In = DAG.getBitcast(MVT::v32i8, In);
12725 SmallVector<SDValue,32> pshufbMask;
12726 for (unsigned i = 0; i < 2; ++i) {
12727 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
12728 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
12729 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
12730 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
12731 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
12732 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
12733 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
12734 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
12735 for (unsigned j = 0; j < 8; ++j)
12736 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
12738 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
12739 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
12740 In = DAG.getBitcast(MVT::v4i64, In);
12742 static const int ShufMask[] = {0, 2, -1, -1};
12743 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
12745 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12746 DAG.getIntPtrConstant(0, DL));
12747 return DAG.getBitcast(VT, In);
12750 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12751 DAG.getIntPtrConstant(0, DL));
12753 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12754 DAG.getIntPtrConstant(4, DL));
12756 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
12757 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
12759 // The PSHUFB mask:
12760 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
12761 -1, -1, -1, -1, -1, -1, -1, -1};
12763 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
12764 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
12765 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
12767 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12768 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12770 // The MOVLHPS Mask:
12771 static const int ShufMask2[] = {0, 1, 4, 5};
12772 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
12773 return DAG.getBitcast(MVT::v8i16, res);
12776 // Handle truncation of V256 to V128 using shuffles.
12777 if (!VT.is128BitVector() || !InVT.is256BitVector())
12780 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
12782 unsigned NumElems = VT.getVectorNumElements();
12783 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
12785 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
12786 // Prepare truncation shuffle mask
12787 for (unsigned i = 0; i != NumElems; ++i)
12788 MaskVec[i] = i * 2;
12789 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
12790 DAG.getUNDEF(NVT), &MaskVec[0]);
12791 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
12792 DAG.getIntPtrConstant(0, DL));
12795 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
12796 SelectionDAG &DAG) const {
12797 assert(!Op.getSimpleValueType().isVector());
12799 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12800 /*IsSigned=*/ true, /*IsReplace=*/ false);
12801 SDValue FIST = Vals.first, StackSlot = Vals.second;
12802 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
12803 if (!FIST.getNode())
12806 if (StackSlot.getNode())
12807 // Load the result.
12808 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12809 FIST, StackSlot, MachinePointerInfo(),
12810 false, false, false, 0);
12812 // The node is the result.
12816 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
12817 SelectionDAG &DAG) const {
12818 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12819 /*IsSigned=*/ false, /*IsReplace=*/ false);
12820 SDValue FIST = Vals.first, StackSlot = Vals.second;
12821 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
12822 if (!FIST.getNode())
12825 if (StackSlot.getNode())
12826 // Load the result.
12827 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12828 FIST, StackSlot, MachinePointerInfo(),
12829 false, false, false, 0);
12831 // The node is the result.
12835 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
12837 MVT VT = Op.getSimpleValueType();
12838 SDValue In = Op.getOperand(0);
12839 MVT SVT = In.getSimpleValueType();
12841 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
12843 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
12844 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
12845 In, DAG.getUNDEF(SVT)));
12848 /// The only differences between FABS and FNEG are the mask and the logic op.
12849 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
12850 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
12851 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
12852 "Wrong opcode for lowering FABS or FNEG.");
12854 bool IsFABS = (Op.getOpcode() == ISD::FABS);
12856 // If this is a FABS and it has an FNEG user, bail out to fold the combination
12857 // into an FNABS. We'll lower the FABS after that if it is still in use.
12859 for (SDNode *User : Op->uses())
12860 if (User->getOpcode() == ISD::FNEG)
12864 MVT VT = Op.getSimpleValueType();
12866 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
12867 // decide if we should generate a 16-byte constant mask when we only need 4 or
12868 // 8 bytes for the scalar case.
12874 if (VT.isVector()) {
12876 EltVT = VT.getVectorElementType();
12877 NumElts = VT.getVectorNumElements();
12879 // There are no scalar bitwise logical SSE/AVX instructions, so we
12880 // generate a 16-byte vector constant and logic op even for the scalar case.
12881 // Using a 16-byte mask allows folding the load of the mask with
12882 // the logic op, so it can save (~4 bytes) on code size.
12883 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
12885 NumElts = (VT == MVT::f64) ? 2 : 4;
12888 unsigned EltBits = EltVT.getSizeInBits();
12889 LLVMContext *Context = DAG.getContext();
12890 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
12892 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
12893 Constant *C = ConstantInt::get(*Context, MaskElt);
12894 C = ConstantVector::getSplat(NumElts, C);
12895 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12896 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
12897 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12899 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12900 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12901 false, false, false, Alignment);
12903 SDValue Op0 = Op.getOperand(0);
12904 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
12906 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
12907 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
12910 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
12912 // For the scalar case extend to a 128-bit vector, perform the logic op,
12913 // and extract the scalar result back out.
12914 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
12915 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
12916 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
12917 DAG.getIntPtrConstant(0, dl));
12920 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
12921 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12922 LLVMContext *Context = DAG.getContext();
12923 SDValue Op0 = Op.getOperand(0);
12924 SDValue Op1 = Op.getOperand(1);
12926 MVT VT = Op.getSimpleValueType();
12927 MVT SrcVT = Op1.getSimpleValueType();
12929 // If second operand is smaller, extend it first.
12930 if (SrcVT.bitsLT(VT)) {
12931 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
12934 // And if it is bigger, shrink it first.
12935 if (SrcVT.bitsGT(VT)) {
12936 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
12940 // At this point the operands and the result should have the same
12941 // type, and that won't be f80 since that is not custom lowered.
12943 const fltSemantics &Sem =
12944 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
12945 const unsigned SizeInBits = VT.getSizeInBits();
12947 SmallVector<Constant *, 4> CV(
12948 VT == MVT::f64 ? 2 : 4,
12949 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
12951 // First, clear all bits but the sign bit from the second operand (sign).
12952 CV[0] = ConstantFP::get(*Context,
12953 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
12954 Constant *C = ConstantVector::get(CV);
12955 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
12956 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
12958 // Perform all logic operations as 16-byte vectors because there are no
12959 // scalar FP logic instructions in SSE. This allows load folding of the
12960 // constants into the logic instructions.
12961 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
12963 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12964 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12965 false, false, false, 16);
12966 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
12967 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
12969 // Next, clear the sign bit from the first operand (magnitude).
12970 // If it's a constant, we can clear it here.
12971 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
12972 APFloat APF = Op0CN->getValueAPF();
12973 // If the magnitude is a positive zero, the sign bit alone is enough.
12974 if (APF.isPosZero())
12975 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
12976 DAG.getIntPtrConstant(0, dl));
12978 CV[0] = ConstantFP::get(*Context, APF);
12980 CV[0] = ConstantFP::get(
12982 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
12984 C = ConstantVector::get(CV);
12985 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
12987 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12988 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12989 false, false, false, 16);
12990 // If the magnitude operand wasn't a constant, we need to AND out the sign.
12991 if (!isa<ConstantFPSDNode>(Op0)) {
12992 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
12993 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
12995 // OR the magnitude value with the sign bit.
12996 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
12997 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
12998 DAG.getIntPtrConstant(0, dl));
13001 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13002 SDValue N0 = Op.getOperand(0);
13004 MVT VT = Op.getSimpleValueType();
13006 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13007 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13008 DAG.getConstant(1, dl, VT));
13009 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13012 // Check whether an OR'd tree is PTEST-able.
13013 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13014 SelectionDAG &DAG) {
13015 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13017 if (!Subtarget->hasSSE41())
13020 if (!Op->hasOneUse())
13023 SDNode *N = Op.getNode();
13026 SmallVector<SDValue, 8> Opnds;
13027 DenseMap<SDValue, unsigned> VecInMap;
13028 SmallVector<SDValue, 8> VecIns;
13029 EVT VT = MVT::Other;
13031 // Recognize a special case where a vector is casted into wide integer to
13033 Opnds.push_back(N->getOperand(0));
13034 Opnds.push_back(N->getOperand(1));
13036 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13037 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13038 // BFS traverse all OR'd operands.
13039 if (I->getOpcode() == ISD::OR) {
13040 Opnds.push_back(I->getOperand(0));
13041 Opnds.push_back(I->getOperand(1));
13042 // Re-evaluate the number of nodes to be traversed.
13043 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13047 // Quit if a non-EXTRACT_VECTOR_ELT
13048 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13051 // Quit if without a constant index.
13052 SDValue Idx = I->getOperand(1);
13053 if (!isa<ConstantSDNode>(Idx))
13056 SDValue ExtractedFromVec = I->getOperand(0);
13057 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13058 if (M == VecInMap.end()) {
13059 VT = ExtractedFromVec.getValueType();
13060 // Quit if not 128/256-bit vector.
13061 if (!VT.is128BitVector() && !VT.is256BitVector())
13063 // Quit if not the same type.
13064 if (VecInMap.begin() != VecInMap.end() &&
13065 VT != VecInMap.begin()->first.getValueType())
13067 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13068 VecIns.push_back(ExtractedFromVec);
13070 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13073 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13074 "Not extracted from 128-/256-bit vector.");
13076 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13078 for (DenseMap<SDValue, unsigned>::const_iterator
13079 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13080 // Quit if not all elements are used.
13081 if (I->second != FullMask)
13085 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13087 // Cast all vectors into TestVT for PTEST.
13088 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13089 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13091 // If more than one full vectors are evaluated, OR them first before PTEST.
13092 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13093 // Each iteration will OR 2 nodes and append the result until there is only
13094 // 1 node left, i.e. the final OR'd value of all vectors.
13095 SDValue LHS = VecIns[Slot];
13096 SDValue RHS = VecIns[Slot + 1];
13097 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13100 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13101 VecIns.back(), VecIns.back());
13104 /// \brief return true if \c Op has a use that doesn't just read flags.
13105 static bool hasNonFlagsUse(SDValue Op) {
13106 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13108 SDNode *User = *UI;
13109 unsigned UOpNo = UI.getOperandNo();
13110 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13111 // Look pass truncate.
13112 UOpNo = User->use_begin().getOperandNo();
13113 User = *User->use_begin();
13116 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13117 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13123 /// Emit nodes that will be selected as "test Op0,Op0", or something
13125 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13126 SelectionDAG &DAG) const {
13127 if (Op.getValueType() == MVT::i1) {
13128 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13129 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13130 DAG.getConstant(0, dl, MVT::i8));
13132 // CF and OF aren't always set the way we want. Determine which
13133 // of these we need.
13134 bool NeedCF = false;
13135 bool NeedOF = false;
13138 case X86::COND_A: case X86::COND_AE:
13139 case X86::COND_B: case X86::COND_BE:
13142 case X86::COND_G: case X86::COND_GE:
13143 case X86::COND_L: case X86::COND_LE:
13144 case X86::COND_O: case X86::COND_NO: {
13145 // Check if we really need to set the
13146 // Overflow flag. If NoSignedWrap is present
13147 // that is not actually needed.
13148 switch (Op->getOpcode()) {
13153 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13154 if (BinNode->Flags.hasNoSignedWrap())
13164 // See if we can use the EFLAGS value from the operand instead of
13165 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13166 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13167 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13168 // Emit a CMP with 0, which is the TEST pattern.
13169 //if (Op.getValueType() == MVT::i1)
13170 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13171 // DAG.getConstant(0, MVT::i1));
13172 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13173 DAG.getConstant(0, dl, Op.getValueType()));
13175 unsigned Opcode = 0;
13176 unsigned NumOperands = 0;
13178 // Truncate operations may prevent the merge of the SETCC instruction
13179 // and the arithmetic instruction before it. Attempt to truncate the operands
13180 // of the arithmetic instruction and use a reduced bit-width instruction.
13181 bool NeedTruncation = false;
13182 SDValue ArithOp = Op;
13183 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13184 SDValue Arith = Op->getOperand(0);
13185 // Both the trunc and the arithmetic op need to have one user each.
13186 if (Arith->hasOneUse())
13187 switch (Arith.getOpcode()) {
13194 NeedTruncation = true;
13200 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13201 // which may be the result of a CAST. We use the variable 'Op', which is the
13202 // non-casted variable when we check for possible users.
13203 switch (ArithOp.getOpcode()) {
13205 // Due to an isel shortcoming, be conservative if this add is likely to be
13206 // selected as part of a load-modify-store instruction. When the root node
13207 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13208 // uses of other nodes in the match, such as the ADD in this case. This
13209 // leads to the ADD being left around and reselected, with the result being
13210 // two adds in the output. Alas, even if none our users are stores, that
13211 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13212 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13213 // climbing the DAG back to the root, and it doesn't seem to be worth the
13215 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13216 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13217 if (UI->getOpcode() != ISD::CopyToReg &&
13218 UI->getOpcode() != ISD::SETCC &&
13219 UI->getOpcode() != ISD::STORE)
13222 if (ConstantSDNode *C =
13223 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13224 // An add of one will be selected as an INC.
13225 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13226 Opcode = X86ISD::INC;
13231 // An add of negative one (subtract of one) will be selected as a DEC.
13232 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13233 Opcode = X86ISD::DEC;
13239 // Otherwise use a regular EFLAGS-setting add.
13240 Opcode = X86ISD::ADD;
13245 // If we have a constant logical shift that's only used in a comparison
13246 // against zero turn it into an equivalent AND. This allows turning it into
13247 // a TEST instruction later.
13248 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13249 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13250 EVT VT = Op.getValueType();
13251 unsigned BitWidth = VT.getSizeInBits();
13252 unsigned ShAmt = Op->getConstantOperandVal(1);
13253 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13255 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13256 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13257 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13258 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13260 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13261 DAG.getConstant(Mask, dl, VT));
13262 DAG.ReplaceAllUsesWith(Op, New);
13268 // If the primary and result isn't used, don't bother using X86ISD::AND,
13269 // because a TEST instruction will be better.
13270 if (!hasNonFlagsUse(Op))
13276 // Due to the ISEL shortcoming noted above, be conservative if this op is
13277 // likely to be selected as part of a load-modify-store instruction.
13278 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13279 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13280 if (UI->getOpcode() == ISD::STORE)
13283 // Otherwise use a regular EFLAGS-setting instruction.
13284 switch (ArithOp.getOpcode()) {
13285 default: llvm_unreachable("unexpected operator!");
13286 case ISD::SUB: Opcode = X86ISD::SUB; break;
13287 case ISD::XOR: Opcode = X86ISD::XOR; break;
13288 case ISD::AND: Opcode = X86ISD::AND; break;
13290 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13291 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13292 if (EFLAGS.getNode())
13295 Opcode = X86ISD::OR;
13309 return SDValue(Op.getNode(), 1);
13315 // If we found that truncation is beneficial, perform the truncation and
13317 if (NeedTruncation) {
13318 EVT VT = Op.getValueType();
13319 SDValue WideVal = Op->getOperand(0);
13320 EVT WideVT = WideVal.getValueType();
13321 unsigned ConvertedOp = 0;
13322 // Use a target machine opcode to prevent further DAGCombine
13323 // optimizations that may separate the arithmetic operations
13324 // from the setcc node.
13325 switch (WideVal.getOpcode()) {
13327 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13328 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13329 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13330 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13331 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13335 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13336 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13337 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13338 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13339 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13345 // Emit a CMP with 0, which is the TEST pattern.
13346 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13347 DAG.getConstant(0, dl, Op.getValueType()));
13349 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13350 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13352 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13353 DAG.ReplaceAllUsesWith(Op, New);
13354 return SDValue(New.getNode(), 1);
13357 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13359 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13360 SDLoc dl, SelectionDAG &DAG) const {
13361 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13362 if (C->getAPIntValue() == 0)
13363 return EmitTest(Op0, X86CC, dl, DAG);
13365 if (Op0.getValueType() == MVT::i1)
13366 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
13369 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13370 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13371 // Do the comparison at i32 if it's smaller, besides the Atom case.
13372 // This avoids subregister aliasing issues. Keep the smaller reference
13373 // if we're optimizing for size, however, as that'll allow better folding
13374 // of memory operations.
13375 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13376 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13377 !Subtarget->isAtom()) {
13378 unsigned ExtendOp =
13379 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13380 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13381 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13383 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13384 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13385 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13387 return SDValue(Sub.getNode(), 1);
13389 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13392 /// Convert a comparison if required by the subtarget.
13393 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13394 SelectionDAG &DAG) const {
13395 // If the subtarget does not support the FUCOMI instruction, floating-point
13396 // comparisons have to be converted.
13397 if (Subtarget->hasCMov() ||
13398 Cmp.getOpcode() != X86ISD::CMP ||
13399 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13400 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13403 // The instruction selector will select an FUCOM instruction instead of
13404 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13405 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13406 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13408 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13409 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13410 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13411 DAG.getConstant(8, dl, MVT::i8));
13412 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13413 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13416 /// The minimum architected relative accuracy is 2^-12. We need one
13417 /// Newton-Raphson step to have a good float result (24 bits of precision).
13418 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13419 DAGCombinerInfo &DCI,
13420 unsigned &RefinementSteps,
13421 bool &UseOneConstNR) const {
13422 EVT VT = Op.getValueType();
13423 const char *RecipOp;
13425 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13426 // TODO: Add support for AVX512 (v16f32).
13427 // It is likely not profitable to do this for f64 because a double-precision
13428 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13429 // instructions: convert to single, rsqrtss, convert back to double, refine
13430 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13431 // along with FMA, this could be a throughput win.
13432 if (VT == MVT::f32 && Subtarget->hasSSE1())
13434 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13435 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13436 RecipOp = "vec-sqrtf";
13440 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13441 if (!Recips.isEnabled(RecipOp))
13444 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13445 UseOneConstNR = false;
13446 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13449 /// The minimum architected relative accuracy is 2^-12. We need one
13450 /// Newton-Raphson step to have a good float result (24 bits of precision).
13451 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13452 DAGCombinerInfo &DCI,
13453 unsigned &RefinementSteps) const {
13454 EVT VT = Op.getValueType();
13455 const char *RecipOp;
13457 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13458 // TODO: Add support for AVX512 (v16f32).
13459 // It is likely not profitable to do this for f64 because a double-precision
13460 // reciprocal estimate with refinement on x86 prior to FMA requires
13461 // 15 instructions: convert to single, rcpss, convert back to double, refine
13462 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13463 // along with FMA, this could be a throughput win.
13464 if (VT == MVT::f32 && Subtarget->hasSSE1())
13466 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13467 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13468 RecipOp = "vec-divf";
13472 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13473 if (!Recips.isEnabled(RecipOp))
13476 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13477 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13480 /// If we have at least two divisions that use the same divisor, convert to
13481 /// multplication by a reciprocal. This may need to be adjusted for a given
13482 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13483 /// This is because we still need one division to calculate the reciprocal and
13484 /// then we need two multiplies by that reciprocal as replacements for the
13485 /// original divisions.
13486 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
13490 static bool isAllOnes(SDValue V) {
13491 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13492 return C && C->isAllOnesValue();
13495 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13496 /// if it's possible.
13497 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13498 SDLoc dl, SelectionDAG &DAG) const {
13499 SDValue Op0 = And.getOperand(0);
13500 SDValue Op1 = And.getOperand(1);
13501 if (Op0.getOpcode() == ISD::TRUNCATE)
13502 Op0 = Op0.getOperand(0);
13503 if (Op1.getOpcode() == ISD::TRUNCATE)
13504 Op1 = Op1.getOperand(0);
13507 if (Op1.getOpcode() == ISD::SHL)
13508 std::swap(Op0, Op1);
13509 if (Op0.getOpcode() == ISD::SHL) {
13510 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13511 if (And00C->getZExtValue() == 1) {
13512 // If we looked past a truncate, check that it's only truncating away
13514 unsigned BitWidth = Op0.getValueSizeInBits();
13515 unsigned AndBitWidth = And.getValueSizeInBits();
13516 if (BitWidth > AndBitWidth) {
13518 DAG.computeKnownBits(Op0, Zeros, Ones);
13519 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13523 RHS = Op0.getOperand(1);
13525 } else if (Op1.getOpcode() == ISD::Constant) {
13526 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13527 uint64_t AndRHSVal = AndRHS->getZExtValue();
13528 SDValue AndLHS = Op0;
13530 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13531 LHS = AndLHS.getOperand(0);
13532 RHS = AndLHS.getOperand(1);
13535 // Use BT if the immediate can't be encoded in a TEST instruction.
13536 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13538 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13542 if (LHS.getNode()) {
13543 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13544 // instruction. Since the shift amount is in-range-or-undefined, we know
13545 // that doing a bittest on the i32 value is ok. We extend to i32 because
13546 // the encoding for the i16 version is larger than the i32 version.
13547 // Also promote i16 to i32 for performance / code size reason.
13548 if (LHS.getValueType() == MVT::i8 ||
13549 LHS.getValueType() == MVT::i16)
13550 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13552 // If the operand types disagree, extend the shift amount to match. Since
13553 // BT ignores high bits (like shifts) we can use anyextend.
13554 if (LHS.getValueType() != RHS.getValueType())
13555 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13557 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13558 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13559 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13560 DAG.getConstant(Cond, dl, MVT::i8), BT);
13566 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13568 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13573 // SSE Condition code mapping:
13582 switch (SetCCOpcode) {
13583 default: llvm_unreachable("Unexpected SETCC condition");
13585 case ISD::SETEQ: SSECC = 0; break;
13587 case ISD::SETGT: Swap = true; // Fallthrough
13589 case ISD::SETOLT: SSECC = 1; break;
13591 case ISD::SETGE: Swap = true; // Fallthrough
13593 case ISD::SETOLE: SSECC = 2; break;
13594 case ISD::SETUO: SSECC = 3; break;
13596 case ISD::SETNE: SSECC = 4; break;
13597 case ISD::SETULE: Swap = true; // Fallthrough
13598 case ISD::SETUGE: SSECC = 5; break;
13599 case ISD::SETULT: Swap = true; // Fallthrough
13600 case ISD::SETUGT: SSECC = 6; break;
13601 case ISD::SETO: SSECC = 7; break;
13603 case ISD::SETONE: SSECC = 8; break;
13606 std::swap(Op0, Op1);
13611 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
13612 // ones, and then concatenate the result back.
13613 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
13614 MVT VT = Op.getSimpleValueType();
13616 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
13617 "Unsupported value type for operation");
13619 unsigned NumElems = VT.getVectorNumElements();
13621 SDValue CC = Op.getOperand(2);
13623 // Extract the LHS vectors
13624 SDValue LHS = Op.getOperand(0);
13625 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
13626 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
13628 // Extract the RHS vectors
13629 SDValue RHS = Op.getOperand(1);
13630 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
13631 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
13633 // Issue the operation on the smaller types and concatenate the result back
13634 MVT EltVT = VT.getVectorElementType();
13635 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
13636 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
13637 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
13638 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
13641 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
13642 SDValue Op0 = Op.getOperand(0);
13643 SDValue Op1 = Op.getOperand(1);
13644 SDValue CC = Op.getOperand(2);
13645 MVT VT = Op.getSimpleValueType();
13648 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
13649 "Unexpected type for boolean compare operation");
13650 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13651 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
13652 DAG.getConstant(-1, dl, VT));
13653 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
13654 DAG.getConstant(-1, dl, VT));
13655 switch (SetCCOpcode) {
13656 default: llvm_unreachable("Unexpected SETCC condition");
13658 // (x == y) -> ~(x ^ y)
13659 return DAG.getNode(ISD::XOR, dl, VT,
13660 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
13661 DAG.getConstant(-1, dl, VT));
13663 // (x != y) -> (x ^ y)
13664 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
13667 // (x > y) -> (x & ~y)
13668 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
13671 // (x < y) -> (~x & y)
13672 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
13675 // (x <= y) -> (~x | y)
13676 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
13679 // (x >=y) -> (x | ~y)
13680 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
13684 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
13685 const X86Subtarget *Subtarget) {
13686 SDValue Op0 = Op.getOperand(0);
13687 SDValue Op1 = Op.getOperand(1);
13688 SDValue CC = Op.getOperand(2);
13689 MVT VT = Op.getSimpleValueType();
13692 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
13693 Op.getValueType().getScalarType() == MVT::i1 &&
13694 "Cannot set masked compare for this operation");
13696 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13698 bool Unsigned = false;
13701 switch (SetCCOpcode) {
13702 default: llvm_unreachable("Unexpected SETCC condition");
13703 case ISD::SETNE: SSECC = 4; break;
13704 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
13705 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
13706 case ISD::SETLT: Swap = true; //fall-through
13707 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
13708 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
13709 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
13710 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
13711 case ISD::SETULE: Unsigned = true; //fall-through
13712 case ISD::SETLE: SSECC = 2; break;
13716 std::swap(Op0, Op1);
13718 return DAG.getNode(Opc, dl, VT, Op0, Op1);
13719 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
13720 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13721 DAG.getConstant(SSECC, dl, MVT::i8));
13724 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
13725 /// operand \p Op1. If non-trivial (for example because it's not constant)
13726 /// return an empty value.
13727 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
13729 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
13733 MVT VT = Op1.getSimpleValueType();
13734 MVT EVT = VT.getVectorElementType();
13735 unsigned n = VT.getVectorNumElements();
13736 SmallVector<SDValue, 8> ULTOp1;
13738 for (unsigned i = 0; i < n; ++i) {
13739 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
13740 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
13743 // Avoid underflow.
13744 APInt Val = Elt->getAPIntValue();
13748 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
13751 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
13754 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
13755 SelectionDAG &DAG) {
13756 SDValue Op0 = Op.getOperand(0);
13757 SDValue Op1 = Op.getOperand(1);
13758 SDValue CC = Op.getOperand(2);
13759 MVT VT = Op.getSimpleValueType();
13760 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13761 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
13766 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
13767 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
13770 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
13771 unsigned Opc = X86ISD::CMPP;
13772 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
13773 assert(VT.getVectorNumElements() <= 16);
13774 Opc = X86ISD::CMPM;
13776 // In the two special cases we can't handle, emit two comparisons.
13779 unsigned CombineOpc;
13780 if (SetCCOpcode == ISD::SETUEQ) {
13781 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
13783 assert(SetCCOpcode == ISD::SETONE);
13784 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
13787 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13788 DAG.getConstant(CC0, dl, MVT::i8));
13789 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13790 DAG.getConstant(CC1, dl, MVT::i8));
13791 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
13793 // Handle all other FP comparisons here.
13794 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13795 DAG.getConstant(SSECC, dl, MVT::i8));
13798 // Break 256-bit integer vector compare into smaller ones.
13799 if (VT.is256BitVector() && !Subtarget->hasInt256())
13800 return Lower256IntVSETCC(Op, DAG);
13802 EVT OpVT = Op1.getValueType();
13803 if (OpVT.getVectorElementType() == MVT::i1)
13804 return LowerBoolVSETCC_AVX512(Op, DAG);
13806 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
13807 if (Subtarget->hasAVX512()) {
13808 if (Op1.getValueType().is512BitVector() ||
13809 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
13810 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
13811 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
13813 // In AVX-512 architecture setcc returns mask with i1 elements,
13814 // But there is no compare instruction for i8 and i16 elements in KNL.
13815 // We are not talking about 512-bit operands in this case, these
13816 // types are illegal.
13818 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
13819 OpVT.getVectorElementType().getSizeInBits() >= 8))
13820 return DAG.getNode(ISD::TRUNCATE, dl, VT,
13821 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
13824 // We are handling one of the integer comparisons here. Since SSE only has
13825 // GT and EQ comparisons for integer, swapping operands and multiple
13826 // operations may be required for some comparisons.
13828 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
13829 bool Subus = false;
13831 switch (SetCCOpcode) {
13832 default: llvm_unreachable("Unexpected SETCC condition");
13833 case ISD::SETNE: Invert = true;
13834 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
13835 case ISD::SETLT: Swap = true;
13836 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
13837 case ISD::SETGE: Swap = true;
13838 case ISD::SETLE: Opc = X86ISD::PCMPGT;
13839 Invert = true; break;
13840 case ISD::SETULT: Swap = true;
13841 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
13842 FlipSigns = true; break;
13843 case ISD::SETUGE: Swap = true;
13844 case ISD::SETULE: Opc = X86ISD::PCMPGT;
13845 FlipSigns = true; Invert = true; break;
13848 // Special case: Use min/max operations for SETULE/SETUGE
13849 MVT VET = VT.getVectorElementType();
13851 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
13852 || (Subtarget->hasSSE2() && (VET == MVT::i8));
13855 switch (SetCCOpcode) {
13857 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
13858 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
13861 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
13864 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
13865 if (!MinMax && hasSubus) {
13866 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
13868 // t = psubus Op0, Op1
13869 // pcmpeq t, <0..0>
13870 switch (SetCCOpcode) {
13872 case ISD::SETULT: {
13873 // If the comparison is against a constant we can turn this into a
13874 // setule. With psubus, setule does not require a swap. This is
13875 // beneficial because the constant in the register is no longer
13876 // destructed as the destination so it can be hoisted out of a loop.
13877 // Only do this pre-AVX since vpcmp* is no longer destructive.
13878 if (Subtarget->hasAVX())
13880 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
13881 if (ULEOp1.getNode()) {
13883 Subus = true; Invert = false; Swap = false;
13887 // Psubus is better than flip-sign because it requires no inversion.
13888 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
13889 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
13893 Opc = X86ISD::SUBUS;
13899 std::swap(Op0, Op1);
13901 // Check that the operation in question is available (most are plain SSE2,
13902 // but PCMPGTQ and PCMPEQQ have different requirements).
13903 if (VT == MVT::v2i64) {
13904 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
13905 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
13907 // First cast everything to the right type.
13908 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
13909 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
13911 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13912 // bits of the inputs before performing those operations. The lower
13913 // compare is always unsigned.
13916 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
13918 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
13919 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
13920 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
13921 Sign, Zero, Sign, Zero);
13923 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
13924 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
13926 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
13927 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
13928 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
13930 // Create masks for only the low parts/high parts of the 64 bit integers.
13931 static const int MaskHi[] = { 1, 1, 3, 3 };
13932 static const int MaskLo[] = { 0, 0, 2, 2 };
13933 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
13934 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
13935 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
13937 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
13938 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
13941 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13943 return DAG.getBitcast(VT, Result);
13946 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
13947 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
13948 // pcmpeqd + pshufd + pand.
13949 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
13951 // First cast everything to the right type.
13952 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
13953 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
13956 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
13958 // Make sure the lower and upper halves are both all-ones.
13959 static const int Mask[] = { 1, 0, 3, 2 };
13960 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
13961 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
13964 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13966 return DAG.getBitcast(VT, Result);
13970 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13971 // bits of the inputs before performing those operations.
13973 EVT EltVT = VT.getVectorElementType();
13974 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
13976 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
13977 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
13980 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
13982 // If the logical-not of the result is required, perform that now.
13984 Result = DAG.getNOT(dl, Result, VT);
13987 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
13990 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
13991 getZeroVector(VT, Subtarget, DAG, dl));
13996 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
13998 MVT VT = Op.getSimpleValueType();
14000 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14002 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14003 && "SetCC type must be 8-bit or 1-bit integer");
14004 SDValue Op0 = Op.getOperand(0);
14005 SDValue Op1 = Op.getOperand(1);
14007 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14009 // Optimize to BT if possible.
14010 // Lower (X & (1 << N)) == 0 to BT(X, N).
14011 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14012 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14013 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14014 Op1.getOpcode() == ISD::Constant &&
14015 cast<ConstantSDNode>(Op1)->isNullValue() &&
14016 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14017 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
14018 if (NewSetCC.getNode()) {
14020 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14025 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14027 if (Op1.getOpcode() == ISD::Constant &&
14028 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
14029 cast<ConstantSDNode>(Op1)->isNullValue()) &&
14030 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14032 // If the input is a setcc, then reuse the input setcc or use a new one with
14033 // the inverted condition.
14034 if (Op0.getOpcode() == X86ISD::SETCC) {
14035 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14036 bool Invert = (CC == ISD::SETNE) ^
14037 cast<ConstantSDNode>(Op1)->isNullValue();
14041 CCode = X86::GetOppositeBranchCondition(CCode);
14042 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14043 DAG.getConstant(CCode, dl, MVT::i8),
14044 Op0.getOperand(1));
14046 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14050 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
14051 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
14052 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14054 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14055 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14058 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14059 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14060 if (X86CC == X86::COND_INVALID)
14063 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14064 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14065 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14066 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14068 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14072 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14073 static bool isX86LogicalCmp(SDValue Op) {
14074 unsigned Opc = Op.getNode()->getOpcode();
14075 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14076 Opc == X86ISD::SAHF)
14078 if (Op.getResNo() == 1 &&
14079 (Opc == X86ISD::ADD ||
14080 Opc == X86ISD::SUB ||
14081 Opc == X86ISD::ADC ||
14082 Opc == X86ISD::SBB ||
14083 Opc == X86ISD::SMUL ||
14084 Opc == X86ISD::UMUL ||
14085 Opc == X86ISD::INC ||
14086 Opc == X86ISD::DEC ||
14087 Opc == X86ISD::OR ||
14088 Opc == X86ISD::XOR ||
14089 Opc == X86ISD::AND))
14092 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14098 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14099 if (V.getOpcode() != ISD::TRUNCATE)
14102 SDValue VOp0 = V.getOperand(0);
14103 unsigned InBits = VOp0.getValueSizeInBits();
14104 unsigned Bits = V.getValueSizeInBits();
14105 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14108 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14109 bool addTest = true;
14110 SDValue Cond = Op.getOperand(0);
14111 SDValue Op1 = Op.getOperand(1);
14112 SDValue Op2 = Op.getOperand(2);
14114 EVT VT = Op1.getValueType();
14117 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14118 // are available or VBLENDV if AVX is available.
14119 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14120 if (Cond.getOpcode() == ISD::SETCC &&
14121 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14122 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14123 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
14124 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14125 int SSECC = translateX86FSETCC(
14126 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14129 if (Subtarget->hasAVX512()) {
14130 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14131 DAG.getConstant(SSECC, DL, MVT::i8));
14132 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14135 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14136 DAG.getConstant(SSECC, DL, MVT::i8));
14138 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14139 // of 3 logic instructions for size savings and potentially speed.
14140 // Unfortunately, there is no scalar form of VBLENDV.
14142 // If either operand is a constant, don't try this. We can expect to
14143 // optimize away at least one of the logic instructions later in that
14144 // case, so that sequence would be faster than a variable blend.
14146 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14147 // uses XMM0 as the selection register. That may need just as many
14148 // instructions as the AND/ANDN/OR sequence due to register moves, so
14151 if (Subtarget->hasAVX() &&
14152 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14154 // Convert to vectors, do a VSELECT, and convert back to scalar.
14155 // All of the conversions should be optimized away.
14157 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14158 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14159 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14160 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14162 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14163 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14165 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14167 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14168 VSel, DAG.getIntPtrConstant(0, DL));
14170 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14171 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14172 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14176 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
14178 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14179 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14180 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14181 Op1Scalar = Op1.getOperand(0);
14183 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14184 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14185 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14186 Op2Scalar = Op2.getOperand(0);
14187 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14188 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14189 Op1Scalar.getValueType(),
14190 Cond, Op1Scalar, Op2Scalar);
14191 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14192 return DAG.getBitcast(VT, newSelect);
14193 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14194 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14195 DAG.getIntPtrConstant(0, DL));
14199 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14200 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14201 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14202 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14203 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14204 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14205 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14207 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14210 if (Cond.getOpcode() == ISD::SETCC) {
14211 SDValue NewCond = LowerSETCC(Cond, DAG);
14212 if (NewCond.getNode())
14216 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14217 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14218 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14219 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14220 if (Cond.getOpcode() == X86ISD::SETCC &&
14221 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14222 isZero(Cond.getOperand(1).getOperand(1))) {
14223 SDValue Cmp = Cond.getOperand(1);
14225 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14227 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14228 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14229 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14231 SDValue CmpOp0 = Cmp.getOperand(0);
14232 // Apply further optimizations for special cases
14233 // (select (x != 0), -1, 0) -> neg & sbb
14234 // (select (x == 0), 0, -1) -> neg & sbb
14235 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14236 if (YC->isNullValue() &&
14237 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14238 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14239 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14240 DAG.getConstant(0, DL,
14241 CmpOp0.getValueType()),
14243 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14244 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14245 SDValue(Neg.getNode(), 1));
14249 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14250 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14251 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14253 SDValue Res = // Res = 0 or -1.
14254 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14255 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14257 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14258 Res = DAG.getNOT(DL, Res, Res.getValueType());
14260 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14261 if (!N2C || !N2C->isNullValue())
14262 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14267 // Look past (and (setcc_carry (cmp ...)), 1).
14268 if (Cond.getOpcode() == ISD::AND &&
14269 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14270 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14271 if (C && C->getAPIntValue() == 1)
14272 Cond = Cond.getOperand(0);
14275 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14276 // setting operand in place of the X86ISD::SETCC.
14277 unsigned CondOpcode = Cond.getOpcode();
14278 if (CondOpcode == X86ISD::SETCC ||
14279 CondOpcode == X86ISD::SETCC_CARRY) {
14280 CC = Cond.getOperand(0);
14282 SDValue Cmp = Cond.getOperand(1);
14283 unsigned Opc = Cmp.getOpcode();
14284 MVT VT = Op.getSimpleValueType();
14286 bool IllegalFPCMov = false;
14287 if (VT.isFloatingPoint() && !VT.isVector() &&
14288 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14289 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14291 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14292 Opc == X86ISD::BT) { // FIXME
14296 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14297 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14298 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14299 Cond.getOperand(0).getValueType() != MVT::i8)) {
14300 SDValue LHS = Cond.getOperand(0);
14301 SDValue RHS = Cond.getOperand(1);
14302 unsigned X86Opcode;
14305 switch (CondOpcode) {
14306 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14307 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14308 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14309 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14310 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14311 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14312 default: llvm_unreachable("unexpected overflowing operator");
14314 if (CondOpcode == ISD::UMULO)
14315 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14318 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14320 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14322 if (CondOpcode == ISD::UMULO)
14323 Cond = X86Op.getValue(2);
14325 Cond = X86Op.getValue(1);
14327 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14332 // Look past the truncate if the high bits are known zero.
14333 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14334 Cond = Cond.getOperand(0);
14336 // We know the result of AND is compared against zero. Try to match
14338 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14339 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14340 if (NewSetCC.getNode()) {
14341 CC = NewSetCC.getOperand(0);
14342 Cond = NewSetCC.getOperand(1);
14349 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14350 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14353 // a < b ? -1 : 0 -> RES = ~setcc_carry
14354 // a < b ? 0 : -1 -> RES = setcc_carry
14355 // a >= b ? -1 : 0 -> RES = setcc_carry
14356 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14357 if (Cond.getOpcode() == X86ISD::SUB) {
14358 Cond = ConvertCmpIfNecessary(Cond, DAG);
14359 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14361 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14362 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14363 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14364 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14366 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14367 return DAG.getNOT(DL, Res, Res.getValueType());
14372 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14373 // widen the cmov and push the truncate through. This avoids introducing a new
14374 // branch during isel and doesn't add any extensions.
14375 if (Op.getValueType() == MVT::i8 &&
14376 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14377 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14378 if (T1.getValueType() == T2.getValueType() &&
14379 // Blacklist CopyFromReg to avoid partial register stalls.
14380 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14381 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14382 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14383 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14387 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14388 // condition is true.
14389 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14390 SDValue Ops[] = { Op2, Op1, CC, Cond };
14391 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14394 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14395 const X86Subtarget *Subtarget,
14396 SelectionDAG &DAG) {
14397 MVT VT = Op->getSimpleValueType(0);
14398 SDValue In = Op->getOperand(0);
14399 MVT InVT = In.getSimpleValueType();
14400 MVT VTElt = VT.getVectorElementType();
14401 MVT InVTElt = InVT.getVectorElementType();
14405 if ((InVTElt == MVT::i1) &&
14406 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14407 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14409 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14410 VTElt.getSizeInBits() <= 16)) ||
14412 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14413 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14415 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14416 VTElt.getSizeInBits() >= 32))))
14417 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14419 unsigned int NumElts = VT.getVectorNumElements();
14421 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14424 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14425 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14426 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14427 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14430 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14431 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14433 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14436 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14438 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14439 if (VT.is512BitVector())
14441 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14444 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14445 const X86Subtarget *Subtarget,
14446 SelectionDAG &DAG) {
14447 SDValue In = Op->getOperand(0);
14448 MVT VT = Op->getSimpleValueType(0);
14449 MVT InVT = In.getSimpleValueType();
14450 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14452 MVT InSVT = InVT.getScalarType();
14453 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
14455 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14457 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14462 // SSE41 targets can use the pmovsx* instructions directly.
14463 if (Subtarget->hasSSE41())
14464 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14466 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
14470 // As SRAI is only available on i16/i32 types, we expand only up to i32
14471 // and handle i64 separately.
14472 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
14473 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
14474 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
14475 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
14476 Curr = DAG.getBitcast(CurrVT, Curr);
14479 SDValue SignExt = Curr;
14480 if (CurrVT != InVT) {
14481 unsigned SignExtShift =
14482 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
14483 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14484 DAG.getConstant(SignExtShift, dl, MVT::i8));
14490 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
14491 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14492 DAG.getConstant(31, dl, MVT::i8));
14493 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
14494 return DAG.getBitcast(VT, Ext);
14500 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14501 SelectionDAG &DAG) {
14502 MVT VT = Op->getSimpleValueType(0);
14503 SDValue In = Op->getOperand(0);
14504 MVT InVT = In.getSimpleValueType();
14507 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
14508 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
14510 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
14511 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
14512 (VT != MVT::v16i16 || InVT != MVT::v16i8))
14515 if (Subtarget->hasInt256())
14516 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14518 // Optimize vectors in AVX mode
14519 // Sign extend v8i16 to v8i32 and
14522 // Divide input vector into two parts
14523 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14524 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14525 // concat the vectors to original VT
14527 unsigned NumElems = InVT.getVectorNumElements();
14528 SDValue Undef = DAG.getUNDEF(InVT);
14530 SmallVector<int,8> ShufMask1(NumElems, -1);
14531 for (unsigned i = 0; i != NumElems/2; ++i)
14534 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
14536 SmallVector<int,8> ShufMask2(NumElems, -1);
14537 for (unsigned i = 0; i != NumElems/2; ++i)
14538 ShufMask2[i] = i + NumElems/2;
14540 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
14542 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
14543 VT.getVectorNumElements()/2);
14545 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
14546 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
14548 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14551 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
14552 // may emit an illegal shuffle but the expansion is still better than scalar
14553 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
14554 // we'll emit a shuffle and a arithmetic shift.
14555 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
14556 // TODO: It is possible to support ZExt by zeroing the undef values during
14557 // the shuffle phase or after the shuffle.
14558 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
14559 SelectionDAG &DAG) {
14560 MVT RegVT = Op.getSimpleValueType();
14561 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
14562 assert(RegVT.isInteger() &&
14563 "We only custom lower integer vector sext loads.");
14565 // Nothing useful we can do without SSE2 shuffles.
14566 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
14568 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
14570 EVT MemVT = Ld->getMemoryVT();
14571 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14572 unsigned RegSz = RegVT.getSizeInBits();
14574 ISD::LoadExtType Ext = Ld->getExtensionType();
14576 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
14577 && "Only anyext and sext are currently implemented.");
14578 assert(MemVT != RegVT && "Cannot extend to the same type");
14579 assert(MemVT.isVector() && "Must load a vector from memory");
14581 unsigned NumElems = RegVT.getVectorNumElements();
14582 unsigned MemSz = MemVT.getSizeInBits();
14583 assert(RegSz > MemSz && "Register size must be greater than the mem size");
14585 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
14586 // The only way in which we have a legal 256-bit vector result but not the
14587 // integer 256-bit operations needed to directly lower a sextload is if we
14588 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
14589 // a 128-bit vector and a normal sign_extend to 256-bits that should get
14590 // correctly legalized. We do this late to allow the canonical form of
14591 // sextload to persist throughout the rest of the DAG combiner -- it wants
14592 // to fold together any extensions it can, and so will fuse a sign_extend
14593 // of an sextload into a sextload targeting a wider value.
14595 if (MemSz == 128) {
14596 // Just switch this to a normal load.
14597 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
14598 "it must be a legal 128-bit vector "
14600 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
14601 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
14602 Ld->isInvariant(), Ld->getAlignment());
14604 assert(MemSz < 128 &&
14605 "Can't extend a type wider than 128 bits to a 256 bit vector!");
14606 // Do an sext load to a 128-bit vector type. We want to use the same
14607 // number of elements, but elements half as wide. This will end up being
14608 // recursively lowered by this routine, but will succeed as we definitely
14609 // have all the necessary features if we're using AVX1.
14611 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
14612 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
14614 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
14615 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
14616 Ld->isNonTemporal(), Ld->isInvariant(),
14617 Ld->getAlignment());
14620 // Replace chain users with the new chain.
14621 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
14622 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
14624 // Finally, do a normal sign-extend to the desired register.
14625 return DAG.getSExtOrTrunc(Load, dl, RegVT);
14628 // All sizes must be a power of two.
14629 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
14630 "Non-power-of-two elements are not custom lowered!");
14632 // Attempt to load the original value using scalar loads.
14633 // Find the largest scalar type that divides the total loaded size.
14634 MVT SclrLoadTy = MVT::i8;
14635 for (MVT Tp : MVT::integer_valuetypes()) {
14636 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
14641 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
14642 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
14644 SclrLoadTy = MVT::f64;
14646 // Calculate the number of scalar loads that we need to perform
14647 // in order to load our vector from memory.
