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 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "Utils/X86ShuffleDecode.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/CodeGen/IntrinsicLowering.h"
32 #include "llvm/CodeGen/MachineFrameInfo.h"
33 #include "llvm/CodeGen/MachineFunction.h"
34 #include "llvm/CodeGen/MachineInstrBuilder.h"
35 #include "llvm/CodeGen/MachineJumpTableInfo.h"
36 #include "llvm/CodeGen/MachineModuleInfo.h"
37 #include "llvm/CodeGen/MachineRegisterInfo.h"
38 #include "llvm/CodeGen/PseudoSourceValue.h"
39 #include "llvm/MC/MCAsmInfo.h"
40 #include "llvm/MC/MCContext.h"
41 #include "llvm/MC/MCExpr.h"
42 #include "llvm/MC/MCSymbol.h"
43 #include "llvm/ADT/BitVector.h"
44 #include "llvm/ADT/SmallSet.h"
45 #include "llvm/ADT/Statistic.h"
46 #include "llvm/ADT/StringExtras.h"
47 #include "llvm/ADT/VectorExtras.h"
48 #include "llvm/Support/CallSite.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/Dwarf.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Support/raw_ostream.h"
55 using namespace dwarf;
57 STATISTIC(NumTailCalls, "Number of tail calls");
59 // Forward declarations.
60 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
63 static SDValue Insert128BitVector(SDValue Result,
69 static SDValue Extract128BitVector(SDValue Vec,
74 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
75 /// sets things up to match to an AVX VEXTRACTF128 instruction or a
76 /// simple subregister reference. Idx is an index in the 128 bits we
77 /// want. It need not be aligned to a 128-bit bounday. That makes
78 /// lowering EXTRACT_VECTOR_ELT operations easier.
79 static SDValue Extract128BitVector(SDValue Vec,
83 EVT VT = Vec.getValueType();
84 assert(VT.getSizeInBits() == 256 && "Unexpected vector size!");
85 EVT ElVT = VT.getVectorElementType();
86 int Factor = VT.getSizeInBits()/128;
87 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
88 VT.getVectorNumElements()/Factor);
90 // Extract from UNDEF is UNDEF.
91 if (Vec.getOpcode() == ISD::UNDEF)
92 return DAG.getNode(ISD::UNDEF, dl, ResultVT);
94 if (isa<ConstantSDNode>(Idx)) {
95 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
97 // Extract the relevant 128 bits. Generate an EXTRACT_SUBVECTOR
98 // we can match to VEXTRACTF128.
99 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
101 // This is the index of the first element of the 128-bit chunk
103 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
106 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
107 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
116 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
117 /// sets things up to match to an AVX VINSERTF128 instruction or a
118 /// simple superregister reference. Idx is an index in the 128 bits
119 /// we want. It need not be aligned to a 128-bit bounday. That makes
120 /// lowering INSERT_VECTOR_ELT operations easier.
121 static SDValue Insert128BitVector(SDValue Result,
126 if (isa<ConstantSDNode>(Idx)) {
127 EVT VT = Vec.getValueType();
128 assert(VT.getSizeInBits() == 128 && "Unexpected vector size!");
130 EVT ElVT = VT.getVectorElementType();
131 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
132 EVT ResultVT = Result.getValueType();
134 // Insert the relevant 128 bits.
135 unsigned ElemsPerChunk = 128/ElVT.getSizeInBits();
137 // This is the index of the first element of the 128-bit chunk
139 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/128)
142 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
143 Result = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
151 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
152 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
153 bool is64Bit = Subtarget->is64Bit();
155 if (Subtarget->isTargetEnvMacho()) {
157 return new X8664_MachoTargetObjectFile();
158 return new TargetLoweringObjectFileMachO();
161 if (Subtarget->isTargetELF())
162 return new TargetLoweringObjectFileELF();
163 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
164 return new TargetLoweringObjectFileCOFF();
165 llvm_unreachable("unknown subtarget type");
168 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
169 : TargetLowering(TM, createTLOF(TM)) {
170 Subtarget = &TM.getSubtarget<X86Subtarget>();
171 X86ScalarSSEf64 = Subtarget->hasXMMInt() || Subtarget->hasAVX();
172 X86ScalarSSEf32 = Subtarget->hasXMM() || Subtarget->hasAVX();
173 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
175 RegInfo = TM.getRegisterInfo();
176 TD = getTargetData();
178 // Set up the TargetLowering object.
179 static MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
181 // X86 is weird, it always uses i8 for shift amounts and setcc results.