14648 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
14650 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
14651 "Can only lower sext loads with a single scalar load!");
14653 unsigned loadRegZize = RegSz;
14654 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
14657 // Represent our vector as a sequence of elements which are the
14658 // largest scalar that we can load.
14659 EVT LoadUnitVecVT = EVT::getVectorVT(
14660 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
14662 // Represent the data using the same element type that is stored in
14663 // memory. In practice, we ''widen'' MemVT.
14665 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
14666 loadRegZize / MemVT.getScalarType().getSizeInBits());
14668 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
14669 "Invalid vector type");
14671 // We can't shuffle using an illegal type.
14672 assert(TLI.isTypeLegal(WideVecVT) &&
14673 "We only lower types that form legal widened vector types");
14675 SmallVector<SDValue, 8> Chains;
14676 SDValue Ptr = Ld->getBasePtr();
14677 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
14678 TLI.getPointerTy(DAG.getDataLayout()));
14679 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
14681 for (unsigned i = 0; i < NumLoads; ++i) {
14682 // Perform a single load.
14683 SDValue ScalarLoad =
14684 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
14685 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
14686 Ld->getAlignment());
14687 Chains.push_back(ScalarLoad.getValue(1));
14688 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
14689 // another round of DAGCombining.
14691 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
14693 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
14694 ScalarLoad, DAG.getIntPtrConstant(i, dl));
14696 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
14699 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
14701 // Bitcast the loaded value to a vector of the original element type, in
14702 // the size of the target vector type.
14703 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
14704 unsigned SizeRatio = RegSz / MemSz;
14706 if (Ext == ISD::SEXTLOAD) {
14707 // If we have SSE4.1, we can directly emit a VSEXT node.
14708 if (Subtarget->hasSSE41()) {
14709 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
14710 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14714 // Otherwise we'll shuffle the small elements in the high bits of the
14715 // larger type and perform an arithmetic shift. If the shift is not legal
14716 // it's better to scalarize.
14717 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
14718 "We can't implement a sext load without an arithmetic right shift!");
14720 // Redistribute the loaded elements into the different locations.
14721 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14722 for (unsigned i = 0; i != NumElems; ++i)
14723 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
14725 SDValue Shuff = DAG.getVectorShuffle(
14726 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14728 Shuff = DAG.getBitcast(RegVT, Shuff);
14730 // Build the arithmetic shift.
14731 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
14732 MemVT.getVectorElementType().getSizeInBits();
14734 DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
14735 DAG.getConstant(Amt, dl, RegVT));
14737 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14741 // Redistribute the loaded elements into the different locations.
14742 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14743 for (unsigned i = 0; i != NumElems; ++i)
14744 ShuffleVec[i * SizeRatio] = i;
14746 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
14747 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14749 // Bitcast to the requested type.
14750 Shuff = DAG.getBitcast(RegVT, Shuff);
14751 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14755 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
14756 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
14757 // from the AND / OR.
14758 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
14759 Opc = Op.getOpcode();
14760 if (Opc != ISD::OR && Opc != ISD::AND)
14762 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14763 Op.getOperand(0).hasOneUse() &&
14764 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
14765 Op.getOperand(1).hasOneUse());
14768 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
14769 // 1 and that the SETCC node has a single use.
14770 static bool isXor1OfSetCC(SDValue Op) {
14771 if (Op.getOpcode() != ISD::XOR)
14773 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
14774 if (N1C && N1C->getAPIntValue() == 1) {
14775 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14776 Op.getOperand(0).hasOneUse();
14781 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
14782 bool addTest = true;
14783 SDValue Chain = Op.getOperand(0);
14784 SDValue Cond = Op.getOperand(1);
14785 SDValue Dest = Op.getOperand(2);
14788 bool Inverted = false;
14790 if (Cond.getOpcode() == ISD::SETCC) {
14791 // Check for setcc([su]{add,sub,mul}o == 0).
14792 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
14793 isa<ConstantSDNode>(Cond.getOperand(1)) &&
14794 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
14795 Cond.getOperand(0).getResNo() == 1 &&
14796 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
14797 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
14798 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
14799 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
14800 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
14801 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
14803 Cond = Cond.getOperand(0);
14805 SDValue NewCond = LowerSETCC(Cond, DAG);
14806 if (NewCond.getNode())
14811 // FIXME: LowerXALUO doesn't handle these!!
14812 else if (Cond.getOpcode() == X86ISD::ADD ||
14813 Cond.getOpcode() == X86ISD::SUB ||
14814 Cond.getOpcode() == X86ISD::SMUL ||
14815 Cond.getOpcode() == X86ISD::UMUL)
14816 Cond = LowerXALUO(Cond, DAG);
14819 // Look pass (and (setcc_carry (cmp ...)), 1).
14820 if (Cond.getOpcode() == ISD::AND &&
14821 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14822 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14823 if (C && C->getAPIntValue() == 1)
14824 Cond = Cond.getOperand(0);
14827 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14828 // setting operand in place of the X86ISD::SETCC.
14829 unsigned CondOpcode = Cond.getOpcode();
14830 if (CondOpcode == X86ISD::SETCC ||
14831 CondOpcode == X86ISD::SETCC_CARRY) {
14832 CC = Cond.getOperand(0);
14834 SDValue Cmp = Cond.getOperand(1);
14835 unsigned Opc = Cmp.getOpcode();
14836 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
14837 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
14841 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
14845 // These can only come from an arithmetic instruction with overflow,
14846 // e.g. SADDO, UADDO.
14847 Cond = Cond.getNode()->getOperand(1);
14853 CondOpcode = Cond.getOpcode();
14854 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14855 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14856 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14857 Cond.getOperand(0).getValueType() != MVT::i8)) {
14858 SDValue LHS = Cond.getOperand(0);
14859 SDValue RHS = Cond.getOperand(1);
14860 unsigned X86Opcode;
14863 // Keep this in sync with LowerXALUO, otherwise we might create redundant
14864 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
14866 switch (CondOpcode) {
14867 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14869 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14871 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
14874 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14875 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14877 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14879 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
14882 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14883 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14884 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14885 default: llvm_unreachable("unexpected overflowing operator");
14888 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
14889 if (CondOpcode == ISD::UMULO)
14890 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14893 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14895 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
14897 if (CondOpcode == ISD::UMULO)
14898 Cond = X86Op.getValue(2);
14900 Cond = X86Op.getValue(1);
14902 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14906 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
14907 SDValue Cmp = Cond.getOperand(0).getOperand(1);
14908 if (CondOpc == ISD::OR) {
14909 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
14910 // two branches instead of an explicit OR instruction with a
14912 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14913 isX86LogicalCmp(Cmp)) {
14914 CC = Cond.getOperand(0).getOperand(0);
14915 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14916 Chain, Dest, CC, Cmp);
14917 CC = Cond.getOperand(1).getOperand(0);
14921 } else { // ISD::AND
14922 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
14923 // two branches instead of an explicit AND instruction with a
14924 // separate test. However, we only do this if this block doesn't
14925 // have a fall-through edge, because this requires an explicit
14926 // jmp when the condition is false.
14927 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14928 isX86LogicalCmp(Cmp) &&
14929 Op.getNode()->hasOneUse()) {
14930 X86::CondCode CCode =
14931 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14932 CCode = X86::GetOppositeBranchCondition(CCode);
14933 CC = DAG.getConstant(CCode, dl, MVT::i8);
14934 SDNode *User = *Op.getNode()->use_begin();
14935 // Look for an unconditional branch following this conditional branch.
14936 // We need this because we need to reverse the successors in order
14937 // to implement FCMP_OEQ.
14938 if (User->getOpcode() == ISD::BR) {
14939 SDValue FalseBB = User->getOperand(1);
14941 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14942 assert(NewBR == User);
14946 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14947 Chain, Dest, CC, Cmp);
14948 X86::CondCode CCode =
14949 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
14950 CCode = X86::GetOppositeBranchCondition(CCode);
14951 CC = DAG.getConstant(CCode, dl, MVT::i8);
14957 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
14958 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
14959 // It should be transformed during dag combiner except when the condition
14960 // is set by a arithmetics with overflow node.
14961 X86::CondCode CCode =
14962 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14963 CCode = X86::GetOppositeBranchCondition(CCode);
14964 CC = DAG.getConstant(CCode, dl, MVT::i8);
14965 Cond = Cond.getOperand(0).getOperand(1);
14967 } else if (Cond.getOpcode() == ISD::SETCC &&
14968 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
14969 // For FCMP_OEQ, we can emit
14970 // two branches instead of an explicit AND instruction with a
14971 // separate test. However, we only do this if this block doesn't
14972 // have a fall-through edge, because this requires an explicit
14973 // jmp when the condition is false.
14974 if (Op.getNode()->hasOneUse()) {
14975 SDNode *User = *Op.getNode()->use_begin();
14976 // Look for an unconditional branch following this conditional branch.
14977 // We need this because we need to reverse the successors in order
14978 // to implement FCMP_OEQ.
14979 if (User->getOpcode() == ISD::BR) {
14980 SDValue FalseBB = User->getOperand(1);
14982 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14983 assert(NewBR == User);
14987 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14988 Cond.getOperand(0), Cond.getOperand(1));
14989 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14990 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14991 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14992 Chain, Dest, CC, Cmp);
14993 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
14998 } else if (Cond.getOpcode() == ISD::SETCC &&
14999 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15000 // For FCMP_UNE, we can emit
15001 // two branches instead of an explicit AND instruction with a
15002 // separate test. However, we only do this if this block doesn't
15003 // have a fall-through edge, because this requires an explicit
15004 // jmp when the condition is false.
15005 if (Op.getNode()->hasOneUse()) {
15006 SDNode *User = *Op.getNode()->use_begin();
15007 // Look for an unconditional branch following this conditional branch.
15008 // We need this because we need to reverse the successors in order
15009 // to implement FCMP_UNE.
15010 if (User->getOpcode() == ISD::BR) {
15011 SDValue FalseBB = User->getOperand(1);
15013 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15014 assert(NewBR == User);
15017 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15018 Cond.getOperand(0), Cond.getOperand(1));
15019 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15020 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15021 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15022 Chain, Dest, CC, Cmp);
15023 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15033 // Look pass the truncate if the high bits are known zero.
15034 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15035 Cond = Cond.getOperand(0);
15037 // We know the result of AND is compared against zero. Try to match
15039 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15040 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
15041 if (NewSetCC.getNode()) {
15042 CC = NewSetCC.getOperand(0);
15043 Cond = NewSetCC.getOperand(1);
15050 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15051 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15052 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15054 Cond = ConvertCmpIfNecessary(Cond, DAG);
15055 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15056 Chain, Dest, CC, Cond);
15059 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15060 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15061 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15062 // that the guard pages used by the OS virtual memory manager are allocated in
15063 // correct sequence.
15065 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15066 SelectionDAG &DAG) const {
15067 MachineFunction &MF = DAG.getMachineFunction();
15068 bool SplitStack = MF.shouldSplitStack();
15069 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15074 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15075 SDNode* Node = Op.getNode();
15077 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15078 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15079 " not tell us which reg is the stack pointer!");
15080 EVT VT = Node->getValueType(0);
15081 SDValue Tmp1 = SDValue(Node, 0);
15082 SDValue Tmp2 = SDValue(Node, 1);
15083 SDValue Tmp3 = Node->getOperand(2);
15084 SDValue Chain = Tmp1.getOperand(0);
15086 // Chain the dynamic stack allocation so that it doesn't modify the stack
15087 // pointer when other instructions are using the stack.
15088 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
15091 SDValue Size = Tmp2.getOperand(1);
15092 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15093 Chain = SP.getValue(1);
15094 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15095 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15096 unsigned StackAlign = TFI.getStackAlignment();
15097 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15098 if (Align > StackAlign)
15099 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
15100 DAG.getConstant(-(uint64_t)Align, dl, VT));
15101 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
15103 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15104 DAG.getIntPtrConstant(0, dl, true), SDValue(),
15107 SDValue Ops[2] = { Tmp1, Tmp2 };
15108 return DAG.getMergeValues(Ops, dl);
15112 SDValue Chain = Op.getOperand(0);
15113 SDValue Size = Op.getOperand(1);
15114 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15115 EVT VT = Op.getNode()->getValueType(0);
15117 bool Is64Bit = Subtarget->is64Bit();
15118 MVT SPTy = getPointerTy(DAG.getDataLayout());
15121 MachineRegisterInfo &MRI = MF.getRegInfo();
15124 // The 64 bit implementation of segmented stacks needs to clobber both r10
15125 // r11. This makes it impossible to use it along with nested parameters.
15126 const Function *F = MF.getFunction();
15128 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15130 if (I->hasNestAttr())
15131 report_fatal_error("Cannot use segmented stacks with functions that "
15132 "have nested arguments.");
15135 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15136 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15137 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15138 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15139 DAG.getRegister(Vreg, SPTy));
15140 SDValue Ops1[2] = { Value, Chain };
15141 return DAG.getMergeValues(Ops1, dl);
15144 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15146 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15147 Flag = Chain.getValue(1);
15148 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15150 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15152 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15153 unsigned SPReg = RegInfo->getStackRegister();
15154 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15155 Chain = SP.getValue(1);
15158 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15159 DAG.getConstant(-(uint64_t)Align, dl, VT));
15160 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15163 SDValue Ops1[2] = { SP, Chain };
15164 return DAG.getMergeValues(Ops1, dl);
15168 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15169 MachineFunction &MF = DAG.getMachineFunction();
15170 auto PtrVT = getPointerTy(MF.getDataLayout());
15171 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15173 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15176 if (!Subtarget->is64Bit() ||
15177 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15178 // vastart just stores the address of the VarArgsFrameIndex slot into the
15179 // memory location argument.
15180 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15181 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15182 MachinePointerInfo(SV), false, false, 0);
15186 // gp_offset (0 - 6 * 8)
15187 // fp_offset (48 - 48 + 8 * 16)
15188 // overflow_arg_area (point to parameters coming in memory).
15190 SmallVector<SDValue, 8> MemOps;
15191 SDValue FIN = Op.getOperand(1);
15193 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15194 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15196 FIN, MachinePointerInfo(SV), false, false, 0);
15197 MemOps.push_back(Store);
15200 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15201 Store = DAG.getStore(Op.getOperand(0), DL,
15202 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15204 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15205 MemOps.push_back(Store);
15207 // Store ptr to overflow_arg_area
15208 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15209 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15210 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15211 MachinePointerInfo(SV, 8),
15213 MemOps.push_back(Store);
15215 // Store ptr to reg_save_area.
15216 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(8, DL));
15217 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15218 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
15219 MachinePointerInfo(SV, 16), false, false, 0);
15220 MemOps.push_back(Store);
15221 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15224 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15225 assert(Subtarget->is64Bit() &&
15226 "LowerVAARG only handles 64-bit va_arg!");
15227 assert(Op.getNode()->getNumOperands() == 4);
15229 MachineFunction &MF = DAG.getMachineFunction();
15230 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
15231 // The Win64 ABI uses char* instead of a structure.
15232 return DAG.expandVAArg(Op.getNode());
15234 SDValue Chain = Op.getOperand(0);
15235 SDValue SrcPtr = Op.getOperand(1);
15236 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15237 unsigned Align = Op.getConstantOperandVal(3);
15240 EVT ArgVT = Op.getNode()->getValueType(0);
15241 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15242 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15245 // Decide which area this value should be read from.
15246 // TODO: Implement the AMD64 ABI in its entirety. This simple
15247 // selection mechanism works only for the basic types.
15248 if (ArgVT == MVT::f80) {
15249 llvm_unreachable("va_arg for f80 not yet implemented");
15250 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15251 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15252 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15253 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15255 llvm_unreachable("Unhandled argument type in LowerVAARG");
15258 if (ArgMode == 2) {
15259 // Sanity Check: Make sure using fp_offset makes sense.
15260 assert(!Subtarget->useSoftFloat() &&
15261 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
15262 Subtarget->hasSSE1());
15265 // Insert VAARG_64 node into the DAG
15266 // VAARG_64 returns two values: Variable Argument Address, Chain
15267 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15268 DAG.getConstant(ArgMode, dl, MVT::i8),
15269 DAG.getConstant(Align, dl, MVT::i32)};
15270 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15271 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15272 VTs, InstOps, MVT::i64,
15273 MachinePointerInfo(SV),
15275 /*Volatile=*/false,
15277 /*WriteMem=*/true);
15278 Chain = VAARG.getValue(1);
15280 // Load the next argument and return it
15281 return DAG.getLoad(ArgVT, dl,
15284 MachinePointerInfo(),
15285 false, false, false, 0);
15288 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15289 SelectionDAG &DAG) {
15290 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
15291 // where a va_list is still an i8*.
15292 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15293 if (Subtarget->isCallingConvWin64(
15294 DAG.getMachineFunction().getFunction()->getCallingConv()))
15295 // Probably a Win64 va_copy.
15296 return DAG.expandVACopy(Op.getNode());
15298 SDValue Chain = Op.getOperand(0);
15299 SDValue DstPtr = Op.getOperand(1);
15300 SDValue SrcPtr = Op.getOperand(2);
15301 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15302 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15305 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15306 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15308 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15311 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15312 // amount is a constant. Takes immediate version of shift as input.
15313 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15314 SDValue SrcOp, uint64_t ShiftAmt,
15315 SelectionDAG &DAG) {
15316 MVT ElementType = VT.getVectorElementType();
15318 // Fold this packed shift into its first operand if ShiftAmt is 0.
15322 // Check for ShiftAmt >= element width
15323 if (ShiftAmt >= ElementType.getSizeInBits()) {
15324 if (Opc == X86ISD::VSRAI)
15325 ShiftAmt = ElementType.getSizeInBits() - 1;
15327 return DAG.getConstant(0, dl, VT);
15330 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15331 && "Unknown target vector shift-by-constant node");
15333 // Fold this packed vector shift into a build vector if SrcOp is a
15334 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15335 if (VT == SrcOp.getSimpleValueType() &&
15336 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15337 SmallVector<SDValue, 8> Elts;
15338 unsigned NumElts = SrcOp->getNumOperands();
15339 ConstantSDNode *ND;
15342 default: llvm_unreachable(nullptr);
15343 case X86ISD::VSHLI:
15344 for (unsigned i=0; i!=NumElts; ++i) {
15345 SDValue CurrentOp = SrcOp->getOperand(i);
15346 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15347 Elts.push_back(CurrentOp);
15350 ND = cast<ConstantSDNode>(CurrentOp);
15351 const APInt &C = ND->getAPIntValue();
15352 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15355 case X86ISD::VSRLI:
15356 for (unsigned i=0; i!=NumElts; ++i) {
15357 SDValue CurrentOp = SrcOp->getOperand(i);
15358 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15359 Elts.push_back(CurrentOp);
15362 ND = cast<ConstantSDNode>(CurrentOp);
15363 const APInt &C = ND->getAPIntValue();
15364 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15367 case X86ISD::VSRAI:
15368 for (unsigned i=0; i!=NumElts; ++i) {
15369 SDValue CurrentOp = SrcOp->getOperand(i);
15370 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15371 Elts.push_back(CurrentOp);
15374 ND = cast<ConstantSDNode>(CurrentOp);
15375 const APInt &C = ND->getAPIntValue();
15376 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15381 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15384 return DAG.getNode(Opc, dl, VT, SrcOp,
15385 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15388 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15389 // may or may not be a constant. Takes immediate version of shift as input.
15390 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15391 SDValue SrcOp, SDValue ShAmt,
15392 SelectionDAG &DAG) {
15393 MVT SVT = ShAmt.getSimpleValueType();
15394 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15396 // Catch shift-by-constant.
15397 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15398 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15399 CShAmt->getZExtValue(), DAG);
15401 // Change opcode to non-immediate version
15403 default: llvm_unreachable("Unknown target vector shift node");
15404 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15405 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15406 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15409 const X86Subtarget &Subtarget =
15410 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15411 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15412 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15413 // Let the shuffle legalizer expand this shift amount node.
15414 SDValue Op0 = ShAmt.getOperand(0);
15415 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15416 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15418 // Need to build a vector containing shift amount.
15419 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15420 SmallVector<SDValue, 4> ShOps;
15421 ShOps.push_back(ShAmt);
15422 if (SVT == MVT::i32) {
15423 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15424 ShOps.push_back(DAG.getUNDEF(SVT));
15426 ShOps.push_back(DAG.getUNDEF(SVT));
15428 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15429 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15432 // The return type has to be a 128-bit type with the same element
15433 // type as the input type.
15434 MVT EltVT = VT.getVectorElementType();
15435 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15437 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15438 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15441 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15442 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15443 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15444 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15445 SDValue PreservedSrc,
15446 const X86Subtarget *Subtarget,
15447 SelectionDAG &DAG) {
15448 EVT VT = Op.getValueType();
15449 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15450 MVT::i1, VT.getVectorNumElements());
15451 SDValue VMask = SDValue();
15452 unsigned OpcodeSelect = ISD::VSELECT;
15455 assert(MaskVT.isSimple() && "invalid mask type");
15457 if (isAllOnes(Mask))
15460 if (MaskVT.bitsGT(Mask.getValueType())) {
15461 EVT newMaskVT = EVT::getIntegerVT(*DAG.getContext(),
15462 MaskVT.getSizeInBits());
15463 VMask = DAG.getBitcast(MaskVT,
15464 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15466 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15467 Mask.getValueType().getSizeInBits());
15468 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15469 // are extracted by EXTRACT_SUBVECTOR.
15470 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15471 DAG.getBitcast(BitcastVT, Mask),
15472 DAG.getIntPtrConstant(0, dl));
15475 switch (Op.getOpcode()) {
15477 case X86ISD::PCMPEQM:
15478 case X86ISD::PCMPGTM:
15480 case X86ISD::CMPMU:
15481 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15482 case X86ISD::VTRUNC:
15483 case X86ISD::VTRUNCS:
15484 case X86ISD::VTRUNCUS:
15485 // We can't use ISD::VSELECT here because it is not always "Legal"
15486 // for the destination type. For example vpmovqb require only AVX512
15487 // and vselect that can operate on byte element type require BWI
15488 OpcodeSelect = X86ISD::SELECT;
15491 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15492 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15493 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
15496 /// \brief Creates an SDNode for a predicated scalar operation.
15497 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
15498 /// The mask is coming as MVT::i8 and it should be truncated
15499 /// to MVT::i1 while lowering masking intrinsics.
15500 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
15501 /// "X86select" instead of "vselect". We just can't create the "vselect" node
15502 /// for a scalar instruction.
15503 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
15504 SDValue PreservedSrc,
15505 const X86Subtarget *Subtarget,
15506 SelectionDAG &DAG) {
15507 if (isAllOnes(Mask))
15510 EVT VT = Op.getValueType();
15512 // The mask should be of type MVT::i1
15513 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
15515 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15516 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15517 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
15520 static int getSEHRegistrationNodeSize(const Function *Fn) {
15521 if (!Fn->hasPersonalityFn())
15522 report_fatal_error(
15523 "querying registration node size for function without personality");
15524 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
15525 // WinEHStatePass for the full struct definition.
15526 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
15527 case EHPersonality::MSVC_X86SEH: return 24;
15528 case EHPersonality::MSVC_CXX: return 16;
15531 report_fatal_error("can only recover FP for MSVC EH personality functions");
15534 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
15535 /// function or when returning to a parent frame after catching an exception, we
15536 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
15537 /// Here's the math:
15538 /// RegNodeBase = EntryEBP - RegNodeSize
15539 /// ParentFP = RegNodeBase - RegNodeFrameOffset
15540 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
15541 /// subtracting the offset (negative on x86) takes us back to the parent FP.
15542 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
15543 SDValue EntryEBP) {
15544 MachineFunction &MF = DAG.getMachineFunction();
15547 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15548 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
15550 // It's possible that the parent function no longer has a personality function
15551 // if the exceptional code was optimized away, in which case we just return
15552 // the incoming EBP.
15553 if (!Fn->hasPersonalityFn())
15556 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
15558 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
15560 MCSymbol *OffsetSym =
15561 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
15562 GlobalValue::getRealLinkageName(Fn->getName()));
15563 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
15564 SDValue RegNodeFrameOffset =
15565 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
15567 // RegNodeBase = EntryEBP - RegNodeSize
15568 // ParentFP = RegNodeBase - RegNodeFrameOffset
15569 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
15570 DAG.getConstant(RegNodeSize, dl, PtrVT));
15571 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
15574 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15575 SelectionDAG &DAG) {
15577 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15578 EVT VT = Op.getValueType();
15579 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
15581 switch(IntrData->Type) {
15582 case INTR_TYPE_1OP:
15583 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
15584 case INTR_TYPE_2OP:
15585 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15587 case INTR_TYPE_2OP_IMM8:
15588 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15589 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
15590 case INTR_TYPE_3OP:
15591 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15592 Op.getOperand(2), Op.getOperand(3));
15593 case INTR_TYPE_4OP:
15594 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15595 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
15596 case INTR_TYPE_1OP_MASK_RM: {
15597 SDValue Src = Op.getOperand(1);
15598 SDValue PassThru = Op.getOperand(2);
15599 SDValue Mask = Op.getOperand(3);
15600 SDValue RoundingMode;
15601 // We allways add rounding mode to the Node.
15602 // If the rounding mode is not specified, we add the
15603 // "current direction" mode.
15604 if (Op.getNumOperands() == 4)
15606 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15608 RoundingMode = Op.getOperand(4);
15609 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15610 if (IntrWithRoundingModeOpcode != 0)
15611 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
15612 X86::STATIC_ROUNDING::CUR_DIRECTION)
15613 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15614 dl, Op.getValueType(), Src, RoundingMode),
15615 Mask, PassThru, Subtarget, DAG);
15616 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
15618 Mask, PassThru, Subtarget, DAG);
15620 case INTR_TYPE_1OP_MASK: {
15621 SDValue Src = Op.getOperand(1);
15622 SDValue PassThru = Op.getOperand(2);
15623 SDValue Mask = Op.getOperand(3);
15624 // We add rounding mode to the Node when
15625 // - RM Opcode is specified and
15626 // - RM is not "current direction".
15627 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15628 if (IntrWithRoundingModeOpcode != 0) {
15629 SDValue Rnd = Op.getOperand(4);
15630 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15631 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15632 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15633 dl, Op.getValueType(),
15635 Mask, PassThru, Subtarget, DAG);
15638 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
15639 Mask, PassThru, Subtarget, DAG);
15641 case INTR_TYPE_SCALAR_MASK_RM: {
15642 SDValue Src1 = Op.getOperand(1);
15643 SDValue Src2 = Op.getOperand(2);
15644 SDValue Src0 = Op.getOperand(3);
15645 SDValue Mask = Op.getOperand(4);
15646 // There are 2 kinds of intrinsics in this group:
15647 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
15648 // (2) With rounding mode and sae - 7 operands.
15649 if (Op.getNumOperands() == 6) {
15650 SDValue Sae = Op.getOperand(5);
15651 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
15652 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
15654 Mask, Src0, Subtarget, DAG);
15656 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
15657 SDValue RoundingMode = Op.getOperand(5);
15658 SDValue Sae = Op.getOperand(6);
15659 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
15660 RoundingMode, Sae),
15661 Mask, Src0, Subtarget, DAG);
15663 case INTR_TYPE_2OP_MASK: {
15664 SDValue Src1 = Op.getOperand(1);
15665 SDValue Src2 = Op.getOperand(2);
15666 SDValue PassThru = Op.getOperand(3);
15667 SDValue Mask = Op.getOperand(4);
15668 // We specify 2 possible opcodes for intrinsics with rounding modes.
15669 // First, we check if the intrinsic may have non-default rounding mode,
15670 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15671 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15672 if (IntrWithRoundingModeOpcode != 0) {
15673 SDValue Rnd = Op.getOperand(5);
15674 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15675 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15676 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15677 dl, Op.getValueType(),
15679 Mask, PassThru, Subtarget, DAG);
15682 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15684 Mask, PassThru, Subtarget, DAG);
15686 case INTR_TYPE_2OP_MASK_RM: {
15687 SDValue Src1 = Op.getOperand(1);
15688 SDValue Src2 = Op.getOperand(2);
15689 SDValue PassThru = Op.getOperand(3);
15690 SDValue Mask = Op.getOperand(4);
15691 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15692 // First, we check if the intrinsic have rounding mode (6 operands),
15693 // if not, we set rounding mode to "current".
15695 if (Op.getNumOperands() == 6)
15696 Rnd = Op.getOperand(5);
15698 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15699 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15701 Mask, PassThru, Subtarget, DAG);
15703 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
15704 SDValue Src1 = Op.getOperand(1);
15705 SDValue Src2 = Op.getOperand(2);
15706 SDValue Src3 = Op.getOperand(3);
15707 SDValue PassThru = Op.getOperand(4);
15708 SDValue Mask = Op.getOperand(5);
15709 SDValue Sae = Op.getOperand(6);
15711 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
15713 Mask, PassThru, Subtarget, DAG);
15715 case INTR_TYPE_3OP_MASK_RM: {
15716 SDValue Src1 = Op.getOperand(1);
15717 SDValue Src2 = Op.getOperand(2);
15718 SDValue Imm = Op.getOperand(3);
15719 SDValue PassThru = Op.getOperand(4);
15720 SDValue Mask = Op.getOperand(5);
15721 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15722 // First, we check if the intrinsic have rounding mode (7 operands),
15723 // if not, we set rounding mode to "current".
15725 if (Op.getNumOperands() == 7)
15726 Rnd = Op.getOperand(6);
15728 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15729 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15730 Src1, Src2, Imm, Rnd),
15731 Mask, PassThru, Subtarget, DAG);
15733 case INTR_TYPE_3OP_IMM8_MASK:
15734 case INTR_TYPE_3OP_MASK: {
15735 SDValue Src1 = Op.getOperand(1);
15736 SDValue Src2 = Op.getOperand(2);
15737 SDValue Src3 = Op.getOperand(3);
15738 SDValue PassThru = Op.getOperand(4);
15739 SDValue Mask = Op.getOperand(5);
15741 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
15742 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
15743 // We specify 2 possible opcodes for intrinsics with rounding modes.
15744 // First, we check if the intrinsic may have non-default rounding mode,
15745 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15746 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15747 if (IntrWithRoundingModeOpcode != 0) {
15748 SDValue Rnd = Op.getOperand(6);
15749 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15750 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15751 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15752 dl, Op.getValueType(),
15753 Src1, Src2, Src3, Rnd),
15754 Mask, PassThru, Subtarget, DAG);
15757 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15759 Mask, PassThru, Subtarget, DAG);
15761 case VPERM_3OP_MASKZ:
15762 case VPERM_3OP_MASK:
15765 case FMA_OP_MASK: {
15766 SDValue Src1 = Op.getOperand(1);
15767 SDValue Src2 = Op.getOperand(2);
15768 SDValue Src3 = Op.getOperand(3);
15769 SDValue Mask = Op.getOperand(4);
15770 EVT VT = Op.getValueType();
15771 SDValue PassThru = SDValue();
15773 // set PassThru element
15774 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
15775 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
15776 else if (IntrData->Type == FMA_OP_MASK3)
15781 // We specify 2 possible opcodes for intrinsics with rounding modes.
15782 // First, we check if the intrinsic may have non-default rounding mode,
15783 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15784 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15785 if (IntrWithRoundingModeOpcode != 0) {
15786 SDValue Rnd = Op.getOperand(5);
15787 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15788 X86::STATIC_ROUNDING::CUR_DIRECTION)
15789 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15790 dl, Op.getValueType(),
15791 Src1, Src2, Src3, Rnd),
15792 Mask, PassThru, Subtarget, DAG);
15794 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
15795 dl, Op.getValueType(),
15797 Mask, PassThru, Subtarget, DAG);
15800 case CMP_MASK_CC: {
15801 // Comparison intrinsics with masks.
15802 // Example of transformation:
15803 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
15804 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
15806 // (v8i1 (insert_subvector undef,
15807 // (v2i1 (and (PCMPEQM %a, %b),
15808 // (extract_subvector
15809 // (v8i1 (bitcast %mask)), 0))), 0))))
15810 EVT VT = Op.getOperand(1).getValueType();
15811 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15812 VT.getVectorNumElements());
15813 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
15814 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15815 Mask.getValueType().getSizeInBits());
15817 if (IntrData->Type == CMP_MASK_CC) {
15818 SDValue CC = Op.getOperand(3);
15819 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
15820 // We specify 2 possible opcodes for intrinsics with rounding modes.
15821 // First, we check if the intrinsic may have non-default rounding mode,
15822 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15823 if (IntrData->Opc1 != 0) {
15824 SDValue Rnd = Op.getOperand(5);
15825 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15826 X86::STATIC_ROUNDING::CUR_DIRECTION)
15827 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
15828 Op.getOperand(2), CC, Rnd);
15830 //default rounding mode
15832 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15833 Op.getOperand(2), CC);
15836 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
15837 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15840 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
15841 DAG.getTargetConstant(0, dl,
15844 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
15845 DAG.getUNDEF(BitcastVT), CmpMask,
15846 DAG.getIntPtrConstant(0, dl));
15847 return DAG.getBitcast(Op.getValueType(), Res);
15849 case COMI: { // Comparison intrinsics
15850 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
15851 SDValue LHS = Op.getOperand(1);
15852 SDValue RHS = Op.getOperand(2);
15853 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
15854 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
15855 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
15856 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15857 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
15858 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15861 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
15862 Op.getOperand(1), Op.getOperand(2), DAG);
15864 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
15865 Op.getSimpleValueType(),
15867 Op.getOperand(2), DAG),
15868 Op.getOperand(4), Op.getOperand(3), Subtarget,
15870 case COMPRESS_EXPAND_IN_REG: {
15871 SDValue Mask = Op.getOperand(3);
15872 SDValue DataToCompress = Op.getOperand(1);
15873 SDValue PassThru = Op.getOperand(2);
15874 if (isAllOnes(Mask)) // return data as is
15875 return Op.getOperand(1);
15877 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15879 Mask, PassThru, Subtarget, DAG);
15882 SDValue Mask = Op.getOperand(3);
15883 EVT VT = Op.getValueType();
15884 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15885 VT.getVectorNumElements());
15886 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15887 Mask.getValueType().getSizeInBits());
15889 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15890 DAG.getBitcast(BitcastVT, Mask),
15891 DAG.getIntPtrConstant(0, dl));
15892 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
15901 default: return SDValue(); // Don't custom lower most intrinsics.
15903 case Intrinsic::x86_avx2_permd:
15904 case Intrinsic::x86_avx2_permps:
15905 // Operands intentionally swapped. Mask is last operand to intrinsic,
15906 // but second operand for node/instruction.
15907 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
15908 Op.getOperand(2), Op.getOperand(1));
15910 // ptest and testp intrinsics. The intrinsic these come from are designed to
15911 // return an integer value, not just an instruction so lower it to the ptest
15912 // or testp pattern and a setcc for the result.
15913 case Intrinsic::x86_sse41_ptestz:
15914 case Intrinsic::x86_sse41_ptestc:
15915 case Intrinsic::x86_sse41_ptestnzc:
15916 case Intrinsic::x86_avx_ptestz_256:
15917 case Intrinsic::x86_avx_ptestc_256:
15918 case Intrinsic::x86_avx_ptestnzc_256:
15919 case Intrinsic::x86_avx_vtestz_ps:
15920 case Intrinsic::x86_avx_vtestc_ps:
15921 case Intrinsic::x86_avx_vtestnzc_ps:
15922 case Intrinsic::x86_avx_vtestz_pd:
15923 case Intrinsic::x86_avx_vtestc_pd:
15924 case Intrinsic::x86_avx_vtestnzc_pd:
15925 case Intrinsic::x86_avx_vtestz_ps_256:
15926 case Intrinsic::x86_avx_vtestc_ps_256:
15927 case Intrinsic::x86_avx_vtestnzc_ps_256:
15928 case Intrinsic::x86_avx_vtestz_pd_256:
15929 case Intrinsic::x86_avx_vtestc_pd_256:
15930 case Intrinsic::x86_avx_vtestnzc_pd_256: {
15931 bool IsTestPacked = false;
15934 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
15935 case Intrinsic::x86_avx_vtestz_ps:
15936 case Intrinsic::x86_avx_vtestz_pd:
15937 case Intrinsic::x86_avx_vtestz_ps_256:
15938 case Intrinsic::x86_avx_vtestz_pd_256:
15939 IsTestPacked = true; // Fallthrough
15940 case Intrinsic::x86_sse41_ptestz:
15941 case Intrinsic::x86_avx_ptestz_256:
15943 X86CC = X86::COND_E;
15945 case Intrinsic::x86_avx_vtestc_ps:
15946 case Intrinsic::x86_avx_vtestc_pd:
15947 case Intrinsic::x86_avx_vtestc_ps_256:
15948 case Intrinsic::x86_avx_vtestc_pd_256:
15949 IsTestPacked = true; // Fallthrough
15950 case Intrinsic::x86_sse41_ptestc:
15951 case Intrinsic::x86_avx_ptestc_256:
15953 X86CC = X86::COND_B;
15955 case Intrinsic::x86_avx_vtestnzc_ps:
15956 case Intrinsic::x86_avx_vtestnzc_pd:
15957 case Intrinsic::x86_avx_vtestnzc_ps_256:
15958 case Intrinsic::x86_avx_vtestnzc_pd_256:
15959 IsTestPacked = true; // Fallthrough
15960 case Intrinsic::x86_sse41_ptestnzc:
15961 case Intrinsic::x86_avx_ptestnzc_256:
15963 X86CC = X86::COND_A;
15967 SDValue LHS = Op.getOperand(1);
15968 SDValue RHS = Op.getOperand(2);
15969 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
15970 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
15971 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15972 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
15973 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15975 case Intrinsic::x86_avx512_kortestz_w:
15976 case Intrinsic::x86_avx512_kortestc_w: {
15977 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
15978 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
15979 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
15980 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15981 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
15982 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
15983 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15986 case Intrinsic::x86_sse42_pcmpistria128:
15987 case Intrinsic::x86_sse42_pcmpestria128:
15988 case Intrinsic::x86_sse42_pcmpistric128:
15989 case Intrinsic::x86_sse42_pcmpestric128:
15990 case Intrinsic::x86_sse42_pcmpistrio128:
15991 case Intrinsic::x86_sse42_pcmpestrio128:
15992 case Intrinsic::x86_sse42_pcmpistris128:
15993 case Intrinsic::x86_sse42_pcmpestris128:
15994 case Intrinsic::x86_sse42_pcmpistriz128:
15995 case Intrinsic::x86_sse42_pcmpestriz128: {
15999 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16000 case Intrinsic::x86_sse42_pcmpistria128:
16001 Opcode = X86ISD::PCMPISTRI;
16002 X86CC = X86::COND_A;
16004 case Intrinsic::x86_sse42_pcmpestria128:
16005 Opcode = X86ISD::PCMPESTRI;
16006 X86CC = X86::COND_A;
16008 case Intrinsic::x86_sse42_pcmpistric128:
16009 Opcode = X86ISD::PCMPISTRI;
16010 X86CC = X86::COND_B;
16012 case Intrinsic::x86_sse42_pcmpestric128:
16013 Opcode = X86ISD::PCMPESTRI;
16014 X86CC = X86::COND_B;
16016 case Intrinsic::x86_sse42_pcmpistrio128:
16017 Opcode = X86ISD::PCMPISTRI;
16018 X86CC = X86::COND_O;
16020 case Intrinsic::x86_sse42_pcmpestrio128:
16021 Opcode = X86ISD::PCMPESTRI;
16022 X86CC = X86::COND_O;
16024 case Intrinsic::x86_sse42_pcmpistris128:
16025 Opcode = X86ISD::PCMPISTRI;
16026 X86CC = X86::COND_S;
16028 case Intrinsic::x86_sse42_pcmpestris128:
16029 Opcode = X86ISD::PCMPESTRI;
16030 X86CC = X86::COND_S;
16032 case Intrinsic::x86_sse42_pcmpistriz128:
16033 Opcode = X86ISD::PCMPISTRI;
16034 X86CC = X86::COND_E;
16036 case Intrinsic::x86_sse42_pcmpestriz128:
16037 Opcode = X86ISD::PCMPESTRI;
16038 X86CC = X86::COND_E;
16041 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16042 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16043 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16044 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16045 DAG.getConstant(X86CC, dl, MVT::i8),
16046 SDValue(PCMP.getNode(), 1));
16047 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16050 case Intrinsic::x86_sse42_pcmpistri128:
16051 case Intrinsic::x86_sse42_pcmpestri128: {
16053 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16054 Opcode = X86ISD::PCMPISTRI;
16056 Opcode = X86ISD::PCMPESTRI;
16058 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16059 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16060 return DAG.getNode(Opcode, dl, VTs, NewOps);
16063 case Intrinsic::x86_seh_lsda: {
16064 // Compute the symbol for the LSDA. We know it'll get emitted later.
16065 MachineFunction &MF = DAG.getMachineFunction();
16066 SDValue Op1 = Op.getOperand(1);
16067 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
16068 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
16069 GlobalValue::getRealLinkageName(Fn->getName()));
16071 // Generate a simple absolute symbol reference. This intrinsic is only
16072 // supported on 32-bit Windows, which isn't PIC.
16073 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
16074 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
16077 case Intrinsic::x86_seh_recoverfp: {
16078 SDValue FnOp = Op.getOperand(1);
16079 SDValue IncomingFPOp = Op.getOperand(2);
16080 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
16081 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
16083 report_fatal_error(
16084 "llvm.x86.seh.recoverfp must take a function as the first argument");
16085 return recoverFramePointer(DAG, Fn, IncomingFPOp);
16088 case Intrinsic::localaddress: {
16089 // Returns one of the stack, base, or frame pointer registers, depending on
16090 // which is used to reference local variables.
16091 MachineFunction &MF = DAG.getMachineFunction();
16092 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16094 if (RegInfo->hasBasePointer(MF))
16095 Reg = RegInfo->getBaseRegister();
16096 else // This function handles the SP or FP case.
16097 Reg = RegInfo->getPtrSizedFrameRegister(MF);
16098 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
16103 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16104 SDValue Src, SDValue Mask, SDValue Base,
16105 SDValue Index, SDValue ScaleOp, SDValue Chain,
16106 const X86Subtarget * Subtarget) {
16108 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16110 llvm_unreachable("Invalid scale type");
16111 unsigned ScaleVal = C->getZExtValue();
16112 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16113 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16115 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16116 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16117 Index.getSimpleValueType().getVectorNumElements());
16119 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16121 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16123 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16124 Mask.getValueType().getSizeInBits());
16126 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16127 // are extracted by EXTRACT_SUBVECTOR.
16128 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16129 DAG.getBitcast(BitcastVT, Mask),
16130 DAG.getIntPtrConstant(0, dl));
16132 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16133 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16134 SDValue Segment = DAG.getRegister(0, MVT::i32);
16135 if (Src.getOpcode() == ISD::UNDEF)
16136 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
16137 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16138 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16139 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16140 return DAG.getMergeValues(RetOps, dl);
16143 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16144 SDValue Src, SDValue Mask, SDValue Base,
16145 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16147 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16149 llvm_unreachable("Invalid scale type");
16150 unsigned ScaleVal = C->getZExtValue();
16151 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16152 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16154 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16155 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16156 SDValue Segment = DAG.getRegister(0, MVT::i32);
16157 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16158 Index.getSimpleValueType().getVectorNumElements());
16160 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16162 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16164 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16165 Mask.getValueType().getSizeInBits());
16167 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16168 // are extracted by EXTRACT_SUBVECTOR.
16169 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16170 DAG.getBitcast(BitcastVT, Mask),
16171 DAG.getIntPtrConstant(0, dl));
16173 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16174 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16175 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16176 return SDValue(Res, 1);
16179 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16180 SDValue Mask, SDValue Base, SDValue Index,
16181 SDValue ScaleOp, SDValue Chain) {
16183 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16184 assert(C && "Invalid scale type");
16185 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16186 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16187 SDValue Segment = DAG.getRegister(0, MVT::i32);
16189 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16191 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16193 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16195 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16196 //SDVTList VTs = DAG.getVTList(MVT::Other);
16197 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16198 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16199 return SDValue(Res, 0);
16202 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16203 // read performance monitor counters (x86_rdpmc).
16204 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16205 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16206 SmallVectorImpl<SDValue> &Results) {
16207 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16208 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16211 // The ECX register is used to select the index of the performance counter
16213 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16215 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16217 // Reads the content of a 64-bit performance counter and returns it in the
16218 // registers EDX:EAX.
16219 if (Subtarget->is64Bit()) {
16220 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16221 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16224 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16225 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16228 Chain = HI.getValue(1);
16230 if (Subtarget->is64Bit()) {
16231 // The EAX register is loaded with the low-order 32 bits. The EDX register
16232 // is loaded with the supported high-order bits of the counter.
16233 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16234 DAG.getConstant(32, DL, MVT::i8));
16235 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16236 Results.push_back(Chain);
16240 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16241 SDValue Ops[] = { LO, HI };
16242 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16243 Results.push_back(Pair);
16244 Results.push_back(Chain);
16247 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16248 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16249 // also used to custom lower READCYCLECOUNTER nodes.
16250 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16251 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16252 SmallVectorImpl<SDValue> &Results) {
16253 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16254 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16257 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16258 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16259 // and the EAX register is loaded with the low-order 32 bits.
16260 if (Subtarget->is64Bit()) {
16261 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16262 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16265 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16266 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16269 SDValue Chain = HI.getValue(1);
16271 if (Opcode == X86ISD::RDTSCP_DAG) {
16272 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16274 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16275 // the ECX register. Add 'ecx' explicitly to the chain.
16276 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16278 // Explicitly store the content of ECX at the location passed in input
16279 // to the 'rdtscp' intrinsic.
16280 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16281 MachinePointerInfo(), false, false, 0);
16284 if (Subtarget->is64Bit()) {
16285 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16286 // the EAX register is loaded with the low-order 32 bits.
16287 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16288 DAG.getConstant(32, DL, MVT::i8));
16289 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16290 Results.push_back(Chain);
16294 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16295 SDValue Ops[] = { LO, HI };
16296 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16297 Results.push_back(Pair);
16298 Results.push_back(Chain);
16301 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16302 SelectionDAG &DAG) {
16303 SmallVector<SDValue, 2> Results;
16305 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16307 return DAG.getMergeValues(Results, DL);
16310 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16311 SelectionDAG &DAG) {
16312 MachineFunction &MF = DAG.getMachineFunction();
16313 const Function *Fn = MF.getFunction();
16315 SDValue Chain = Op.getOperand(0);
16317 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16318 "using llvm.x86.seh.restoreframe requires a frame pointer");
16320 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16321 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16323 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16324 unsigned FrameReg =
16325 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16326 unsigned SPReg = RegInfo->getStackRegister();
16327 unsigned SlotSize = RegInfo->getSlotSize();
16329 // Get incoming EBP.
16330 SDValue IncomingEBP =
16331 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16333 // SP is saved in the first field of every registration node, so load
16334 // [EBP-RegNodeSize] into SP.
16335 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16336 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16337 DAG.getConstant(-RegNodeSize, dl, VT));
16339 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16340 false, VT.getScalarSizeInBits() / 8);
16341 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16343 if (!RegInfo->needsStackRealignment(MF)) {
16344 // Adjust EBP to point back to the original frame position.
16345 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16346 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16348 assert(RegInfo->hasBasePointer(MF) &&
16349 "functions with Win32 EH must use frame or base pointer register");
16351 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16352 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16353 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16355 // Reload the spilled EBP value, now that the stack and base pointers are
16357 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16358 X86FI->setHasSEHFramePtrSave(true);
16359 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16360 X86FI->setSEHFramePtrSaveIndex(FI);
16361 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16362 MachinePointerInfo(), false, false, false,
16363 VT.getScalarSizeInBits() / 8);
16364 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16370 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16371 /// return truncate Store/MaskedStore Node
16372 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16376 SDValue Mask = Op.getOperand(4);
16377 SDValue DataToTruncate = Op.getOperand(3);
16378 SDValue Addr = Op.getOperand(2);
16379 SDValue Chain = Op.getOperand(0);
16381 EVT VT = DataToTruncate.getValueType();
16382 EVT SVT = EVT::getVectorVT(*DAG.getContext(),
16383 ElementType, VT.getVectorNumElements());
16385 if (isAllOnes(Mask)) // return just a truncate store
16386 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
16387 MachinePointerInfo(), SVT, false, false,
16388 SVT.getScalarSizeInBits()/8);
16390 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
16391 MVT::i1, VT.getVectorNumElements());
16392 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16393 Mask.getValueType().getSizeInBits());
16394 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16395 // are extracted by EXTRACT_SUBVECTOR.
16396 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16397 DAG.getBitcast(BitcastVT, Mask),
16398 DAG.getIntPtrConstant(0, dl));
16400 MachineMemOperand *MMO = DAG.getMachineFunction().
16401 getMachineMemOperand(MachinePointerInfo(),
16402 MachineMemOperand::MOStore, SVT.getStoreSize(),
16403 SVT.getScalarSizeInBits()/8);
16405 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
16406 VMask, SVT, MMO, true);
16409 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16410 SelectionDAG &DAG) {
16411 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
16413 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
16415 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
16416 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
16421 switch(IntrData->Type) {
16423 llvm_unreachable("Unknown Intrinsic Type");
16427 // Emit the node with the right value type.
16428 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
16429 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16431 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
16432 // Otherwise return the value from Rand, which is always 0, casted to i32.
16433 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
16434 DAG.getConstant(1, dl, Op->getValueType(1)),
16435 DAG.getConstant(X86::COND_B, dl, MVT::i32),
16436 SDValue(Result.getNode(), 1) };
16437 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
16438 DAG.getVTList(Op->getValueType(1), MVT::Glue),
16441 // Return { result, isValid, chain }.
16442 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
16443 SDValue(Result.getNode(), 2));
16446 //gather(v1, mask, index, base, scale);
16447 SDValue Chain = Op.getOperand(0);
16448 SDValue Src = Op.getOperand(2);
16449 SDValue Base = Op.getOperand(3);
16450 SDValue Index = Op.getOperand(4);
16451 SDValue Mask = Op.getOperand(5);
16452 SDValue Scale = Op.getOperand(6);
16453 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
16457 //scatter(base, mask, index, v1, scale);
16458 SDValue Chain = Op.getOperand(0);
16459 SDValue Base = Op.getOperand(2);
16460 SDValue Mask = Op.getOperand(3);
16461 SDValue Index = Op.getOperand(4);
16462 SDValue Src = Op.getOperand(5);
16463 SDValue Scale = Op.getOperand(6);
16464 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
16468 SDValue Hint = Op.getOperand(6);
16469 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
16470 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
16471 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
16472 SDValue Chain = Op.getOperand(0);
16473 SDValue Mask = Op.getOperand(2);
16474 SDValue Index = Op.getOperand(3);
16475 SDValue Base = Op.getOperand(4);
16476 SDValue Scale = Op.getOperand(5);
16477 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
16479 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
16481 SmallVector<SDValue, 2> Results;
16482 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
16484 return DAG.getMergeValues(Results, dl);
16486 // Read Performance Monitoring Counters.
16488 SmallVector<SDValue, 2> Results;
16489 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
16490 return DAG.getMergeValues(Results, dl);
16492 // XTEST intrinsics.
16494 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16495 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16496 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16497 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
16499 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
16500 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
16501 Ret, SDValue(InTrans.getNode(), 1));
16505 SmallVector<SDValue, 2> Results;
16506 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16507 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
16508 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
16509 DAG.getConstant(-1, dl, MVT::i8));
16510 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
16511 Op.getOperand(4), GenCF.getValue(1));
16512 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
16513 Op.getOperand(5), MachinePointerInfo(),
16515 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16516 DAG.getConstant(X86::COND_B, dl, MVT::i8),
16518 Results.push_back(SetCC);
16519 Results.push_back(Store);
16520 return DAG.getMergeValues(Results, dl);
16522 case COMPRESS_TO_MEM: {
16524 SDValue Mask = Op.getOperand(4);
16525 SDValue DataToCompress = Op.getOperand(3);
16526 SDValue Addr = Op.getOperand(2);
16527 SDValue Chain = Op.getOperand(0);
16529 EVT VT = DataToCompress.getValueType();
16530 if (isAllOnes(Mask)) // return just a store
16531 return DAG.getStore(Chain, dl, DataToCompress, Addr,
16532 MachinePointerInfo(), false, false,
16533 VT.getScalarSizeInBits()/8);
16535 SDValue Compressed =
16536 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
16537 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
16538 return DAG.getStore(Chain, dl, Compressed, Addr,
16539 MachinePointerInfo(), false, false,
16540 VT.getScalarSizeInBits()/8);
16542 case TRUNCATE_TO_MEM_VI8:
16543 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
16544 case TRUNCATE_TO_MEM_VI16:
16545 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
16546 case TRUNCATE_TO_MEM_VI32:
16547 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
16548 case EXPAND_FROM_MEM: {
16550 SDValue Mask = Op.getOperand(4);
16551 SDValue PassThru = Op.getOperand(3);
16552 SDValue Addr = Op.getOperand(2);
16553 SDValue Chain = Op.getOperand(0);
16554 EVT VT = Op.getValueType();
16556 if (isAllOnes(Mask)) // return just a load
16557 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
16558 false, VT.getScalarSizeInBits()/8);
16560 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
16561 false, false, false,
16562 VT.getScalarSizeInBits()/8);
16564 SDValue Results[] = {
16565 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
16566 Mask, PassThru, Subtarget, DAG), Chain};
16567 return DAG.getMergeValues(Results, dl);
16572 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
16573 SelectionDAG &DAG) const {
16574 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
16575 MFI->setReturnAddressIsTaken(true);
16577 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
16580 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16582 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16585 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
16586 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16587 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
16588 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16589 DAG.getNode(ISD::ADD, dl, PtrVT,
16590 FrameAddr, Offset),
16591 MachinePointerInfo(), false, false, false, 0);
16594 // Just load the return address.
16595 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
16596 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16597 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
16600 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
16601 MachineFunction &MF = DAG.getMachineFunction();
16602 MachineFrameInfo *MFI = MF.getFrameInfo();
16603 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
16604 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16605 EVT VT = Op.getValueType();
16607 MFI->setFrameAddressIsTaken(true);
16609 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
16610 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
16611 // is not possible to crawl up the stack without looking at the unwind codes
16613 int FrameAddrIndex = FuncInfo->getFAIndex();
16614 if (!FrameAddrIndex) {
16615 // Set up a frame object for the return address.
16616 unsigned SlotSize = RegInfo->getSlotSize();
16617 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
16618 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
16619 FuncInfo->setFAIndex(FrameAddrIndex);
16621 return DAG.getFrameIndex(FrameAddrIndex, VT);
16624 unsigned FrameReg =
16625 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16626 SDLoc dl(Op); // FIXME probably not meaningful
16627 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16628 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
16629 (FrameReg == X86::EBP && VT == MVT::i32)) &&
16630 "Invalid Frame Register!");
16631 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
16633 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
16634 MachinePointerInfo(),
16635 false, false, false, 0);
16639 // FIXME? Maybe this could be a TableGen attribute on some registers and
16640 // this table could be generated automatically from RegInfo.
16641 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
16642 SelectionDAG &DAG) const {
16643 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16644 const MachineFunction &MF = DAG.getMachineFunction();
16646 unsigned Reg = StringSwitch<unsigned>(RegName)
16647 .Case("esp", X86::ESP)
16648 .Case("rsp", X86::RSP)
16649 .Case("ebp", X86::EBP)
16650 .Case("rbp", X86::RBP)
16653 if (Reg == X86::EBP || Reg == X86::RBP) {
16654 if (!TFI.hasFP(MF))
16655 report_fatal_error("register " + StringRef(RegName) +
16656 " is allocatable: function has no frame pointer");
16659 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16660 unsigned FrameReg =
16661 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16662 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
16663 "Invalid Frame Register!");
16671 report_fatal_error("Invalid register name global variable");
16674 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
16675 SelectionDAG &DAG) const {
16676 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16677 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
16680 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
16681 SDValue Chain = Op.getOperand(0);
16682 SDValue Offset = Op.getOperand(1);
16683 SDValue Handler = Op.getOperand(2);
16686 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16687 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16688 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
16689 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
16690 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
16691 "Invalid Frame Register!");
16692 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
16693 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
16695 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
16696 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
16698 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
16699 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
16701 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
16703 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
16704 DAG.getRegister(StoreAddrReg, PtrVT));
16707 SDValue X86TargetLowering::LowerCATCHRET(SDValue Op, SelectionDAG &DAG) const {
16708 SDValue Chain = Op.getOperand(0);
16709 SDValue Dest = Op.getOperand(1);
16712 MVT PtrVT = getPointerTy(DAG.getDataLayout());
16713 unsigned ReturnReg = (PtrVT == MVT::i64 ? X86::RAX : X86::EAX);
16715 // Load the address of the destination block.
16716 MachineBasicBlock *DestMBB = cast<BasicBlockSDNode>(Dest)->getBasicBlock();
16717 SDValue BlockPtr = DAG.getMCSymbol(DestMBB->getSymbol(), PtrVT);
16718 unsigned WrapperKind =
16719 Subtarget->isPICStyleRIPRel() ? X86ISD::WrapperRIP : X86ISD::Wrapper;
16720 SDValue WrappedPtr = DAG.getNode(WrapperKind, DL, PtrVT, BlockPtr);
16721 Chain = DAG.getCopyToReg(Chain, DL, ReturnReg, WrappedPtr);
16722 return DAG.getNode(X86ISD::CATCHRET, DL, MVT::Other, Chain,
16723 DAG.getRegister(ReturnReg, PtrVT));
16726 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
16727 SelectionDAG &DAG) const {
16729 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
16730 DAG.getVTList(MVT::i32, MVT::Other),
16731 Op.getOperand(0), Op.getOperand(1));
16734 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
16735 SelectionDAG &DAG) const {
16737 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
16738 Op.getOperand(0), Op.getOperand(1));
16741 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
16742 return Op.getOperand(0);
16745 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
16746 SelectionDAG &DAG) const {
16747 SDValue Root = Op.getOperand(0);
16748 SDValue Trmp = Op.getOperand(1); // trampoline
16749 SDValue FPtr = Op.getOperand(2); // nested function
16750 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
16753 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
16754 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
16756 if (Subtarget->is64Bit()) {
16757 SDValue OutChains[6];
16759 // Large code-model.
16760 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
16761 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
16763 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
16764 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
16766 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
16768 // Load the pointer to the nested function into R11.
16769 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
16770 SDValue Addr = Trmp;
16771 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16772 Addr, MachinePointerInfo(TrmpAddr),
16775 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16776 DAG.getConstant(2, dl, MVT::i64));
16777 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
16778 MachinePointerInfo(TrmpAddr, 2),
16781 // Load the 'nest' parameter value into R10.
16782 // R10 is specified in X86CallingConv.td
16783 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
16784 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16785 DAG.getConstant(10, dl, MVT::i64));
16786 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16787 Addr, MachinePointerInfo(TrmpAddr, 10),
16790 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16791 DAG.getConstant(12, dl, MVT::i64));
16792 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
16793 MachinePointerInfo(TrmpAddr, 12),
16796 // Jump to the nested function.
16797 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
16798 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16799 DAG.getConstant(20, dl, MVT::i64));
16800 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16801 Addr, MachinePointerInfo(TrmpAddr, 20),
16804 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
16805 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16806 DAG.getConstant(22, dl, MVT::i64));
16807 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
16808 Addr, MachinePointerInfo(TrmpAddr, 22),
16811 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16813 const Function *Func =
16814 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
16815 CallingConv::ID CC = Func->getCallingConv();
16820 llvm_unreachable("Unsupported calling convention");
16821 case CallingConv::C:
16822 case CallingConv::X86_StdCall: {
16823 // Pass 'nest' parameter in ECX.
16824 // Must be kept in sync with X86CallingConv.td
16825 NestReg = X86::ECX;
16827 // Check that ECX wasn't needed by an 'inreg' parameter.
16828 FunctionType *FTy = Func->getFunctionType();
16829 const AttributeSet &Attrs = Func->getAttributes();
16831 if (!Attrs.isEmpty() && !Func->isVarArg()) {
16832 unsigned InRegCount = 0;
16835 for (FunctionType::param_iterator I = FTy->param_begin(),
16836 E = FTy->param_end(); I != E; ++I, ++Idx)
16837 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
16838 auto &DL = DAG.getDataLayout();
16839 // FIXME: should only count parameters that are lowered to integers.
16840 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
16843 if (InRegCount > 2) {
16844 report_fatal_error("Nest register in use - reduce number of inreg"
16850 case CallingConv::X86_FastCall:
16851 case CallingConv::X86_ThisCall:
16852 case CallingConv::Fast:
16853 // Pass 'nest' parameter in EAX.
16854 // Must be kept in sync with X86CallingConv.td
16855 NestReg = X86::EAX;
16859 SDValue OutChains[4];
16860 SDValue Addr, Disp;
16862 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16863 DAG.getConstant(10, dl, MVT::i32));
16864 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
16866 // This is storing the opcode for MOV32ri.
16867 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
16868 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
16869 OutChains[0] = DAG.getStore(Root, dl,
16870 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
16871 Trmp, MachinePointerInfo(TrmpAddr),
16874 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16875 DAG.getConstant(1, dl, MVT::i32));
16876 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
16877 MachinePointerInfo(TrmpAddr, 1),
16880 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
16881 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16882 DAG.getConstant(5, dl, MVT::i32));
16883 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
16884 Addr, MachinePointerInfo(TrmpAddr, 5),
16887 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16888 DAG.getConstant(6, dl, MVT::i32));
16889 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
16890 MachinePointerInfo(TrmpAddr, 6),
16893 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16897 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
16898 SelectionDAG &DAG) const {
16900 The rounding mode is in bits 11:10 of FPSR, and has the following
16902 00 Round to nearest
16907 FLT_ROUNDS, on the other hand, expects the following:
16914 To perform the conversion, we do:
16915 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
16918 MachineFunction &MF = DAG.getMachineFunction();
16919 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16920 unsigned StackAlignment = TFI.getStackAlignment();
16921 MVT VT = Op.getSimpleValueType();
16924 // Save FP Control Word to stack slot
16925 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
16926 SDValue StackSlot =
16927 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
16929 MachineMemOperand *MMO =
16930 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
16931 MachineMemOperand::MOStore, 2, 2);
16933 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
16934 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
16935 DAG.getVTList(MVT::Other),
16936 Ops, MVT::i16, MMO);
16938 // Load FP Control Word from stack slot
16939 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
16940 MachinePointerInfo(), false, false, false, 0);
16942 // Transform as necessary
16944 DAG.getNode(ISD::SRL, DL, MVT::i16,
16945 DAG.getNode(ISD::AND, DL, MVT::i16,
16946 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
16947 DAG.getConstant(11, DL, MVT::i8));
16949 DAG.getNode(ISD::SRL, DL, MVT::i16,
16950 DAG.getNode(ISD::AND, DL, MVT::i16,
16951 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
16952 DAG.getConstant(9, DL, MVT::i8));
16955 DAG.getNode(ISD::AND, DL, MVT::i16,
16956 DAG.getNode(ISD::ADD, DL, MVT::i16,
16957 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
16958 DAG.getConstant(1, DL, MVT::i16)),
16959 DAG.getConstant(3, DL, MVT::i16));
16961 return DAG.getNode((VT.getSizeInBits() < 16 ?
16962 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
16965 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
16966 MVT VT = Op.getSimpleValueType();
16968 unsigned NumBits = VT.getSizeInBits();
16971 Op = Op.getOperand(0);
16972 if (VT == MVT::i8) {
16973 // Zero extend to i32 since there is not an i8 bsr.
16975 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16978 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
16979 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16980 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16982 // If src is zero (i.e. bsr sets ZF), returns NumBits.
16985 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
16986 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16989 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
16991 // Finally xor with NumBits-1.
16992 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16993 DAG.getConstant(NumBits - 1, dl, OpVT));
16996 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17000 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
17001 MVT VT = Op.getSimpleValueType();
17003 unsigned NumBits = VT.getSizeInBits();
17006 Op = Op.getOperand(0);
17007 if (VT == MVT::i8) {
17008 // Zero extend to i32 since there is not an i8 bsr.
17010 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17013 // Issue a bsr (scan bits in reverse).
17014 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17015 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17017 // And xor with NumBits-1.
17018 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17019 DAG.getConstant(NumBits - 1, dl, OpVT));
17022 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17026 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
17027 MVT VT = Op.getSimpleValueType();
17028 unsigned NumBits = VT.getSizeInBits();
17030 Op = Op.getOperand(0);
17032 // Issue a bsf (scan bits forward) which also sets EFLAGS.
17033 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17034 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
17036 // If src is zero (i.e. bsf sets ZF), returns NumBits.
17039 DAG.getConstant(NumBits, dl, VT),
17040 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17043 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
17046 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
17047 // ones, and then concatenate the result back.
17048 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
17049 MVT VT = Op.getSimpleValueType();
17051 assert(VT.is256BitVector() && VT.isInteger() &&
17052 "Unsupported value type for operation");
17054 unsigned NumElems = VT.getVectorNumElements();
17057 // Extract the LHS vectors
17058 SDValue LHS = Op.getOperand(0);
17059 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17060 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17062 // Extract the RHS vectors
17063 SDValue RHS = Op.getOperand(1);
17064 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
17065 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
17067 MVT EltVT = VT.getVectorElementType();
17068 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17070 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17071 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
17072 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
17075 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
17076 if (Op.getValueType() == MVT::i1)
17077 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17078 Op.getOperand(0), Op.getOperand(1));
17079 assert(Op.getSimpleValueType().is256BitVector() &&
17080 Op.getSimpleValueType().isInteger() &&
17081 "Only handle AVX 256-bit vector integer operation");
17082 return Lower256IntArith(Op, DAG);
17085 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
17086 if (Op.getValueType() == MVT::i1)
17087 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17088 Op.getOperand(0), Op.getOperand(1));
17089 assert(Op.getSimpleValueType().is256BitVector() &&
17090 Op.getSimpleValueType().isInteger() &&
17091 "Only handle AVX 256-bit vector integer operation");
17092 return Lower256IntArith(Op, DAG);
17095 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
17096 assert(Op.getSimpleValueType().is256BitVector() &&
17097 Op.getSimpleValueType().isInteger() &&
17098 "Only handle AVX 256-bit vector integer operation");
17099 return Lower256IntArith(Op, DAG);
17102 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
17103 SelectionDAG &DAG) {
17105 MVT VT = Op.getSimpleValueType();
17108 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
17110 // Decompose 256-bit ops into smaller 128-bit ops.
17111 if (VT.is256BitVector() && !Subtarget->hasInt256())
17112 return Lower256IntArith(Op, DAG);
17114 SDValue A = Op.getOperand(0);
17115 SDValue B = Op.getOperand(1);
17117 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
17118 // pairs, multiply and truncate.
17119 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
17120 if (Subtarget->hasInt256()) {
17121 if (VT == MVT::v32i8) {
17122 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
17123 SDValue Lo = DAG.getIntPtrConstant(0, dl);
17124 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
17125 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
17126 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
17127 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
17128 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
17129 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17130 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
17131 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
17134 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
17135 return DAG.getNode(
17136 ISD::TRUNCATE, dl, VT,
17137 DAG.getNode(ISD::MUL, dl, ExVT,
17138 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
17139 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
17142 assert(VT == MVT::v16i8 &&
17143 "Pre-AVX2 support only supports v16i8 multiplication");
17144 MVT ExVT = MVT::v8i16;
17146 // Extract the lo parts and sign extend to i16
17148 if (Subtarget->hasSSE41()) {
17149 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
17150 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
17152 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
17153 -1, 4, -1, 5, -1, 6, -1, 7};
17154 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17155 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17156 ALo = DAG.getBitcast(ExVT, ALo);
17157 BLo = DAG.getBitcast(ExVT, BLo);
17158 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
17159 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
17162 // Extract the hi parts and sign extend to i16
17164 if (Subtarget->hasSSE41()) {
17165 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
17166 -1, -1, -1, -1, -1, -1, -1, -1};
17167 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17168 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17169 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
17170 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
17172 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
17173 -1, 12, -1, 13, -1, 14, -1, 15};
17174 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17175 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17176 AHi = DAG.getBitcast(ExVT, AHi);
17177 BHi = DAG.getBitcast(ExVT, BHi);
17178 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
17179 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
17182 // Multiply, mask the lower 8bits of the lo/hi results and pack
17183 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
17184 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
17185 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
17186 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
17187 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17190 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
17191 if (VT == MVT::v4i32) {
17192 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
17193 "Should not custom lower when pmuldq is available!");
17195 // Extract the odd parts.
17196 static const int UnpackMask[] = { 1, -1, 3, -1 };
17197 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
17198 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
17200 // Multiply the even parts.
17201 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
17202 // Now multiply odd parts.
17203 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
17205 Evens = DAG.getBitcast(VT, Evens);
17206 Odds = DAG.getBitcast(VT, Odds);
17208 // Merge the two vectors back together with a shuffle. This expands into 2
17210 static const int ShufMask[] = { 0, 4, 2, 6 };
17211 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
17214 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
17215 "Only know how to lower V2I64/V4I64/V8I64 multiply");
17217 // Ahi = psrlqi(a, 32);
17218 // Bhi = psrlqi(b, 32);
17220 // AloBlo = pmuludq(a, b);
17221 // AloBhi = pmuludq(a, Bhi);
17222 // AhiBlo = pmuludq(Ahi, b);
17224 // AloBhi = psllqi(AloBhi, 32);
17225 // AhiBlo = psllqi(AhiBlo, 32);
17226 // return AloBlo + AloBhi + AhiBlo;
17228 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
17229 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17231 SDValue AhiBlo = Ahi;
17232 SDValue AloBhi = Bhi;
17233 // Bit cast to 32-bit vectors for MULUDQ
17234 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17235 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17236 A = DAG.getBitcast(MulVT, A);
17237 B = DAG.getBitcast(MulVT, B);
17238 Ahi = DAG.getBitcast(MulVT, Ahi);
17239 Bhi = DAG.getBitcast(MulVT, Bhi);
17241 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17242 // After shifting right const values the result may be all-zero.
17243 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17244 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17245 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17247 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17248 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17249 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17252 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17253 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17256 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17257 assert(Subtarget->isTargetWin64() && "Unexpected target");
17258 EVT VT = Op.getValueType();
17259 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17260 "Unexpected return type for lowering");
17264 switch (Op->getOpcode()) {
17265 default: llvm_unreachable("Unexpected request for libcall!");
17266 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17267 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17268 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17269 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17270 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17271 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17275 SDValue InChain = DAG.getEntryNode();
17277 TargetLowering::ArgListTy Args;
17278 TargetLowering::ArgListEntry Entry;
17279 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17280 EVT ArgVT = Op->getOperand(i).getValueType();
17281 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17282 "Unexpected argument type for lowering");
17283 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17284 Entry.Node = StackPtr;
17285 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17287 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17288 Entry.Ty = PointerType::get(ArgTy,0);
17289 Entry.isSExt = false;
17290 Entry.isZExt = false;
17291 Args.push_back(Entry);
17294 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
17295 getPointerTy(DAG.getDataLayout()));
17297 TargetLowering::CallLoweringInfo CLI(DAG);
17298 CLI.setDebugLoc(dl).setChain(InChain)
17299 .setCallee(getLibcallCallingConv(LC),
17300 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
17301 Callee, std::move(Args), 0)
17302 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
17304 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
17305 return DAG.getBitcast(VT, CallInfo.first);
17308 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
17309 SelectionDAG &DAG) {
17310 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
17311 EVT VT = Op0.getValueType();
17314 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
17315 (VT == MVT::v8i32 && Subtarget->hasInt256()));
17317 // PMULxD operations multiply each even value (starting at 0) of LHS with
17318 // the related value of RHS and produce a widen result.
17319 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17320 // => <2 x i64> <ae|cg>
17322 // In other word, to have all the results, we need to perform two PMULxD:
17323 // 1. one with the even values.
17324 // 2. one with the odd values.
17325 // To achieve #2, with need to place the odd values at an even position.
17327 // Place the odd value at an even position (basically, shift all values 1
17328 // step to the left):
17329 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
17330 // <a|b|c|d> => <b|undef|d|undef>
17331 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
17332 // <e|f|g|h> => <f|undef|h|undef>
17333 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
17335 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
17337 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
17338 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
17340 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
17341 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17342 // => <2 x i64> <ae|cg>
17343 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
17344 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
17345 // => <2 x i64> <bf|dh>
17346 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
17348 // Shuffle it back into the right order.
17349 SDValue Highs, Lows;
17350 if (VT == MVT::v8i32) {
17351 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
17352 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17353 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
17354 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17356 const int HighMask[] = {1, 5, 3, 7};
17357 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17358 const int LowMask[] = {0, 4, 2, 6};
17359 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17362 // If we have a signed multiply but no PMULDQ fix up the high parts of a
17363 // unsigned multiply.
17364 if (IsSigned && !Subtarget->hasSSE41()) {
17365 SDValue ShAmt = DAG.getConstant(
17367 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
17368 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
17369 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
17370 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
17371 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
17373 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
17374 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
17377 // The first result of MUL_LOHI is actually the low value, followed by the
17379 SDValue Ops[] = {Lows, Highs};
17380 return DAG.getMergeValues(Ops, dl);
17383 // Return true if the required (according to Opcode) shift-imm form is natively
17384 // supported by the Subtarget
17385 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
17387 if (VT.getScalarSizeInBits() < 16)
17390 if (VT.is512BitVector() &&
17391 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
17394 bool LShift = VT.is128BitVector() ||
17395 (VT.is256BitVector() && Subtarget->hasInt256());
17397 bool AShift = LShift && (Subtarget->hasVLX() ||
17398 (VT != MVT::v2i64 && VT != MVT::v4i64));
17399 return (Opcode == ISD::SRA) ? AShift : LShift;
17402 // The shift amount is a variable, but it is the same for all vector lanes.
17403 // These instructions are defined together with shift-immediate.
17405 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
17407 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
17410 // Return true if the required (according to Opcode) variable-shift form is
17411 // natively supported by the Subtarget
17412 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
17415 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
17418 // vXi16 supported only on AVX-512, BWI
17419 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
17422 if (VT.is512BitVector() || Subtarget->hasVLX())
17425 bool LShift = VT.is128BitVector() || VT.is256BitVector();
17426 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
17427 return (Opcode == ISD::SRA) ? AShift : LShift;
17430 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
17431 const X86Subtarget *Subtarget) {
17432 MVT VT = Op.getSimpleValueType();
17434 SDValue R = Op.getOperand(0);
17435 SDValue Amt = Op.getOperand(1);
17437 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17438 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17440 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
17441 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
17442 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
17443 SDValue Ex = DAG.getBitcast(ExVT, R);
17445 if (ShiftAmt >= 32) {
17446 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
17448 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
17449 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17450 ShiftAmt - 32, DAG);
17451 if (VT == MVT::v2i64)
17452 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
17453 if (VT == MVT::v4i64)
17454 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17455 {9, 1, 11, 3, 13, 5, 15, 7});
17457 // SRA upper i32, SHL whole i64 and select lower i32.
17458 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17461 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
17462 Lower = DAG.getBitcast(ExVT, Lower);
17463 if (VT == MVT::v2i64)
17464 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
17465 if (VT == MVT::v4i64)
17466 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17467 {8, 1, 10, 3, 12, 5, 14, 7});
17469 return DAG.getBitcast(VT, Ex);
17472 // Optimize shl/srl/sra with constant shift amount.
17473 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
17474 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
17475 uint64_t ShiftAmt = ShiftConst->getZExtValue();
17477 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17478 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17480 // i64 SRA needs to be performed as partial shifts.
17481 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
17482 Op.getOpcode() == ISD::SRA)
17483 return ArithmeticShiftRight64(ShiftAmt);
17485 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
17486 unsigned NumElts = VT.getVectorNumElements();
17487 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
17489 if (Op.getOpcode() == ISD::SHL) {
17490 // Simple i8 add case
17492 return DAG.getNode(ISD::ADD, dl, VT, R, R);
17494 // Make a large shift.
17495 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
17497 SHL = DAG.getBitcast(VT, SHL);
17498 // Zero out the rightmost bits.
17499 SmallVector<SDValue, 32> V(
17500 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
17501 return DAG.getNode(ISD::AND, dl, VT, SHL,
17502 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17504 if (Op.getOpcode() == ISD::SRL) {
17505 // Make a large shift.
17506 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
17508 SRL = DAG.getBitcast(VT, SRL);
17509 // Zero out the leftmost bits.
17510 SmallVector<SDValue, 32> V(
17511 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
17512 return DAG.getNode(ISD::AND, dl, VT, SRL,
17513 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17515 if (Op.getOpcode() == ISD::SRA) {
17516 if (ShiftAmt == 7) {
17517 // ashr(R, 7) === cmp_slt(R, 0)
17518 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
17519 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
17522 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
17523 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17524 SmallVector<SDValue, 32> V(NumElts,
17525 DAG.getConstant(128 >> ShiftAmt, dl,
17527 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
17528 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
17529 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
17532 llvm_unreachable("Unknown shift opcode.");
17537 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17538 if (!Subtarget->is64Bit() &&
17539 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
17541 // Peek through any splat that was introduced for i64 shift vectorization.
17542 int SplatIndex = -1;
17543 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
17544 if (SVN->isSplat()) {
17545 SplatIndex = SVN->getSplatIndex();
17546 Amt = Amt.getOperand(0);
17547 assert(SplatIndex < (int)VT.getVectorNumElements() &&
17548 "Splat shuffle referencing second operand");
17551 if (Amt.getOpcode() != ISD::BITCAST ||
17552 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
17555 Amt = Amt.getOperand(0);
17556 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17557 VT.getVectorNumElements();
17558 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
17559 uint64_t ShiftAmt = 0;
17560 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
17561 for (unsigned i = 0; i != Ratio; ++i) {
17562 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
17566 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
17569 // Check remaining shift amounts (if not a splat).
17570 if (SplatIndex < 0) {
17571 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17572 uint64_t ShAmt = 0;
17573 for (unsigned j = 0; j != Ratio; ++j) {
17574 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
17578 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
17580 if (ShAmt != ShiftAmt)
17585 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17586 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17588 if (Op.getOpcode() == ISD::SRA)
17589 return ArithmeticShiftRight64(ShiftAmt);
17595 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
17596 const X86Subtarget* Subtarget) {
17597 MVT VT = Op.getSimpleValueType();
17599 SDValue R = Op.getOperand(0);
17600 SDValue Amt = Op.getOperand(1);
17602 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17603 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17605 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
17606 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
17608 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
17610 EVT EltVT = VT.getVectorElementType();
17612 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
17613 // Check if this build_vector node is doing a splat.
17614 // If so, then set BaseShAmt equal to the splat value.
17615 BaseShAmt = BV->getSplatValue();
17616 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
17617 BaseShAmt = SDValue();
17619 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
17620 Amt = Amt.getOperand(0);
17622 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
17623 if (SVN && SVN->isSplat()) {
17624 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
17625 SDValue InVec = Amt.getOperand(0);
17626 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
17627 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
17628 "Unexpected shuffle index found!");
17629 BaseShAmt = InVec.getOperand(SplatIdx);
17630 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
17631 if (ConstantSDNode *C =
17632 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
17633 if (C->getZExtValue() == SplatIdx)
17634 BaseShAmt = InVec.getOperand(1);
17639 // Avoid introducing an extract element from a shuffle.
17640 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
17641 DAG.getIntPtrConstant(SplatIdx, dl));
17645 if (BaseShAmt.getNode()) {
17646 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
17647 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
17648 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
17649 else if (EltVT.bitsLT(MVT::i32))
17650 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
17652 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
17656 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17657 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
17658 Amt.getOpcode() == ISD::BITCAST &&
17659 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
17660 Amt = Amt.getOperand(0);
17661 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17662 VT.getVectorNumElements();
17663 std::vector<SDValue> Vals(Ratio);
17664 for (unsigned i = 0; i != Ratio; ++i)
17665 Vals[i] = Amt.getOperand(i);
17666 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17667 for (unsigned j = 0; j != Ratio; ++j)
17668 if (Vals[j] != Amt.getOperand(i + j))
17672 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
17673 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
17678 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
17679 SelectionDAG &DAG) {
17680 MVT VT = Op.getSimpleValueType();
17682 SDValue R = Op.getOperand(0);
17683 SDValue Amt = Op.getOperand(1);
17685 assert(VT.isVector() && "Custom lowering only for vector shifts!");
17686 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
17688 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
17691 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
17694 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
17697 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
17698 // shifts per-lane and then shuffle the partial results back together.
17699 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
17700 // Splat the shift amounts so the scalar shifts above will catch it.
17701 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
17702 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
17703 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
17704 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
17705 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
17708 // i64 vector arithmetic shift can be emulated with the transform:
17709 // M = lshr(SIGN_BIT, Amt)
17710 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
17711 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
17712 Op.getOpcode() == ISD::SRA) {
17713 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
17714 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
17715 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17716 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
17717 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
17721 // If possible, lower this packed shift into a vector multiply instead of
17722 // expanding it into a sequence of scalar shifts.
17723 // Do this only if the vector shift count is a constant build_vector.
17724 if (Op.getOpcode() == ISD::SHL &&
17725 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
17726 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
17727 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17728 SmallVector<SDValue, 8> Elts;
17729 EVT SVT = VT.getScalarType();
17730 unsigned SVTBits = SVT.getSizeInBits();
17731 const APInt &One = APInt(SVTBits, 1);
17732 unsigned NumElems = VT.getVectorNumElements();
17734 for (unsigned i=0; i !=NumElems; ++i) {
17735 SDValue Op = Amt->getOperand(i);
17736 if (Op->getOpcode() == ISD::UNDEF) {
17737 Elts.push_back(Op);
17741 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
17742 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
17743 uint64_t ShAmt = C.getZExtValue();
17744 if (ShAmt >= SVTBits) {
17745 Elts.push_back(DAG.getUNDEF(SVT));
17748 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
17750 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
17751 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
17754 // Lower SHL with variable shift amount.
17755 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
17756 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
17758 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
17759 DAG.getConstant(0x3f800000U, dl, VT));
17760 Op = DAG.getBitcast(MVT::v4f32, Op);
17761 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
17762 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
17765 // If possible, lower this shift as a sequence of two shifts by
17766 // constant plus a MOVSS/MOVSD instead of scalarizing it.
17768 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
17770 // Could be rewritten as:
17771 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
17773 // The advantage is that the two shifts from the example would be
17774 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
17775 // the vector shift into four scalar shifts plus four pairs of vector
17777 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
17778 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17779 unsigned TargetOpcode = X86ISD::MOVSS;
17780 bool CanBeSimplified;
17781 // The splat value for the first packed shift (the 'X' from the example).
17782 SDValue Amt1 = Amt->getOperand(0);
17783 // The splat value for the second packed shift (the 'Y' from the example).
17784 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
17785 Amt->getOperand(2);
17787 // See if it is possible to replace this node with a sequence of
17788 // two shifts followed by a MOVSS/MOVSD
17789 if (VT == MVT::v4i32) {
17790 // Check if it is legal to use a MOVSS.
17791 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
17792 Amt2 == Amt->getOperand(3);
17793 if (!CanBeSimplified) {
17794 // Otherwise, check if we can still simplify this node using a MOVSD.
17795 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
17796 Amt->getOperand(2) == Amt->getOperand(3);
17797 TargetOpcode = X86ISD::MOVSD;
17798 Amt2 = Amt->getOperand(2);
17801 // Do similar checks for the case where the machine value type
17803 CanBeSimplified = Amt1 == Amt->getOperand(1);
17804 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
17805 CanBeSimplified = Amt2 == Amt->getOperand(i);
17807 if (!CanBeSimplified) {
17808 TargetOpcode = X86ISD::MOVSD;
17809 CanBeSimplified = true;
17810 Amt2 = Amt->getOperand(4);
17811 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
17812 CanBeSimplified = Amt1 == Amt->getOperand(i);
17813 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
17814 CanBeSimplified = Amt2 == Amt->getOperand(j);
17818 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
17819 isa<ConstantSDNode>(Amt2)) {
17820 // Replace this node with two shifts followed by a MOVSS/MOVSD.
17821 EVT CastVT = MVT::v4i32;
17823 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
17824 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
17826 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
17827 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
17828 if (TargetOpcode == X86ISD::MOVSD)
17829 CastVT = MVT::v2i64;
17830 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
17831 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
17832 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
17834 return DAG.getBitcast(VT, Result);
17838 // v4i32 Non Uniform Shifts.
17839 // If the shift amount is constant we can shift each lane using the SSE2
17840 // immediate shifts, else we need to zero-extend each lane to the lower i64
17841 // and shift using the SSE2 variable shifts.
17842 // The separate results can then be blended together.
17843 if (VT == MVT::v4i32) {
17844 unsigned Opc = Op.getOpcode();
17845 SDValue Amt0, Amt1, Amt2, Amt3;
17846 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17847 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
17848 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
17849 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
17850 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
17852 // ISD::SHL is handled above but we include it here for completeness.
17855 llvm_unreachable("Unknown target vector shift node");
17857 Opc = X86ISD::VSHL;
17860 Opc = X86ISD::VSRL;
17863 Opc = X86ISD::VSRA;
17866 // The SSE2 shifts use the lower i64 as the same shift amount for
17867 // all lanes and the upper i64 is ignored. These shuffle masks
17868 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
17869 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
17870 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
17871 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
17872 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
17873 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
17876 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
17877 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
17878 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
17879 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
17880 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
17881 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
17882 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
17885 if (VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget->hasInt256())) {
17886 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
17887 unsigned ShiftOpcode = Op->getOpcode();
17889 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
17890 // On SSE41 targets we make use of the fact that VSELECT lowers
17891 // to PBLENDVB which selects bytes based just on the sign bit.
17892 if (Subtarget->hasSSE41()) {
17893 V0 = DAG.getBitcast(VT, V0);
17894 V1 = DAG.getBitcast(VT, V1);
17895 Sel = DAG.getBitcast(VT, Sel);
17896 return DAG.getBitcast(SelVT,
17897 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
17899 // On pre-SSE41 targets we test for the sign bit by comparing to
17900 // zero - a negative value will set all bits of the lanes to true
17901 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
17902 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
17903 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
17904 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
17907 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
17908 // We can safely do this using i16 shifts as we're only interested in
17909 // the 3 lower bits of each byte.
17910 Amt = DAG.getBitcast(ExtVT, Amt);
17911 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
17912 Amt = DAG.getBitcast(VT, Amt);
17914 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
17915 // r = VSELECT(r, shift(r, 4), a);
17917 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
17918 R = SignBitSelect(VT, Amt, M, R);
17921 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17923 // r = VSELECT(r, shift(r, 2), a);
17924 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
17925 R = SignBitSelect(VT, Amt, M, R);
17928 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17930 // return VSELECT(r, shift(r, 1), a);
17931 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
17932 R = SignBitSelect(VT, Amt, M, R);
17936 if (Op->getOpcode() == ISD::SRA) {
17937 // For SRA we need to unpack each byte to the higher byte of a i16 vector
17938 // so we can correctly sign extend. We don't care what happens to the
17940 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
17941 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
17942 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
17943 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
17944 ALo = DAG.getBitcast(ExtVT, ALo);
17945 AHi = DAG.getBitcast(ExtVT, AHi);
17946 RLo = DAG.getBitcast(ExtVT, RLo);
17947 RHi = DAG.getBitcast(ExtVT, RHi);
17949 // r = VSELECT(r, shift(r, 4), a);
17950 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17951 DAG.getConstant(4, dl, ExtVT));
17952 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17953 DAG.getConstant(4, dl, ExtVT));
17954 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17955 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17958 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
17959 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
17961 // r = VSELECT(r, shift(r, 2), a);
17962 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17963 DAG.getConstant(2, dl, ExtVT));
17964 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17965 DAG.getConstant(2, dl, ExtVT));
17966 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17967 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17970 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
17971 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
17973 // r = VSELECT(r, shift(r, 1), a);
17974 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17975 DAG.getConstant(1, dl, ExtVT));
17976 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17977 DAG.getConstant(1, dl, ExtVT));
17978 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17979 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17981 // Logical shift the result back to the lower byte, leaving a zero upper
17983 // meaning that we can safely pack with PACKUSWB.
17985 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
17987 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
17988 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17992 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
17993 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
17994 // solution better.
17995 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
17996 MVT ExtVT = MVT::v8i32;
17998 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
17999 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
18000 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
18001 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18002 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
18005 if (Subtarget->hasInt256() && VT == MVT::v16i16) {
18006 MVT ExtVT = MVT::v8i32;
18007 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18008 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
18009 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
18010 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
18011 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
18012 ALo = DAG.getBitcast(ExtVT, ALo);
18013 AHi = DAG.getBitcast(ExtVT, AHi);
18014 RLo = DAG.getBitcast(ExtVT, RLo);
18015 RHi = DAG.getBitcast(ExtVT, RHi);
18016 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
18017 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
18018 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
18019 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
18020 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
18023 if (VT == MVT::v8i16) {
18024 unsigned ShiftOpcode = Op->getOpcode();
18026 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
18027 // On SSE41 targets we make use of the fact that VSELECT lowers
18028 // to PBLENDVB which selects bytes based just on the sign bit.
18029 if (Subtarget->hasSSE41()) {
18030 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
18031 V0 = DAG.getBitcast(ExtVT, V0);
18032 V1 = DAG.getBitcast(ExtVT, V1);
18033 Sel = DAG.getBitcast(ExtVT, Sel);
18034 return DAG.getBitcast(
18035 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
18037 // On pre-SSE41 targets we splat the sign bit - a negative value will
18038 // set all bits of the lanes to true and VSELECT uses that in
18039 // its OR(AND(V0,C),AND(V1,~C)) lowering.
18041 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
18042 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
18045 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
18046 if (Subtarget->hasSSE41()) {
18047 // On SSE41 targets we need to replicate the shift mask in both
18048 // bytes for PBLENDVB.
18051 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
18052 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
18054 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
18057 // r = VSELECT(r, shift(r, 8), a);
18058 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
18059 R = SignBitSelect(Amt, M, R);
18062 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18064 // r = VSELECT(r, shift(r, 4), a);
18065 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18066 R = SignBitSelect(Amt, M, R);
18069 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18071 // r = VSELECT(r, shift(r, 2), a);
18072 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18073 R = SignBitSelect(Amt, M, R);
18076 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18078 // return VSELECT(r, shift(r, 1), a);
18079 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18080 R = SignBitSelect(Amt, M, R);
18084 // Decompose 256-bit shifts into smaller 128-bit shifts.