182 setBooleanContents(ZeroOrOneBooleanContent);
184 // For 64-bit since we have so many registers use the ILP scheduler, for
185 // 32-bit code use the register pressure specific scheduling.
186 if (Subtarget->is64Bit())
187 setSchedulingPreference(Sched::ILP);
189 setSchedulingPreference(Sched::RegPressure);
190 setStackPointerRegisterToSaveRestore(X86StackPtr);
192 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
193 // Setup Windows compiler runtime calls.
194 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
195 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
196 setLibcallName(RTLIB::SREM_I64, "_allrem");
197 setLibcallName(RTLIB::UREM_I64, "_aullrem");
198 setLibcallName(RTLIB::MUL_I64, "_allmul");
199 setLibcallName(RTLIB::FPTOUINT_F64_I64, "_ftol2");
200 setLibcallName(RTLIB::FPTOUINT_F32_I64, "_ftol2");
201 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
202 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
203 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
204 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
205 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
206 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::C);
207 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::C);
210 if (Subtarget->isTargetDarwin()) {
211 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
212 setUseUnderscoreSetJmp(false);
213 setUseUnderscoreLongJmp(false);
214 } else if (Subtarget->isTargetMingw()) {
215 // MS runtime is weird: it exports _setjmp, but longjmp!
216 setUseUnderscoreSetJmp(true);
217 setUseUnderscoreLongJmp(false);
219 setUseUnderscoreSetJmp(true);
220 setUseUnderscoreLongJmp(true);
223 // Set up the register classes.
224 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
225 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
226 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
227 if (Subtarget->is64Bit())
228 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
230 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
232 // We don't accept any truncstore of integer registers.
233 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
234 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
235 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
236 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
237 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
238 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
240 // SETOEQ and SETUNE require checking two conditions.
241 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
242 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
243 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
244 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
245 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
246 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
248 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
250 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
251 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
252 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
254 if (Subtarget->is64Bit()) {
255 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
256 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
257 } else if (!UseSoftFloat) {
258 // We have an algorithm for SSE2->double, and we turn this into a
259 // 64-bit FILD followed by conditional FADD for other targets.
260 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
261 // We have an algorithm for SSE2, and we turn this into a 64-bit
262 // FILD for other targets.
263 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
266 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
268 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
269 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
272 // SSE has no i16 to fp conversion, only i32
273 if (X86ScalarSSEf32) {
274 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
275 // f32 and f64 cases are Legal, f80 case is not
276 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
278 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
279 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
282 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
283 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
286 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
287 // are Legal, f80 is custom lowered.
288 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
289 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
291 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
293 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
294 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
296 if (X86ScalarSSEf32) {
297 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
298 // f32 and f64 cases are Legal, f80 case is not
299 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
301 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
302 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
305 // Handle FP_TO_UINT by promoting the destination to a larger signed
307 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
308 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
309 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
311 if (Subtarget->is64Bit()) {
312 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
313 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
314 } else if (!UseSoftFloat) {
315 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
316 // Expand FP_TO_UINT into a select.
317 // FIXME: We would like to use a Custom expander here eventually to do
318 // the optimal thing for SSE vs. the default expansion in the legalizer.
319 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
321 // With SSE3 we can use fisttpll to convert to a signed i64; without
322 // SSE, we're stuck with a fistpll.
323 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
326 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
327 if (!X86ScalarSSEf64) {
328 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
329 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
330 if (Subtarget->is64Bit()) {
331 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
332 // Without SSE, i64->f64 goes through memory.
333 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
337 // Scalar integer divide and remainder are lowered to use operations that
338 // produce two results, to match the available instructions. This exposes
339 // the two-result form to trivial CSE, which is able to combine x/y and x%y
340 // into a single instruction.
342 // Scalar integer multiply-high is also lowered to use two-result
343 // operations, to match the available instructions. However, plain multiply
344 // (low) operations are left as Legal, as there are single-result
345 // instructions for this in x86. Using the two-result multiply instructions
346 // when both high and low results are needed must be arranged by dagcombine.