18085 if (VT.is256BitVector()) {
18086 unsigned NumElems = VT.getVectorNumElements();
18087 MVT EltVT = VT.getVectorElementType();
18088 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18090 // Extract the two vectors
18091 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18092 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18094 // Recreate the shift amount vectors
18095 SDValue Amt1, Amt2;
18096 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18097 // Constant shift amount
18098 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
18099 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
18100 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
18102 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18103 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
18105 // Variable shift amount
18106 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
18107 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
18110 // Issue new vector shifts for the smaller types
18111 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
18112 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
18114 // Concatenate the result back
18115 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
18121 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
18122 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
18123 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
18124 // looks for this combo and may remove the "setcc" instruction if the "setcc"
18125 // has only one use.
18126 SDNode *N = Op.getNode();
18127 SDValue LHS = N->getOperand(0);
18128 SDValue RHS = N->getOperand(1);
18129 unsigned BaseOp = 0;
18132 switch (Op.getOpcode()) {
18133 default: llvm_unreachable("Unknown ovf instruction!");
18135 // A subtract of one will be selected as a INC. Note that INC doesn't
18136 // set CF, so we can't do this for UADDO.
18137 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18139 BaseOp = X86ISD::INC;
18140 Cond = X86::COND_O;
18143 BaseOp = X86ISD::ADD;
18144 Cond = X86::COND_O;
18147 BaseOp = X86ISD::ADD;
18148 Cond = X86::COND_B;
18151 // A subtract of one will be selected as a DEC. Note that DEC doesn't
18152 // set CF, so we can't do this for USUBO.
18153 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18155 BaseOp = X86ISD::DEC;
18156 Cond = X86::COND_O;
18159 BaseOp = X86ISD::SUB;
18160 Cond = X86::COND_O;
18163 BaseOp = X86ISD::SUB;
18164 Cond = X86::COND_B;
18167 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
18168 Cond = X86::COND_O;
18170 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
18171 if (N->getValueType(0) == MVT::i8) {
18172 BaseOp = X86ISD::UMUL8;
18173 Cond = X86::COND_O;
18176 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
18178 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
18181 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
18182 DAG.getConstant(X86::COND_O, DL, MVT::i32),
18183 SDValue(Sum.getNode(), 2));
18185 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18189 // Also sets EFLAGS.
18190 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
18191 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
18194 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
18195 DAG.getConstant(Cond, DL, MVT::i32),
18196 SDValue(Sum.getNode(), 1));
18198 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18201 /// Returns true if the operand type is exactly twice the native width, and
18202 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
18203 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
18204 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
18205 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
18206 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
18209 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
18210 else if (OpWidth == 128)
18211 return Subtarget->hasCmpxchg16b();
18216 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
18217 return needsCmpXchgNb(SI->getValueOperand()->getType());
18220 // Note: this turns large loads into lock cmpxchg8b/16b.
18221 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
18222 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
18223 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
18224 return needsCmpXchgNb(PTy->getElementType());
18227 TargetLoweringBase::AtomicRMWExpansionKind
18228 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18229 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18230 Type *MemType = AI->getType();
18232 // If the operand is too big, we must see if cmpxchg8/16b is available
18233 // and default to library calls otherwise.
18234 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
18235 return needsCmpXchgNb(MemType) ? AtomicRMWExpansionKind::CmpXChg
18236 : AtomicRMWExpansionKind::None;
18239 AtomicRMWInst::BinOp Op = AI->getOperation();
18242 llvm_unreachable("Unknown atomic operation");
18243 case AtomicRMWInst::Xchg:
18244 case AtomicRMWInst::Add:
18245 case AtomicRMWInst::Sub:
18246 // It's better to use xadd, xsub or xchg for these in all cases.
18247 return AtomicRMWExpansionKind::None;
18248 case AtomicRMWInst::Or:
18249 case AtomicRMWInst::And:
18250 case AtomicRMWInst::Xor:
18251 // If the atomicrmw's result isn't actually used, we can just add a "lock"
18252 // prefix to a normal instruction for these operations.
18253 return !AI->use_empty() ? AtomicRMWExpansionKind::CmpXChg
18254 : AtomicRMWExpansionKind::None;
18255 case AtomicRMWInst::Nand:
18256 case AtomicRMWInst::Max:
18257 case AtomicRMWInst::Min:
18258 case AtomicRMWInst::UMax:
18259 case AtomicRMWInst::UMin:
18260 // These always require a non-trivial set of data operations on x86. We must
18261 // use a cmpxchg loop.
18262 return AtomicRMWExpansionKind::CmpXChg;
18266 static bool hasMFENCE(const X86Subtarget& Subtarget) {
18267 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
18268 // no-sse2). There isn't any reason to disable it if the target processor
18270 return Subtarget.hasSSE2() || Subtarget.is64Bit();
18274 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
18275 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18276 Type *MemType = AI->getType();
18277 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
18278 // there is no benefit in turning such RMWs into loads, and it is actually
18279 // harmful as it introduces a mfence.
18280 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
18283 auto Builder = IRBuilder<>(AI);
18284 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
18285 auto SynchScope = AI->getSynchScope();
18286 // We must restrict the ordering to avoid generating loads with Release or
18287 // ReleaseAcquire orderings.
18288 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
18289 auto Ptr = AI->getPointerOperand();
18291 // Before the load we need a fence. Here is an example lifted from
18292 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
18295 // x.store(1, relaxed);
18296 // r1 = y.fetch_add(0, release);
18298 // y.fetch_add(42, acquire);
18299 // r2 = x.load(relaxed);
18300 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
18301 // lowered to just a load without a fence. A mfence flushes the store buffer,
18302 // making the optimization clearly correct.
18303 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
18304 // otherwise, we might be able to be more aggressive on relaxed idempotent
18305 // rmw. In practice, they do not look useful, so we don't try to be
18306 // especially clever.
18307 if (SynchScope == SingleThread)
18308 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
18309 // the IR level, so we must wrap it in an intrinsic.
18312 if (!hasMFENCE(*Subtarget))
18313 // FIXME: it might make sense to use a locked operation here but on a
18314 // different cache-line to prevent cache-line bouncing. In practice it
18315 // is probably a small win, and x86 processors without mfence are rare
18316 // enough that we do not bother.
18320 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
18321 Builder.CreateCall(MFence, {});
18323 // Finally we can emit the atomic load.
18324 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
18325 AI->getType()->getPrimitiveSizeInBits());
18326 Loaded->setAtomic(Order, SynchScope);
18327 AI->replaceAllUsesWith(Loaded);
18328 AI->eraseFromParent();
18332 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
18333 SelectionDAG &DAG) {
18335 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
18336 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
18337 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
18338 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
18340 // The only fence that needs an instruction is a sequentially-consistent
18341 // cross-thread fence.
18342 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
18343 if (hasMFENCE(*Subtarget))
18344 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
18346 SDValue Chain = Op.getOperand(0);
18347 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
18349 DAG.getRegister(X86::ESP, MVT::i32), // Base
18350 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
18351 DAG.getRegister(0, MVT::i32), // Index
18352 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
18353 DAG.getRegister(0, MVT::i32), // Segment.
18357 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
18358 return SDValue(Res, 0);
18361 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
18362 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
18365 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
18366 SelectionDAG &DAG) {
18367 MVT T = Op.getSimpleValueType();
18371 switch(T.SimpleTy) {
18372 default: llvm_unreachable("Invalid value type!");
18373 case MVT::i8: Reg = X86::AL; size = 1; break;
18374 case MVT::i16: Reg = X86::AX; size = 2; break;
18375 case MVT::i32: Reg = X86::EAX; size = 4; break;
18377 assert(Subtarget->is64Bit() && "Node not type legal!");
18378 Reg = X86::RAX; size = 8;
18381 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
18382 Op.getOperand(2), SDValue());
18383 SDValue Ops[] = { cpIn.getValue(0),
18386 DAG.getTargetConstant(size, DL, MVT::i8),
18387 cpIn.getValue(1) };
18388 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
18389 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
18390 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
18394 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
18395 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
18396 MVT::i32, cpOut.getValue(2));
18397 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
18398 DAG.getConstant(X86::COND_E, DL, MVT::i8),
18401 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
18402 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
18403 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
18407 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
18408 SelectionDAG &DAG) {
18409 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
18410 MVT DstVT = Op.getSimpleValueType();
18412 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
18413 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18414 if (DstVT != MVT::f64)
18415 // This conversion needs to be expanded.
18418 SDValue InVec = Op->getOperand(0);
18420 unsigned NumElts = SrcVT.getVectorNumElements();
18421 EVT SVT = SrcVT.getVectorElementType();
18423 // Widen the vector in input in the case of MVT::v2i32.
18424 // Example: from MVT::v2i32 to MVT::v4i32.
18425 SmallVector<SDValue, 16> Elts;
18426 for (unsigned i = 0, e = NumElts; i != e; ++i)
18427 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
18428 DAG.getIntPtrConstant(i, dl)));
18430 // Explicitly mark the extra elements as Undef.
18431 Elts.append(NumElts, DAG.getUNDEF(SVT));
18433 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
18434 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
18435 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
18436 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
18437 DAG.getIntPtrConstant(0, dl));
18440 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
18441 Subtarget->hasMMX() && "Unexpected custom BITCAST");
18442 assert((DstVT == MVT::i64 ||
18443 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
18444 "Unexpected custom BITCAST");
18445 // i64 <=> MMX conversions are Legal.
18446 if (SrcVT==MVT::i64 && DstVT.isVector())
18448 if (DstVT==MVT::i64 && SrcVT.isVector())
18450 // MMX <=> MMX conversions are Legal.
18451 if (SrcVT.isVector() && DstVT.isVector())
18453 // All other conversions need to be expanded.
18457 /// Compute the horizontal sum of bytes in V for the elements of VT.
18459 /// Requires V to be a byte vector and VT to be an integer vector type with
18460 /// wider elements than V's type. The width of the elements of VT determines
18461 /// how many bytes of V are summed horizontally to produce each element of the
18463 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
18464 const X86Subtarget *Subtarget,
18465 SelectionDAG &DAG) {
18467 MVT ByteVecVT = V.getSimpleValueType();
18468 MVT EltVT = VT.getVectorElementType();
18469 int NumElts = VT.getVectorNumElements();
18470 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
18471 "Expected value to have byte element type.");
18472 assert(EltVT != MVT::i8 &&
18473 "Horizontal byte sum only makes sense for wider elements!");
18474 unsigned VecSize = VT.getSizeInBits();
18475 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
18477 // PSADBW instruction horizontally add all bytes and leave the result in i64
18478 // chunks, thus directly computes the pop count for v2i64 and v4i64.
18479 if (EltVT == MVT::i64) {
18480 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18481 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
18482 return DAG.getBitcast(VT, V);
18485 if (EltVT == MVT::i32) {
18486 // We unpack the low half and high half into i32s interleaved with zeros so
18487 // that we can use PSADBW to horizontally sum them. The most useful part of
18488 // this is that it lines up the results of two PSADBW instructions to be
18489 // two v2i64 vectors which concatenated are the 4 population counts. We can
18490 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
18491 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
18492 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
18493 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
18495 // Do the horizontal sums into two v2i64s.
18496 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18497 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18498 DAG.getBitcast(ByteVecVT, Low), Zeros);
18499 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18500 DAG.getBitcast(ByteVecVT, High), Zeros);
18502 // Merge them together.
18503 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
18504 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
18505 DAG.getBitcast(ShortVecVT, Low),
18506 DAG.getBitcast(ShortVecVT, High));
18508 return DAG.getBitcast(VT, V);
18511 // The only element type left is i16.
18512 assert(EltVT == MVT::i16 && "Unknown how to handle type");
18514 // To obtain pop count for each i16 element starting from the pop count for
18515 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
18516 // right by 8. It is important to shift as i16s as i8 vector shift isn't
18517 // directly supported.
18518 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
18519 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
18520 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18521 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
18522 DAG.getBitcast(ByteVecVT, V));
18523 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18526 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
18527 const X86Subtarget *Subtarget,
18528 SelectionDAG &DAG) {
18529 MVT VT = Op.getSimpleValueType();
18530 MVT EltVT = VT.getVectorElementType();
18531 unsigned VecSize = VT.getSizeInBits();
18533 // Implement a lookup table in register by using an algorithm based on:
18534 // http://wm.ite.pl/articles/sse-popcount.html
18536 // The general idea is that every lower byte nibble in the input vector is an
18537 // index into a in-register pre-computed pop count table. We then split up the
18538 // input vector in two new ones: (1) a vector with only the shifted-right
18539 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
18540 // masked out higher ones) for each byte. PSHUB is used separately with both
18541 // to index the in-register table. Next, both are added and the result is a
18542 // i8 vector where each element contains the pop count for input byte.
18544 // To obtain the pop count for elements != i8, we follow up with the same
18545 // approach and use additional tricks as described below.
18547 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
18548 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
18549 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
18550 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
18552 int NumByteElts = VecSize / 8;
18553 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
18554 SDValue In = DAG.getBitcast(ByteVecVT, Op);
18555 SmallVector<SDValue, 16> LUTVec;
18556 for (int i = 0; i < NumByteElts; ++i)
18557 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
18558 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
18559 SmallVector<SDValue, 16> Mask0F(NumByteElts,
18560 DAG.getConstant(0x0F, DL, MVT::i8));
18561 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
18564 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
18565 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
18566 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
18569 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
18571 // The input vector is used as the shuffle mask that index elements into the
18572 // LUT. After counting low and high nibbles, add the vector to obtain the
18573 // final pop count per i8 element.
18574 SDValue HighPopCnt =
18575 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
18576 SDValue LowPopCnt =
18577 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
18578 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
18580 if (EltVT == MVT::i8)
18583 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
18586 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
18587 const X86Subtarget *Subtarget,
18588 SelectionDAG &DAG) {
18589 MVT VT = Op.getSimpleValueType();
18590 assert(VT.is128BitVector() &&
18591 "Only 128-bit vector bitmath lowering supported.");
18593 int VecSize = VT.getSizeInBits();
18594 MVT EltVT = VT.getVectorElementType();
18595 int Len = EltVT.getSizeInBits();
18597 // This is the vectorized version of the "best" algorithm from
18598 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
18599 // with a minor tweak to use a series of adds + shifts instead of vector
18600 // multiplications. Implemented for all integer vector types. We only use
18601 // this when we don't have SSSE3 which allows a LUT-based lowering that is
18602 // much faster, even faster than using native popcnt instructions.
18604 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
18605 MVT VT = V.getSimpleValueType();
18606 SmallVector<SDValue, 32> Shifters(
18607 VT.getVectorNumElements(),
18608 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
18609 return DAG.getNode(OpCode, DL, VT, V,
18610 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
18612 auto GetMask = [&](SDValue V, APInt Mask) {
18613 MVT VT = V.getSimpleValueType();
18614 SmallVector<SDValue, 32> Masks(
18615 VT.getVectorNumElements(),
18616 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
18617 return DAG.getNode(ISD::AND, DL, VT, V,
18618 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
18621 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
18622 // x86, so set the SRL type to have elements at least i16 wide. This is
18623 // correct because all of our SRLs are followed immediately by a mask anyways
18624 // that handles any bits that sneak into the high bits of the byte elements.
18625 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
18629 // v = v - ((v >> 1) & 0x55555555...)
18631 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
18632 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
18633 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
18635 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
18636 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
18637 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
18638 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
18639 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
18641 // v = (v + (v >> 4)) & 0x0F0F0F0F...
18642 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
18643 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
18644 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
18646 // At this point, V contains the byte-wise population count, and we are
18647 // merely doing a horizontal sum if necessary to get the wider element
18649 if (EltVT == MVT::i8)
18652 return LowerHorizontalByteSum(
18653 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
18657 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18658 SelectionDAG &DAG) {
18659 MVT VT = Op.getSimpleValueType();
18660 // FIXME: Need to add AVX-512 support here!
18661 assert((VT.is256BitVector() || VT.is128BitVector()) &&
18662 "Unknown CTPOP type to handle");
18663 SDLoc DL(Op.getNode());
18664 SDValue Op0 = Op.getOperand(0);
18666 if (!Subtarget->hasSSSE3()) {
18667 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
18668 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
18669 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
18672 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
18673 unsigned NumElems = VT.getVectorNumElements();
18675 // Extract each 128-bit vector, compute pop count and concat the result.
18676 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
18677 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
18679 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
18680 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
18681 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
18684 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
18687 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18688 SelectionDAG &DAG) {
18689 assert(Op.getValueType().isVector() &&
18690 "We only do custom lowering for vector population count.");
18691 return LowerVectorCTPOP(Op, Subtarget, DAG);
18694 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
18695 SDNode *Node = Op.getNode();
18697 EVT T = Node->getValueType(0);
18698 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
18699 DAG.getConstant(0, dl, T), Node->getOperand(2));
18700 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
18701 cast<AtomicSDNode>(Node)->getMemoryVT(),
18702 Node->getOperand(0),
18703 Node->getOperand(1), negOp,
18704 cast<AtomicSDNode>(Node)->getMemOperand(),
18705 cast<AtomicSDNode>(Node)->getOrdering(),
18706 cast<AtomicSDNode>(Node)->getSynchScope());
18709 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
18710 SDNode *Node = Op.getNode();
18712 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
18714 // Convert seq_cst store -> xchg
18715 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
18716 // FIXME: On 32-bit, store -> fist or movq would be more efficient
18717 // (The only way to get a 16-byte store is cmpxchg16b)
18718 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
18719 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
18720 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
18721 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
18722 cast<AtomicSDNode>(Node)->getMemoryVT(),
18723 Node->getOperand(0),
18724 Node->getOperand(1), Node->getOperand(2),
18725 cast<AtomicSDNode>(Node)->getMemOperand(),
18726 cast<AtomicSDNode>(Node)->getOrdering(),
18727 cast<AtomicSDNode>(Node)->getSynchScope());
18728 return Swap.getValue(1);
18730 // Other atomic stores have a simple pattern.
18734 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
18735 EVT VT = Op.getNode()->getSimpleValueType(0);
18737 // Let legalize expand this if it isn't a legal type yet.
18738 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
18741 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18744 bool ExtraOp = false;
18745 switch (Op.getOpcode()) {
18746 default: llvm_unreachable("Invalid code");
18747 case ISD::ADDC: Opc = X86ISD::ADD; break;
18748 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
18749 case ISD::SUBC: Opc = X86ISD::SUB; break;
18750 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
18754 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18756 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18757 Op.getOperand(1), Op.getOperand(2));
18760 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
18761 SelectionDAG &DAG) {
18762 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
18764 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
18765 // which returns the values as { float, float } (in XMM0) or
18766 // { double, double } (which is returned in XMM0, XMM1).
18768 SDValue Arg = Op.getOperand(0);
18769 EVT ArgVT = Arg.getValueType();
18770 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18772 TargetLowering::ArgListTy Args;
18773 TargetLowering::ArgListEntry Entry;
18777 Entry.isSExt = false;
18778 Entry.isZExt = false;
18779 Args.push_back(Entry);
18781 bool isF64 = ArgVT == MVT::f64;
18782 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
18783 // the small struct {f32, f32} is returned in (eax, edx). For f64,
18784 // the results are returned via SRet in memory.
18785 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
18786 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18788 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
18790 Type *RetTy = isF64
18791 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
18792 : (Type*)VectorType::get(ArgTy, 4);
18794 TargetLowering::CallLoweringInfo CLI(DAG);
18795 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
18796 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
18798 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
18801 // Returned in xmm0 and xmm1.
18802 return CallResult.first;
18804 // Returned in bits 0:31 and 32:64 xmm0.
18805 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
18806 CallResult.first, DAG.getIntPtrConstant(0, dl));
18807 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
18808 CallResult.first, DAG.getIntPtrConstant(1, dl));
18809 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
18810 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
18813 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
18814 SelectionDAG &DAG) {
18815 assert(Subtarget->hasAVX512() &&
18816 "MGATHER/MSCATTER are supported on AVX-512 arch only");
18818 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
18819 EVT VT = N->getValue().getValueType();
18820 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
18823 // X86 scatter kills mask register, so its type should be added to
18824 // the list of return values
18825 if (N->getNumValues() == 1) {
18826 SDValue Index = N->getIndex();
18827 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
18828 !Index.getValueType().is512BitVector())
18829 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
18831 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
18832 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
18833 N->getOperand(3), Index };
18835 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
18836 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
18837 return SDValue(NewScatter.getNode(), 0);
18842 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
18843 SelectionDAG &DAG) {
18844 assert(Subtarget->hasAVX512() &&
18845 "MGATHER/MSCATTER are supported on AVX-512 arch only");
18847 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
18848 EVT VT = Op.getValueType();
18849 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
18852 SDValue Index = N->getIndex();
18853 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
18854 !Index.getValueType().is512BitVector()) {
18855 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
18856 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
18857 N->getOperand(3), Index };
18858 DAG.UpdateNodeOperands(N, Ops);
18863 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
18864 SelectionDAG &DAG) const {
18865 // TODO: Eventually, the lowering of these nodes should be informed by or
18866 // deferred to the GC strategy for the function in which they appear. For
18867 // now, however, they must be lowered to something. Since they are logically
18868 // no-ops in the case of a null GC strategy (or a GC strategy which does not
18869 // require special handling for these nodes), lower them as literal NOOPs for
18871 SmallVector<SDValue, 2> Ops;
18873 Ops.push_back(Op.getOperand(0));
18874 if (Op->getGluedNode())
18875 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
18878 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
18879 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
18884 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
18885 SelectionDAG &DAG) const {
18886 // TODO: Eventually, the lowering of these nodes should be informed by or
18887 // deferred to the GC strategy for the function in which they appear. For
18888 // now, however, they must be lowered to something. Since they are logically
18889 // no-ops in the case of a null GC strategy (or a GC strategy which does not
18890 // require special handling for these nodes), lower them as literal NOOPs for
18892 SmallVector<SDValue, 2> Ops;
18894 Ops.push_back(Op.getOperand(0));
18895 if (Op->getGluedNode())
18896 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
18899 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
18900 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
18905 /// LowerOperation - Provide custom lowering hooks for some operations.
18907 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
18908 switch (Op.getOpcode()) {
18909 default: llvm_unreachable("Should not custom lower this!");
18910 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
18911 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
18912 return LowerCMP_SWAP(Op, Subtarget, DAG);
18913 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
18914 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
18915 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
18916 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
18917 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
18918 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
18919 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
18920 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
18921 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
18922 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
18923 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
18924 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
18925 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
18926 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
18927 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
18928 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
18929 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
18930 case ISD::SHL_PARTS:
18931 case ISD::SRA_PARTS:
18932 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
18933 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
18934 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
18935 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
18936 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
18937 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
18938 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
18939 case ISD::SIGN_EXTEND_VECTOR_INREG:
18940 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
18941 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
18942 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
18943 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
18944 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
18946 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
18947 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
18948 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
18949 case ISD::SETCC: return LowerSETCC(Op, DAG);
18950 case ISD::SELECT: return LowerSELECT(Op, DAG);
18951 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
18952 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
18953 case ISD::VASTART: return LowerVASTART(Op, DAG);
18954 case ISD::VAARG: return LowerVAARG(Op, DAG);
18955 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
18956 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
18957 case ISD::INTRINSIC_VOID:
18958 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
18959 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
18960 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
18961 case ISD::FRAME_TO_ARGS_OFFSET:
18962 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
18963 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
18964 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
18965 case ISD::CATCHRET: return LowerCATCHRET(Op, DAG);
18966 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
18967 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
18968 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
18969 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
18970 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
18971 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
18972 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
18973 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
18974 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
18975 case ISD::UMUL_LOHI:
18976 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
18979 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
18985 case ISD::UMULO: return LowerXALUO(Op, DAG);
18986 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
18987 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
18991 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
18992 case ISD::ADD: return LowerADD(Op, DAG);
18993 case ISD::SUB: return LowerSUB(Op, DAG);
18997 case ISD::UMIN: return LowerMINMAX(Op, DAG);
18998 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
18999 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
19000 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
19001 case ISD::GC_TRANSITION_START:
19002 return LowerGC_TRANSITION_START(Op, DAG);
19003 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
19007 /// ReplaceNodeResults - Replace a node with an illegal result type
19008 /// with a new node built out of custom code.
19009 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
19010 SmallVectorImpl<SDValue>&Results,
19011 SelectionDAG &DAG) const {
19013 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19014 switch (N->getOpcode()) {
19016 llvm_unreachable("Do not know how to custom type legalize this operation!");
19017 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
19018 case X86ISD::FMINC:
19020 case X86ISD::FMAXC:
19021 case X86ISD::FMAX: {
19022 EVT VT = N->getValueType(0);
19023 if (VT != MVT::v2f32)
19024 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
19025 SDValue UNDEF = DAG.getUNDEF(VT);
19026 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19027 N->getOperand(0), UNDEF);
19028 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19029 N->getOperand(1), UNDEF);
19030 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
19033 case ISD::SIGN_EXTEND_INREG:
19038 // We don't want to expand or promote these.
19045 case ISD::UDIVREM: {
19046 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
19047 Results.push_back(V);
19050 case ISD::FP_TO_SINT:
19051 case ISD::FP_TO_UINT: {
19052 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
19054 std::pair<SDValue,SDValue> Vals =
19055 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
19056 SDValue FIST = Vals.first, StackSlot = Vals.second;
19057 if (FIST.getNode()) {
19058 EVT VT = N->getValueType(0);
19059 // Return a load from the stack slot.
19060 if (StackSlot.getNode())
19061 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
19062 MachinePointerInfo(),
19063 false, false, false, 0));
19065 Results.push_back(FIST);
19069 case ISD::UINT_TO_FP: {
19070 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19071 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
19072 N->getValueType(0) != MVT::v2f32)
19074 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
19076 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
19078 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
19079 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
19080 DAG.getBitcast(MVT::v2i64, VBias));
19081 Or = DAG.getBitcast(MVT::v2f64, Or);
19082 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
19083 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
19086 case ISD::FP_ROUND: {
19087 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
19089 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
19090 Results.push_back(V);
19093 case ISD::FP_EXTEND: {
19094 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
19095 // No other ValueType for FP_EXTEND should reach this point.
19096 assert(N->getValueType(0) == MVT::v2f32 &&
19097 "Do not know how to legalize this Node");
19100 case ISD::INTRINSIC_W_CHAIN: {
19101 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
19103 default : llvm_unreachable("Do not know how to custom type "
19104 "legalize this intrinsic operation!");
19105 case Intrinsic::x86_rdtsc:
19106 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19108 case Intrinsic::x86_rdtscp:
19109 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
19111 case Intrinsic::x86_rdpmc:
19112 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
19115 case ISD::READCYCLECOUNTER: {
19116 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19119 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
19120 EVT T = N->getValueType(0);
19121 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
19122 bool Regs64bit = T == MVT::i128;
19123 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
19124 SDValue cpInL, cpInH;
19125 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19126 DAG.getConstant(0, dl, HalfT));
19127 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19128 DAG.getConstant(1, dl, HalfT));
19129 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
19130 Regs64bit ? X86::RAX : X86::EAX,
19132 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
19133 Regs64bit ? X86::RDX : X86::EDX,
19134 cpInH, cpInL.getValue(1));
19135 SDValue swapInL, swapInH;
19136 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19137 DAG.getConstant(0, dl, HalfT));
19138 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19139 DAG.getConstant(1, dl, HalfT));
19140 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
19141 Regs64bit ? X86::RBX : X86::EBX,
19142 swapInL, cpInH.getValue(1));
19143 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
19144 Regs64bit ? X86::RCX : X86::ECX,
19145 swapInH, swapInL.getValue(1));
19146 SDValue Ops[] = { swapInH.getValue(0),
19148 swapInH.getValue(1) };
19149 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19150 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
19151 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
19152 X86ISD::LCMPXCHG8_DAG;
19153 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
19154 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
19155 Regs64bit ? X86::RAX : X86::EAX,
19156 HalfT, Result.getValue(1));
19157 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
19158 Regs64bit ? X86::RDX : X86::EDX,
19159 HalfT, cpOutL.getValue(2));
19160 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
19162 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
19163 MVT::i32, cpOutH.getValue(2));
19165 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
19166 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
19167 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
19169 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
19170 Results.push_back(Success);
19171 Results.push_back(EFLAGS.getValue(1));
19174 case ISD::ATOMIC_SWAP:
19175 case ISD::ATOMIC_LOAD_ADD:
19176 case ISD::ATOMIC_LOAD_SUB:
19177 case ISD::ATOMIC_LOAD_AND:
19178 case ISD::ATOMIC_LOAD_OR:
19179 case ISD::ATOMIC_LOAD_XOR:
19180 case ISD::ATOMIC_LOAD_NAND:
19181 case ISD::ATOMIC_LOAD_MIN:
19182 case ISD::ATOMIC_LOAD_MAX:
19183 case ISD::ATOMIC_LOAD_UMIN:
19184 case ISD::ATOMIC_LOAD_UMAX:
19185 case ISD::ATOMIC_LOAD: {
19186 // Delegate to generic TypeLegalization. Situations we can really handle
19187 // should have already been dealt with by AtomicExpandPass.cpp.
19190 case ISD::BITCAST: {
19191 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19192 EVT DstVT = N->getValueType(0);
19193 EVT SrcVT = N->getOperand(0)->getValueType(0);
19195 if (SrcVT != MVT::f64 ||
19196 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19199 unsigned NumElts = DstVT.getVectorNumElements();
19200 EVT SVT = DstVT.getVectorElementType();
19201 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19202 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19203 MVT::v2f64, N->getOperand(0));
19204 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
19206 if (ExperimentalVectorWideningLegalization) {
19207 // If we are legalizing vectors by widening, we already have the desired
19208 // legal vector type, just return it.
19209 Results.push_back(ToVecInt);
19213 SmallVector<SDValue, 8> Elts;
19214 for (unsigned i = 0, e = NumElts; i != e; ++i)
19215 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19216 ToVecInt, DAG.getIntPtrConstant(i, dl)));
19218 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19223 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19224 switch ((X86ISD::NodeType)Opcode) {
19225 case X86ISD::FIRST_NUMBER: break;
19226 case X86ISD::BSF: return "X86ISD::BSF";
19227 case X86ISD::BSR: return "X86ISD::BSR";
19228 case X86ISD::SHLD: return "X86ISD::SHLD";
19229 case X86ISD::SHRD: return "X86ISD::SHRD";
19230 case X86ISD::FAND: return "X86ISD::FAND";
19231 case X86ISD::FANDN: return "X86ISD::FANDN";
19232 case X86ISD::FOR: return "X86ISD::FOR";
19233 case X86ISD::FXOR: return "X86ISD::FXOR";
19234 case X86ISD::FILD: return "X86ISD::FILD";
19235 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
19236 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
19237 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
19238 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
19239 case X86ISD::FLD: return "X86ISD::FLD";
19240 case X86ISD::FST: return "X86ISD::FST";
19241 case X86ISD::CALL: return "X86ISD::CALL";
19242 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
19243 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
19244 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
19245 case X86ISD::BT: return "X86ISD::BT";
19246 case X86ISD::CMP: return "X86ISD::CMP";
19247 case X86ISD::COMI: return "X86ISD::COMI";
19248 case X86ISD::UCOMI: return "X86ISD::UCOMI";
19249 case X86ISD::CMPM: return "X86ISD::CMPM";
19250 case X86ISD::CMPMU: return "X86ISD::CMPMU";
19251 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
19252 case X86ISD::SETCC: return "X86ISD::SETCC";
19253 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
19254 case X86ISD::FSETCC: return "X86ISD::FSETCC";
19255 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
19256 case X86ISD::CMOV: return "X86ISD::CMOV";
19257 case X86ISD::BRCOND: return "X86ISD::BRCOND";
19258 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
19259 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
19260 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
19261 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
19262 case X86ISD::Wrapper: return "X86ISD::Wrapper";
19263 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
19264 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
19265 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
19266 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
19267 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
19268 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
19269 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
19270 case X86ISD::PINSRB: return "X86ISD::PINSRB";
19271 case X86ISD::PINSRW: return "X86ISD::PINSRW";
19272 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
19273 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
19274 case X86ISD::ANDNP: return "X86ISD::ANDNP";
19275 case X86ISD::PSIGN: return "X86ISD::PSIGN";
19276 case X86ISD::BLENDI: return "X86ISD::BLENDI";
19277 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
19278 case X86ISD::ADDUS: return "X86ISD::ADDUS";
19279 case X86ISD::SUBUS: return "X86ISD::SUBUS";
19280 case X86ISD::HADD: return "X86ISD::HADD";
19281 case X86ISD::HSUB: return "X86ISD::HSUB";
19282 case X86ISD::FHADD: return "X86ISD::FHADD";
19283 case X86ISD::FHSUB: return "X86ISD::FHSUB";
19284 case X86ISD::ABS: return "X86ISD::ABS";
19285 case X86ISD::FMAX: return "X86ISD::FMAX";
19286 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
19287 case X86ISD::FMIN: return "X86ISD::FMIN";
19288 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
19289 case X86ISD::FMAXC: return "X86ISD::FMAXC";
19290 case X86ISD::FMINC: return "X86ISD::FMINC";
19291 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
19292 case X86ISD::FRCP: return "X86ISD::FRCP";
19293 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
19294 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
19295 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
19296 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
19297 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
19298 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
19299 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
19300 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
19301 case X86ISD::CATCHRET: return "X86ISD::CATCHRET";
19302 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
19303 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
19304 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
19305 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
19306 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
19307 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
19308 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
19309 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
19310 case X86ISD::VZEXT: return "X86ISD::VZEXT";
19311 case X86ISD::VSEXT: return "X86ISD::VSEXT";
19312 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
19313 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
19314 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
19315 case X86ISD::VINSERT: return "X86ISD::VINSERT";
19316 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
19317 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
19318 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
19319 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
19320 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
19321 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
19322 case X86ISD::VSHL: return "X86ISD::VSHL";
19323 case X86ISD::VSRL: return "X86ISD::VSRL";
19324 case X86ISD::VSRA: return "X86ISD::VSRA";
19325 case X86ISD::VSHLI: return "X86ISD::VSHLI";
19326 case X86ISD::VSRLI: return "X86ISD::VSRLI";
19327 case X86ISD::VSRAI: return "X86ISD::VSRAI";
19328 case X86ISD::CMPP: return "X86ISD::CMPP";
19329 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
19330 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
19331 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
19332 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
19333 case X86ISD::ADD: return "X86ISD::ADD";
19334 case X86ISD::SUB: return "X86ISD::SUB";
19335 case X86ISD::ADC: return "X86ISD::ADC";
19336 case X86ISD::SBB: return "X86ISD::SBB";
19337 case X86ISD::SMUL: return "X86ISD::SMUL";
19338 case X86ISD::UMUL: return "X86ISD::UMUL";
19339 case X86ISD::SMUL8: return "X86ISD::SMUL8";
19340 case X86ISD::UMUL8: return "X86ISD::UMUL8";
19341 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
19342 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
19343 case X86ISD::INC: return "X86ISD::INC";
19344 case X86ISD::DEC: return "X86ISD::DEC";
19345 case X86ISD::OR: return "X86ISD::OR";
19346 case X86ISD::XOR: return "X86ISD::XOR";
19347 case X86ISD::AND: return "X86ISD::AND";
19348 case X86ISD::BEXTR: return "X86ISD::BEXTR";
19349 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
19350 case X86ISD::PTEST: return "X86ISD::PTEST";
19351 case X86ISD::TESTP: return "X86ISD::TESTP";
19352 case X86ISD::TESTM: return "X86ISD::TESTM";
19353 case X86ISD::TESTNM: return "X86ISD::TESTNM";
19354 case X86ISD::KORTEST: return "X86ISD::KORTEST";
19355 case X86ISD::KTEST: return "X86ISD::KTEST";
19356 case X86ISD::PACKSS: return "X86ISD::PACKSS";
19357 case X86ISD::PACKUS: return "X86ISD::PACKUS";
19358 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
19359 case X86ISD::VALIGN: return "X86ISD::VALIGN";
19360 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
19361 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
19362 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
19363 case X86ISD::SHUFP: return "X86ISD::SHUFP";
19364 case X86ISD::SHUF128: return "X86ISD::SHUF128";
19365 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
19366 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
19367 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
19368 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
19369 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
19370 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
19371 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
19372 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
19373 case X86ISD::MOVSD: return "X86ISD::MOVSD";
19374 case X86ISD::MOVSS: return "X86ISD::MOVSS";
19375 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
19376 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
19377 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
19378 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
19379 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
19380 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
19381 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
19382 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
19383 case X86ISD::VPERMV: return "X86ISD::VPERMV";
19384 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
19385 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
19386 case X86ISD::VPERMI: return "X86ISD::VPERMI";
19387 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
19388 case X86ISD::VRANGE: return "X86ISD::VRANGE";
19389 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
19390 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
19391 case X86ISD::PSADBW: return "X86ISD::PSADBW";
19392 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
19393 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
19394 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
19395 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
19396 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
19397 case X86ISD::MFENCE: return "X86ISD::MFENCE";
19398 case X86ISD::SFENCE: return "X86ISD::SFENCE";
19399 case X86ISD::LFENCE: return "X86ISD::LFENCE";
19400 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
19401 case X86ISD::SAHF: return "X86ISD::SAHF";
19402 case X86ISD::RDRAND: return "X86ISD::RDRAND";
19403 case X86ISD::RDSEED: return "X86ISD::RDSEED";
19404 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
19405 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
19406 case X86ISD::FMADD: return "X86ISD::FMADD";
19407 case X86ISD::FMSUB: return "X86ISD::FMSUB";
19408 case X86ISD::FNMADD: return "X86ISD::FNMADD";
19409 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
19410 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
19411 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
19412 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
19413 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
19414 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
19415 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
19416 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
19417 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
19418 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
19419 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
19420 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
19421 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
19422 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
19423 case X86ISD::XTEST: return "X86ISD::XTEST";
19424 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
19425 case X86ISD::EXPAND: return "X86ISD::EXPAND";
19426 case X86ISD::SELECT: return "X86ISD::SELECT";
19427 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
19428 case X86ISD::RCP28: return "X86ISD::RCP28";
19429 case X86ISD::EXP2: return "X86ISD::EXP2";
19430 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
19431 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
19432 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
19433 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
19434 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
19435 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
19436 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
19437 case X86ISD::SCALEF: return "X86ISD::SCALEF";
19438 case X86ISD::ADDS: return "X86ISD::ADDS";
19439 case X86ISD::SUBS: return "X86ISD::SUBS";
19440 case X86ISD::AVG: return "X86ISD::AVG";
19441 case X86ISD::MULHRS: return "X86ISD::MULHRS";
19442 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
19443 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
19444 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
19445 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
19450 // isLegalAddressingMode - Return true if the addressing mode represented
19451 // by AM is legal for this target, for a load/store of the specified type.
19452 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
19453 const AddrMode &AM, Type *Ty,
19454 unsigned AS) const {
19455 // X86 supports extremely general addressing modes.
19456 CodeModel::Model M = getTargetMachine().getCodeModel();
19457 Reloc::Model R = getTargetMachine().getRelocationModel();
19459 // X86 allows a sign-extended 32-bit immediate field as a displacement.
19460 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
19465 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
19467 // If a reference to this global requires an extra load, we can't fold it.
19468 if (isGlobalStubReference(GVFlags))
19471 // If BaseGV requires a register for the PIC base, we cannot also have a
19472 // BaseReg specified.
19473 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
19476 // If lower 4G is not available, then we must use rip-relative addressing.
19477 if ((M != CodeModel::Small || R != Reloc::Static) &&
19478 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
19482 switch (AM.Scale) {
19488 // These scales always work.
19493 // These scales are formed with basereg+scalereg. Only accept if there is
19498 default: // Other stuff never works.
19505 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
19506 unsigned Bits = Ty->getScalarSizeInBits();
19508 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
19509 // particularly cheaper than those without.
19513 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
19514 // variable shifts just as cheap as scalar ones.
19515 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
19518 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
19519 // fully general vector.
19523 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
19524 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19526 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
19527 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
19528 return NumBits1 > NumBits2;
19531 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
19532 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19535 if (!isTypeLegal(EVT::getEVT(Ty1)))
19538 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
19540 // Assuming the caller doesn't have a zeroext or signext return parameter,
19541 // truncation all the way down to i1 is valid.
19545 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
19546 return isInt<32>(Imm);
19549 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
19550 // Can also use sub to handle negated immediates.
19551 return isInt<32>(Imm);
19554 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
19555 if (!VT1.isInteger() || !VT2.isInteger())
19557 unsigned NumBits1 = VT1.getSizeInBits();
19558 unsigned NumBits2 = VT2.getSizeInBits();
19559 return NumBits1 > NumBits2;
19562 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
19563 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19564 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
19567 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
19568 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19569 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
19572 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
19573 EVT VT1 = Val.getValueType();
19574 if (isZExtFree(VT1, VT2))
19577 if (Val.getOpcode() != ISD::LOAD)
19580 if (!VT1.isSimple() || !VT1.isInteger() ||
19581 !VT2.isSimple() || !VT2.isInteger())
19584 switch (VT1.getSimpleVT().SimpleTy) {
19589 // X86 has 8, 16, and 32-bit zero-extending loads.
19596 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
19599 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
19600 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
19603 VT = VT.getScalarType();
19605 if (!VT.isSimple())
19608 switch (VT.getSimpleVT().SimpleTy) {
19619 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
19620 // i16 instructions are longer (0x66 prefix) and potentially slower.
19621 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
19624 /// isShuffleMaskLegal - Targets can use this to indicate that they only
19625 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
19626 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
19627 /// are assumed to be legal.
19629 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
19631 if (!VT.isSimple())
19634 // Not for i1 vectors
19635 if (VT.getScalarType() == MVT::i1)
19638 // Very little shuffling can be done for 64-bit vectors right now.
19639 if (VT.getSizeInBits() == 64)
19642 // We only care that the types being shuffled are legal. The lowering can
19643 // handle any possible shuffle mask that results.
19644 return isTypeLegal(VT.getSimpleVT());
19648 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
19650 // Just delegate to the generic legality, clear masks aren't special.
19651 return isShuffleMaskLegal(Mask, VT);
19654 //===----------------------------------------------------------------------===//
19655 // X86 Scheduler Hooks
19656 //===----------------------------------------------------------------------===//
19658 /// Utility function to emit xbegin specifying the start of an RTM region.
19659 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
19660 const TargetInstrInfo *TII) {
19661 DebugLoc DL = MI->getDebugLoc();
19663 const BasicBlock *BB = MBB->getBasicBlock();
19664 MachineFunction::iterator I = MBB;
19667 // For the v = xbegin(), we generate
19678 MachineBasicBlock *thisMBB = MBB;
19679 MachineFunction *MF = MBB->getParent();
19680 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
19681 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
19682 MF->insert(I, mainMBB);
19683 MF->insert(I, sinkMBB);
19685 // Transfer the remainder of BB and its successor edges to sinkMBB.
19686 sinkMBB->splice(sinkMBB->begin(), MBB,
19687 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19688 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
19692 // # fallthrough to mainMBB
19693 // # abortion to sinkMBB
19694 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
19695 thisMBB->addSuccessor(mainMBB);
19696 thisMBB->addSuccessor(sinkMBB);
19700 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
19701 mainMBB->addSuccessor(sinkMBB);
19704 // EAX is live into the sinkMBB
19705 sinkMBB->addLiveIn(X86::EAX);
19706 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19707 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19710 MI->eraseFromParent();
19714 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
19715 // or XMM0_V32I8 in AVX all of this code can be replaced with that
19716 // in the .td file.
19717 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
19718 const TargetInstrInfo *TII) {
19720 switch (MI->getOpcode()) {
19721 default: llvm_unreachable("illegal opcode!");
19722 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
19723 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
19724 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
19725 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
19726 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
19727 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
19728 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
19729 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
19732 DebugLoc dl = MI->getDebugLoc();
19733 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
19735 unsigned NumArgs = MI->getNumOperands();
19736 for (unsigned i = 1; i < NumArgs; ++i) {
19737 MachineOperand &Op = MI->getOperand(i);
19738 if (!(Op.isReg() && Op.isImplicit()))
19739 MIB.addOperand(Op);
19741 if (MI->hasOneMemOperand())
19742 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
19744 BuildMI(*BB, MI, dl,
19745 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19746 .addReg(X86::XMM0);
19748 MI->eraseFromParent();
19752 // FIXME: Custom handling because TableGen doesn't support multiple implicit
19753 // defs in an instruction pattern
19754 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
19755 const TargetInstrInfo *TII) {
19757 switch (MI->getOpcode()) {
19758 default: llvm_unreachable("illegal opcode!");
19759 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
19760 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
19761 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
19762 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
19763 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
19764 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
19765 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
19766 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
19769 DebugLoc dl = MI->getDebugLoc();
19770 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
19772 unsigned NumArgs = MI->getNumOperands(); // remove the results
19773 for (unsigned i = 1; i < NumArgs; ++i) {
19774 MachineOperand &Op = MI->getOperand(i);
19775 if (!(Op.isReg() && Op.isImplicit()))
19776 MIB.addOperand(Op);
19778 if (MI->hasOneMemOperand())
19779 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
19781 BuildMI(*BB, MI, dl,
19782 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19785 MI->eraseFromParent();
19789 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
19790 const X86Subtarget *Subtarget) {
19791 DebugLoc dl = MI->getDebugLoc();
19792 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19793 // Address into RAX/EAX, other two args into ECX, EDX.
19794 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
19795 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
19796 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
19797 for (int i = 0; i < X86::AddrNumOperands; ++i)
19798 MIB.addOperand(MI->getOperand(i));
19800 unsigned ValOps = X86::AddrNumOperands;
19801 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
19802 .addReg(MI->getOperand(ValOps).getReg());
19803 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
19804 .addReg(MI->getOperand(ValOps+1).getReg());
19806 // The instruction doesn't actually take any operands though.