347 for (unsigned i = 0, e = 4; i != e; ++i) {
349 setOperationAction(ISD::MULHS, VT, Expand);
350 setOperationAction(ISD::MULHU, VT, Expand);
351 setOperationAction(ISD::SDIV, VT, Expand);
352 setOperationAction(ISD::UDIV, VT, Expand);
353 setOperationAction(ISD::SREM, VT, Expand);
354 setOperationAction(ISD::UREM, VT, Expand);
356 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
357 setOperationAction(ISD::ADDC, VT, Custom);
358 setOperationAction(ISD::ADDE, VT, Custom);
359 setOperationAction(ISD::SUBC, VT, Custom);
360 setOperationAction(ISD::SUBE, VT, Custom);
363 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
364 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
365 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
366 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
367 if (Subtarget->is64Bit())
368 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
369 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
370 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
371 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
372 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
373 setOperationAction(ISD::FREM , MVT::f32 , Expand);
374 setOperationAction(ISD::FREM , MVT::f64 , Expand);
375 setOperationAction(ISD::FREM , MVT::f80 , Expand);
376 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
378 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
379 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
380 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
381 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
382 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
383 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
384 if (Subtarget->is64Bit()) {
385 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
386 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
389 if (Subtarget->hasPOPCNT()) {
390 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
392 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
393 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
394 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
395 if (Subtarget->is64Bit())
396 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
399 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
400 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
402 // These should be promoted to a larger select which is supported.
403 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
404 // X86 wants to expand cmov itself.
405 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
406 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
407 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
408 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
409 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
410 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
411 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
412 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
413 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
414 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
415 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
416 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
417 if (Subtarget->is64Bit()) {
418 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
419 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
421 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
424 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
425 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
426 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
427 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
428 if (Subtarget->is64Bit())
429 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
430 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
431 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
432 if (Subtarget->is64Bit()) {
433 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
434 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
435 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
436 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
437 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
439 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
440 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
441 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
442 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
443 if (Subtarget->is64Bit()) {
444 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
445 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
446 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
449 if (Subtarget->hasXMM())
450 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
452 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
453 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
455 // On X86 and X86-64, atomic operations are lowered to locked instructions.
456 // Locked instructions, in turn, have implicit fence semantics (all memory
457 // operations are flushed before issuing the locked instruction, and they
458 // are not buffered), so we can fold away the common pattern of
459 // fence-atomic-fence.
460 setShouldFoldAtomicFences(true);
462 // Expand certain atomics
463 for (unsigned i = 0, e = 4; i != e; ++i) {
465 setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
466 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
469 if (!Subtarget->is64Bit()) {
470 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
471 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
472 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
473 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
474 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
475 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
476 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
479 // FIXME - use subtarget debug flags
480 if (!Subtarget->isTargetDarwin() &&
481 !Subtarget->isTargetELF() &&
482 !Subtarget->isTargetCygMing()) {
483 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
486 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
487 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
488 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
489 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
490 if (Subtarget->is64Bit()) {
491 setExceptionPointerRegister(X86::RAX);
492 setExceptionSelectorRegister(X86::RDX);
494 setExceptionPointerRegister(X86::EAX);
495 setExceptionSelectorRegister(X86::EDX);
497 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
498 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
500 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
502 setOperationAction(ISD::TRAP, MVT::Other, Legal);
504 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
505 setOperationAction(ISD::VASTART , MVT::Other, Custom);
506 setOperationAction(ISD::VAEND , MVT::Other, Expand);
507 if (Subtarget->is64Bit()) {
508 setOperationAction(ISD::VAARG , MVT::Other, Custom);
509 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
511 setOperationAction(ISD::VAARG , MVT::Other, Expand);
512 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
515 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
516 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
517 setOperationAction(ISD::DYNAMIC_STACKALLOC,
518 (Subtarget->is64Bit() ? MVT::i64 : MVT::i32),
519 (Subtarget->isTargetCOFF()
520 && !Subtarget->isTargetEnvMacho()
523 if (!UseSoftFloat && X86ScalarSSEf64) {
524 // f32 and f64 use SSE.
525 // Set up the FP register classes.
526 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
527 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
529 // Use ANDPD to simulate FABS.
530 setOperationAction(ISD::FABS , MVT::f64, Custom);
531 setOperationAction(ISD::FABS , MVT::f32, Custom);
533 // Use XORP to simulate FNEG.
534 setOperationAction(ISD::FNEG , MVT::f64, Custom);
535 setOperationAction(ISD::FNEG , MVT::f32, Custom);
537 // Use ANDPD and ORPD to simulate FCOPYSIGN.
538 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
539 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
541 // Lower this to FGETSIGNx86 plus an AND.