19807 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
19809 MI->eraseFromParent(); // The pseudo is gone now.
19813 MachineBasicBlock *
19814 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
19815 MachineBasicBlock *MBB) const {
19816 // Emit va_arg instruction on X86-64.
19818 // Operands to this pseudo-instruction:
19819 // 0 ) Output : destination address (reg)
19820 // 1-5) Input : va_list address (addr, i64mem)
19821 // 6 ) ArgSize : Size (in bytes) of vararg type
19822 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
19823 // 8 ) Align : Alignment of type
19824 // 9 ) EFLAGS (implicit-def)
19826 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
19827 static_assert(X86::AddrNumOperands == 5,
19828 "VAARG_64 assumes 5 address operands");
19830 unsigned DestReg = MI->getOperand(0).getReg();
19831 MachineOperand &Base = MI->getOperand(1);
19832 MachineOperand &Scale = MI->getOperand(2);
19833 MachineOperand &Index = MI->getOperand(3);
19834 MachineOperand &Disp = MI->getOperand(4);
19835 MachineOperand &Segment = MI->getOperand(5);
19836 unsigned ArgSize = MI->getOperand(6).getImm();
19837 unsigned ArgMode = MI->getOperand(7).getImm();
19838 unsigned Align = MI->getOperand(8).getImm();
19840 // Memory Reference
19841 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
19842 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19843 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19845 // Machine Information
19846 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19847 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
19848 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
19849 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
19850 DebugLoc DL = MI->getDebugLoc();
19852 // struct va_list {
19855 // i64 overflow_area (address)
19856 // i64 reg_save_area (address)
19858 // sizeof(va_list) = 24
19859 // alignment(va_list) = 8
19861 unsigned TotalNumIntRegs = 6;
19862 unsigned TotalNumXMMRegs = 8;
19863 bool UseGPOffset = (ArgMode == 1);
19864 bool UseFPOffset = (ArgMode == 2);
19865 unsigned MaxOffset = TotalNumIntRegs * 8 +
19866 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
19868 /* Align ArgSize to a multiple of 8 */
19869 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
19870 bool NeedsAlign = (Align > 8);
19872 MachineBasicBlock *thisMBB = MBB;
19873 MachineBasicBlock *overflowMBB;
19874 MachineBasicBlock *offsetMBB;
19875 MachineBasicBlock *endMBB;
19877 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
19878 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
19879 unsigned OffsetReg = 0;
19881 if (!UseGPOffset && !UseFPOffset) {
19882 // If we only pull from the overflow region, we don't create a branch.
19883 // We don't need to alter control flow.
19884 OffsetDestReg = 0; // unused
19885 OverflowDestReg = DestReg;
19887 offsetMBB = nullptr;
19888 overflowMBB = thisMBB;
19891 // First emit code to check if gp_offset (or fp_offset) is below the bound.
19892 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
19893 // If not, pull from overflow_area. (branch to overflowMBB)
19898 // offsetMBB overflowMBB
19903 // Registers for the PHI in endMBB
19904 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
19905 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
19907 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
19908 MachineFunction *MF = MBB->getParent();
19909 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19910 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19911 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19913 MachineFunction::iterator MBBIter = MBB;
19916 // Insert the new basic blocks
19917 MF->insert(MBBIter, offsetMBB);
19918 MF->insert(MBBIter, overflowMBB);
19919 MF->insert(MBBIter, endMBB);
19921 // Transfer the remainder of MBB and its successor edges to endMBB.
19922 endMBB->splice(endMBB->begin(), thisMBB,
19923 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
19924 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
19926 // Make offsetMBB and overflowMBB successors of thisMBB
19927 thisMBB->addSuccessor(offsetMBB);
19928 thisMBB->addSuccessor(overflowMBB);
19930 // endMBB is a successor of both offsetMBB and overflowMBB
19931 offsetMBB->addSuccessor(endMBB);
19932 overflowMBB->addSuccessor(endMBB);
19934 // Load the offset value into a register
19935 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
19936 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
19940 .addDisp(Disp, UseFPOffset ? 4 : 0)
19941 .addOperand(Segment)
19942 .setMemRefs(MMOBegin, MMOEnd);
19944 // Check if there is enough room left to pull this argument.
19945 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
19947 .addImm(MaxOffset + 8 - ArgSizeA8);
19949 // Branch to "overflowMBB" if offset >= max
19950 // Fall through to "offsetMBB" otherwise
19951 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
19952 .addMBB(overflowMBB);
19955 // In offsetMBB, emit code to use the reg_save_area.
19957 assert(OffsetReg != 0);
19959 // Read the reg_save_area address.
19960 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
19961 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
19966 .addOperand(Segment)
19967 .setMemRefs(MMOBegin, MMOEnd);
19969 // Zero-extend the offset
19970 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
19971 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
19974 .addImm(X86::sub_32bit);
19976 // Add the offset to the reg_save_area to get the final address.
19977 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
19978 .addReg(OffsetReg64)
19979 .addReg(RegSaveReg);
19981 // Compute the offset for the next argument
19982 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
19983 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
19985 .addImm(UseFPOffset ? 16 : 8);
19987 // Store it back into the va_list.
19988 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
19992 .addDisp(Disp, UseFPOffset ? 4 : 0)
19993 .addOperand(Segment)
19994 .addReg(NextOffsetReg)
19995 .setMemRefs(MMOBegin, MMOEnd);
19998 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
20003 // Emit code to use overflow area
20006 // Load the overflow_area address into a register.
20007 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
20008 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
20013 .addOperand(Segment)
20014 .setMemRefs(MMOBegin, MMOEnd);
20016 // If we need to align it, do so. Otherwise, just copy the address
20017 // to OverflowDestReg.
20019 // Align the overflow address
20020 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
20021 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
20023 // aligned_addr = (addr + (align-1)) & ~(align-1)
20024 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
20025 .addReg(OverflowAddrReg)
20028 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
20030 .addImm(~(uint64_t)(Align-1));
20032 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
20033 .addReg(OverflowAddrReg);
20036 // Compute the next overflow address after this argument.
20037 // (the overflow address should be kept 8-byte aligned)
20038 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
20039 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
20040 .addReg(OverflowDestReg)
20041 .addImm(ArgSizeA8);
20043 // Store the new overflow address.
20044 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
20049 .addOperand(Segment)
20050 .addReg(NextAddrReg)
20051 .setMemRefs(MMOBegin, MMOEnd);
20053 // If we branched, emit the PHI to the front of endMBB.
20055 BuildMI(*endMBB, endMBB->begin(), DL,
20056 TII->get(X86::PHI), DestReg)
20057 .addReg(OffsetDestReg).addMBB(offsetMBB)
20058 .addReg(OverflowDestReg).addMBB(overflowMBB);
20061 // Erase the pseudo instruction
20062 MI->eraseFromParent();
20067 MachineBasicBlock *
20068 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
20070 MachineBasicBlock *MBB) const {
20071 // Emit code to save XMM registers to the stack. The ABI says that the
20072 // number of registers to save is given in %al, so it's theoretically
20073 // possible to do an indirect jump trick to avoid saving all of them,
20074 // however this code takes a simpler approach and just executes all
20075 // of the stores if %al is non-zero. It's less code, and it's probably
20076 // easier on the hardware branch predictor, and stores aren't all that
20077 // expensive anyway.
20079 // Create the new basic blocks. One block contains all the XMM stores,
20080 // and one block is the final destination regardless of whether any
20081 // stores were performed.
20082 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20083 MachineFunction *F = MBB->getParent();
20084 MachineFunction::iterator MBBIter = MBB;
20086 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
20087 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
20088 F->insert(MBBIter, XMMSaveMBB);
20089 F->insert(MBBIter, EndMBB);
20091 // Transfer the remainder of MBB and its successor edges to EndMBB.
20092 EndMBB->splice(EndMBB->begin(), MBB,
20093 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20094 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
20096 // The original block will now fall through to the XMM save block.
20097 MBB->addSuccessor(XMMSaveMBB);
20098 // The XMMSaveMBB will fall through to the end block.
20099 XMMSaveMBB->addSuccessor(EndMBB);
20101 // Now add the instructions.
20102 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20103 DebugLoc DL = MI->getDebugLoc();
20105 unsigned CountReg = MI->getOperand(0).getReg();
20106 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
20107 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
20109 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
20110 // If %al is 0, branch around the XMM save block.
20111 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
20112 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
20113 MBB->addSuccessor(EndMBB);
20116 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
20117 // that was just emitted, but clearly shouldn't be "saved".
20118 assert((MI->getNumOperands() <= 3 ||
20119 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
20120 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
20121 && "Expected last argument to be EFLAGS");
20122 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
20123 // In the XMM save block, save all the XMM argument registers.
20124 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
20125 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
20126 MachineMemOperand *MMO = F->getMachineMemOperand(
20127 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
20128 MachineMemOperand::MOStore,
20129 /*Size=*/16, /*Align=*/16);
20130 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
20131 .addFrameIndex(RegSaveFrameIndex)
20132 .addImm(/*Scale=*/1)
20133 .addReg(/*IndexReg=*/0)
20134 .addImm(/*Disp=*/Offset)
20135 .addReg(/*Segment=*/0)
20136 .addReg(MI->getOperand(i).getReg())
20137 .addMemOperand(MMO);
20140 MI->eraseFromParent(); // The pseudo instruction is gone now.
20145 // The EFLAGS operand of SelectItr might be missing a kill marker
20146 // because there were multiple uses of EFLAGS, and ISel didn't know
20147 // which to mark. Figure out whether SelectItr should have had a
20148 // kill marker, and set it if it should. Returns the correct kill
20150 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
20151 MachineBasicBlock* BB,
20152 const TargetRegisterInfo* TRI) {
20153 // Scan forward through BB for a use/def of EFLAGS.
20154 MachineBasicBlock::iterator miI(std::next(SelectItr));
20155 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
20156 const MachineInstr& mi = *miI;
20157 if (mi.readsRegister(X86::EFLAGS))
20159 if (mi.definesRegister(X86::EFLAGS))
20160 break; // Should have kill-flag - update below.
20163 // If we hit the end of the block, check whether EFLAGS is live into a
20165 if (miI == BB->end()) {
20166 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
20167 sEnd = BB->succ_end();
20168 sItr != sEnd; ++sItr) {
20169 MachineBasicBlock* succ = *sItr;
20170 if (succ->isLiveIn(X86::EFLAGS))
20175 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
20176 // out. SelectMI should have a kill flag on EFLAGS.
20177 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
20181 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
20182 // together with other CMOV pseudo-opcodes into a single basic-block with
20183 // conditional jump around it.
20184 static bool isCMOVPseudo(MachineInstr *MI) {
20185 switch (MI->getOpcode()) {
20186 case X86::CMOV_FR32:
20187 case X86::CMOV_FR64:
20188 case X86::CMOV_GR8:
20189 case X86::CMOV_GR16:
20190 case X86::CMOV_GR32:
20191 case X86::CMOV_RFP32:
20192 case X86::CMOV_RFP64:
20193 case X86::CMOV_RFP80:
20194 case X86::CMOV_V2F64:
20195 case X86::CMOV_V2I64:
20196 case X86::CMOV_V4F32:
20197 case X86::CMOV_V4F64:
20198 case X86::CMOV_V4I64:
20199 case X86::CMOV_V16F32:
20200 case X86::CMOV_V8F32:
20201 case X86::CMOV_V8F64:
20202 case X86::CMOV_V8I64:
20203 case X86::CMOV_V8I1:
20204 case X86::CMOV_V16I1:
20205 case X86::CMOV_V32I1:
20206 case X86::CMOV_V64I1:
20214 MachineBasicBlock *
20215 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20216 MachineBasicBlock *BB) const {
20217 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20218 DebugLoc DL = MI->getDebugLoc();
20220 // To "insert" a SELECT_CC instruction, we actually have to insert the
20221 // diamond control-flow pattern. The incoming instruction knows the
20222 // destination vreg to set, the condition code register to branch on, the
20223 // true/false values to select between, and a branch opcode to use.
20224 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20225 MachineFunction::iterator It = BB;
20231 // cmpTY ccX, r1, r2
20233 // fallthrough --> copy0MBB
20234 MachineBasicBlock *thisMBB = BB;
20235 MachineFunction *F = BB->getParent();
20237 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
20238 // as described above, by inserting a BB, and then making a PHI at the join
20239 // point to select the true and false operands of the CMOV in the PHI.
20241 // The code also handles two different cases of multiple CMOV opcodes
20245 // In this case, there are multiple CMOVs in a row, all which are based on
20246 // the same condition setting (or the exact opposite condition setting).
20247 // In this case we can lower all the CMOVs using a single inserted BB, and
20248 // then make a number of PHIs at the join point to model the CMOVs. The only
20249 // trickiness here, is that in a case like:
20251 // t2 = CMOV cond1 t1, f1
20252 // t3 = CMOV cond1 t2, f2
20254 // when rewriting this into PHIs, we have to perform some renaming on the
20255 // temps since you cannot have a PHI operand refer to a PHI result earlier
20256 // in the same block. The "simple" but wrong lowering would be:
20258 // t2 = PHI t1(BB1), f1(BB2)
20259 // t3 = PHI t2(BB1), f2(BB2)
20261 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
20262 // renaming is to note that on the path through BB1, t2 is really just a
20263 // copy of t1, and do that renaming, properly generating:
20265 // t2 = PHI t1(BB1), f1(BB2)
20266 // t3 = PHI t1(BB1), f2(BB2)
20268 // Case 2, we lower cascaded CMOVs such as
20270 // (CMOV (CMOV F, T, cc1), T, cc2)
20272 // to two successives branches. For that, we look for another CMOV as the
20273 // following instruction.
20275 // Without this, we would add a PHI between the two jumps, which ends up
20276 // creating a few copies all around. For instance, for
20278 // (sitofp (zext (fcmp une)))
20280 // we would generate:
20282 // ucomiss %xmm1, %xmm0
20283 // movss <1.0f>, %xmm0
20284 // movaps %xmm0, %xmm1
20286 // xorps %xmm1, %xmm1
20289 // movaps %xmm1, %xmm0
20293 // because this custom-inserter would have generated:
20305 // A: X = ...; Y = ...
20307 // C: Z = PHI [X, A], [Y, B]
20309 // E: PHI [X, C], [Z, D]
20311 // If we lower both CMOVs in a single step, we can instead generate:
20323 // A: X = ...; Y = ...
20325 // E: PHI [X, A], [X, C], [Y, D]
20327 // Which, in our sitofp/fcmp example, gives us something like:
20329 // ucomiss %xmm1, %xmm0
20330 // movss <1.0f>, %xmm0
20333 // xorps %xmm0, %xmm0
20337 MachineInstr *CascadedCMOV = nullptr;
20338 MachineInstr *LastCMOV = MI;
20339 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
20340 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
20341 MachineBasicBlock::iterator NextMIIt =
20342 std::next(MachineBasicBlock::iterator(MI));
20344 // Check for case 1, where there are multiple CMOVs with the same condition
20345 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
20346 // number of jumps the most.
20348 if (isCMOVPseudo(MI)) {
20349 // See if we have a string of CMOVS with the same condition.
20350 while (NextMIIt != BB->end() &&
20351 isCMOVPseudo(NextMIIt) &&
20352 (NextMIIt->getOperand(3).getImm() == CC ||
20353 NextMIIt->getOperand(3).getImm() == OppCC)) {
20354 LastCMOV = &*NextMIIt;
20359 // This checks for case 2, but only do this if we didn't already find
20360 // case 1, as indicated by LastCMOV == MI.
20361 if (LastCMOV == MI &&
20362 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
20363 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
20364 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
20365 CascadedCMOV = &*NextMIIt;
20368 MachineBasicBlock *jcc1MBB = nullptr;
20370 // If we have a cascaded CMOV, we lower it to two successive branches to
20371 // the same block. EFLAGS is used by both, so mark it as live in the second.
20372 if (CascadedCMOV) {
20373 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
20374 F->insert(It, jcc1MBB);
20375 jcc1MBB->addLiveIn(X86::EFLAGS);
20378 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
20379 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
20380 F->insert(It, copy0MBB);
20381 F->insert(It, sinkMBB);
20383 // If the EFLAGS register isn't dead in the terminator, then claim that it's
20384 // live into the sink and copy blocks.
20385 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
20387 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
20388 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
20389 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
20390 copy0MBB->addLiveIn(X86::EFLAGS);
20391 sinkMBB->addLiveIn(X86::EFLAGS);
20394 // Transfer the remainder of BB and its successor edges to sinkMBB.
20395 sinkMBB->splice(sinkMBB->begin(), BB,
20396 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
20397 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
20399 // Add the true and fallthrough blocks as its successors.
20400 if (CascadedCMOV) {
20401 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
20402 BB->addSuccessor(jcc1MBB);
20404 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
20405 // jump to the sinkMBB.
20406 jcc1MBB->addSuccessor(copy0MBB);
20407 jcc1MBB->addSuccessor(sinkMBB);
20409 BB->addSuccessor(copy0MBB);
20412 // The true block target of the first (or only) branch is always sinkMBB.
20413 BB->addSuccessor(sinkMBB);
20415 // Create the conditional branch instruction.
20416 unsigned Opc = X86::GetCondBranchFromCond(CC);
20417 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
20419 if (CascadedCMOV) {
20420 unsigned Opc2 = X86::GetCondBranchFromCond(
20421 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
20422 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
20426 // %FalseValue = ...
20427 // # fallthrough to sinkMBB
20428 copy0MBB->addSuccessor(sinkMBB);
20431 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
20433 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
20434 MachineBasicBlock::iterator MIItEnd =
20435 std::next(MachineBasicBlock::iterator(LastCMOV));
20436 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
20437 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
20438 MachineInstrBuilder MIB;
20440 // As we are creating the PHIs, we have to be careful if there is more than
20441 // one. Later CMOVs may reference the results of earlier CMOVs, but later
20442 // PHIs have to reference the individual true/false inputs from earlier PHIs.
20443 // That also means that PHI construction must work forward from earlier to
20444 // later, and that the code must maintain a mapping from earlier PHI's
20445 // destination registers, and the registers that went into the PHI.
20447 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
20448 unsigned DestReg = MIIt->getOperand(0).getReg();
20449 unsigned Op1Reg = MIIt->getOperand(1).getReg();
20450 unsigned Op2Reg = MIIt->getOperand(2).getReg();
20452 // If this CMOV we are generating is the opposite condition from
20453 // the jump we generated, then we have to swap the operands for the
20454 // PHI that is going to be generated.
20455 if (MIIt->getOperand(3).getImm() == OppCC)
20456 std::swap(Op1Reg, Op2Reg);
20458 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
20459 Op1Reg = RegRewriteTable[Op1Reg].first;
20461 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
20462 Op2Reg = RegRewriteTable[Op2Reg].second;
20464 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
20465 TII->get(X86::PHI), DestReg)
20466 .addReg(Op1Reg).addMBB(copy0MBB)
20467 .addReg(Op2Reg).addMBB(thisMBB);
20469 // Add this PHI to the rewrite table.
20470 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
20473 // If we have a cascaded CMOV, the second Jcc provides the same incoming
20474 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
20475 if (CascadedCMOV) {
20476 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
20477 // Copy the PHI result to the register defined by the second CMOV.
20478 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
20479 DL, TII->get(TargetOpcode::COPY),
20480 CascadedCMOV->getOperand(0).getReg())
20481 .addReg(MI->getOperand(0).getReg());
20482 CascadedCMOV->eraseFromParent();
20485 // Now remove the CMOV(s).
20486 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
20487 (MIIt++)->eraseFromParent();
20492 MachineBasicBlock *
20493 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
20494 MachineBasicBlock *BB) const {
20495 // Combine the following atomic floating-point modification pattern:
20496 // a.store(reg OP a.load(acquire), release)
20497 // Transform them into:
20498 // OPss (%gpr), %xmm
20499 // movss %xmm, (%gpr)
20500 // Or sd equivalent for 64-bit operations.
20502 switch (MI->getOpcode()) {
20503 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
20504 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
20505 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
20507 const X86InstrInfo *TII = Subtarget->getInstrInfo();
20508 DebugLoc DL = MI->getDebugLoc();
20509 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
20510 unsigned MSrc = MI->getOperand(0).getReg();
20511 unsigned VSrc = MI->getOperand(5).getReg();
20512 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
20513 .addReg(/*Base=*/MSrc)
20514 .addImm(/*Scale=*/1)
20515 .addReg(/*Index=*/0)
20518 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
20519 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
20521 .addReg(/*Base=*/MSrc)
20522 .addImm(/*Scale=*/1)
20523 .addReg(/*Index=*/0)
20524 .addImm(/*Disp=*/0)
20525 .addReg(/*Segment=*/0);
20526 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
20527 MI->eraseFromParent(); // The pseudo instruction is gone now.
20531 MachineBasicBlock *
20532 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
20533 MachineBasicBlock *BB) const {
20534 MachineFunction *MF = BB->getParent();
20535 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20536 DebugLoc DL = MI->getDebugLoc();
20537 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20539 assert(MF->shouldSplitStack());
20541 const bool Is64Bit = Subtarget->is64Bit();
20542 const bool IsLP64 = Subtarget->isTarget64BitLP64();
20544 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
20545 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
20548 // ... [Till the alloca]
20549 // If stacklet is not large enough, jump to mallocMBB
20552 // Allocate by subtracting from RSP
20553 // Jump to continueMBB
20556 // Allocate by call to runtime
20560 // [rest of original BB]
20563 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20564 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20565 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20567 MachineRegisterInfo &MRI = MF->getRegInfo();
20568 const TargetRegisterClass *AddrRegClass =
20569 getRegClassFor(getPointerTy(MF->getDataLayout()));
20571 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20572 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20573 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
20574 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
20575 sizeVReg = MI->getOperand(1).getReg(),
20576 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
20578 MachineFunction::iterator MBBIter = BB;
20581 MF->insert(MBBIter, bumpMBB);
20582 MF->insert(MBBIter, mallocMBB);
20583 MF->insert(MBBIter, continueMBB);
20585 continueMBB->splice(continueMBB->begin(), BB,
20586 std::next(MachineBasicBlock::iterator(MI)), BB->end());
20587 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
20589 // Add code to the main basic block to check if the stack limit has been hit,
20590 // and if so, jump to mallocMBB otherwise to bumpMBB.
20591 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
20592 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
20593 .addReg(tmpSPVReg).addReg(sizeVReg);
20594 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
20595 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
20596 .addReg(SPLimitVReg);
20597 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
20599 // bumpMBB simply decreases the stack pointer, since we know the current
20600 // stacklet has enough space.
20601 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
20602 .addReg(SPLimitVReg);
20603 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
20604 .addReg(SPLimitVReg);
20605 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20607 // Calls into a routine in libgcc to allocate more space from the heap.
20608 const uint32_t *RegMask =
20609 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
20611 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
20613 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20614 .addExternalSymbol("__morestack_allocate_stack_space")
20615 .addRegMask(RegMask)
20616 .addReg(X86::RDI, RegState::Implicit)
20617 .addReg(X86::RAX, RegState::ImplicitDefine);
20618 } else if (Is64Bit) {
20619 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
20621 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20622 .addExternalSymbol("__morestack_allocate_stack_space")
20623 .addRegMask(RegMask)
20624 .addReg(X86::EDI, RegState::Implicit)
20625 .addReg(X86::EAX, RegState::ImplicitDefine);
20627 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
20629 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
20630 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
20631 .addExternalSymbol("__morestack_allocate_stack_space")
20632 .addRegMask(RegMask)
20633 .addReg(X86::EAX, RegState::ImplicitDefine);
20637 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
20640 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
20641 .addReg(IsLP64 ? X86::RAX : X86::EAX);
20642 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20644 // Set up the CFG correctly.
20645 BB->addSuccessor(bumpMBB);
20646 BB->addSuccessor(mallocMBB);
20647 mallocMBB->addSuccessor(continueMBB);
20648 bumpMBB->addSuccessor(continueMBB);
20650 // Take care of the PHI nodes.
20651 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
20652 MI->getOperand(0).getReg())
20653 .addReg(mallocPtrVReg).addMBB(mallocMBB)
20654 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
20656 // Delete the original pseudo instruction.
20657 MI->eraseFromParent();
20660 return continueMBB;
20663 MachineBasicBlock *
20664 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
20665 MachineBasicBlock *BB) const {
20666 DebugLoc DL = MI->getDebugLoc();
20668 assert(!Subtarget->isTargetMachO());
20670 Subtarget->getFrameLowering()->emitStackProbeCall(*BB->getParent(), *BB, MI,
20673 MI->eraseFromParent(); // The pseudo instruction is gone now.
20677 MachineBasicBlock *
20678 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
20679 MachineBasicBlock *BB) const {
20680 // This is pretty easy. We're taking the value that we received from
20681 // our load from the relocation, sticking it in either RDI (x86-64)
20682 // or EAX and doing an indirect call. The return value will then
20683 // be in the normal return register.
20684 MachineFunction *F = BB->getParent();
20685 const X86InstrInfo *TII = Subtarget->getInstrInfo();
20686 DebugLoc DL = MI->getDebugLoc();
20688 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
20689 assert(MI->getOperand(3).isGlobal() && "This should be a global");
20691 // Get a register mask for the lowered call.
20692 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
20693 // proper register mask.
20694 const uint32_t *RegMask =
20695 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
20696 if (Subtarget->is64Bit()) {
20697 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20698 TII->get(X86::MOV64rm), X86::RDI)
20700 .addImm(0).addReg(0)
20701 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20702 MI->getOperand(3).getTargetFlags())
20704 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
20705 addDirectMem(MIB, X86::RDI);
20706 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
20707 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
20708 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20709 TII->get(X86::MOV32rm), X86::EAX)
20711 .addImm(0).addReg(0)
20712 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20713 MI->getOperand(3).getTargetFlags())
20715 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20716 addDirectMem(MIB, X86::EAX);
20717 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20719 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20720 TII->get(X86::MOV32rm), X86::EAX)
20721 .addReg(TII->getGlobalBaseReg(F))
20722 .addImm(0).addReg(0)
20723 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20724 MI->getOperand(3).getTargetFlags())
20726 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20727 addDirectMem(MIB, X86::EAX);
20728 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20731 MI->eraseFromParent(); // The pseudo instruction is gone now.
20735 MachineBasicBlock *
20736 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
20737 MachineBasicBlock *MBB) const {
20738 DebugLoc DL = MI->getDebugLoc();
20739 MachineFunction *MF = MBB->getParent();
20740 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20741 MachineRegisterInfo &MRI = MF->getRegInfo();
20743 const BasicBlock *BB = MBB->getBasicBlock();
20744 MachineFunction::iterator I = MBB;
20747 // Memory Reference
20748 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20749 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20752 unsigned MemOpndSlot = 0;
20754 unsigned CurOp = 0;
20756 DstReg = MI->getOperand(CurOp++).getReg();
20757 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
20758 assert(RC->hasType(MVT::i32) && "Invalid destination!");
20759 unsigned mainDstReg = MRI.createVirtualRegister(RC);
20760 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
20762 MemOpndSlot = CurOp;
20764 MVT PVT = getPointerTy(MF->getDataLayout());
20765 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
20766 "Invalid Pointer Size!");
20768 // For v = setjmp(buf), we generate
20771 // buf[LabelOffset] = restoreMBB
20772 // SjLjSetup restoreMBB
20778 // v = phi(main, restore)
20781 // if base pointer being used, load it from frame
20784 MachineBasicBlock *thisMBB = MBB;
20785 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20786 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20787 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
20788 MF->insert(I, mainMBB);
20789 MF->insert(I, sinkMBB);
20790 MF->push_back(restoreMBB);
20792 MachineInstrBuilder MIB;
20794 // Transfer the remainder of BB and its successor edges to sinkMBB.
20795 sinkMBB->splice(sinkMBB->begin(), MBB,
20796 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20797 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20800 unsigned PtrStoreOpc = 0;
20801 unsigned LabelReg = 0;
20802 const int64_t LabelOffset = 1 * PVT.getStoreSize();
20803 Reloc::Model RM = MF->getTarget().getRelocationModel();
20804 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
20805 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
20807 // Prepare IP either in reg or imm.
20808 if (!UseImmLabel) {
20809 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
20810 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
20811 LabelReg = MRI.createVirtualRegister(PtrRC);
20812 if (Subtarget->is64Bit()) {
20813 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
20817 .addMBB(restoreMBB)
20820 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
20821 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
20822 .addReg(XII->getGlobalBaseReg(MF))
20825 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
20829 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
20831 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
20832 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20833 if (i == X86::AddrDisp)
20834 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
20836 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
20839 MIB.addReg(LabelReg);
20841 MIB.addMBB(restoreMBB);
20842 MIB.setMemRefs(MMOBegin, MMOEnd);
20844 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
20845 .addMBB(restoreMBB);
20847 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
20848 MIB.addRegMask(RegInfo->getNoPreservedMask());
20849 thisMBB->addSuccessor(mainMBB);
20850 thisMBB->addSuccessor(restoreMBB);
20854 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
20855 mainMBB->addSuccessor(sinkMBB);
20858 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20859 TII->get(X86::PHI), DstReg)
20860 .addReg(mainDstReg).addMBB(mainMBB)
20861 .addReg(restoreDstReg).addMBB(restoreMBB);
20864 if (RegInfo->hasBasePointer(*MF)) {
20865 const bool Uses64BitFramePtr =
20866 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
20867 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
20868 X86FI->setRestoreBasePointer(MF);
20869 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
20870 unsigned BasePtr = RegInfo->getBaseRegister();
20871 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
20872 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
20873 FramePtr, true, X86FI->getRestoreBasePointerOffset())
20874 .setMIFlag(MachineInstr::FrameSetup);
20876 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
20877 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
20878 restoreMBB->addSuccessor(sinkMBB);
20880 MI->eraseFromParent();
20884 MachineBasicBlock *
20885 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
20886 MachineBasicBlock *MBB) const {
20887 DebugLoc DL = MI->getDebugLoc();
20888 MachineFunction *MF = MBB->getParent();
20889 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20890 MachineRegisterInfo &MRI = MF->getRegInfo();
20892 // Memory Reference
20893 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20894 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20896 MVT PVT = getPointerTy(MF->getDataLayout());
20897 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
20898 "Invalid Pointer Size!");
20900 const TargetRegisterClass *RC =
20901 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
20902 unsigned Tmp = MRI.createVirtualRegister(RC);
20903 // Since FP is only updated here but NOT referenced, it's treated as GPR.
20904 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
20905 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
20906 unsigned SP = RegInfo->getStackRegister();
20908 MachineInstrBuilder MIB;
20910 const int64_t LabelOffset = 1 * PVT.getStoreSize();
20911 const int64_t SPOffset = 2 * PVT.getStoreSize();
20913 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
20914 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
20917 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
20918 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
20919 MIB.addOperand(MI->getOperand(i));
20920 MIB.setMemRefs(MMOBegin, MMOEnd);
20922 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
20923 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20924 if (i == X86::AddrDisp)
20925 MIB.addDisp(MI->getOperand(i), LabelOffset);
20927 MIB.addOperand(MI->getOperand(i));
20929 MIB.setMemRefs(MMOBegin, MMOEnd);
20931 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
20932 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20933 if (i == X86::AddrDisp)
20934 MIB.addDisp(MI->getOperand(i), SPOffset);
20936 MIB.addOperand(MI->getOperand(i));
20938 MIB.setMemRefs(MMOBegin, MMOEnd);
20940 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
20942 MI->eraseFromParent();
20946 // Replace 213-type (isel default) FMA3 instructions with 231-type for
20947 // accumulator loops. Writing back to the accumulator allows the coalescer
20948 // to remove extra copies in the loop.
20949 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
20950 MachineBasicBlock *
20951 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
20952 MachineBasicBlock *MBB) const {
20953 MachineOperand &AddendOp = MI->getOperand(3);
20955 // Bail out early if the addend isn't a register - we can't switch these.
20956 if (!AddendOp.isReg())
20959 MachineFunction &MF = *MBB->getParent();
20960 MachineRegisterInfo &MRI = MF.getRegInfo();
20962 // Check whether the addend is defined by a PHI:
20963 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
20964 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
20965 if (!AddendDef.isPHI())
20968 // Look for the following pattern:
20970 // %addend = phi [%entry, 0], [%loop, %result]
20972 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
20976 // %addend = phi [%entry, 0], [%loop, %result]
20978 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
20980 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
20981 assert(AddendDef.getOperand(i).isReg());
20982 MachineOperand PHISrcOp = AddendDef.getOperand(i);
20983 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
20984 if (&PHISrcInst == MI) {
20985 // Found a matching instruction.
20986 unsigned NewFMAOpc = 0;
20987 switch (MI->getOpcode()) {
20988 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
20989 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
20990 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
20991 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
20992 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
20993 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
20994 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
20995 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
20996 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
20997 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
20998 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
20999 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21000 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21001 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21002 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21003 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21004 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21005 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21006 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21007 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21009 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21010 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21011 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21012 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21013 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21014 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21015 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21016 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21017 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21018 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21019 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21020 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21021 default: llvm_unreachable("Unrecognized FMA variant.");
21024 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21025 MachineInstrBuilder MIB =
21026 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21027 .addOperand(MI->getOperand(0))
21028 .addOperand(MI->getOperand(3))
21029 .addOperand(MI->getOperand(2))
21030 .addOperand(MI->getOperand(1));
21031 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21032 MI->eraseFromParent();
21039 MachineBasicBlock *
21040 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21041 MachineBasicBlock *BB) const {
21042 switch (MI->getOpcode()) {
21043 default: llvm_unreachable("Unexpected instr type to insert");
21044 case X86::TAILJMPd64:
21045 case X86::TAILJMPr64:
21046 case X86::TAILJMPm64:
21047 case X86::TAILJMPd64_REX:
21048 case X86::TAILJMPr64_REX:
21049 case X86::TAILJMPm64_REX:
21050 llvm_unreachable("TAILJMP64 would not be touched here.");
21051 case X86::TCRETURNdi64:
21052 case X86::TCRETURNri64:
21053 case X86::TCRETURNmi64:
21055 case X86::WIN_ALLOCA:
21056 return EmitLoweredWinAlloca(MI, BB);
21057 case X86::SEG_ALLOCA_32:
21058 case X86::SEG_ALLOCA_64:
21059 return EmitLoweredSegAlloca(MI, BB);
21060 case X86::TLSCall_32:
21061 case X86::TLSCall_64:
21062 return EmitLoweredTLSCall(MI, BB);
21063 case X86::CMOV_FR32:
21064 case X86::CMOV_FR64:
21065 case X86::CMOV_GR8:
21066 case X86::CMOV_GR16:
21067 case X86::CMOV_GR32:
21068 case X86::CMOV_RFP32:
21069 case X86::CMOV_RFP64:
21070 case X86::CMOV_RFP80:
21071 case X86::CMOV_V2F64:
21072 case X86::CMOV_V2I64:
21073 case X86::CMOV_V4F32:
21074 case X86::CMOV_V4F64:
21075 case X86::CMOV_V4I64:
21076 case X86::CMOV_V16F32:
21077 case X86::CMOV_V8F32:
21078 case X86::CMOV_V8F64:
21079 case X86::CMOV_V8I64:
21080 case X86::CMOV_V8I1:
21081 case X86::CMOV_V16I1:
21082 case X86::CMOV_V32I1:
21083 case X86::CMOV_V64I1:
21084 return EmitLoweredSelect(MI, BB);
21086 case X86::RELEASE_FADD32mr:
21087 case X86::RELEASE_FADD64mr:
21088 return EmitLoweredAtomicFP(MI, BB);
21090 case X86::FP32_TO_INT16_IN_MEM:
21091 case X86::FP32_TO_INT32_IN_MEM:
21092 case X86::FP32_TO_INT64_IN_MEM:
21093 case X86::FP64_TO_INT16_IN_MEM:
21094 case X86::FP64_TO_INT32_IN_MEM:
21095 case X86::FP64_TO_INT64_IN_MEM:
21096 case X86::FP80_TO_INT16_IN_MEM:
21097 case X86::FP80_TO_INT32_IN_MEM:
21098 case X86::FP80_TO_INT64_IN_MEM: {
21099 MachineFunction *F = BB->getParent();
21100 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21101 DebugLoc DL = MI->getDebugLoc();
21103 // Change the floating point control register to use "round towards zero"
21104 // mode when truncating to an integer value.
21105 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21106 addFrameReference(BuildMI(*BB, MI, DL,
21107 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21109 // Load the old value of the high byte of the control word...
21111 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21112 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21115 // Set the high part to be round to zero...
21116 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21119 // Reload the modified control word now...
21120 addFrameReference(BuildMI(*BB, MI, DL,
21121 TII->get(X86::FLDCW16m)), CWFrameIdx);
21123 // Restore the memory image of control word to original value
21124 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21127 // Get the X86 opcode to use.
21129 switch (MI->getOpcode()) {
21130 default: llvm_unreachable("illegal opcode!");
21131 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21132 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
21133 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
21134 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
21135 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
21136 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
21137 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
21138 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
21139 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
21143 MachineOperand &Op = MI->getOperand(0);
21145 AM.BaseType = X86AddressMode::RegBase;
21146 AM.Base.Reg = Op.getReg();
21148 AM.BaseType = X86AddressMode::FrameIndexBase;
21149 AM.Base.FrameIndex = Op.getIndex();
21151 Op = MI->getOperand(1);
21153 AM.Scale = Op.getImm();
21154 Op = MI->getOperand(2);
21156 AM.IndexReg = Op.getImm();
21157 Op = MI->getOperand(3);
21158 if (Op.isGlobal()) {
21159 AM.GV = Op.getGlobal();
21161 AM.Disp = Op.getImm();
21163 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21164 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21166 // Reload the original control word now.
21167 addFrameReference(BuildMI(*BB, MI, DL,
21168 TII->get(X86::FLDCW16m)), CWFrameIdx);
21170 MI->eraseFromParent(); // The pseudo instruction is gone now.
21173 // String/text processing lowering.
21174 case X86::PCMPISTRM128REG:
21175 case X86::VPCMPISTRM128REG:
21176 case X86::PCMPISTRM128MEM:
21177 case X86::VPCMPISTRM128MEM:
21178 case X86::PCMPESTRM128REG:
21179 case X86::VPCMPESTRM128REG:
21180 case X86::PCMPESTRM128MEM:
21181 case X86::VPCMPESTRM128MEM:
21182 assert(Subtarget->hasSSE42() &&
21183 "Target must have SSE4.2 or AVX features enabled");
21184 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
21186 // String/text processing lowering.
21187 case X86::PCMPISTRIREG:
21188 case X86::VPCMPISTRIREG:
21189 case X86::PCMPISTRIMEM:
21190 case X86::VPCMPISTRIMEM:
21191 case X86::PCMPESTRIREG:
21192 case X86::VPCMPESTRIREG:
21193 case X86::PCMPESTRIMEM:
21194 case X86::VPCMPESTRIMEM:
21195 assert(Subtarget->hasSSE42() &&
21196 "Target must have SSE4.2 or AVX features enabled");
21197 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
21199 // Thread synchronization.
21201 return EmitMonitor(MI, BB, Subtarget);
21205 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
21207 case X86::VASTART_SAVE_XMM_REGS:
21208 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
21210 case X86::VAARG_64:
21211 return EmitVAARG64WithCustomInserter(MI, BB);
21213 case X86::EH_SjLj_SetJmp32:
21214 case X86::EH_SjLj_SetJmp64:
21215 return emitEHSjLjSetJmp(MI, BB);
21217 case X86::EH_SjLj_LongJmp32:
21218 case X86::EH_SjLj_LongJmp64:
21219 return emitEHSjLjLongJmp(MI, BB);
21221 case TargetOpcode::STATEPOINT:
21222 // As an implementation detail, STATEPOINT shares the STACKMAP format at
21223 // this point in the process. We diverge later.
21224 return emitPatchPoint(MI, BB);
21226 case TargetOpcode::STACKMAP:
21227 case TargetOpcode::PATCHPOINT:
21228 return emitPatchPoint(MI, BB);
21230 case X86::VFMADDPDr213r:
21231 case X86::VFMADDPSr213r:
21232 case X86::VFMADDSDr213r:
21233 case X86::VFMADDSSr213r:
21234 case X86::VFMSUBPDr213r:
21235 case X86::VFMSUBPSr213r:
21236 case X86::VFMSUBSDr213r:
21237 case X86::VFMSUBSSr213r:
21238 case X86::VFNMADDPDr213r:
21239 case X86::VFNMADDPSr213r:
21240 case X86::VFNMADDSDr213r:
21241 case X86::VFNMADDSSr213r:
21242 case X86::VFNMSUBPDr213r:
21243 case X86::VFNMSUBPSr213r:
21244 case X86::VFNMSUBSDr213r:
21245 case X86::VFNMSUBSSr213r:
21246 case X86::VFMADDSUBPDr213r:
21247 case X86::VFMADDSUBPSr213r:
21248 case X86::VFMSUBADDPDr213r:
21249 case X86::VFMSUBADDPSr213r:
21250 case X86::VFMADDPDr213rY:
21251 case X86::VFMADDPSr213rY:
21252 case X86::VFMSUBPDr213rY:
21253 case X86::VFMSUBPSr213rY:
21254 case X86::VFNMADDPDr213rY:
21255 case X86::VFNMADDPSr213rY:
21256 case X86::VFNMSUBPDr213rY:
21257 case X86::VFNMSUBPSr213rY:
21258 case X86::VFMADDSUBPDr213rY:
21259 case X86::VFMADDSUBPSr213rY:
21260 case X86::VFMSUBADDPDr213rY:
21261 case X86::VFMSUBADDPSr213rY:
21262 return emitFMA3Instr(MI, BB);
21266 //===----------------------------------------------------------------------===//
21267 // X86 Optimization Hooks
21268 //===----------------------------------------------------------------------===//
21270 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
21273 const SelectionDAG &DAG,
21274 unsigned Depth) const {
21275 unsigned BitWidth = KnownZero.getBitWidth();
21276 unsigned Opc = Op.getOpcode();
21277 assert((Opc >= ISD::BUILTIN_OP_END ||
21278 Opc == ISD::INTRINSIC_WO_CHAIN ||
21279 Opc == ISD::INTRINSIC_W_CHAIN ||
21280 Opc == ISD::INTRINSIC_VOID) &&
21281 "Should use MaskedValueIsZero if you don't know whether Op"
21282 " is a target node!");
21284 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
21298 // These nodes' second result is a boolean.
21299 if (Op.getResNo() == 0)
21302 case X86ISD::SETCC:
21303 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
21305 case ISD::INTRINSIC_WO_CHAIN: {
21306 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
21307 unsigned NumLoBits = 0;
21310 case Intrinsic::x86_sse_movmsk_ps:
21311 case Intrinsic::x86_avx_movmsk_ps_256:
21312 case Intrinsic::x86_sse2_movmsk_pd:
21313 case Intrinsic::x86_avx_movmsk_pd_256:
21314 case Intrinsic::x86_mmx_pmovmskb:
21315 case Intrinsic::x86_sse2_pmovmskb_128:
21316 case Intrinsic::x86_avx2_pmovmskb: {
21317 // High bits of movmskp{s|d}, pmovmskb are known zero.
21319 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
21320 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
21321 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
21322 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
21323 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
21324 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
21325 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
21326 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
21328 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
21337 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
21339 const SelectionDAG &,
21340 unsigned Depth) const {
21341 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
21342 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
21343 return Op.getValueType().getScalarType().getSizeInBits();
21349 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
21350 /// node is a GlobalAddress + offset.
21351 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
21352 const GlobalValue* &GA,
21353 int64_t &Offset) const {
21354 if (N->getOpcode() == X86ISD::Wrapper) {
21355 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
21356 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
21357 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
21361 return TargetLowering::isGAPlusOffset(N, GA, Offset);
21364 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
21365 /// same as extracting the high 128-bit part of 256-bit vector and then
21366 /// inserting the result into the low part of a new 256-bit vector
21367 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
21368 EVT VT = SVOp->getValueType(0);
21369 unsigned NumElems = VT.getVectorNumElements();
21371 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21372 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
21373 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21374 SVOp->getMaskElt(j) >= 0)
21380 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
21381 /// same as extracting the low 128-bit part of 256-bit vector and then
21382 /// inserting the result into the high part of a new 256-bit vector
21383 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
21384 EVT VT = SVOp->getValueType(0);
21385 unsigned NumElems = VT.getVectorNumElements();
21387 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21388 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
21389 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21390 SVOp->getMaskElt(j) >= 0)
21396 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
21397 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
21398 TargetLowering::DAGCombinerInfo &DCI,
21399 const X86Subtarget* Subtarget) {
21401 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
21402 SDValue V1 = SVOp->getOperand(0);
21403 SDValue V2 = SVOp->getOperand(1);
21404 EVT VT = SVOp->getValueType(0);
21405 unsigned NumElems = VT.getVectorNumElements();
21407 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
21408 V2.getOpcode() == ISD::CONCAT_VECTORS) {
21412 // V UNDEF BUILD_VECTOR UNDEF
21414 // CONCAT_VECTOR CONCAT_VECTOR
21417 // RESULT: V + zero extended
21419 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
21420 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
21421 V1.getOperand(1).getOpcode() != ISD::UNDEF)
21424 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
21427 // To match the shuffle mask, the first half of the mask should
21428 // be exactly the first vector, and all the rest a splat with the
21429 // first element of the second one.
21430 for (unsigned i = 0; i != NumElems/2; ++i)
21431 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
21432 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
21435 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
21436 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
21437 if (Ld->hasNUsesOfValue(1, 0)) {
21438 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
21439 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
21441 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
21443 Ld->getPointerInfo(),
21444 Ld->getAlignment(),
21445 false/*isVolatile*/, true/*ReadMem*/,
21446 false/*WriteMem*/);
21448 // Make sure the newly-created LOAD is in the same position as Ld in
21449 // terms of dependency. We create a TokenFactor for Ld and ResNode,
21450 // and update uses of Ld's output chain to use the TokenFactor.
21451 if (Ld->hasAnyUseOfValue(1)) {
21452 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
21453 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
21454 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
21455 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
21456 SDValue(ResNode.getNode(), 1));
21459 return DAG.getBitcast(VT, ResNode);
21463 // Emit a zeroed vector and insert the desired subvector on its
21465 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
21466 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
21467 return DCI.CombineTo(N, InsV);
21470 //===--------------------------------------------------------------------===//
21471 // Combine some shuffles into subvector extracts and inserts:
21474 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21475 if (isShuffleHigh128VectorInsertLow(SVOp)) {
21476 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
21477 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
21478 return DCI.CombineTo(N, InsV);
21481 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21482 if (isShuffleLow128VectorInsertHigh(SVOp)) {
21483 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
21484 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
21485 return DCI.CombineTo(N, InsV);
21491 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
21494 /// This is the leaf of the recursive combinine below. When we have found some
21495 /// chain of single-use x86 shuffle instructions and accumulated the combined
21496 /// shuffle mask represented by them, this will try to pattern match that mask
21497 /// into either a single instruction if there is a special purpose instruction
21498 /// for this operation, or into a PSHUFB instruction which is a fully general
21499 /// instruction but should only be used to replace chains over a certain depth.
21500 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
21501 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
21502 TargetLowering::DAGCombinerInfo &DCI,
21503 const X86Subtarget *Subtarget) {
21504 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
21506 // Find the operand that enters the chain. Note that multiple uses are OK
21507 // here, we're not going to remove the operand we find.
21508 SDValue Input = Op.getOperand(0);
21509 while (Input.getOpcode() == ISD::BITCAST)
21510 Input = Input.getOperand(0);
21512 MVT VT = Input.getSimpleValueType();
21513 MVT RootVT = Root.getSimpleValueType();
21516 // Just remove no-op shuffle masks.
21517 if (Mask.size() == 1) {
21518 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
21523 // Use the float domain if the operand type is a floating point type.
21524 bool FloatDomain = VT.isFloatingPoint();
21526 // For floating point shuffles, we don't have free copies in the shuffle
21527 // instructions or the ability to load as part of the instruction, so
21528 // canonicalize their shuffles to UNPCK or MOV variants.
21530 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
21531 // vectors because it can have a load folded into it that UNPCK cannot. This
21532 // doesn't preclude something switching to the shorter encoding post-RA.
21534 // FIXME: Should teach these routines about AVX vector widths.
21535 if (FloatDomain && VT.getSizeInBits() == 128) {
21536 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
21537 bool Lo = Mask.equals({0, 0});
21540 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
21541 // is no slower than UNPCKLPD but has the option to fold the input operand
21542 // into even an unaligned memory load.
21543 if (Lo && Subtarget->hasSSE3()) {
21544 Shuffle = X86ISD::MOVDDUP;
21545 ShuffleVT = MVT::v2f64;
21547 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
21548 // than the UNPCK variants.
21549 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
21550 ShuffleVT = MVT::v4f32;
21552 if (Depth == 1 && Root->getOpcode() == Shuffle)
21553 return false; // Nothing to do!
21554 Op = DAG.getBitcast(ShuffleVT, Input);
21555 DCI.AddToWorklist(Op.getNode());
21556 if (Shuffle == X86ISD::MOVDDUP)
21557 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21559 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21560 DCI.AddToWorklist(Op.getNode());
21561 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21565 if (Subtarget->hasSSE3() &&
21566 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
21567 bool Lo = Mask.equals({0, 0, 2, 2});
21568 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
21569 MVT ShuffleVT = MVT::v4f32;
21570 if (Depth == 1 && Root->getOpcode() == Shuffle)
21571 return false; // Nothing to do!
21572 Op = DAG.getBitcast(ShuffleVT, Input);
21573 DCI.AddToWorklist(Op.getNode());
21574 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21575 DCI.AddToWorklist(Op.getNode());
21576 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21580 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
21581 bool Lo = Mask.equals({0, 0, 1, 1});
21582 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21583 MVT ShuffleVT = MVT::v4f32;
21584 if (Depth == 1 && Root->getOpcode() == Shuffle)
21585 return false; // Nothing to do!
21586 Op = DAG.getBitcast(ShuffleVT, Input);
21587 DCI.AddToWorklist(Op.getNode());
21588 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21589 DCI.AddToWorklist(Op.getNode());
21590 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21596 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
21597 // variants as none of these have single-instruction variants that are
21598 // superior to the UNPCK formulation.
21599 if (!FloatDomain && VT.getSizeInBits() == 128 &&
21600 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
21601 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
21602 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
21604 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
21605 bool Lo = Mask[0] == 0;
21606 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21607 if (Depth == 1 && Root->getOpcode() == Shuffle)
21608 return false; // Nothing to do!
21610 switch (Mask.size()) {
21612 ShuffleVT = MVT::v8i16;
21615 ShuffleVT = MVT::v16i8;
21618 llvm_unreachable("Impossible mask size!");
21620 Op = DAG.getBitcast(ShuffleVT, Input);
21621 DCI.AddToWorklist(Op.getNode());
21622 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21623 DCI.AddToWorklist(Op.getNode());
21624 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21629 // Don't try to re-form single instruction chains under any circumstances now
21630 // that we've done encoding canonicalization for them.
21634 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
21635 // can replace them with a single PSHUFB instruction profitably. Intel's
21636 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
21637 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
21638 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
21639 SmallVector<SDValue, 16> PSHUFBMask;
21640 int NumBytes = VT.getSizeInBits() / 8;
21641 int Ratio = NumBytes / Mask.size();
21642 for (int i = 0; i < NumBytes; ++i) {
21643 if (Mask[i / Ratio] == SM_SentinelUndef) {
21644 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
21647 int M = Mask[i / Ratio] != SM_SentinelZero
21648 ? Ratio * Mask[i / Ratio] + i % Ratio
21650 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
21652 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
21653 Op = DAG.getBitcast(ByteVT, Input);
21654 DCI.AddToWorklist(Op.getNode());
21655 SDValue PSHUFBMaskOp =
21656 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
21657 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
21658 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
21659 DCI.AddToWorklist(Op.getNode());
21660 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21665 // Failed to find any combines.
21669 /// \brief Fully generic combining of x86 shuffle instructions.
21671 /// This should be the last combine run over the x86 shuffle instructions. Once
21672 /// they have been fully optimized, this will recursively consider all chains
21673 /// of single-use shuffle instructions, build a generic model of the cumulative
21674 /// shuffle operation, and check for simpler instructions which implement this
21675 /// operation. We use this primarily for two purposes:
21677 /// 1) Collapse generic shuffles to specialized single instructions when
21678 /// equivalent. In most cases, this is just an encoding size win, but
21679 /// sometimes we will collapse multiple generic shuffles into a single
21680 /// special-purpose shuffle.
21681 /// 2) Look for sequences of shuffle instructions with 3 or more total
21682 /// instructions, and replace them with the slightly more expensive SSSE3
21683 /// PSHUFB instruction if available. We do this as the last combining step
21684 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
21685 /// a suitable short sequence of other instructions. The PHUFB will either
21686 /// use a register or have to read from memory and so is slightly (but only
21687 /// slightly) more expensive than the other shuffle instructions.
21689 /// Because this is inherently a quadratic operation (for each shuffle in
21690 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
21691 /// This should never be an issue in practice as the shuffle lowering doesn't
21692 /// produce sequences of more than 8 instructions.
21694 /// FIXME: We will currently miss some cases where the redundant shuffling
21695 /// would simplify under the threshold for PSHUFB formation because of
21696 /// combine-ordering. To fix this, we should do the redundant instruction
21697 /// combining in this recursive walk.
21698 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
21699 ArrayRef<int> RootMask,
21700 int Depth, bool HasPSHUFB,
21702 TargetLowering::DAGCombinerInfo &DCI,
21703 const X86Subtarget *Subtarget) {
21704 // Bound the depth of our recursive combine because this is ultimately
21705 // quadratic in nature.
21709 // Directly rip through bitcasts to find the underlying operand.
21710 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
21711 Op = Op.getOperand(0);
21713 MVT VT = Op.getSimpleValueType();
21714 if (!VT.isVector())
21715 return false; // Bail if we hit a non-vector.
21717 assert(Root.getSimpleValueType().isVector() &&
21718 "Shuffles operate on vector types!");
21719 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
21720 "Can only combine shuffles of the same vector register size.");
21722 if (!isTargetShuffle(Op.getOpcode()))
21724 SmallVector<int, 16> OpMask;
21726 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
21727 // We only can combine unary shuffles which we can decode the mask for.
21728 if (!HaveMask || !IsUnary)
21731 assert(VT.getVectorNumElements() == OpMask.size() &&
21732 "Different mask size from vector size!");
21733 assert(((RootMask.size() > OpMask.size() &&
21734 RootMask.size() % OpMask.size() == 0) ||
21735 (OpMask.size() > RootMask.size() &&
21736 OpMask.size() % RootMask.size() == 0) ||
21737 OpMask.size() == RootMask.size()) &&
21738 "The smaller number of elements must divide the larger.");
21739 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
21740 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
21741 assert(((RootRatio == 1 && OpRatio == 1) ||
21742 (RootRatio == 1) != (OpRatio == 1)) &&
21743 "Must not have a ratio for both incoming and op masks!");
21745 SmallVector<int, 16> Mask;
21746 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
21748 // Merge this shuffle operation's mask into our accumulated mask. Note that
21749 // this shuffle's mask will be the first applied to the input, followed by the
21750 // root mask to get us all the way to the root value arrangement. The reason
21751 // for this order is that we are recursing up the operation chain.
21752 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
21753 int RootIdx = i / RootRatio;
21754 if (RootMask[RootIdx] < 0) {
21755 // This is a zero or undef lane, we're done.
21756 Mask.push_back(RootMask[RootIdx]);
21760 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
21761 int OpIdx = RootMaskedIdx / OpRatio;
21762 if (OpMask[OpIdx] < 0) {
21763 // The incoming lanes are zero or undef, it doesn't matter which ones we
21765 Mask.push_back(OpMask[OpIdx]);
21769 // Ok, we have non-zero lanes, map them through.
21770 Mask.push_back(OpMask[OpIdx] * OpRatio +
21771 RootMaskedIdx % OpRatio);
21774 // See if we can recurse into the operand to combine more things.
21775 switch (Op.getOpcode()) {
21776 case X86ISD::PSHUFB:
21778 case X86ISD::PSHUFD:
21779 case X86ISD::PSHUFHW:
21780 case X86ISD::PSHUFLW:
21781 if (Op.getOperand(0).hasOneUse() &&
21782 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
21783 HasPSHUFB, DAG, DCI, Subtarget))
21787 case X86ISD::UNPCKL:
21788 case X86ISD::UNPCKH:
21789 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
21790 // We can't check for single use, we have to check that this shuffle is the only user.
21791 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
21792 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
21793 HasPSHUFB, DAG, DCI, Subtarget))
21798 // Minor canonicalization of the accumulated shuffle mask to make it easier
21799 // to match below. All this does is detect masks with squential pairs of
21800 // elements, and shrink them to the half-width mask. It does this in a loop
21801 // so it will reduce the size of the mask to the minimal width mask which
21802 // performs an equivalent shuffle.
21803 SmallVector<int, 16> WidenedMask;
21804 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
21805 Mask = std::move(WidenedMask);
21806 WidenedMask.clear();
21809 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
21813 /// \brief Get the PSHUF-style mask from PSHUF node.
21815 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
21816 /// PSHUF-style masks that can be reused with such instructions.
21817 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
21818 MVT VT = N.getSimpleValueType();
21819 SmallVector<int, 4> Mask;
21821 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
21825 // If we have more than 128-bits, only the low 128-bits of shuffle mask
21826 // matter. Check that the upper masks are repeats and remove them.
21827 if (VT.getSizeInBits() > 128) {
21828 int LaneElts = 128 / VT.getScalarSizeInBits();
21830 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
21831 for (int j = 0; j < LaneElts; ++j)
21832 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
21833 "Mask doesn't repeat in high 128-bit lanes!");
21835 Mask.resize(LaneElts);
21838 switch (N.getOpcode()) {
21839 case X86ISD::PSHUFD:
21841 case X86ISD::PSHUFLW:
21844 case X86ISD::PSHUFHW:
21845 Mask.erase(Mask.begin(), Mask.begin() + 4);
21846 for (int &M : Mask)
21850 llvm_unreachable("No valid shuffle instruction found!");
21854 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
21856 /// We walk up the chain and look for a combinable shuffle, skipping over
21857 /// shuffles that we could hoist this shuffle's transformation past without
21858 /// altering anything.
21860 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
21862 TargetLowering::DAGCombinerInfo &DCI) {
21863 assert(N.getOpcode() == X86ISD::PSHUFD &&
21864 "Called with something other than an x86 128-bit half shuffle!");
21867 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
21868 // of the shuffles in the chain so that we can form a fresh chain to replace
21870 SmallVector<SDValue, 8> Chain;
21871 SDValue V = N.getOperand(0);
21872 for (; V.hasOneUse(); V = V.getOperand(0)) {
21873 switch (V.getOpcode()) {
21875 return SDValue(); // Nothing combined!
21878 // Skip bitcasts as we always know the type for the target specific
21882 case X86ISD::PSHUFD:
21883 // Found another dword shuffle.
21886 case X86ISD::PSHUFLW:
21887 // Check that the low words (being shuffled) are the identity in the
21888 // dword shuffle, and the high words are self-contained.
21889 if (Mask[0] != 0 || Mask[1] != 1 ||
21890 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
21893 Chain.push_back(V);
21896 case X86ISD::PSHUFHW:
21897 // Check that the high words (being shuffled) are the identity in the
21898 // dword shuffle, and the low words are self-contained.
21899 if (Mask[2] != 2 || Mask[3] != 3 ||
21900 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
21903 Chain.push_back(V);
21906 case X86ISD::UNPCKL:
21907 case X86ISD::UNPCKH:
21908 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
21909 // shuffle into a preceding word shuffle.
21910 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
21911 V.getSimpleValueType().getScalarType() != MVT::i16)
21914 // Search for a half-shuffle which we can combine with.
21915 unsigned CombineOp =
21916 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
21917 if (V.getOperand(0) != V.getOperand(1) ||
21918 !V->isOnlyUserOf(V.getOperand(0).getNode()))
21920 Chain.push_back(V);
21921 V = V.getOperand(0);
21923 switch (V.getOpcode()) {
21925 return SDValue(); // Nothing to combine.
21927 case X86ISD::PSHUFLW:
21928 case X86ISD::PSHUFHW:
21929 if (V.getOpcode() == CombineOp)
21932 Chain.push_back(V);
21936 V = V.getOperand(0);
21940 } while (V.hasOneUse());
21943 // Break out of the loop if we break out of the switch.
21947 if (!V.hasOneUse())
21948 // We fell out of the loop without finding a viable combining instruction.
21951 // Merge this node's mask and our incoming mask.
21952 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
21953 for (int &M : Mask)
21955 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
21956 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
21958 // Rebuild the chain around this new shuffle.
21959 while (!Chain.empty()) {
21960 SDValue W = Chain.pop_back_val();
21962 if (V.getValueType() != W.getOperand(0).getValueType())
21963 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
21965 switch (W.getOpcode()) {
21967 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
21969 case X86ISD::UNPCKL:
21970 case X86ISD::UNPCKH:
21971 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
21974 case X86ISD::PSHUFD:
21975 case X86ISD::PSHUFLW:
21976 case X86ISD::PSHUFHW:
21977 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
21981 if (V.getValueType() != N.getValueType())
21982 V = DAG.getBitcast(N.getValueType(), V);
21984 // Return the new chain to replace N.
21988 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
21990 /// We walk up the chain, skipping shuffles of the other half and looking
21991 /// through shuffles which switch halves trying to find a shuffle of the same
21992 /// pair of dwords.
21993 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
21995 TargetLowering::DAGCombinerInfo &DCI) {
21997 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
21998 "Called with something other than an x86 128-bit half shuffle!");
22000 unsigned CombineOpcode = N.getOpcode();
22002 // Walk up a single-use chain looking for a combinable shuffle.
22003 SDValue V = N.getOperand(0);
22004 for (; V.hasOneUse(); V = V.getOperand(0)) {
22005 switch (V.getOpcode()) {
22007 return false; // Nothing combined!
22010 // Skip bitcasts as we always know the type for the target specific
22014 case X86ISD::PSHUFLW:
22015 case X86ISD::PSHUFHW:
22016 if (V.getOpcode() == CombineOpcode)
22019 // Other-half shuffles are no-ops.
22022 // Break out of the loop if we break out of the switch.
22026 if (!V.hasOneUse())
22027 // We fell out of the loop without finding a viable combining instruction.
22030 // Combine away the bottom node as its shuffle will be accumulated into
22031 // a preceding shuffle.
22032 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22034 // Record the old value.
22037 // Merge this node's mask and our incoming mask (adjusted to account for all
22038 // the pshufd instructions encountered).
22039 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22040 for (int &M : Mask)
22042 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22043 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22045 // Check that the shuffles didn't cancel each other out. If not, we need to
22046 // combine to the new one.
22048 // Replace the combinable shuffle with the combined one, updating all users
22049 // so that we re-evaluate the chain here.
22050 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22055 /// \brief Try to combine x86 target specific shuffles.
22056 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22057 TargetLowering::DAGCombinerInfo &DCI,
22058 const X86Subtarget *Subtarget) {
22060 MVT VT = N.getSimpleValueType();
22061 SmallVector<int, 4> Mask;
22063 switch (N.getOpcode()) {
22064 case X86ISD::PSHUFD:
22065 case X86ISD::PSHUFLW:
22066 case X86ISD::PSHUFHW:
22067 Mask = getPSHUFShuffleMask(N);
22068 assert(Mask.size() == 4);
22074 // Nuke no-op shuffles that show up after combining.
22075 if (isNoopShuffleMask(Mask))
22076 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22078 // Look for simplifications involving one or two shuffle instructions.
22079 SDValue V = N.getOperand(0);
22080 switch (N.getOpcode()) {
22083 case X86ISD::PSHUFLW:
22084 case X86ISD::PSHUFHW:
22085 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
22087 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22088 return SDValue(); // We combined away this shuffle, so we're done.
22090 // See if this reduces to a PSHUFD which is no more expensive and can
22091 // combine with more operations. Note that it has to at least flip the
22092 // dwords as otherwise it would have been removed as a no-op.
22093 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
22094 int DMask[] = {0, 1, 2, 3};
22095 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22096 DMask[DOffset + 0] = DOffset + 1;
22097 DMask[DOffset + 1] = DOffset + 0;
22098 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
22099 V = DAG.getBitcast(DVT, V);
22100 DCI.AddToWorklist(V.getNode());
22101 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
22102 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
22103 DCI.AddToWorklist(V.getNode());
22104 return DAG.getBitcast(VT, V);
22107 // Look for shuffle patterns which can be implemented as a single unpack.
22108 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22109 // only works when we have a PSHUFD followed by two half-shuffles.
22110 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22111 (V.getOpcode() == X86ISD::PSHUFLW ||
22112 V.getOpcode() == X86ISD::PSHUFHW) &&
22113 V.getOpcode() != N.getOpcode() &&
22115 SDValue D = V.getOperand(0);
22116 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22117 D = D.getOperand(0);
22118 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22119 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22120 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22121 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22122 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22124 for (int i = 0; i < 4; ++i) {
22125 WordMask[i + NOffset] = Mask[i] + NOffset;
22126 WordMask[i + VOffset] = VMask[i] + VOffset;
22128 // Map the word mask through the DWord mask.
22130 for (int i = 0; i < 8; ++i)
22131 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
22132 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22133 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
22134 // We can replace all three shuffles with an unpack.
22135 V = DAG.getBitcast(VT, D.getOperand(0));
22136 DCI.AddToWorklist(V.getNode());
22137 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
22146 case X86ISD::PSHUFD:
22147 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
22156 /// \brief Try to combine a shuffle into a target-specific add-sub node.
22158 /// We combine this directly on the abstract vector shuffle nodes so it is
22159 /// easier to generically match. We also insert dummy vector shuffle nodes for
22160 /// the operands which explicitly discard the lanes which are unused by this
22161 /// operation to try to flow through the rest of the combiner the fact that
22162 /// they're unused.
22163 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
22165 EVT VT = N->getValueType(0);
22167 // We only handle target-independent shuffles.
22168 // FIXME: It would be easy and harmless to use the target shuffle mask
22169 // extraction tool to support more.
22170 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
22173 auto *SVN = cast<ShuffleVectorSDNode>(N);
22174 ArrayRef<int> Mask = SVN->getMask();
22175 SDValue V1 = N->getOperand(0);
22176 SDValue V2 = N->getOperand(1);
22178 // We require the first shuffle operand to be the SUB node, and the second to
22179 // be the ADD node.
22180 // FIXME: We should support the commuted patterns.
22181 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
22184 // If there are other uses of these operations we can't fold them.
22185 if (!V1->hasOneUse() || !V2->hasOneUse())
22188 // Ensure that both operations have the same operands. Note that we can
22189 // commute the FADD operands.
22190 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
22191 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
22192 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
22195 // We're looking for blends between FADD and FSUB nodes. We insist on these
22196 // nodes being lined up in a specific expected pattern.
22197 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
22198 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
22199 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
22202 // Only specific types are legal at this point, assert so we notice if and
22203 // when these change.
22204 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
22205 VT == MVT::v4f64) &&
22206 "Unknown vector type encountered!");
22208 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
22211 /// PerformShuffleCombine - Performs several different shuffle combines.
22212 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
22213 TargetLowering::DAGCombinerInfo &DCI,
22214 const X86Subtarget *Subtarget) {
22216 SDValue N0 = N->getOperand(0);
22217 SDValue N1 = N->getOperand(1);
22218 EVT VT = N->getValueType(0);
22220 // Don't create instructions with illegal types after legalize types has run.
22221 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22222 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
22225 // If we have legalized the vector types, look for blends of FADD and FSUB
22226 // nodes that we can fuse into an ADDSUB node.
22227 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
22228 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
22231 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
22232 if (Subtarget->hasFp256() && VT.is256BitVector() &&
22233 N->getOpcode() == ISD::VECTOR_SHUFFLE)
22234 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
22236 // During Type Legalization, when promoting illegal vector types,
22237 // the backend might introduce new shuffle dag nodes and bitcasts.
22239 // This code performs the following transformation:
22240 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
22241 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
22243 // We do this only if both the bitcast and the BINOP dag nodes have
22244 // one use. Also, perform this transformation only if the new binary
22245 // operation is legal. This is to avoid introducing dag nodes that
22246 // potentially need to be further expanded (or custom lowered) into a
22247 // less optimal sequence of dag nodes.
22248 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
22249 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
22250 N0.getOpcode() == ISD::BITCAST) {
22251 SDValue BC0 = N0.getOperand(0);
22252 EVT SVT = BC0.getValueType();
22253 unsigned Opcode = BC0.getOpcode();
22254 unsigned NumElts = VT.getVectorNumElements();
22256 if (BC0.hasOneUse() && SVT.isVector() &&
22257 SVT.getVectorNumElements() * 2 == NumElts &&
22258 TLI.isOperationLegal(Opcode, VT)) {
22259 bool CanFold = false;
22271 unsigned SVTNumElts = SVT.getVectorNumElements();
22272 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22273 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
22274 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
22275 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
22276 CanFold = SVOp->getMaskElt(i) < 0;
22279 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
22280 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
22281 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
22282 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
22287 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
22288 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
22289 // consecutive, non-overlapping, and in the right order.
22290 SmallVector<SDValue, 16> Elts;
22291 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
22292 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
22294 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
22297 if (isTargetShuffle(N->getOpcode())) {
22299 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
22300 if (Shuffle.getNode())
22303 // Try recursively combining arbitrary sequences of x86 shuffle
22304 // instructions into higher-order shuffles. We do this after combining
22305 // specific PSHUF instruction sequences into their minimal form so that we
22306 // can evaluate how many specialized shuffle instructions are involved in
22307 // a particular chain.
22308 SmallVector<int, 1> NonceMask; // Just a placeholder.
22309 NonceMask.push_back(0);
22310 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
22311 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
22313 return SDValue(); // This routine will use CombineTo to replace N.
22319 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
22320 /// specific shuffle of a load can be folded into a single element load.
22321 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
22322 /// shuffles have been custom lowered so we need to handle those here.
22323 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
22324 TargetLowering::DAGCombinerInfo &DCI) {
22325 if (DCI.isBeforeLegalizeOps())
22328 SDValue InVec = N->getOperand(0);
22329 SDValue EltNo = N->getOperand(1);
22331 if (!isa<ConstantSDNode>(EltNo))
22334 EVT OriginalVT = InVec.getValueType();
22336 if (InVec.getOpcode() == ISD::BITCAST) {
22337 // Don't duplicate a load with other uses.
22338 if (!InVec.hasOneUse())
22340 EVT BCVT = InVec.getOperand(0).getValueType();
22341 if (!BCVT.isVector() ||
22342 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
22344 InVec = InVec.getOperand(0);
22347 EVT CurrentVT = InVec.getValueType();
22349 if (!isTargetShuffle(InVec.getOpcode()))
22352 // Don't duplicate a load with other uses.
22353 if (!InVec.hasOneUse())
22356 SmallVector<int, 16> ShuffleMask;
22358 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
22359 ShuffleMask, UnaryShuffle))
22362 // Select the input vector, guarding against out of range extract vector.
22363 unsigned NumElems = CurrentVT.getVectorNumElements();
22364 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
22365 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
22366 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
22367 : InVec.getOperand(1);
22369 // If inputs to shuffle are the same for both ops, then allow 2 uses
22370 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
22371 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
22373 if (LdNode.getOpcode() == ISD::BITCAST) {
22374 // Don't duplicate a load with other uses.
22375 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
22378 AllowedUses = 1; // only allow 1 load use if we have a bitcast
22379 LdNode = LdNode.getOperand(0);
22382 if (!ISD::isNormalLoad(LdNode.getNode()))
22385 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
22387 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
22390 EVT EltVT = N->getValueType(0);
22391 // If there's a bitcast before the shuffle, check if the load type and
22392 // alignment is valid.
22393 unsigned Align = LN0->getAlignment();
22394 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22395 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
22396 EltVT.getTypeForEVT(*DAG.getContext()));
22398 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
22401 // All checks match so transform back to vector_shuffle so that DAG combiner
22402 // can finish the job
22405 // Create shuffle node taking into account the case that its a unary shuffle
22406 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
22407 : InVec.getOperand(1);
22408 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
22409 InVec.getOperand(0), Shuffle,
22411 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
22412 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
22416 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
22417 /// special and don't usually play with other vector types, it's better to
22418 /// handle them early to be sure we emit efficient code by avoiding
22419 /// store-load conversions.
22420 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
22421 if (N->getValueType(0) != MVT::x86mmx ||
22422 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
22423 N->getOperand(0)->getValueType(0) != MVT::v2i32)
22426 SDValue V = N->getOperand(0);
22427 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
22428 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
22429 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
22430 N->getValueType(0), V.getOperand(0));
22435 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
22436 /// generation and convert it from being a bunch of shuffles and extracts
22437 /// into a somewhat faster sequence. For i686, the best sequence is apparently
22438 /// storing the value and loading scalars back, while for x64 we should
22439 /// use 64-bit extracts and shifts.
22440 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
22441 TargetLowering::DAGCombinerInfo &DCI) {
22442 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
22445 SDValue InputVector = N->getOperand(0);
22446 SDLoc dl(InputVector);
22447 // Detect mmx to i32 conversion through a v2i32 elt extract.
22448 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
22449 N->getValueType(0) == MVT::i32 &&
22450 InputVector.getValueType() == MVT::v2i32) {
22452 // The bitcast source is a direct mmx result.
22453 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
22454 if (MMXSrc.getValueType() == MVT::x86mmx)
22455 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22456 N->getValueType(0),
22457 InputVector.getNode()->getOperand(0));
22459 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
22460 SDValue MMXSrcOp = MMXSrc.getOperand(0);
22461 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
22462 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
22463 MMXSrcOp.getOpcode() == ISD::BITCAST &&
22464 MMXSrcOp.getValueType() == MVT::v1i64 &&
22465 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
22466 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22467 N->getValueType(0),
22468 MMXSrcOp.getOperand(0));
22471 EVT VT = N->getValueType(0);
22473 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
22474 InputVector.getOpcode() == ISD::BITCAST &&
22475 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
22476 uint64_t ExtractedElt =
22477 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
22478 uint64_t InputValue =
22479 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
22480 uint64_t Res = (InputValue >> ExtractedElt) & 1;
22481 return DAG.getConstant(Res, dl, MVT::i1);
22483 // Only operate on vectors of 4 elements, where the alternative shuffling
22484 // gets to be more expensive.
22485 if (InputVector.getValueType() != MVT::v4i32)
22488 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
22489 // single use which is a sign-extend or zero-extend, and all elements are
22491 SmallVector<SDNode *, 4> Uses;
22492 unsigned ExtractedElements = 0;
22493 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
22494 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
22495 if (UI.getUse().getResNo() != InputVector.getResNo())
22498 SDNode *Extract = *UI;
22499 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
22502 if (Extract->getValueType(0) != MVT::i32)
22504 if (!Extract->hasOneUse())
22506 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
22507 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
22509 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
22512 // Record which element was extracted.
22513 ExtractedElements |=
22514 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
22516 Uses.push_back(Extract);
22519 // If not all the elements were used, this may not be worthwhile.
22520 if (ExtractedElements != 15)
22523 // Ok, we've now decided to do the transformation.
22524 // If 64-bit shifts are legal, use the extract-shift sequence,
22525 // otherwise bounce the vector off the cache.
22526 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22529 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
22530 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
22531 auto &DL = DAG.getDataLayout();
22532 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
22533 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22534 DAG.getConstant(0, dl, VecIdxTy));
22535 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22536 DAG.getConstant(1, dl, VecIdxTy));
22538 SDValue ShAmt = DAG.getConstant(
22539 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
22540 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
22541 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22542 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
22543 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
22544 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22545 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
22547 // Store the value to a temporary stack slot.
22548 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
22549 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
22550 MachinePointerInfo(), false, false, 0);
22552 EVT ElementType = InputVector.getValueType().getVectorElementType();
22553 unsigned EltSize = ElementType.getSizeInBits() / 8;
22555 // Replace each use (extract) with a load of the appropriate element.
22556 for (unsigned i = 0; i < 4; ++i) {
22557 uint64_t Offset = EltSize * i;
22558 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
22559 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
22561 SDValue ScalarAddr =
22562 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
22564 // Load the scalar.
22565 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
22566 ScalarAddr, MachinePointerInfo(),
22567 false, false, false, 0);
22572 // Replace the extracts
22573 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
22574 UE = Uses.end(); UI != UE; ++UI) {
22575 SDNode *Extract = *UI;
22577 SDValue Idx = Extract->getOperand(1);
22578 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
22579 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
22582 // The replacement was made in place; don't return anything.
22587 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
22588 const X86Subtarget *Subtarget) {
22590 SDValue Cond = N->getOperand(0);
22591 SDValue LHS = N->getOperand(1);
22592 SDValue RHS = N->getOperand(2);
22594 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
22595 SDValue CondSrc = Cond->getOperand(0);
22596 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
22597 Cond = CondSrc->getOperand(0);
22600 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
22603 // A vselect where all conditions and data are constants can be optimized into
22604 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
22605 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
22606 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
22609 unsigned MaskValue = 0;
22610 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
22613 MVT VT = N->getSimpleValueType(0);
22614 unsigned NumElems = VT.getVectorNumElements();
22615 SmallVector<int, 8> ShuffleMask(NumElems, -1);
22616 for (unsigned i = 0; i < NumElems; ++i) {
22617 // Be sure we emit undef where we can.
22618 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
22619 ShuffleMask[i] = -1;
22621 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
22624 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22625 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
22627 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
22630 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
22632 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
22633 TargetLowering::DAGCombinerInfo &DCI,
22634 const X86Subtarget *Subtarget) {
22636 SDValue Cond = N->getOperand(0);
22637 // Get the LHS/RHS of the select.
22638 SDValue LHS = N->getOperand(1);
22639 SDValue RHS = N->getOperand(2);
22640 EVT VT = LHS.getValueType();
22641 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22643 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
22644 // instructions match the semantics of the common C idiom x<y?x:y but not
22645 // x<=y?x:y, because of how they handle negative zero (which can be
22646 // ignored in unsafe-math mode).
22647 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
22648 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
22649 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
22650 (Subtarget->hasSSE2() ||
22651 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
22652 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22654 unsigned Opcode = 0;
22655 // Check for x CC y ? x : y.
22656 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22657 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22661 // Converting this to a min would handle NaNs incorrectly, and swapping
22662 // the operands would cause it to handle comparisons between positive
22663 // and negative zero incorrectly.
22664 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22665 if (!DAG.getTarget().Options.UnsafeFPMath &&
22666 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22668 std::swap(LHS, RHS);
22670 Opcode = X86ISD::FMIN;
22673 // Converting this to a min would handle comparisons between positive
22674 // and negative zero incorrectly.
22675 if (!DAG.getTarget().Options.UnsafeFPMath &&
22676 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22678 Opcode = X86ISD::FMIN;
22681 // Converting this to a min would handle both negative zeros and NaNs
22682 // incorrectly, but we can swap the operands to fix both.
22683 std::swap(LHS, RHS);
22687 Opcode = X86ISD::FMIN;
22691 // Converting this to a max would handle comparisons between positive
22692 // and negative zero incorrectly.
22693 if (!DAG.getTarget().Options.UnsafeFPMath &&
22694 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22696 Opcode = X86ISD::FMAX;
22699 // Converting this to a max would handle NaNs incorrectly, and swapping
22700 // the operands would cause it to handle comparisons between positive
22701 // and negative zero incorrectly.
22702 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22703 if (!DAG.getTarget().Options.UnsafeFPMath &&
22704 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22706 std::swap(LHS, RHS);
22708 Opcode = X86ISD::FMAX;
22711 // Converting this to a max would handle both negative zeros and NaNs
22712 // incorrectly, but we can swap the operands to fix both.
22713 std::swap(LHS, RHS);
22717 Opcode = X86ISD::FMAX;
22720 // Check for x CC y ? y : x -- a min/max with reversed arms.
22721 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
22722 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
22726 // Converting this to a min would handle comparisons between positive
22727 // and negative zero incorrectly, and swapping the operands would
22728 // cause it to handle NaNs incorrectly.
22729 if (!DAG.getTarget().Options.UnsafeFPMath &&
22730 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
22731 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22733 std::swap(LHS, RHS);
22735 Opcode = X86ISD::FMIN;
22738 // Converting this to a min would handle NaNs incorrectly.
22739 if (!DAG.getTarget().Options.UnsafeFPMath &&
22740 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
22742 Opcode = X86ISD::FMIN;
22745 // Converting this to a min would handle both negative zeros and NaNs
22746 // incorrectly, but we can swap the operands to fix both.
22747 std::swap(LHS, RHS);
22751 Opcode = X86ISD::FMIN;
22755 // Converting this to a max would handle NaNs incorrectly.
22756 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22758 Opcode = X86ISD::FMAX;
22761 // Converting this to a max would handle comparisons between positive
22762 // and negative zero incorrectly, and swapping the operands would
22763 // cause it to handle NaNs incorrectly.
22764 if (!DAG.getTarget().Options.UnsafeFPMath &&
22765 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
22766 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22768 std::swap(LHS, RHS);
22770 Opcode = X86ISD::FMAX;
22773 // Converting this to a max would handle both negative zeros and NaNs
22774 // incorrectly, but we can swap the operands to fix both.
22775 std::swap(LHS, RHS);
22779 Opcode = X86ISD::FMAX;
22785 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
22788 EVT CondVT = Cond.getValueType();
22789 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
22790 CondVT.getVectorElementType() == MVT::i1) {
22791 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
22792 // lowering on KNL. In this case we convert it to
22793 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
22794 // The same situation for all 128 and 256-bit vectors of i8 and i16.
22795 // Since SKX these selects have a proper lowering.
22796 EVT OpVT = LHS.getValueType();
22797 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
22798 (OpVT.getVectorElementType() == MVT::i8 ||
22799 OpVT.getVectorElementType() == MVT::i16) &&
22800 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
22801 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
22802 DCI.AddToWorklist(Cond.getNode());
22803 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
22806 // If this is a select between two integer constants, try to do some
22808 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
22809 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
22810 // Don't do this for crazy integer types.
22811 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
22812 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
22813 // so that TrueC (the true value) is larger than FalseC.
22814 bool NeedsCondInvert = false;
22816 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
22817 // Efficiently invertible.
22818 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
22819 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
22820 isa<ConstantSDNode>(Cond.getOperand(1))))) {
22821 NeedsCondInvert = true;
22822 std::swap(TrueC, FalseC);
22825 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
22826 if (FalseC->getAPIntValue() == 0 &&
22827 TrueC->getAPIntValue().isPowerOf2()) {
22828 if (NeedsCondInvert) // Invert the condition if needed.
22829 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22830 DAG.getConstant(1, DL, Cond.getValueType()));
22832 // Zero extend the condition if needed.
22833 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
22835 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
22836 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
22837 DAG.getConstant(ShAmt, DL, MVT::i8));
22840 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
22841 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
22842 if (NeedsCondInvert) // Invert the condition if needed.
22843 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22844 DAG.getConstant(1, DL, Cond.getValueType()));
22846 // Zero extend the condition if needed.
22847 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
22848 FalseC->getValueType(0), Cond);
22849 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22850 SDValue(FalseC, 0));
22853 // Optimize cases that will turn into an LEA instruction. This requires
22854 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
22855 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
22856 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
22857 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
22859 bool isFastMultiplier = false;
22861 switch ((unsigned char)Diff) {
22863 case 1: // result = add base, cond
22864 case 2: // result = lea base( , cond*2)
22865 case 3: // result = lea base(cond, cond*2)
22866 case 4: // result = lea base( , cond*4)
22867 case 5: // result = lea base(cond, cond*4)
22868 case 8: // result = lea base( , cond*8)
22869 case 9: // result = lea base(cond, cond*8)
22870 isFastMultiplier = true;
22875 if (isFastMultiplier) {
22876 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
22877 if (NeedsCondInvert) // Invert the condition if needed.
22878 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22879 DAG.getConstant(1, DL, Cond.getValueType()));
22881 // Zero extend the condition if needed.
22882 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
22884 // Scale the condition by the difference.
22886 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
22887 DAG.getConstant(Diff, DL,
22888 Cond.getValueType()));
22890 // Add the base if non-zero.
22891 if (FalseC->getAPIntValue() != 0)
22892 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22893 SDValue(FalseC, 0));
22900 // Canonicalize max and min:
22901 // (x > y) ? x : y -> (x >= y) ? x : y
22902 // (x < y) ? x : y -> (x <= y) ? x : y
22903 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
22904 // the need for an extra compare
22905 // against zero. e.g.
22906 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
22908 // testl %edi, %edi
22910 // cmovgl %edi, %eax
22914 // cmovsl %eax, %edi
22915 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
22916 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22917 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22918 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22923 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
22924 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
22925 Cond.getOperand(0), Cond.getOperand(1), NewCC);
22926 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
22931 // Early exit check
22932 if (!TLI.isTypeLegal(VT))
22935 // Match VSELECTs into subs with unsigned saturation.
22936 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
22937 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
22938 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
22939 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
22940 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22942 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
22943 // left side invert the predicate to simplify logic below.
22945 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
22947 CC = ISD::getSetCCInverse(CC, true);
22948 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
22952 if (Other.getNode() && Other->getNumOperands() == 2 &&
22953 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
22954 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
22955 SDValue CondRHS = Cond->getOperand(1);
22957 // Look for a general sub with unsigned saturation first.
22958 // x >= y ? x-y : 0 --> subus x, y
22959 // x > y ? x-y : 0 --> subus x, y
22960 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
22961 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
22962 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
22964 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
22965 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
22966 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
22967 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
22968 // If the RHS is a constant we have to reverse the const
22969 // canonicalization.
22970 // x > C-1 ? x+-C : 0 --> subus x, C
22971 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
22972 CondRHSConst->getAPIntValue() ==
22973 (-OpRHSConst->getAPIntValue() - 1))
22974 return DAG.getNode(
22975 X86ISD::SUBUS, DL, VT, OpLHS,
22976 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
22978 // Another special case: If C was a sign bit, the sub has been
22979 // canonicalized into a xor.
22980 // FIXME: Would it be better to use computeKnownBits to determine
22981 // whether it's safe to decanonicalize the xor?
22982 // x s< 0 ? x^C : 0 --> subus x, C
22983 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
22984 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
22985 OpRHSConst->getAPIntValue().isSignBit())
22986 // Note that we have to rebuild the RHS constant here to ensure we
22987 // don't rely on particular values of undef lanes.
22988 return DAG.getNode(
22989 X86ISD::SUBUS, DL, VT, OpLHS,
22990 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
22995 // Simplify vector selection if condition value type matches vselect
22997 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
22998 assert(Cond.getValueType().isVector() &&
22999 "vector select expects a vector selector!");
23001 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23002 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23004 // Try invert the condition if true value is not all 1s and false value
23006 if (!TValIsAllOnes && !FValIsAllZeros &&
23007 // Check if the selector will be produced by CMPP*/PCMP*
23008 Cond.getOpcode() == ISD::SETCC &&
23009 // Check if SETCC has already been promoted
23010 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
23012 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23013 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23015 if (TValIsAllZeros || FValIsAllOnes) {
23016 SDValue CC = Cond.getOperand(2);
23017 ISD::CondCode NewCC =
23018 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23019 Cond.getOperand(0).getValueType().isInteger());
23020 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23021 std::swap(LHS, RHS);
23022 TValIsAllOnes = FValIsAllOnes;
23023 FValIsAllZeros = TValIsAllZeros;
23027 if (TValIsAllOnes || FValIsAllZeros) {
23030 if (TValIsAllOnes && FValIsAllZeros)
23032 else if (TValIsAllOnes)
23034 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
23035 else if (FValIsAllZeros)
23036 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
23037 DAG.getBitcast(CondVT, LHS));
23039 return DAG.getBitcast(VT, Ret);
23043 // We should generate an X86ISD::BLENDI from a vselect if its argument
23044 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
23045 // constants. This specific pattern gets generated when we split a
23046 // selector for a 512 bit vector in a machine without AVX512 (but with
23047 // 256-bit vectors), during legalization:
23049 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
23051 // Iff we find this pattern and the build_vectors are built from
23052 // constants, we translate the vselect into a shuffle_vector that we
23053 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
23054 if ((N->getOpcode() == ISD::VSELECT ||
23055 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
23056 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
23057 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
23058 if (Shuffle.getNode())
23062 // If this is a *dynamic* select (non-constant condition) and we can match
23063 // this node with one of the variable blend instructions, restructure the
23064 // condition so that the blends can use the high bit of each element and use
23065 // SimplifyDemandedBits to simplify the condition operand.
23066 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
23067 !DCI.isBeforeLegalize() &&
23068 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
23069 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
23071 // Don't optimize vector selects that map to mask-registers.
23075 // We can only handle the cases where VSELECT is directly legal on the
23076 // subtarget. We custom lower VSELECT nodes with constant conditions and
23077 // this makes it hard to see whether a dynamic VSELECT will correctly
23078 // lower, so we both check the operation's status and explicitly handle the
23079 // cases where a *dynamic* blend will fail even though a constant-condition
23080 // blend could be custom lowered.
23081 // FIXME: We should find a better way to handle this class of problems.
23082 // Potentially, we should combine constant-condition vselect nodes
23083 // pre-legalization into shuffles and not mark as many types as custom
23085 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
23087 // FIXME: We don't support i16-element blends currently. We could and
23088 // should support them by making *all* the bits in the condition be set
23089 // rather than just the high bit and using an i8-element blend.
23090 if (VT.getScalarType() == MVT::i16)
23092 // Dynamic blending was only available from SSE4.1 onward.
23093 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
23095 // Byte blends are only available in AVX2
23096 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
23097 !Subtarget->hasAVX2())
23100 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
23101 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
23103 APInt KnownZero, KnownOne;
23104 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
23105 DCI.isBeforeLegalizeOps());
23106 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
23107 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
23109 // If we changed the computation somewhere in the DAG, this change
23110 // will affect all users of Cond.
23111 // Make sure it is fine and update all the nodes so that we do not
23112 // use the generic VSELECT anymore. Otherwise, we may perform
23113 // wrong optimizations as we messed up with the actual expectation
23114 // for the vector boolean values.
23115 if (Cond != TLO.Old) {
23116 // Check all uses of that condition operand to check whether it will be
23117 // consumed by non-BLEND instructions, which may depend on all bits are
23119 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23121 if (I->getOpcode() != ISD::VSELECT)
23122 // TODO: Add other opcodes eventually lowered into BLEND.
23125 // Update all the users of the condition, before committing the change,
23126 // so that the VSELECT optimizations that expect the correct vector
23127 // boolean value will not be triggered.
23128 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23130 DAG.ReplaceAllUsesOfValueWith(
23132 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
23133 Cond, I->getOperand(1), I->getOperand(2)));
23134 DCI.CommitTargetLoweringOpt(TLO);
23137 // At this point, only Cond is changed. Change the condition
23138 // just for N to keep the opportunity to optimize all other
23139 // users their own way.
23140 DAG.ReplaceAllUsesOfValueWith(
23142 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
23143 TLO.New, N->getOperand(1), N->getOperand(2)));
23151 // Check whether a boolean test is testing a boolean value generated by
23152 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
23155 // Simplify the following patterns:
23156 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
23157 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
23158 // to (Op EFLAGS Cond)
23160 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
23161 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
23162 // to (Op EFLAGS !Cond)
23164 // where Op could be BRCOND or CMOV.
23166 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
23167 // Quit if not CMP and SUB with its value result used.
23168 if (Cmp.getOpcode() != X86ISD::CMP &&
23169 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
23172 // Quit if not used as a boolean value.
23173 if (CC != X86::COND_E && CC != X86::COND_NE)
23176 // Check CMP operands. One of them should be 0 or 1 and the other should be
23177 // an SetCC or extended from it.
23178 SDValue Op1 = Cmp.getOperand(0);
23179 SDValue Op2 = Cmp.getOperand(1);
23182 const ConstantSDNode* C = nullptr;
23183 bool needOppositeCond = (CC == X86::COND_E);
23184 bool checkAgainstTrue = false; // Is it a comparison against 1?
23186 if ((C = dyn_cast<ConstantSDNode>(Op1)))
23188 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
23190 else // Quit if all operands are not constants.
23193 if (C->getZExtValue() == 1) {
23194 needOppositeCond = !needOppositeCond;
23195 checkAgainstTrue = true;
23196 } else if (C->getZExtValue() != 0)
23197 // Quit if the constant is neither 0 or 1.
23200 bool truncatedToBoolWithAnd = false;
23201 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
23202 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
23203 SetCC.getOpcode() == ISD::TRUNCATE ||
23204 SetCC.getOpcode() == ISD::AND) {
23205 if (SetCC.getOpcode() == ISD::AND) {
23207 ConstantSDNode *CS;
23208 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
23209 CS->getZExtValue() == 1)
23211 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
23212 CS->getZExtValue() == 1)
23216 SetCC = SetCC.getOperand(OpIdx);
23217 truncatedToBoolWithAnd = true;
23219 SetCC = SetCC.getOperand(0);
23222 switch (SetCC.getOpcode()) {
23223 case X86ISD::SETCC_CARRY:
23224 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
23225 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
23226 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
23227 // truncated to i1 using 'and'.
23228 if (checkAgainstTrue && !truncatedToBoolWithAnd)
23230 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
23231 "Invalid use of SETCC_CARRY!");
23233 case X86ISD::SETCC:
23234 // Set the condition code or opposite one if necessary.
23235 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
23236 if (needOppositeCond)
23237 CC = X86::GetOppositeBranchCondition(CC);
23238 return SetCC.getOperand(1);
23239 case X86ISD::CMOV: {
23240 // Check whether false/true value has canonical one, i.e. 0 or 1.
23241 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
23242 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
23243 // Quit if true value is not a constant.
23246 // Quit if false value is not a constant.
23248 SDValue Op = SetCC.getOperand(0);
23249 // Skip 'zext' or 'trunc' node.
23250 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
23251 Op.getOpcode() == ISD::TRUNCATE)
23252 Op = Op.getOperand(0);
23253 // A special case for rdrand/rdseed, where 0 is set if false cond is
23255 if ((Op.getOpcode() != X86ISD::RDRAND &&
23256 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
23259 // Quit if false value is not the constant 0 or 1.
23260 bool FValIsFalse = true;
23261 if (FVal && FVal->getZExtValue() != 0) {
23262 if (FVal->getZExtValue() != 1)
23264 // If FVal is 1, opposite cond is needed.
23265 needOppositeCond = !needOppositeCond;
23266 FValIsFalse = false;
23268 // Quit if TVal is not the constant opposite of FVal.
23269 if (FValIsFalse && TVal->getZExtValue() != 1)
23271 if (!FValIsFalse && TVal->getZExtValue() != 0)
23273 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
23274 if (needOppositeCond)
23275 CC = X86::GetOppositeBranchCondition(CC);
23276 return SetCC.getOperand(3);
23283 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
23285 /// (X86or (X86setcc) (X86setcc))
23286 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
23287 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
23288 X86::CondCode &CC1, SDValue &Flags,
23290 if (Cond->getOpcode() == X86ISD::CMP) {
23291 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
23292 if (!CondOp1C || !CondOp1C->isNullValue())
23295 Cond = Cond->getOperand(0);
23300 SDValue SetCC0, SetCC1;
23301 switch (Cond->getOpcode()) {
23302 default: return false;
23309 SetCC0 = Cond->getOperand(0);
23310 SetCC1 = Cond->getOperand(1);
23314 // Make sure we have SETCC nodes, using the same flags value.
23315 if (SetCC0.getOpcode() != X86ISD::SETCC ||
23316 SetCC1.getOpcode() != X86ISD::SETCC ||
23317 SetCC0->getOperand(1) != SetCC1->getOperand(1))
23320 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
23321 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
23322 Flags = SetCC0->getOperand(1);
23326 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
23327 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
23328 TargetLowering::DAGCombinerInfo &DCI,
23329 const X86Subtarget *Subtarget) {
23332 // If the flag operand isn't dead, don't touch this CMOV.
23333 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
23336 SDValue FalseOp = N->getOperand(0);
23337 SDValue TrueOp = N->getOperand(1);
23338 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
23339 SDValue Cond = N->getOperand(3);
23341 if (CC == X86::COND_E || CC == X86::COND_NE) {
23342 switch (Cond.getOpcode()) {
23346 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
23347 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
23348 return (CC == X86::COND_E) ? FalseOp : TrueOp;
23354 Flags = checkBoolTestSetCCCombine(Cond, CC);
23355 if (Flags.getNode() &&
23356 // Extra check as FCMOV only supports a subset of X86 cond.
23357 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
23358 SDValue Ops[] = { FalseOp, TrueOp,
23359 DAG.getConstant(CC, DL, MVT::i8), Flags };
23360 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23363 // If this is a select between two integer constants, try to do some
23364 // optimizations. Note that the operands are ordered the opposite of SELECT
23366 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
23367 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
23368 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
23369 // larger than FalseC (the false value).
23370 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
23371 CC = X86::GetOppositeBranchCondition(CC);
23372 std::swap(TrueC, FalseC);
23373 std::swap(TrueOp, FalseOp);
23376 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
23377 // This is efficient for any integer data type (including i8/i16) and
23379 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
23380 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23381 DAG.getConstant(CC, DL, MVT::i8), Cond);
23383 // Zero extend the condition if needed.
23384 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
23386 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23387 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
23388 DAG.getConstant(ShAmt, DL, MVT::i8));
23389 if (N->getNumValues() == 2) // Dead flag value?
23390 return DCI.CombineTo(N, Cond, SDValue());
23394 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
23395 // for any integer data type, including i8/i16.
23396 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23397 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23398 DAG.getConstant(CC, DL, MVT::i8), Cond);
23400 // Zero extend the condition if needed.
23401 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23402 FalseC->getValueType(0), Cond);
23403 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23404 SDValue(FalseC, 0));
23406 if (N->getNumValues() == 2) // Dead flag value?
23407 return DCI.CombineTo(N, Cond, SDValue());
23411 // Optimize cases that will turn into an LEA instruction. This requires
23412 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23413 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23414 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23415 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23417 bool isFastMultiplier = false;
23419 switch ((unsigned char)Diff) {
23421 case 1: // result = add base, cond
23422 case 2: // result = lea base( , cond*2)
23423 case 3: // result = lea base(cond, cond*2)
23424 case 4: // result = lea base( , cond*4)
23425 case 5: // result = lea base(cond, cond*4)
23426 case 8: // result = lea base( , cond*8)
23427 case 9: // result = lea base(cond, cond*8)
23428 isFastMultiplier = true;
23433 if (isFastMultiplier) {
23434 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23435 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23436 DAG.getConstant(CC, DL, MVT::i8), Cond);
23437 // Zero extend the condition if needed.
23438 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23440 // Scale the condition by the difference.
23442 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23443 DAG.getConstant(Diff, DL, Cond.getValueType()));
23445 // Add the base if non-zero.
23446 if (FalseC->getAPIntValue() != 0)
23447 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23448 SDValue(FalseC, 0));
23449 if (N->getNumValues() == 2) // Dead flag value?
23450 return DCI.CombineTo(N, Cond, SDValue());
23457 // Handle these cases:
23458 // (select (x != c), e, c) -> select (x != c), e, x),
23459 // (select (x == c), c, e) -> select (x == c), x, e)
23460 // where the c is an integer constant, and the "select" is the combination
23461 // of CMOV and CMP.
23463 // The rationale for this change is that the conditional-move from a constant
23464 // needs two instructions, however, conditional-move from a register needs
23465 // only one instruction.
23467 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
23468 // some instruction-combining opportunities. This opt needs to be
23469 // postponed as late as possible.
23471 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
23472 // the DCI.xxxx conditions are provided to postpone the optimization as
23473 // late as possible.
23475 ConstantSDNode *CmpAgainst = nullptr;
23476 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
23477 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
23478 !isa<ConstantSDNode>(Cond.getOperand(0))) {
23480 if (CC == X86::COND_NE &&
23481 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
23482 CC = X86::GetOppositeBranchCondition(CC);
23483 std::swap(TrueOp, FalseOp);
23486 if (CC == X86::COND_E &&
23487 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
23488 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
23489 DAG.getConstant(CC, DL, MVT::i8), Cond };
23490 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
23495 // Fold and/or of setcc's to double CMOV:
23496 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
23497 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
23499 // This combine lets us generate:
23500 // cmovcc1 (jcc1 if we don't have CMOV)
23506 // cmovne (jne if we don't have CMOV)
23507 // When we can't use the CMOV instruction, it might increase branch
23509 // When we can use CMOV, or when there is no mispredict, this improves
23510 // throughput and reduces register pressure.
23512 if (CC == X86::COND_NE) {
23514 X86::CondCode CC0, CC1;
23516 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
23518 std::swap(FalseOp, TrueOp);
23519 CC0 = X86::GetOppositeBranchCondition(CC0);
23520 CC1 = X86::GetOppositeBranchCondition(CC1);
23523 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
23525 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
23526 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
23527 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23528 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
23536 /// PerformMulCombine - Optimize a single multiply with constant into two
23537 /// in order to implement it with two cheaper instructions, e.g.
23538 /// LEA + SHL, LEA + LEA.
23539 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
23540 TargetLowering::DAGCombinerInfo &DCI) {
23541 // An imul is usually smaller than the alternative sequence.
23542 if (DAG.getMachineFunction().getFunction()->optForMinSize())
23545 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
23548 EVT VT = N->getValueType(0);
23549 if (VT != MVT::i64 && VT != MVT::i32)
23552 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
23555 uint64_t MulAmt = C->getZExtValue();
23556 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
23559 uint64_t MulAmt1 = 0;
23560 uint64_t MulAmt2 = 0;
23561 if ((MulAmt % 9) == 0) {
23563 MulAmt2 = MulAmt / 9;
23564 } else if ((MulAmt % 5) == 0) {
23566 MulAmt2 = MulAmt / 5;
23567 } else if ((MulAmt % 3) == 0) {
23569 MulAmt2 = MulAmt / 3;
23572 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
23575 if (isPowerOf2_64(MulAmt2) &&
23576 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
23577 // If second multiplifer is pow2, issue it first. We want the multiply by
23578 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
23580 std::swap(MulAmt1, MulAmt2);
23583 if (isPowerOf2_64(MulAmt1))
23584 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
23585 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
23587 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
23588 DAG.getConstant(MulAmt1, DL, VT));
23590 if (isPowerOf2_64(MulAmt2))
23591 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
23592 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
23594 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
23595 DAG.getConstant(MulAmt2, DL, VT));
23597 // Do not add new nodes to DAG combiner worklist.
23598 DCI.CombineTo(N, NewMul, false);
23603 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
23604 SDValue N0 = N->getOperand(0);
23605 SDValue N1 = N->getOperand(1);
23606 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
23607 EVT VT = N0.getValueType();
23609 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
23610 // since the result of setcc_c is all zero's or all ones.
23611 if (VT.isInteger() && !VT.isVector() &&
23612 N1C && N0.getOpcode() == ISD::AND &&
23613 N0.getOperand(1).getOpcode() == ISD::Constant) {
23614 SDValue N00 = N0.getOperand(0);
23615 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
23616 APInt ShAmt = N1C->getAPIntValue();
23617 Mask = Mask.shl(ShAmt);
23618 bool MaskOK = false;
23619 // We can handle cases concerning bit-widening nodes containing setcc_c if
23620 // we carefully interrogate the mask to make sure we are semantics
23622 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
23623 // of the underlying setcc_c operation if the setcc_c was zero extended.
23624 // Consider the following example:
23625 // zext(setcc_c) -> i32 0x0000FFFF
23626 // c1 -> i32 0x0000FFFF
23627 // c2 -> i32 0x00000001
23628 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
23629 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
23630 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23632 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
23633 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
23635 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
23636 N00.getOpcode() == ISD::ANY_EXTEND) &&
23637 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
23638 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
23640 if (MaskOK && Mask != 0) {
23642 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
23646 // Hardware support for vector shifts is sparse which makes us scalarize the
23647 // vector operations in many cases. Also, on sandybridge ADD is faster than
23649 // (shl V, 1) -> add V,V
23650 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
23651 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
23652 assert(N0.getValueType().isVector() && "Invalid vector shift type");
23653 // We shift all of the values by one. In many cases we do not have
23654 // hardware support for this operation. This is better expressed as an ADD
23656 if (N1SplatC->getAPIntValue() == 1)
23657 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
23663 /// \brief Returns a vector of 0s if the node in input is a vector logical
23664 /// shift by a constant amount which is known to be bigger than or equal
23665 /// to the vector element size in bits.
23666 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
23667 const X86Subtarget *Subtarget) {
23668 EVT VT = N->getValueType(0);
23670 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
23671 (!Subtarget->hasInt256() ||
23672 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
23675 SDValue Amt = N->getOperand(1);
23677 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
23678 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
23679 APInt ShiftAmt = AmtSplat->getAPIntValue();
23680 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
23682 // SSE2/AVX2 logical shifts always return a vector of 0s
23683 // if the shift amount is bigger than or equal to
23684 // the element size. The constant shift amount will be
23685 // encoded as a 8-bit immediate.
23686 if (ShiftAmt.trunc(8).uge(MaxAmount))
23687 return getZeroVector(VT, Subtarget, DAG, DL);
23693 /// PerformShiftCombine - Combine shifts.
23694 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
23695 TargetLowering::DAGCombinerInfo &DCI,
23696 const X86Subtarget *Subtarget) {
23697 if (N->getOpcode() == ISD::SHL)
23698 if (SDValue V = PerformSHLCombine(N, DAG))
23701 // Try to fold this logical shift into a zero vector.
23702 if (N->getOpcode() != ISD::SRA)
23703 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
23709 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
23710 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
23711 // and friends. Likewise for OR -> CMPNEQSS.
23712 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
23713 TargetLowering::DAGCombinerInfo &DCI,
23714 const X86Subtarget *Subtarget) {
23717 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
23718 // we're requiring SSE2 for both.
23719 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
23720 SDValue N0 = N->getOperand(0);
23721 SDValue N1 = N->getOperand(1);
23722 SDValue CMP0 = N0->getOperand(1);
23723 SDValue CMP1 = N1->getOperand(1);
23726 // The SETCCs should both refer to the same CMP.
23727 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
23730 SDValue CMP00 = CMP0->getOperand(0);
23731 SDValue CMP01 = CMP0->getOperand(1);
23732 EVT VT = CMP00.getValueType();
23734 if (VT == MVT::f32 || VT == MVT::f64) {
23735 bool ExpectingFlags = false;
23736 // Check for any users that want flags:
23737 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
23738 !ExpectingFlags && UI != UE; ++UI)
23739 switch (UI->getOpcode()) {
23744 ExpectingFlags = true;
23746 case ISD::CopyToReg:
23747 case ISD::SIGN_EXTEND:
23748 case ISD::ZERO_EXTEND:
23749 case ISD::ANY_EXTEND:
23753 if (!ExpectingFlags) {
23754 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
23755 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
23757 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
23758 X86::CondCode tmp = cc0;
23763 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
23764 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
23765 // FIXME: need symbolic constants for these magic numbers.
23766 // See X86ATTInstPrinter.cpp:printSSECC().
23767 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
23768 if (Subtarget->hasAVX512()) {
23769 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
23771 DAG.getConstant(x86cc, DL, MVT::i8));
23772 if (N->getValueType(0) != MVT::i1)
23773 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
23777 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
23778 CMP00.getValueType(), CMP00, CMP01,
23779 DAG.getConstant(x86cc, DL,
23782 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
23783 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
23785 if (is64BitFP && !Subtarget->is64Bit()) {
23786 // On a 32-bit target, we cannot bitcast the 64-bit float to a
23787 // 64-bit integer, since that's not a legal type. Since
23788 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
23789 // bits, but can do this little dance to extract the lowest 32 bits
23790 // and work with those going forward.
23791 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
23793 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
23794 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
23795 Vector32, DAG.getIntPtrConstant(0, DL));
23799 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
23800 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
23801 DAG.getConstant(1, DL, IntVT));
23802 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
23804 return OneBitOfTruth;
23812 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
23813 /// so it can be folded inside ANDNP.
23814 static bool CanFoldXORWithAllOnes(const SDNode *N) {
23815 EVT VT = N->getValueType(0);
23817 // Match direct AllOnes for 128 and 256-bit vectors
23818 if (ISD::isBuildVectorAllOnes(N))
23821 // Look through a bit convert.
23822 if (N->getOpcode() == ISD::BITCAST)
23823 N = N->getOperand(0).getNode();
23825 // Sometimes the operand may come from a insert_subvector building a 256-bit
23827 if (VT.is256BitVector() &&
23828 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
23829 SDValue V1 = N->getOperand(0);
23830 SDValue V2 = N->getOperand(1);
23832 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
23833 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
23834 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
23835 ISD::isBuildVectorAllOnes(V2.getNode()))
23842 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
23843 // register. In most cases we actually compare or select YMM-sized registers
23844 // and mixing the two types creates horrible code. This method optimizes
23845 // some of the transition sequences.
23846 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
23847 TargetLowering::DAGCombinerInfo &DCI,
23848 const X86Subtarget *Subtarget) {
23849 EVT VT = N->getValueType(0);
23850 if (!VT.is256BitVector())
23853 assert((N->getOpcode() == ISD::ANY_EXTEND ||
23854 N->getOpcode() == ISD::ZERO_EXTEND ||
23855 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
23857 SDValue Narrow = N->getOperand(0);
23858 EVT NarrowVT = Narrow->getValueType(0);
23859 if (!NarrowVT.is128BitVector())
23862 if (Narrow->getOpcode() != ISD::XOR &&
23863 Narrow->getOpcode() != ISD::AND &&
23864 Narrow->getOpcode() != ISD::OR)
23867 SDValue N0 = Narrow->getOperand(0);
23868 SDValue N1 = Narrow->getOperand(1);
23871 // The Left side has to be a trunc.
23872 if (N0.getOpcode() != ISD::TRUNCATE)
23875 // The type of the truncated inputs.
23876 EVT WideVT = N0->getOperand(0)->getValueType(0);
23880 // The right side has to be a 'trunc' or a constant vector.
23881 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
23882 ConstantSDNode *RHSConstSplat = nullptr;
23883 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
23884 RHSConstSplat = RHSBV->getConstantSplatNode();
23885 if (!RHSTrunc && !RHSConstSplat)
23888 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23890 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
23893 // Set N0 and N1 to hold the inputs to the new wide operation.
23894 N0 = N0->getOperand(0);
23895 if (RHSConstSplat) {
23896 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
23897 SDValue(RHSConstSplat, 0));
23898 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
23899 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
23900 } else if (RHSTrunc) {
23901 N1 = N1->getOperand(0);
23904 // Generate the wide operation.
23905 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
23906 unsigned Opcode = N->getOpcode();
23908 case ISD::ANY_EXTEND:
23910 case ISD::ZERO_EXTEND: {
23911 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
23912 APInt Mask = APInt::getAllOnesValue(InBits);
23913 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
23914 return DAG.getNode(ISD::AND, DL, VT,
23915 Op, DAG.getConstant(Mask, DL, VT));
23917 case ISD::SIGN_EXTEND:
23918 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
23919 Op, DAG.getValueType(NarrowVT));
23921 llvm_unreachable("Unexpected opcode");
23925 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
23926 TargetLowering::DAGCombinerInfo &DCI,
23927 const X86Subtarget *Subtarget) {
23928 SDValue N0 = N->getOperand(0);
23929 SDValue N1 = N->getOperand(1);
23932 // A vector zext_in_reg may be represented as a shuffle,
23933 // feeding into a bitcast (this represents anyext) feeding into
23934 // an and with a mask.
23935 // We'd like to try to combine that into a shuffle with zero
23936 // plus a bitcast, removing the and.
23937 if (N0.getOpcode() != ISD::BITCAST ||
23938 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
23941 // The other side of the AND should be a splat of 2^C, where C
23942 // is the number of bits in the source type.
23943 if (N1.getOpcode() == ISD::BITCAST)
23944 N1 = N1.getOperand(0);
23945 if (N1.getOpcode() != ISD::BUILD_VECTOR)
23947 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
23949 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
23950 EVT SrcType = Shuffle->getValueType(0);
23952 // We expect a single-source shuffle
23953 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
23956 unsigned SrcSize = SrcType.getScalarSizeInBits();
23958 APInt SplatValue, SplatUndef;
23959 unsigned SplatBitSize;
23961 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
23962 SplatBitSize, HasAnyUndefs))
23965 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
23966 // Make sure the splat matches the mask we expect
23967 if (SplatBitSize > ResSize ||
23968 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
23971 // Make sure the input and output size make sense
23972 if (SrcSize >= ResSize || ResSize % SrcSize)
23975 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
23976 // The number of u's between each two values depends on the ratio between
23977 // the source and dest type.
23978 unsigned ZextRatio = ResSize / SrcSize;
23979 bool IsZext = true;
23980 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
23981 if (i % ZextRatio) {
23982 if (Shuffle->getMaskElt(i) > 0) {
23988 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
23989 // Expected element number
23999 // Ok, perform the transformation - replace the shuffle with
24000 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
24001 // (instead of undef) where the k elements come from the zero vector.
24002 SmallVector<int, 8> Mask;
24003 unsigned NumElems = SrcType.getVectorNumElements();
24004 for (unsigned i = 0; i < NumElems; ++i)
24006 Mask.push_back(NumElems);
24008 Mask.push_back(i / ZextRatio);
24010 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
24011 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
24012 return DAG.getBitcast(N0.getValueType(), NewShuffle);
24015 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24016 TargetLowering::DAGCombinerInfo &DCI,
24017 const X86Subtarget *Subtarget) {
24018 if (DCI.isBeforeLegalizeOps())
24021 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
24024 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24027 EVT VT = N->getValueType(0);
24028 SDValue N0 = N->getOperand(0);
24029 SDValue N1 = N->getOperand(1);
24032 // Create BEXTR instructions
24033 // BEXTR is ((X >> imm) & (2**size-1))
24034 if (VT == MVT::i32 || VT == MVT::i64) {
24035 // Check for BEXTR.
24036 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24037 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24038 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24039 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24040 if (MaskNode && ShiftNode) {
24041 uint64_t Mask = MaskNode->getZExtValue();
24042 uint64_t Shift = ShiftNode->getZExtValue();
24043 if (isMask_64(Mask)) {
24044 uint64_t MaskSize = countPopulation(Mask);
24045 if (Shift + MaskSize <= VT.getSizeInBits())
24046 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
24047 DAG.getConstant(Shift | (MaskSize << 8), DL,
24056 // Want to form ANDNP nodes:
24057 // 1) In the hopes of then easily combining them with OR and AND nodes
24058 // to form PBLEND/PSIGN.
24059 // 2) To match ANDN packed intrinsics
24060 if (VT != MVT::v2i64 && VT != MVT::v4i64)
24063 // Check LHS for vnot
24064 if (N0.getOpcode() == ISD::XOR &&
24065 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
24066 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
24067 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
24069 // Check RHS for vnot
24070 if (N1.getOpcode() == ISD::XOR &&
24071 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
24072 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
24073 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
24078 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
24079 TargetLowering::DAGCombinerInfo &DCI,
24080 const X86Subtarget *Subtarget) {
24081 if (DCI.isBeforeLegalizeOps())
24084 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24087 SDValue N0 = N->getOperand(0);
24088 SDValue N1 = N->getOperand(1);
24089 EVT VT = N->getValueType(0);
24091 // look for psign/blend
24092 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
24093 if (!Subtarget->hasSSSE3() ||
24094 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
24097 // Canonicalize pandn to RHS
24098 if (N0.getOpcode() == X86ISD::ANDNP)
24100 // or (and (m, y), (pandn m, x))
24101 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
24102 SDValue Mask = N1.getOperand(0);
24103 SDValue X = N1.getOperand(1);
24105 if (N0.getOperand(0) == Mask)
24106 Y = N0.getOperand(1);
24107 if (N0.getOperand(1) == Mask)
24108 Y = N0.getOperand(0);
24110 // Check to see if the mask appeared in both the AND and ANDNP and
24114 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
24115 // Look through mask bitcast.
24116 if (Mask.getOpcode() == ISD::BITCAST)
24117 Mask = Mask.getOperand(0);
24118 if (X.getOpcode() == ISD::BITCAST)
24119 X = X.getOperand(0);
24120 if (Y.getOpcode() == ISD::BITCAST)
24121 Y = Y.getOperand(0);
24123 EVT MaskVT = Mask.getValueType();
24125 // Validate that the Mask operand is a vector sra node.
24126 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
24127 // there is no psrai.b
24128 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
24129 unsigned SraAmt = ~0;
24130 if (Mask.getOpcode() == ISD::SRA) {
24131 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
24132 if (auto *AmtConst = AmtBV->getConstantSplatNode())
24133 SraAmt = AmtConst->getZExtValue();
24134 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
24135 SDValue SraC = Mask.getOperand(1);
24136 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
24138 if ((SraAmt + 1) != EltBits)
24143 // Now we know we at least have a plendvb with the mask val. See if
24144 // we can form a psignb/w/d.
24145 // psign = x.type == y.type == mask.type && y = sub(0, x);
24146 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
24147 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
24148 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
24149 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
24150 "Unsupported VT for PSIGN");
24151 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
24152 return DAG.getBitcast(VT, Mask);
24154 // PBLENDVB only available on SSE 4.1
24155 if (!Subtarget->hasSSE41())
24158 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
24160 X = DAG.getBitcast(BlendVT, X);
24161 Y = DAG.getBitcast(BlendVT, Y);
24162 Mask = DAG.getBitcast(BlendVT, Mask);
24163 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
24164 return DAG.getBitcast(VT, Mask);
24168 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
24171 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
24172 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
24174 // SHLD/SHRD instructions have lower register pressure, but on some
24175 // platforms they have higher latency than the equivalent
24176 // series of shifts/or that would otherwise be generated.
24177 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
24178 // have higher latencies and we are not optimizing for size.
24179 if (!OptForSize && Subtarget->isSHLDSlow())
24182 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
24184 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
24186 if (!N0.hasOneUse() || !N1.hasOneUse())
24189 SDValue ShAmt0 = N0.getOperand(1);
24190 if (ShAmt0.getValueType() != MVT::i8)
24192 SDValue ShAmt1 = N1.getOperand(1);
24193 if (ShAmt1.getValueType() != MVT::i8)
24195 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
24196 ShAmt0 = ShAmt0.getOperand(0);
24197 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
24198 ShAmt1 = ShAmt1.getOperand(0);
24201 unsigned Opc = X86ISD::SHLD;
24202 SDValue Op0 = N0.getOperand(0);
24203 SDValue Op1 = N1.getOperand(0);
24204 if (ShAmt0.getOpcode() == ISD::SUB) {
24205 Opc = X86ISD::SHRD;
24206 std::swap(Op0, Op1);
24207 std::swap(ShAmt0, ShAmt1);
24210 unsigned Bits = VT.getSizeInBits();
24211 if (ShAmt1.getOpcode() == ISD::SUB) {
24212 SDValue Sum = ShAmt1.getOperand(0);
24213 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
24214 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
24215 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
24216 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
24217 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
24218 return DAG.getNode(Opc, DL, VT,
24220 DAG.getNode(ISD::TRUNCATE, DL,
24223 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
24224 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
24226 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
24227 return DAG.getNode(Opc, DL, VT,
24228 N0.getOperand(0), N1.getOperand(0),
24229 DAG.getNode(ISD::TRUNCATE, DL,
24236 // Generate NEG and CMOV for integer abs.
24237 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
24238 EVT VT = N->getValueType(0);
24240 // Since X86 does not have CMOV for 8-bit integer, we don't convert
24241 // 8-bit integer abs to NEG and CMOV.
24242 if (VT.isInteger() && VT.getSizeInBits() == 8)
24245 SDValue N0 = N->getOperand(0);
24246 SDValue N1 = N->getOperand(1);
24249 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
24250 // and change it to SUB and CMOV.
24251 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
24252 N0.getOpcode() == ISD::ADD &&
24253 N0.getOperand(1) == N1 &&
24254 N1.getOpcode() == ISD::SRA &&
24255 N1.getOperand(0) == N0.getOperand(0))
24256 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
24257 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
24258 // Generate SUB & CMOV.
24259 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
24260 DAG.getConstant(0, DL, VT), N0.getOperand(0));
24262 SDValue Ops[] = { N0.getOperand(0), Neg,
24263 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
24264 SDValue(Neg.getNode(), 1) };
24265 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
24270 // Try to turn tests against the signbit in the form of:
24271 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
24274 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
24275 // This is only worth doing if the output type is i8.
24276 if (N->getValueType(0) != MVT::i8)
24279 SDValue N0 = N->getOperand(0);
24280 SDValue N1 = N->getOperand(1);
24282 // We should be performing an xor against a truncated shift.
24283 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
24286 // Make sure we are performing an xor against one.
24287 if (!isa<ConstantSDNode>(N1) || !cast<ConstantSDNode>(N1)->isOne())
24290 // SetCC on x86 zero extends so only act on this if it's a logical shift.
24291 SDValue Shift = N0.getOperand(0);
24292 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
24295 // Make sure we are truncating from one of i16, i32 or i64.
24296 EVT ShiftTy = Shift.getValueType();
24297 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
24300 // Make sure the shift amount extracts the sign bit.
24301 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
24302 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
24305 // Create a greater-than comparison against -1.
24306 // N.B. Using SETGE against 0 works but we want a canonical looking
24307 // comparison, using SETGT matches up with what TranslateX86CC.
24309 SDValue ShiftOp = Shift.getOperand(0);
24310 EVT ShiftOpTy = ShiftOp.getValueType();
24311 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
24312 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
24316 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
24317 TargetLowering::DAGCombinerInfo &DCI,
24318 const X86Subtarget *Subtarget) {
24319 if (DCI.isBeforeLegalizeOps())
24322 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
24325 if (Subtarget->hasCMov())
24326 if (SDValue RV = performIntegerAbsCombine(N, DAG))
24332 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
24333 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
24334 TargetLowering::DAGCombinerInfo &DCI,
24335 const X86Subtarget *Subtarget) {
24336 LoadSDNode *Ld = cast<LoadSDNode>(N);
24337 EVT RegVT = Ld->getValueType(0);
24338 EVT MemVT = Ld->getMemoryVT();
24340 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24342 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
24343 // into two 16-byte operations.
24344 ISD::LoadExtType Ext = Ld->getExtensionType();
24346 unsigned AddressSpace = Ld->getAddressSpace();
24347 unsigned Alignment = Ld->getAlignment();
24348 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
24349 Ext == ISD::NON_EXTLOAD &&
24350 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
24351 AddressSpace, Alignment, &Fast) && !Fast) {
24352 unsigned NumElems = RegVT.getVectorNumElements();
24356 SDValue Ptr = Ld->getBasePtr();
24357 SDValue Increment =
24358 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24360 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
24362 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24363 Ld->getPointerInfo(), Ld->isVolatile(),
24364 Ld->isNonTemporal(), Ld->isInvariant(),
24366 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24367 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24368 Ld->getPointerInfo(), Ld->isVolatile(),
24369 Ld->isNonTemporal(), Ld->isInvariant(),
24370 std::min(16U, Alignment));
24371 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
24373 Load2.getValue(1));
24375 SDValue NewVec = DAG.getUNDEF(RegVT);
24376 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
24377 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
24378 return DCI.CombineTo(N, NewVec, TF, true);
24384 /// PerformMLOADCombine - Resolve extending loads
24385 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
24386 TargetLowering::DAGCombinerInfo &DCI,
24387 const X86Subtarget *Subtarget) {
24388 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
24389 if (Mld->getExtensionType() != ISD::SEXTLOAD)
24392 EVT VT = Mld->getValueType(0);
24393 unsigned NumElems = VT.getVectorNumElements();
24394 EVT LdVT = Mld->getMemoryVT();
24397 assert(LdVT != VT && "Cannot extend to the same type");
24398 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
24399 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
24400 // From, To sizes and ElemCount must be pow of two
24401 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24402 "Unexpected size for extending masked load");
24404 unsigned SizeRatio = ToSz / FromSz;
24405 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
24407 // Create a type on which we perform the shuffle
24408 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24409 LdVT.getScalarType(), NumElems*SizeRatio);
24410 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24412 // Convert Src0 value
24413 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
24414 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
24415 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24416 for (unsigned i = 0; i != NumElems; ++i)
24417 ShuffleVec[i] = i * SizeRatio;
24419 // Can't shuffle using an illegal type.
24420 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24421 && "WideVecVT should be legal");
24422 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
24423 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
24425 // Prepare the new mask
24427 SDValue Mask = Mld->getMask();
24428 if (Mask.getValueType() == VT) {
24429 // Mask and original value have the same type
24430 NewMask = DAG.getBitcast(WideVecVT, Mask);
24431 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24432 for (unsigned i = 0; i != NumElems; ++i)
24433 ShuffleVec[i] = i * SizeRatio;
24434 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24435 ShuffleVec[i] = NumElems*SizeRatio;
24436 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24437 DAG.getConstant(0, dl, WideVecVT),
24441 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24442 unsigned WidenNumElts = NumElems*SizeRatio;
24443 unsigned MaskNumElts = VT.getVectorNumElements();
24444 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24447 unsigned NumConcat = WidenNumElts / MaskNumElts;
24448 SmallVector<SDValue, 16> Ops(NumConcat);
24449 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24451 for (unsigned i = 1; i != NumConcat; ++i)
24454 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24457 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
24458 Mld->getBasePtr(), NewMask, WideSrc0,
24459 Mld->getMemoryVT(), Mld->getMemOperand(),
24461 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
24462 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
24465 /// PerformMSTORECombine - Resolve truncating stores
24466 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
24467 const X86Subtarget *Subtarget) {
24468 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
24469 if (!Mst->isTruncatingStore())
24472 EVT VT = Mst->getValue().getValueType();
24473 unsigned NumElems = VT.getVectorNumElements();
24474 EVT StVT = Mst->getMemoryVT();
24477 assert(StVT != VT && "Cannot truncate to the same type");
24478 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24479 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24481 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24483 // The truncating store is legal in some cases. For example
24484 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24485 // are designated for truncate store.
24486 // In this case we don't need any further transformations.
24487 if (TLI.isTruncStoreLegal(VT, StVT))
24490 // From, To sizes and ElemCount must be pow of two
24491 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24492 "Unexpected size for truncating masked store");
24493 // We are going to use the original vector elt for storing.
24494 // Accumulated smaller vector elements must be a multiple of the store size.
24495 assert (((NumElems * FromSz) % ToSz) == 0 &&
24496 "Unexpected ratio for truncating masked store");
24498 unsigned SizeRatio = FromSz / ToSz;
24499 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24501 // Create a type on which we perform the shuffle
24502 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24503 StVT.getScalarType(), NumElems*SizeRatio);
24505 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24507 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
24508 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24509 for (unsigned i = 0; i != NumElems; ++i)
24510 ShuffleVec[i] = i * SizeRatio;
24512 // Can't shuffle using an illegal type.
24513 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24514 && "WideVecVT should be legal");
24516 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24517 DAG.getUNDEF(WideVecVT),
24521 SDValue Mask = Mst->getMask();
24522 if (Mask.getValueType() == VT) {
24523 // Mask and original value have the same type
24524 NewMask = DAG.getBitcast(WideVecVT, Mask);
24525 for (unsigned i = 0; i != NumElems; ++i)
24526 ShuffleVec[i] = i * SizeRatio;
24527 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24528 ShuffleVec[i] = NumElems*SizeRatio;
24529 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24530 DAG.getConstant(0, dl, WideVecVT),
24534 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24535 unsigned WidenNumElts = NumElems*SizeRatio;
24536 unsigned MaskNumElts = VT.getVectorNumElements();
24537 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24540 unsigned NumConcat = WidenNumElts / MaskNumElts;
24541 SmallVector<SDValue, 16> Ops(NumConcat);
24542 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24544 for (unsigned i = 1; i != NumConcat; ++i)
24547 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24550 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
24551 NewMask, StVT, Mst->getMemOperand(), false);
24553 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
24554 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
24555 const X86Subtarget *Subtarget) {
24556 StoreSDNode *St = cast<StoreSDNode>(N);
24557 EVT VT = St->getValue().getValueType();
24558 EVT StVT = St->getMemoryVT();
24560 SDValue StoredVal = St->getOperand(1);
24561 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24563 // If we are saving a concatenation of two XMM registers and 32-byte stores
24564 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
24566 unsigned AddressSpace = St->getAddressSpace();
24567 unsigned Alignment = St->getAlignment();
24568 if (VT.is256BitVector() && StVT == VT &&
24569 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
24570 AddressSpace, Alignment, &Fast) && !Fast) {
24571 unsigned NumElems = VT.getVectorNumElements();
24575 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
24576 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
24579 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24580 SDValue Ptr0 = St->getBasePtr();
24581 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
24583 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
24584 St->getPointerInfo(), St->isVolatile(),
24585 St->isNonTemporal(), Alignment);
24586 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
24587 St->getPointerInfo(), St->isVolatile(),
24588 St->isNonTemporal(),
24589 std::min(16U, Alignment));
24590 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
24593 // Optimize trunc store (of multiple scalars) to shuffle and store.
24594 // First, pack all of the elements in one place. Next, store to memory
24595 // in fewer chunks.
24596 if (St->isTruncatingStore() && VT.isVector()) {
24597 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24598 unsigned NumElems = VT.getVectorNumElements();
24599 assert(StVT != VT && "Cannot truncate to the same type");
24600 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24601 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24603 // The truncating store is legal in some cases. For example
24604 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24605 // are designated for truncate store.
24606 // In this case we don't need any further transformations.
24607 if (TLI.isTruncStoreLegal(VT, StVT))
24610 // From, To sizes and ElemCount must be pow of two
24611 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
24612 // We are going to use the original vector elt for storing.
24613 // Accumulated smaller vector elements must be a multiple of the store size.
24614 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
24616 unsigned SizeRatio = FromSz / ToSz;
24618 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24620 // Create a type on which we perform the shuffle
24621 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24622 StVT.getScalarType(), NumElems*SizeRatio);
24624 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24626 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
24627 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
24628 for (unsigned i = 0; i != NumElems; ++i)
24629 ShuffleVec[i] = i * SizeRatio;
24631 // Can't shuffle using an illegal type.
24632 if (!TLI.isTypeLegal(WideVecVT))
24635 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24636 DAG.getUNDEF(WideVecVT),
24638 // At this point all of the data is stored at the bottom of the
24639 // register. We now need to save it to mem.
24641 // Find the largest store unit
24642 MVT StoreType = MVT::i8;
24643 for (MVT Tp : MVT::integer_valuetypes()) {
24644 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
24648 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
24649 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
24650 (64 <= NumElems * ToSz))
24651 StoreType = MVT::f64;
24653 // Bitcast the original vector into a vector of store-size units
24654 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
24655 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
24656 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
24657 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
24658 SmallVector<SDValue, 8> Chains;
24659 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
24660 TLI.getPointerTy(DAG.getDataLayout()));
24661 SDValue Ptr = St->getBasePtr();
24663 // Perform one or more big stores into memory.
24664 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
24665 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
24666 StoreType, ShuffWide,
24667 DAG.getIntPtrConstant(i, dl));
24668 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
24669 St->getPointerInfo(), St->isVolatile(),
24670 St->isNonTemporal(), St->getAlignment());
24671 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24672 Chains.push_back(Ch);
24675 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
24678 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
24679 // the FP state in cases where an emms may be missing.
24680 // A preferable solution to the general problem is to figure out the right
24681 // places to insert EMMS. This qualifies as a quick hack.
24683 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
24684 if (VT.getSizeInBits() != 64)
24687 const Function *F = DAG.getMachineFunction().getFunction();
24688 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
24690 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
24691 if ((VT.isVector() ||
24692 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
24693 isa<LoadSDNode>(St->getValue()) &&
24694 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
24695 St->getChain().hasOneUse() && !St->isVolatile()) {
24696 SDNode* LdVal = St->getValue().getNode();
24697 LoadSDNode *Ld = nullptr;
24698 int TokenFactorIndex = -1;
24699 SmallVector<SDValue, 8> Ops;
24700 SDNode* ChainVal = St->getChain().getNode();
24701 // Must be a store of a load. We currently handle two cases: the load
24702 // is a direct child, and it's under an intervening TokenFactor. It is
24703 // possible to dig deeper under nested TokenFactors.
24704 if (ChainVal == LdVal)
24705 Ld = cast<LoadSDNode>(St->getChain());
24706 else if (St->getValue().hasOneUse() &&
24707 ChainVal->getOpcode() == ISD::TokenFactor) {
24708 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
24709 if (ChainVal->getOperand(i).getNode() == LdVal) {
24710 TokenFactorIndex = i;
24711 Ld = cast<LoadSDNode>(St->getValue());
24713 Ops.push_back(ChainVal->getOperand(i));
24717 if (!Ld || !ISD::isNormalLoad(Ld))
24720 // If this is not the MMX case, i.e. we are just turning i64 load/store
24721 // into f64 load/store, avoid the transformation if there are multiple
24722 // uses of the loaded value.
24723 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
24728 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
24729 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
24731 if (Subtarget->is64Bit() || F64IsLegal) {
24732 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
24733 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
24734 Ld->getPointerInfo(), Ld->isVolatile(),
24735 Ld->isNonTemporal(), Ld->isInvariant(),
24736 Ld->getAlignment());
24737 SDValue NewChain = NewLd.getValue(1);
24738 if (TokenFactorIndex != -1) {
24739 Ops.push_back(NewChain);
24740 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24742 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
24743 St->getPointerInfo(),
24744 St->isVolatile(), St->isNonTemporal(),
24745 St->getAlignment());
24748 // Otherwise, lower to two pairs of 32-bit loads / stores.
24749 SDValue LoAddr = Ld->getBasePtr();
24750 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
24751 DAG.getConstant(4, LdDL, MVT::i32));
24753 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
24754 Ld->getPointerInfo(),
24755 Ld->isVolatile(), Ld->isNonTemporal(),
24756 Ld->isInvariant(), Ld->getAlignment());
24757 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
24758 Ld->getPointerInfo().getWithOffset(4),
24759 Ld->isVolatile(), Ld->isNonTemporal(),
24761 MinAlign(Ld->getAlignment(), 4));
24763 SDValue NewChain = LoLd.getValue(1);
24764 if (TokenFactorIndex != -1) {
24765 Ops.push_back(LoLd);
24766 Ops.push_back(HiLd);
24767 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24770 LoAddr = St->getBasePtr();
24771 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
24772 DAG.getConstant(4, StDL, MVT::i32));
24774 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
24775 St->getPointerInfo(),
24776 St->isVolatile(), St->isNonTemporal(),
24777 St->getAlignment());
24778 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
24779 St->getPointerInfo().getWithOffset(4),
24781 St->isNonTemporal(),
24782 MinAlign(St->getAlignment(), 4));
24783 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
24786 // This is similar to the above case, but here we handle a scalar 64-bit
24787 // integer store that is extracted from a vector on a 32-bit target.
24788 // If we have SSE2, then we can treat it like a floating-point double
24789 // to get past legalization. The execution dependencies fixup pass will
24790 // choose the optimal machine instruction for the store if this really is
24791 // an integer or v2f32 rather than an f64.
24792 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
24793 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
24794 SDValue OldExtract = St->getOperand(1);
24795 SDValue ExtOp0 = OldExtract.getOperand(0);
24796 unsigned VecSize = ExtOp0.getValueSizeInBits();
24797 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
24798 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
24799 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
24800 BitCast, OldExtract.getOperand(1));
24801 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
24802 St->getPointerInfo(), St->isVolatile(),
24803 St->isNonTemporal(), St->getAlignment());
24809 /// Return 'true' if this vector operation is "horizontal"
24810 /// and return the operands for the horizontal operation in LHS and RHS. A
24811 /// horizontal operation performs the binary operation on successive elements
24812 /// of its first operand, then on successive elements of its second operand,
24813 /// returning the resulting values in a vector. For example, if
24814 /// A = < float a0, float a1, float a2, float a3 >
24816 /// B = < float b0, float b1, float b2, float b3 >
24817 /// then the result of doing a horizontal operation on A and B is
24818 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
24819 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
24820 /// A horizontal-op B, for some already available A and B, and if so then LHS is
24821 /// set to A, RHS to B, and the routine returns 'true'.
24822 /// Note that the binary operation should have the property that if one of the
24823 /// operands is UNDEF then the result is UNDEF.
24824 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
24825 // Look for the following pattern: if
24826 // A = < float a0, float a1, float a2, float a3 >
24827 // B = < float b0, float b1, float b2, float b3 >
24829 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
24830 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
24831 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
24832 // which is A horizontal-op B.
24834 // At least one of the operands should be a vector shuffle.
24835 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
24836 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
24839 MVT VT = LHS.getSimpleValueType();
24841 assert((VT.is128BitVector() || VT.is256BitVector()) &&
24842 "Unsupported vector type for horizontal add/sub");
24844 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
24845 // operate independently on 128-bit lanes.
24846 unsigned NumElts = VT.getVectorNumElements();
24847 unsigned NumLanes = VT.getSizeInBits()/128;
24848 unsigned NumLaneElts = NumElts / NumLanes;
24849 assert((NumLaneElts % 2 == 0) &&
24850 "Vector type should have an even number of elements in each lane");
24851 unsigned HalfLaneElts = NumLaneElts/2;
24853 // View LHS in the form
24854 // LHS = VECTOR_SHUFFLE A, B, LMask
24855 // If LHS is not a shuffle then pretend it is the shuffle
24856 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
24857 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
24860 SmallVector<int, 16> LMask(NumElts);
24861 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
24862 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
24863 A = LHS.getOperand(0);
24864 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
24865 B = LHS.getOperand(1);
24866 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
24867 std::copy(Mask.begin(), Mask.end(), LMask.begin());
24869 if (LHS.getOpcode() != ISD::UNDEF)
24871 for (unsigned i = 0; i != NumElts; ++i)
24875 // Likewise, view RHS in the form
24876 // RHS = VECTOR_SHUFFLE C, D, RMask
24878 SmallVector<int, 16> RMask(NumElts);
24879 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
24880 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
24881 C = RHS.getOperand(0);
24882 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
24883 D = RHS.getOperand(1);
24884 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
24885 std::copy(Mask.begin(), Mask.end(), RMask.begin());
24887 if (RHS.getOpcode() != ISD::UNDEF)
24889 for (unsigned i = 0; i != NumElts; ++i)
24893 // Check that the shuffles are both shuffling the same vectors.
24894 if (!(A == C && B == D) && !(A == D && B == C))
24897 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
24898 if (!A.getNode() && !B.getNode())
24901 // If A and B occur in reverse order in RHS, then "swap" them (which means
24902 // rewriting the mask).
24904 ShuffleVectorSDNode::commuteMask(RMask);
24906 // At this point LHS and RHS are equivalent to
24907 // LHS = VECTOR_SHUFFLE A, B, LMask
24908 // RHS = VECTOR_SHUFFLE A, B, RMask
24909 // Check that the masks correspond to performing a horizontal operation.
24910 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
24911 for (unsigned i = 0; i != NumLaneElts; ++i) {
24912 int LIdx = LMask[i+l], RIdx = RMask[i+l];
24914 // Ignore any UNDEF components.
24915 if (LIdx < 0 || RIdx < 0 ||
24916 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
24917 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
24920 // Check that successive elements are being operated on. If not, this is
24921 // not a horizontal operation.
24922 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
24923 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
24924 if (!(LIdx == Index && RIdx == Index + 1) &&
24925 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
24930 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
24931 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
24935 /// Do target-specific dag combines on floating point adds.
24936 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
24937 const X86Subtarget *Subtarget) {
24938 EVT VT = N->getValueType(0);
24939 SDValue LHS = N->getOperand(0);
24940 SDValue RHS = N->getOperand(1);
24942 // Try to synthesize horizontal adds from adds of shuffles.
24943 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
24944 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
24945 isHorizontalBinOp(LHS, RHS, true))
24946 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
24950 /// Do target-specific dag combines on floating point subs.
24951 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
24952 const X86Subtarget *Subtarget) {
24953 EVT VT = N->getValueType(0);
24954 SDValue LHS = N->getOperand(0);
24955 SDValue RHS = N->getOperand(1);
24957 // Try to synthesize horizontal subs from subs of shuffles.
24958 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
24959 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
24960 isHorizontalBinOp(LHS, RHS, false))
24961 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
24965 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
24966 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
24967 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
24969 // F[X]OR(0.0, x) -> x
24970 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
24971 if (C->getValueAPF().isPosZero())
24972 return N->getOperand(1);
24974 // F[X]OR(x, 0.0) -> x
24975 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
24976 if (C->getValueAPF().isPosZero())
24977 return N->getOperand(0);
24981 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
24982 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
24983 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
24985 // Only perform optimizations if UnsafeMath is used.
24986 if (!DAG.getTarget().Options.UnsafeFPMath)
24989 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
24990 // into FMINC and FMAXC, which are Commutative operations.
24991 unsigned NewOp = 0;
24992 switch (N->getOpcode()) {
24993 default: llvm_unreachable("unknown opcode");
24994 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
24995 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
24998 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
24999 N->getOperand(0), N->getOperand(1));
25002 /// Do target-specific dag combines on X86ISD::FAND nodes.
25003 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25004 // FAND(0.0, x) -> 0.0
25005 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25006 if (C->getValueAPF().isPosZero())
25007 return N->getOperand(0);
25009 // FAND(x, 0.0) -> 0.0
25010 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25011 if (C->getValueAPF().isPosZero())
25012 return N->getOperand(1);
25017 /// Do target-specific dag combines on X86ISD::FANDN nodes
25018 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25019 // FANDN(0.0, x) -> x
25020 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25021 if (C->getValueAPF().isPosZero())
25022 return N->getOperand(1);
25024 // FANDN(x, 0.0) -> 0.0
25025 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25026 if (C->getValueAPF().isPosZero())
25027 return N->getOperand(1);
25032 static SDValue PerformBTCombine(SDNode *N,
25034 TargetLowering::DAGCombinerInfo &DCI) {
25035 // BT ignores high bits in the bit index operand.
25036 SDValue Op1 = N->getOperand(1);
25037 if (Op1.hasOneUse()) {
25038 unsigned BitWidth = Op1.getValueSizeInBits();
25039 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25040 APInt KnownZero, KnownOne;
25041 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
25042 !DCI.isBeforeLegalizeOps());
25043 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25044 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
25045 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
25046 DCI.CommitTargetLoweringOpt(TLO);
25051 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
25052 SDValue Op = N->getOperand(0);
25053 if (Op.getOpcode() == ISD::BITCAST)
25054 Op = Op.getOperand(0);
25055 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
25056 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
25057 VT.getVectorElementType().getSizeInBits() ==
25058 OpVT.getVectorElementType().getSizeInBits()) {
25059 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
25064 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
25065 const X86Subtarget *Subtarget) {
25066 EVT VT = N->getValueType(0);
25067 if (!VT.isVector())
25070 SDValue N0 = N->getOperand(0);
25071 SDValue N1 = N->getOperand(1);
25072 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
25075 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
25076 // both SSE and AVX2 since there is no sign-extended shift right
25077 // operation on a vector with 64-bit elements.
25078 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
25079 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
25080 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
25081 N0.getOpcode() == ISD::SIGN_EXTEND)) {
25082 SDValue N00 = N0.getOperand(0);
25084 // EXTLOAD has a better solution on AVX2,
25085 // it may be replaced with X86ISD::VSEXT node.
25086 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
25087 if (!ISD::isNormalLoad(N00.getNode()))
25090 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
25091 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
25093 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
25099 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
25100 TargetLowering::DAGCombinerInfo &DCI,
25101 const X86Subtarget *Subtarget) {
25102 SDValue N0 = N->getOperand(0);
25103 EVT VT = N->getValueType(0);
25104 EVT SVT = VT.getScalarType();
25105 EVT InVT = N0.getValueType();
25106 EVT InSVT = InVT.getScalarType();
25109 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
25110 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
25111 // This exposes the sext to the sdivrem lowering, so that it directly extends
25112 // from AH (which we otherwise need to do contortions to access).
25113 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
25114 InVT == MVT::i8 && VT == MVT::i32) {
25115 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25116 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
25117 N0.getOperand(0), N0.getOperand(1));
25118 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25119 return R.getValue(1);
25122 if (!DCI.isBeforeLegalizeOps()) {
25123 if (InVT == MVT::i1) {
25124 SDValue Zero = DAG.getConstant(0, DL, VT);
25126 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
25127 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
25132 if (VT.isVector() && Subtarget->hasSSE2()) {
25133 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
25134 EVT InVT = N.getValueType();
25135 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
25136 Size / InVT.getScalarSizeInBits());
25137 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
25138 DAG.getUNDEF(InVT));
25140 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
25143 // If target-size is less than 128-bits, extend to a type that would extend
25144 // to 128 bits, extend that and extract the original target vector.
25145 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
25146 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25147 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25148 unsigned Scale = 128 / VT.getSizeInBits();
25150 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
25151 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
25152 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
25153 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
25154 DAG.getIntPtrConstant(0, DL));
25157 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
25158 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
25159 if (VT.getSizeInBits() == 128 &&
25160 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25161 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25162 SDValue ExOp = ExtendVecSize(DL, N0, 128);
25163 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
25166 // On pre-AVX2 targets, split into 128-bit nodes of
25167 // ISD::SIGN_EXTEND_VECTOR_INREG.
25168 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
25169 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25170 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25171 unsigned NumVecs = VT.getSizeInBits() / 128;
25172 unsigned NumSubElts = 128 / SVT.getSizeInBits();
25173 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
25174 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
25176 SmallVector<SDValue, 8> Opnds;
25177 for (unsigned i = 0, Offset = 0; i != NumVecs;
25178 ++i, Offset += NumSubElts) {
25179 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
25180 DAG.getIntPtrConstant(Offset, DL));
25181 SrcVec = ExtendVecSize(DL, SrcVec, 128);
25182 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
25183 Opnds.push_back(SrcVec);
25185 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
25189 if (!Subtarget->hasFp256())
25192 if (VT.isVector() && VT.getSizeInBits() == 256)
25193 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
25199 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
25200 const X86Subtarget* Subtarget) {
25202 EVT VT = N->getValueType(0);
25204 // Let legalize expand this if it isn't a legal type yet.
25205 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
25208 EVT ScalarVT = VT.getScalarType();
25209 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
25210 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
25211 !Subtarget->hasAVX512()))
25214 SDValue A = N->getOperand(0);
25215 SDValue B = N->getOperand(1);
25216 SDValue C = N->getOperand(2);
25218 bool NegA = (A.getOpcode() == ISD::FNEG);
25219 bool NegB = (B.getOpcode() == ISD::FNEG);
25220 bool NegC = (C.getOpcode() == ISD::FNEG);
25222 // Negative multiplication when NegA xor NegB
25223 bool NegMul = (NegA != NegB);
25225 A = A.getOperand(0);
25227 B = B.getOperand(0);
25229 C = C.getOperand(0);
25233 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
25235 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
25237 return DAG.getNode(Opcode, dl, VT, A, B, C);
25240 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
25241 TargetLowering::DAGCombinerInfo &DCI,
25242 const X86Subtarget *Subtarget) {
25243 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
25244 // (and (i32 x86isd::setcc_carry), 1)
25245 // This eliminates the zext. This transformation is necessary because
25246 // ISD::SETCC is always legalized to i8.
25248 SDValue N0 = N->getOperand(0);
25249 EVT VT = N->getValueType(0);
25251 if (N0.getOpcode() == ISD::AND &&
25253 N0.getOperand(0).hasOneUse()) {
25254 SDValue N00 = N0.getOperand(0);
25255 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25256 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
25257 if (!C || C->getZExtValue() != 1)
25259 return DAG.getNode(ISD::AND, dl, VT,
25260 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
25261 N00.getOperand(0), N00.getOperand(1)),
25262 DAG.getConstant(1, dl, VT));
25266 if (N0.getOpcode() == ISD::TRUNCATE &&
25268 N0.getOperand(0).hasOneUse()) {
25269 SDValue N00 = N0.getOperand(0);
25270 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25271 return DAG.getNode(ISD::AND, dl, VT,
25272 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
25273 N00.getOperand(0), N00.getOperand(1)),
25274 DAG.getConstant(1, dl, VT));
25278 if (VT.is256BitVector())
25279 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
25282 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
25283 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
25284 // This exposes the zext to the udivrem lowering, so that it directly extends
25285 // from AH (which we otherwise need to do contortions to access).
25286 if (N0.getOpcode() == ISD::UDIVREM &&
25287 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
25288 (VT == MVT::i32 || VT == MVT::i64)) {
25289 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25290 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
25291 N0.getOperand(0), N0.getOperand(1));
25292 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25293 return R.getValue(1);
25299 // Optimize x == -y --> x+y == 0
25300 // x != -y --> x+y != 0
25301 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
25302 const X86Subtarget* Subtarget) {
25303 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
25304 SDValue LHS = N->getOperand(0);
25305 SDValue RHS = N->getOperand(1);
25306 EVT VT = N->getValueType(0);
25309 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
25310 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
25311 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
25312 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
25313 LHS.getOperand(1));
25314 return DAG.getSetCC(DL, N->getValueType(0), addV,
25315 DAG.getConstant(0, DL, addV.getValueType()), CC);
25317 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
25318 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
25319 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
25320 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
25321 RHS.getOperand(1));
25322 return DAG.getSetCC(DL, N->getValueType(0), addV,
25323 DAG.getConstant(0, DL, addV.getValueType()), CC);
25326 if (VT.getScalarType() == MVT::i1 &&
25327 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
25329 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25330 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25331 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25333 if (!IsSEXT0 || !IsVZero1) {
25334 // Swap the operands and update the condition code.
25335 std::swap(LHS, RHS);
25336 CC = ISD::getSetCCSwappedOperands(CC);
25338 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25339 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25340 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25343 if (IsSEXT0 && IsVZero1) {
25344 assert(VT == LHS.getOperand(0).getValueType() &&
25345 "Uexpected operand type");
25346 if (CC == ISD::SETGT)
25347 return DAG.getConstant(0, DL, VT);
25348 if (CC == ISD::SETLE)
25349 return DAG.getConstant(1, DL, VT);
25350 if (CC == ISD::SETEQ || CC == ISD::SETGE)
25351 return DAG.getNOT(DL, LHS.getOperand(0), VT);
25353 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
25354 "Unexpected condition code!");
25355 return LHS.getOperand(0);
25362 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
25363 SelectionDAG &DAG) {
25365 MVT VT = Load->getSimpleValueType(0);
25366 MVT EVT = VT.getVectorElementType();
25367 SDValue Addr = Load->getOperand(1);
25368 SDValue NewAddr = DAG.getNode(
25369 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
25370 DAG.getConstant(Index * EVT.getStoreSize(), dl,
25371 Addr.getSimpleValueType()));
25374 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
25375 DAG.getMachineFunction().getMachineMemOperand(
25376 Load->getMemOperand(), 0, EVT.getStoreSize()));
25380 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
25381 const X86Subtarget *Subtarget) {
25383 MVT VT = N->getOperand(1)->getSimpleValueType(0);
25384 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
25385 "X86insertps is only defined for v4x32");
25387 SDValue Ld = N->getOperand(1);
25388 if (MayFoldLoad(Ld)) {
25389 // Extract the countS bits from the immediate so we can get the proper
25390 // address when narrowing the vector load to a specific element.
25391 // When the second source op is a memory address, insertps doesn't use
25392 // countS and just gets an f32 from that address.
25393 unsigned DestIndex =
25394 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
25396 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
25398 // Create this as a scalar to vector to match the instruction pattern.
25399 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
25400 // countS bits are ignored when loading from memory on insertps, which
25401 // means we don't need to explicitly set them to 0.
25402 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
25403 LoadScalarToVector, N->getOperand(2));
25408 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
25409 SDValue V0 = N->getOperand(0);
25410 SDValue V1 = N->getOperand(1);
25412 EVT VT = N->getValueType(0);
25414 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
25415 // operands and changing the mask to 1. This saves us a bunch of
25416 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
25417 // x86InstrInfo knows how to commute this back after instruction selection
25418 // if it would help register allocation.
25420 // TODO: If optimizing for size or a processor that doesn't suffer from
25421 // partial register update stalls, this should be transformed into a MOVSD
25422 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
25424 if (VT == MVT::v2f64)
25425 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
25426 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
25427 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
25428 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
25434 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
25435 // as "sbb reg,reg", since it can be extended without zext and produces
25436 // an all-ones bit which is more useful than 0/1 in some cases.
25437 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
25440 return DAG.getNode(ISD::AND, DL, VT,
25441 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25442 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25444 DAG.getConstant(1, DL, VT));
25445 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
25446 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
25447 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25448 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25452 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
25453 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
25454 TargetLowering::DAGCombinerInfo &DCI,
25455 const X86Subtarget *Subtarget) {
25457 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
25458 SDValue EFLAGS = N->getOperand(1);
25460 if (CC == X86::COND_A) {
25461 // Try to convert COND_A into COND_B in an attempt to facilitate
25462 // materializing "setb reg".
25464 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
25465 // cannot take an immediate as its first operand.
25467 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
25468 EFLAGS.getValueType().isInteger() &&
25469 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
25470 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
25471 EFLAGS.getNode()->getVTList(),
25472 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
25473 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
25474 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
25478 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
25479 // a zext and produces an all-ones bit which is more useful than 0/1 in some
25481 if (CC == X86::COND_B)
25482 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
25484 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25485 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25486 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
25492 // Optimize branch condition evaluation.
25494 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
25495 TargetLowering::DAGCombinerInfo &DCI,
25496 const X86Subtarget *Subtarget) {
25498 SDValue Chain = N->getOperand(0);
25499 SDValue Dest = N->getOperand(1);
25500 SDValue EFLAGS = N->getOperand(3);
25501 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
25503 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25504 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25505 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
25512 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
25513 SelectionDAG &DAG) {
25514 // Take advantage of vector comparisons producing 0 or -1 in each lane to
25515 // optimize away operation when it's from a constant.
25517 // The general transformation is:
25518 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
25519 // AND(VECTOR_CMP(x,y), constant2)
25520 // constant2 = UNARYOP(constant)
25522 // Early exit if this isn't a vector operation, the operand of the
25523 // unary operation isn't a bitwise AND, or if the sizes of the operations
25524 // aren't the same.
25525 EVT VT = N->getValueType(0);
25526 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
25527 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
25528 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
25531 // Now check that the other operand of the AND is a constant. We could
25532 // make the transformation for non-constant splats as well, but it's unclear
25533 // that would be a benefit as it would not eliminate any operations, just
25534 // perform one more step in scalar code before moving to the vector unit.
25535 if (BuildVectorSDNode *BV =
25536 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
25537 // Bail out if the vector isn't a constant.
25538 if (!BV->isConstant())
25541 // Everything checks out. Build up the new and improved node.
25543 EVT IntVT = BV->getValueType(0);
25544 // Create a new constant of the appropriate type for the transformed
25546 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
25547 // The AND node needs bitcasts to/from an integer vector type around it.
25548 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
25549 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
25550 N->getOperand(0)->getOperand(0), MaskConst);
25551 SDValue Res = DAG.getBitcast(VT, NewAnd);
25558 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25559 const X86Subtarget *Subtarget) {
25560 SDValue Op0 = N->getOperand(0);
25561 EVT VT = N->getValueType(0);
25562 EVT InVT = Op0.getValueType();
25563 EVT InSVT = InVT.getScalarType();
25564 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25566 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
25567 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
25568 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25570 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25571 InVT.getVectorNumElements());
25572 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
25574 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
25575 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
25577 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25583 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25584 const X86Subtarget *Subtarget) {
25585 // First try to optimize away the conversion entirely when it's
25586 // conditionally from a constant. Vectors only.
25587 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
25590 // Now move on to more general possibilities.
25591 SDValue Op0 = N->getOperand(0);
25592 EVT VT = N->getValueType(0);
25593 EVT InVT = Op0.getValueType();
25594 EVT InSVT = InVT.getScalarType();
25596 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
25597 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
25598 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25600 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25601 InVT.getVectorNumElements());
25602 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
25603 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25606 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
25607 // a 32-bit target where SSE doesn't support i64->FP operations.
25608 if (Op0.getOpcode() == ISD::LOAD) {
25609 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
25610 EVT LdVT = Ld->getValueType(0);
25612 // This transformation is not supported if the result type is f16
25613 if (VT == MVT::f16)
25616 if (!Ld->isVolatile() && !VT.isVector() &&
25617 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
25618 !Subtarget->is64Bit() && LdVT == MVT::i64) {
25619 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
25620 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
25621 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
25628 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
25629 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
25630 X86TargetLowering::DAGCombinerInfo &DCI) {
25631 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
25632 // the result is either zero or one (depending on the input carry bit).
25633 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
25634 if (X86::isZeroNode(N->getOperand(0)) &&
25635 X86::isZeroNode(N->getOperand(1)) &&
25636 // We don't have a good way to replace an EFLAGS use, so only do this when
25638 SDValue(N, 1).use_empty()) {
25640 EVT VT = N->getValueType(0);
25641 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
25642 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
25643 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
25644 DAG.getConstant(X86::COND_B, DL,
25647 DAG.getConstant(1, DL, VT));
25648 return DCI.CombineTo(N, Res1, CarryOut);
25654 // fold (add Y, (sete X, 0)) -> adc 0, Y
25655 // (add Y, (setne X, 0)) -> sbb -1, Y
25656 // (sub (sete X, 0), Y) -> sbb 0, Y
25657 // (sub (setne X, 0), Y) -> adc -1, Y
25658 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
25661 // Look through ZExts.
25662 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
25663 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
25666 SDValue SetCC = Ext.getOperand(0);
25667 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
25670 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
25671 if (CC != X86::COND_E && CC != X86::COND_NE)
25674 SDValue Cmp = SetCC.getOperand(1);
25675 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
25676 !X86::isZeroNode(Cmp.getOperand(1)) ||
25677 !Cmp.getOperand(0).getValueType().isInteger())
25680 SDValue CmpOp0 = Cmp.getOperand(0);
25681 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
25682 DAG.getConstant(1, DL, CmpOp0.getValueType()));
25684 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
25685 if (CC == X86::COND_NE)
25686 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
25687 DL, OtherVal.getValueType(), OtherVal,
25688 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
25690 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
25691 DL, OtherVal.getValueType(), OtherVal,
25692 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
25695 /// PerformADDCombine - Do target-specific dag combines on integer adds.
25696 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
25697 const X86Subtarget *Subtarget) {
25698 EVT VT = N->getValueType(0);
25699 SDValue Op0 = N->getOperand(0);
25700 SDValue Op1 = N->getOperand(1);
25702 // Try to synthesize horizontal adds from adds of shuffles.
25703 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25704 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25705 isHorizontalBinOp(Op0, Op1, true))
25706 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
25708 return OptimizeConditionalInDecrement(N, DAG);
25711 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
25712 const X86Subtarget *Subtarget) {
25713 SDValue Op0 = N->getOperand(0);
25714 SDValue Op1 = N->getOperand(1);
25716 // X86 can't encode an immediate LHS of a sub. See if we can push the
25717 // negation into a preceding instruction.
25718 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
25719 // If the RHS of the sub is a XOR with one use and a constant, invert the
25720 // immediate. Then add one to the LHS of the sub so we can turn
25721 // X-Y -> X+~Y+1, saving one register.
25722 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
25723 isa<ConstantSDNode>(Op1.getOperand(1))) {
25724 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
25725 EVT VT = Op0.getValueType();
25726 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
25728 DAG.getConstant(~XorC, SDLoc(Op1), VT));
25729 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
25730 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
25734 // Try to synthesize horizontal adds from adds of shuffles.
25735 EVT VT = N->getValueType(0);
25736 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25737 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25738 isHorizontalBinOp(Op0, Op1, true))
25739 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
25741 return OptimizeConditionalInDecrement(N, DAG);
25744 /// performVZEXTCombine - Performs build vector combines
25745 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
25746 TargetLowering::DAGCombinerInfo &DCI,
25747 const X86Subtarget *Subtarget) {
25749 MVT VT = N->getSimpleValueType(0);
25750 SDValue Op = N->getOperand(0);
25751 MVT OpVT = Op.getSimpleValueType();
25752 MVT OpEltVT = OpVT.getVectorElementType();
25753 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
25755 // (vzext (bitcast (vzext (x)) -> (vzext x)
25757 while (V.getOpcode() == ISD::BITCAST)
25758 V = V.getOperand(0);
25760 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
25761 MVT InnerVT = V.getSimpleValueType();
25762 MVT InnerEltVT = InnerVT.getVectorElementType();
25764 // If the element sizes match exactly, we can just do one larger vzext. This
25765 // is always an exact type match as vzext operates on integer types.
25766 if (OpEltVT == InnerEltVT) {
25767 assert(OpVT == InnerVT && "Types must match for vzext!");
25768 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
25771 // The only other way we can combine them is if only a single element of the
25772 // inner vzext is used in the input to the outer vzext.
25773 if (InnerEltVT.getSizeInBits() < InputBits)
25776 // In this case, the inner vzext is completely dead because we're going to
25777 // only look at bits inside of the low element. Just do the outer vzext on
25778 // a bitcast of the input to the inner.
25779 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
25782 // Check if we can bypass extracting and re-inserting an element of an input
25783 // vector. Essentially:
25784 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
25785 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
25786 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
25787 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
25788 SDValue ExtractedV = V.getOperand(0);
25789 SDValue OrigV = ExtractedV.getOperand(0);
25790 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
25791 if (ExtractIdx->getZExtValue() == 0) {
25792 MVT OrigVT = OrigV.getSimpleValueType();
25793 // Extract a subvector if necessary...
25794 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
25795 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
25796 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
25797 OrigVT.getVectorNumElements() / Ratio);
25798 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
25799 DAG.getIntPtrConstant(0, DL));
25801 Op = DAG.getBitcast(OpVT, OrigV);
25802 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
25809 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
25810 DAGCombinerInfo &DCI) const {
25811 SelectionDAG &DAG = DCI.DAG;
25812 switch (N->getOpcode()) {
25814 case ISD::EXTRACT_VECTOR_ELT:
25815 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
25818 case X86ISD::SHRUNKBLEND:
25819 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
25820 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
25821 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
25822 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
25823 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
25824 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
25825 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
25828 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
25829 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
25830 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
25831 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
25832 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
25833 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
25834 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
25835 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
25836 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
25837 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
25838 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
25839 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
25841 case X86ISD::FOR: return PerformFORCombine(N, DAG);
25843 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
25844 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
25845 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
25846 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
25847 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
25848 case ISD::ANY_EXTEND:
25849 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
25850 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
25851 case ISD::SIGN_EXTEND_INREG:
25852 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
25853 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
25854 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
25855 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
25856 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
25857 case X86ISD::SHUFP: // Handle all target specific shuffles
25858 case X86ISD::PALIGNR:
25859 case X86ISD::UNPCKH:
25860 case X86ISD::UNPCKL:
25861 case X86ISD::MOVHLPS:
25862 case X86ISD::MOVLHPS:
25863 case X86ISD::PSHUFB:
25864 case X86ISD::PSHUFD:
25865 case X86ISD::PSHUFHW:
25866 case X86ISD::PSHUFLW:
25867 case X86ISD::MOVSS:
25868 case X86ISD::MOVSD:
25869 case X86ISD::VPERMILPI:
25870 case X86ISD::VPERM2X128:
25871 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
25872 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
25873 case X86ISD::INSERTPS: {
25874 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
25875 return PerformINSERTPSCombine(N, DAG, Subtarget);
25878 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
25884 /// isTypeDesirableForOp - Return true if the target has native support for
25885 /// the specified value type and it is 'desirable' to use the type for the
25886 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
25887 /// instruction encodings are longer and some i16 instructions are slow.
25888 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
25889 if (!isTypeLegal(VT))
25891 if (VT != MVT::i16)
25898 case ISD::SIGN_EXTEND:
25899 case ISD::ZERO_EXTEND:
25900 case ISD::ANY_EXTEND:
25913 /// IsDesirableToPromoteOp - This method query the target whether it is
25914 /// beneficial for dag combiner to promote the specified node. If true, it
25915 /// should return the desired promotion type by reference.
25916 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
25917 EVT VT = Op.getValueType();
25918 if (VT != MVT::i16)
25921 bool Promote = false;
25922 bool Commute = false;
25923 switch (Op.getOpcode()) {
25926 LoadSDNode *LD = cast<LoadSDNode>(Op);
25927 // If the non-extending load has a single use and it's not live out, then it
25928 // might be folded.
25929 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
25930 Op.hasOneUse()*/) {
25931 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
25932 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
25933 // The only case where we'd want to promote LOAD (rather then it being
25934 // promoted as an operand is when it's only use is liveout.
25935 if (UI->getOpcode() != ISD::CopyToReg)
25942 case ISD::SIGN_EXTEND:
25943 case ISD::ZERO_EXTEND:
25944 case ISD::ANY_EXTEND:
25949 SDValue N0 = Op.getOperand(0);
25950 // Look out for (store (shl (load), x)).
25951 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
25964 SDValue N0 = Op.getOperand(0);
25965 SDValue N1 = Op.getOperand(1);
25966 if (!Commute && MayFoldLoad(N1))
25968 // Avoid disabling potential load folding opportunities.
25969 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
25971 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
25981 //===----------------------------------------------------------------------===//
25982 // X86 Inline Assembly Support
25983 //===----------------------------------------------------------------------===//
25985 // Helper to match a string separated by whitespace.
25986 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
25987 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
25989 for (StringRef Piece : Pieces) {
25990 if (!S.startswith(Piece)) // Check if the piece matches.
25993 S = S.substr(Piece.size());
25994 StringRef::size_type Pos = S.find_first_not_of(" \t");
25995 if (Pos == 0) // We matched a prefix.
26004 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
26006 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
26007 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
26008 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
26009 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
26011 if (AsmPieces.size() == 3)
26013 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
26020 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
26021 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
26023 std::string AsmStr = IA->getAsmString();
26025 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
26026 if (!Ty || Ty->getBitWidth() % 16 != 0)
26029 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
26030 SmallVector<StringRef, 4> AsmPieces;
26031 SplitString(AsmStr, AsmPieces, ";\n");
26033 switch (AsmPieces.size()) {
26034 default: return false;
26036 // FIXME: this should verify that we are targeting a 486 or better. If not,
26037 // we will turn this bswap into something that will be lowered to logical
26038 // ops instead of emitting the bswap asm. For now, we don't support 486 or
26039 // lower so don't worry about this.
26041 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
26042 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
26043 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
26044 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
26045 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
26046 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
26047 // No need to check constraints, nothing other than the equivalent of
26048 // "=r,0" would be valid here.
26049 return IntrinsicLowering::LowerToByteSwap(CI);
26052 // rorw $$8, ${0:w} --> llvm.bswap.i16
26053 if (CI->getType()->isIntegerTy(16) &&
26054 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26055 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
26056 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
26058 StringRef ConstraintsStr = IA->getConstraintString();
26059 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26060 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26061 if (clobbersFlagRegisters(AsmPieces))
26062 return IntrinsicLowering::LowerToByteSwap(CI);
26066 if (CI->getType()->isIntegerTy(32) &&
26067 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26068 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
26069 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
26070 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
26072 StringRef ConstraintsStr = IA->getConstraintString();
26073 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26074 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26075 if (clobbersFlagRegisters(AsmPieces))
26076 return IntrinsicLowering::LowerToByteSwap(CI);
26079 if (CI->getType()->isIntegerTy(64)) {
26080 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
26081 if (Constraints.size() >= 2 &&
26082 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
26083 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
26084 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
26085 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
26086 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
26087 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
26088 return IntrinsicLowering::LowerToByteSwap(CI);
26096 /// getConstraintType - Given a constraint letter, return the type of
26097 /// constraint it is for this target.
26098 X86TargetLowering::ConstraintType
26099 X86TargetLowering::getConstraintType(StringRef Constraint) const {
26100 if (Constraint.size() == 1) {
26101 switch (Constraint[0]) {
26112 return C_RegisterClass;
26136 return TargetLowering::getConstraintType(Constraint);
26139 /// Examine constraint type and operand type and determine a weight value.
26140 /// This object must already have been set up with the operand type
26141 /// and the current alternative constraint selected.
26142 TargetLowering::ConstraintWeight
26143 X86TargetLowering::getSingleConstraintMatchWeight(
26144 AsmOperandInfo &info, const char *constraint) const {
26145 ConstraintWeight weight = CW_Invalid;
26146 Value *CallOperandVal = info.CallOperandVal;
26147 // If we don't have a value, we can't do a match,
26148 // but allow it at the lowest weight.
26149 if (!CallOperandVal)
26151 Type *type = CallOperandVal->getType();
26152 // Look at the constraint type.
26153 switch (*constraint) {
26155 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
26166 if (CallOperandVal->getType()->isIntegerTy())
26167 weight = CW_SpecificReg;
26172 if (type->isFloatingPointTy())
26173 weight = CW_SpecificReg;
26176 if (type->isX86_MMXTy() && Subtarget->hasMMX())
26177 weight = CW_SpecificReg;
26181 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
26182 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
26183 weight = CW_Register;
26186 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
26187 if (C->getZExtValue() <= 31)
26188 weight = CW_Constant;
26192 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26193 if (C->getZExtValue() <= 63)
26194 weight = CW_Constant;
26198 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26199 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
26200 weight = CW_Constant;
26204 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26205 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
26206 weight = CW_Constant;
26210 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26211 if (C->getZExtValue() <= 3)
26212 weight = CW_Constant;
26216 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26217 if (C->getZExtValue() <= 0xff)
26218 weight = CW_Constant;
26223 if (isa<ConstantFP>(CallOperandVal)) {
26224 weight = CW_Constant;
26228 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26229 if ((C->getSExtValue() >= -0x80000000LL) &&
26230 (C->getSExtValue() <= 0x7fffffffLL))
26231 weight = CW_Constant;
26235 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26236 if (C->getZExtValue() <= 0xffffffff)
26237 weight = CW_Constant;
26244 /// LowerXConstraint - try to replace an X constraint, which matches anything,
26245 /// with another that has more specific requirements based on the type of the
26246 /// corresponding operand.
26247 const char *X86TargetLowering::
26248 LowerXConstraint(EVT ConstraintVT) const {
26249 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
26250 // 'f' like normal targets.
26251 if (ConstraintVT.isFloatingPoint()) {
26252 if (Subtarget->hasSSE2())
26254 if (Subtarget->hasSSE1())
26258 return TargetLowering::LowerXConstraint(ConstraintVT);
26261 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
26262 /// vector. If it is invalid, don't add anything to Ops.
26263 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
26264 std::string &Constraint,
26265 std::vector<SDValue>&Ops,
26266 SelectionDAG &DAG) const {
26269 // Only support length 1 constraints for now.
26270 if (Constraint.length() > 1) return;
26272 char ConstraintLetter = Constraint[0];
26273 switch (ConstraintLetter) {
26276 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26277 if (C->getZExtValue() <= 31) {
26278 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26279 Op.getValueType());
26285 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26286 if (C->getZExtValue() <= 63) {
26287 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26288 Op.getValueType());
26294 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26295 if (isInt<8>(C->getSExtValue())) {
26296 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26297 Op.getValueType());
26303 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26304 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
26305 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
26306 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
26307 Op.getValueType());
26313 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26314 if (C->getZExtValue() <= 3) {
26315 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26316 Op.getValueType());
26322 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26323 if (C->getZExtValue() <= 255) {
26324 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26325 Op.getValueType());
26331 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26332 if (C->getZExtValue() <= 127) {
26333 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26334 Op.getValueType());
26340 // 32-bit signed value
26341 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26342 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26343 C->getSExtValue())) {
26344 // Widen to 64 bits here to get it sign extended.
26345 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
26348 // FIXME gcc accepts some relocatable values here too, but only in certain
26349 // memory models; it's complicated.
26354 // 32-bit unsigned value
26355 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26356 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26357 C->getZExtValue())) {
26358 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26359 Op.getValueType());
26363 // FIXME gcc accepts some relocatable values here too, but only in certain
26364 // memory models; it's complicated.
26368 // Literal immediates are always ok.
26369 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
26370 // Widen to 64 bits here to get it sign extended.
26371 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
26375 // In any sort of PIC mode addresses need to be computed at runtime by
26376 // adding in a register or some sort of table lookup. These can't
26377 // be used as immediates.
26378 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
26381 // If we are in non-pic codegen mode, we allow the address of a global (with
26382 // an optional displacement) to be used with 'i'.
26383 GlobalAddressSDNode *GA = nullptr;
26384 int64_t Offset = 0;
26386 // Match either (GA), (GA+C), (GA+C1+C2), etc.
26388 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
26389 Offset += GA->getOffset();
26391 } else if (Op.getOpcode() == ISD::ADD) {
26392 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26393 Offset += C->getZExtValue();
26394 Op = Op.getOperand(0);
26397 } else if (Op.getOpcode() == ISD::SUB) {
26398 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26399 Offset += -C->getZExtValue();
26400 Op = Op.getOperand(0);
26405 // Otherwise, this isn't something we can handle, reject it.
26409 const GlobalValue *GV = GA->getGlobal();
26410 // If we require an extra load to get this address, as in PIC mode, we
26411 // can't accept it.
26412 if (isGlobalStubReference(
26413 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
26416 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
26417 GA->getValueType(0), Offset);
26422 if (Result.getNode()) {
26423 Ops.push_back(Result);
26426 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
26429 std::pair<unsigned, const TargetRegisterClass *>
26430 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
26431 StringRef Constraint,
26433 // First, see if this is a constraint that directly corresponds to an LLVM
26435 if (Constraint.size() == 1) {
26436 // GCC Constraint Letters
26437 switch (Constraint[0]) {
26439 // TODO: Slight differences here in allocation order and leaving
26440 // RIP in the class. Do they matter any more here than they do
26441 // in the normal allocation?
26442 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
26443 if (Subtarget->is64Bit()) {
26444 if (VT == MVT::i32 || VT == MVT::f32)
26445 return std::make_pair(0U, &X86::GR32RegClass);
26446 if (VT == MVT::i16)
26447 return std::make_pair(0U, &X86::GR16RegClass);
26448 if (VT == MVT::i8 || VT == MVT::i1)
26449 return std::make_pair(0U, &X86::GR8RegClass);
26450 if (VT == MVT::i64 || VT == MVT::f64)
26451 return std::make_pair(0U, &X86::GR64RegClass);
26454 // 32-bit fallthrough
26455 case 'Q': // Q_REGS
26456 if (VT == MVT::i32 || VT == MVT::f32)
26457 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
26458 if (VT == MVT::i16)
26459 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
26460 if (VT == MVT::i8 || VT == MVT::i1)
26461 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
26462 if (VT == MVT::i64)
26463 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
26465 case 'r': // GENERAL_REGS
26466 case 'l': // INDEX_REGS
26467 if (VT == MVT::i8 || VT == MVT::i1)
26468 return std::make_pair(0U, &X86::GR8RegClass);
26469 if (VT == MVT::i16)
26470 return std::make_pair(0U, &X86::GR16RegClass);
26471 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
26472 return std::make_pair(0U, &X86::GR32RegClass);
26473 return std::make_pair(0U, &X86::GR64RegClass);
26474 case 'R': // LEGACY_REGS
26475 if (VT == MVT::i8 || VT == MVT::i1)
26476 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
26477 if (VT == MVT::i16)
26478 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
26479 if (VT == MVT::i32 || !Subtarget->is64Bit())
26480 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
26481 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
26482 case 'f': // FP Stack registers.
26483 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
26484 // value to the correct fpstack register class.
26485 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
26486 return std::make_pair(0U, &X86::RFP32RegClass);
26487 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
26488 return std::make_pair(0U, &X86::RFP64RegClass);
26489 return std::make_pair(0U, &X86::RFP80RegClass);
26490 case 'y': // MMX_REGS if MMX allowed.
26491 if (!Subtarget->hasMMX()) break;
26492 return std::make_pair(0U, &X86::VR64RegClass);
26493 case 'Y': // SSE_REGS if SSE2 allowed
26494 if (!Subtarget->hasSSE2()) break;
26496 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
26497 if (!Subtarget->hasSSE1()) break;
26499 switch (VT.SimpleTy) {
26501 // Scalar SSE types.
26504 return std::make_pair(0U, &X86::FR32RegClass);
26507 return std::make_pair(0U, &X86::FR64RegClass);
26515 return std::make_pair(0U, &X86::VR128RegClass);
26523 return std::make_pair(0U, &X86::VR256RegClass);
26528 return std::make_pair(0U, &X86::VR512RegClass);
26534 // Use the default implementation in TargetLowering to convert the register
26535 // constraint into a member of a register class.
26536 std::pair<unsigned, const TargetRegisterClass*> Res;
26537 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
26539 // Not found as a standard register?
26541 // Map st(0) -> st(7) -> ST0
26542 if (Constraint.size() == 7 && Constraint[0] == '{' &&
26543 tolower(Constraint[1]) == 's' &&
26544 tolower(Constraint[2]) == 't' &&
26545 Constraint[3] == '(' &&
26546 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
26547 Constraint[5] == ')' &&
26548 Constraint[6] == '}') {
26550 Res.first = X86::FP0+Constraint[4]-'0';
26551 Res.second = &X86::RFP80RegClass;
26555 // GCC allows "st(0)" to be called just plain "st".
26556 if (StringRef("{st}").equals_lower(Constraint)) {
26557 Res.first = X86::FP0;
26558 Res.second = &X86::RFP80RegClass;
26563 if (StringRef("{flags}").equals_lower(Constraint)) {
26564 Res.first = X86::EFLAGS;
26565 Res.second = &X86::CCRRegClass;
26569 // 'A' means EAX + EDX.
26570 if (Constraint == "A") {
26571 Res.first = X86::EAX;
26572 Res.second = &X86::GR32_ADRegClass;
26578 // Otherwise, check to see if this is a register class of the wrong value
26579 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
26580 // turn into {ax},{dx}.
26581 // MVT::Other is used to specify clobber names.
26582 if (Res.second->hasType(VT) || VT == MVT::Other)
26583 return Res; // Correct type already, nothing to do.
26585 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
26586 // return "eax". This should even work for things like getting 64bit integer
26587 // registers when given an f64 type.
26588 const TargetRegisterClass *Class = Res.second;
26589 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
26590 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
26591 unsigned Size = VT.getSizeInBits();
26592 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
26593 : Size == 16 ? MVT::i16
26594 : Size == 32 ? MVT::i32
26595 : Size == 64 ? MVT::i64
26597 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
26599 Res.first = DestReg;
26600 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
26601 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
26602 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
26603 : &X86::GR64RegClass;
26604 assert(Res.second->contains(Res.first) && "Register in register class");
26606 // No register found/type mismatch.
26608 Res.second = nullptr;
26610 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
26611 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
26612 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
26613 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
26614 Class == &X86::VR512RegClass) {
26615 // Handle references to XMM physical registers that got mapped into the
26616 // wrong class. This can happen with constraints like {xmm0} where the
26617 // target independent register mapper will just pick the first match it can
26618 // find, ignoring the required type.
26620 if (VT == MVT::f32 || VT == MVT::i32)
26621 Res.second = &X86::FR32RegClass;
26622 else if (VT == MVT::f64 || VT == MVT::i64)
26623 Res.second = &X86::FR64RegClass;
26624 else if (X86::VR128RegClass.hasType(VT))
26625 Res.second = &X86::VR128RegClass;
26626 else if (X86::VR256RegClass.hasType(VT))
26627 Res.second = &X86::VR256RegClass;
26628 else if (X86::VR512RegClass.hasType(VT))
26629 Res.second = &X86::VR512RegClass;
26631 // Type mismatch and not a clobber: Return an error;
26633 Res.second = nullptr;
26640 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
26641 const AddrMode &AM, Type *Ty,
26642 unsigned AS) const {
26643 // Scaling factors are not free at all.
26644 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
26645 // will take 2 allocations in the out of order engine instead of 1
26646 // for plain addressing mode, i.e. inst (reg1).
26648 // vaddps (%rsi,%drx), %ymm0, %ymm1
26649 // Requires two allocations (one for the load, one for the computation)
26651 // vaddps (%rsi), %ymm0, %ymm1
26652 // Requires just 1 allocation, i.e., freeing allocations for other operations
26653 // and having less micro operations to execute.
26655 // For some X86 architectures, this is even worse because for instance for
26656 // stores, the complex addressing mode forces the instruction to use the
26657 // "load" ports instead of the dedicated "store" port.
26658 // E.g., on Haswell:
26659 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
26660 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
26661 if (isLegalAddressingMode(DL, AM, Ty, AS))
26662 // Scale represents reg2 * scale, thus account for 1
26663 // as soon as we use a second register.
26664 return AM.Scale != 0;
26668 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
26669 // Integer division on x86 is expensive. However, when aggressively optimizing
26670 // for code size, we prefer to use a div instruction, as it is usually smaller
26671 // than the alternative sequence.
26672 // The exception to this is vector division. Since x86 doesn't have vector
26673 // integer division, leaving the division as-is is a loss even in terms of
26674 // size, because it will have to be scalarized, while the alternative code
26675 // sequence can be performed in vector form.
26676 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
26677 Attribute::MinSize);
26678 return OptSize && !VT.isVector();