542 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
543 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
545 // We don't support sin/cos/fmod
546 setOperationAction(ISD::FSIN , MVT::f64, Expand);
547 setOperationAction(ISD::FCOS , MVT::f64, Expand);
548 setOperationAction(ISD::FSIN , MVT::f32, Expand);
549 setOperationAction(ISD::FCOS , MVT::f32, Expand);
551 // Expand FP immediates into loads from the stack, except for the special
553 addLegalFPImmediate(APFloat(+0.0)); // xorpd
554 addLegalFPImmediate(APFloat(+0.0f)); // xorps
555 } else if (!UseSoftFloat && X86ScalarSSEf32) {
556 // Use SSE for f32, x87 for f64.
557 // Set up the FP register classes.
558 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
559 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
561 // Use ANDPS to simulate FABS.
562 setOperationAction(ISD::FABS , MVT::f32, Custom);
564 // Use XORP to simulate FNEG.
565 setOperationAction(ISD::FNEG , MVT::f32, Custom);
567 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
569 // Use ANDPS and ORPS to simulate FCOPYSIGN.
570 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
571 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
573 // We don't support sin/cos/fmod
574 setOperationAction(ISD::FSIN , MVT::f32, Expand);
575 setOperationAction(ISD::FCOS , MVT::f32, Expand);
577 // Special cases we handle for FP constants.
578 addLegalFPImmediate(APFloat(+0.0f)); // xorps
579 addLegalFPImmediate(APFloat(+0.0)); // FLD0
580 addLegalFPImmediate(APFloat(+1.0)); // FLD1
581 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
582 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
585 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
586 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
588 } else if (!UseSoftFloat) {
589 // f32 and f64 in x87.
590 // Set up the FP register classes.
591 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
592 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
594 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
595 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
596 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
597 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
600 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
601 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
603 addLegalFPImmediate(APFloat(+0.0)); // FLD0
604 addLegalFPImmediate(APFloat(+1.0)); // FLD1
605 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
606 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
607 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
608 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
609 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
610 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
613 // We don't support FMA.
614 setOperationAction(ISD::FMA, MVT::f64, Expand);
615 setOperationAction(ISD::FMA, MVT::f32, Expand);
617 // Long double always uses X87.
619 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
620 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
621 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
623 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
624 addLegalFPImmediate(TmpFlt); // FLD0
626 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
629 APFloat TmpFlt2(+1.0);
630 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
632 addLegalFPImmediate(TmpFlt2); // FLD1
633 TmpFlt2.changeSign();
634 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
638 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
639 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
642 setOperationAction(ISD::FMA, MVT::f80, Expand);
645 // Always use a library call for pow.
646 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
647 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
648 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
650 setOperationAction(ISD::FLOG, MVT::f80, Expand);
651 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
652 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
653 setOperationAction(ISD::FEXP, MVT::f80, Expand);
654 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
656 // First set operation action for all vector types to either promote
657 // (for widening) or expand (for scalarization). Then we will selectively
658 // turn on ones that can be effectively codegen'd.
659 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
660 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
661 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
662 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
663 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
664 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
665 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
666 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
667 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
668 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
669 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
670 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
671 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
672 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
673 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
674 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
675 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
676 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
677 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
678 setOperationAction(ISD::INSERT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
679 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
680 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
681 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
682 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
683 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
684 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
685 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
686 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
687 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
688 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
689 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
690 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
691 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
692 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
693 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
694 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
695 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
696 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
697 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
698 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
699 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
700 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
701 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
702 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
703 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
704 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
705 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
706 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
707 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
708 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
709 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
710 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
711 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
712 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
713 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
714 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
715 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
716 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
717 setTruncStoreAction((MVT::SimpleValueType)VT,
718 (MVT::SimpleValueType)InnerVT, Expand);
719 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
720 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
721 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
724 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
725 // with -msoft-float, disable use of MMX as well.
726 if (!UseSoftFloat && Subtarget->hasMMX()) {
727 addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass);
728 // No operations on x86mmx supported, everything uses intrinsics.
731 // MMX-sized vectors (other than x86mmx) are expected to be expanded
732 // into smaller operations.
733 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
734 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
735 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
736 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
737 setOperationAction(ISD::AND, MVT::v8i8, Expand);
738 setOperationAction(ISD::AND, MVT::v4i16, Expand);
739 setOperationAction(ISD::AND, MVT::v2i32, Expand);
740 setOperationAction(ISD::AND, MVT::v1i64, Expand);
741 setOperationAction(ISD::OR, MVT::v8i8, Expand);
742 setOperationAction(ISD::OR, MVT::v4i16, Expand);
743 setOperationAction(ISD::OR, MVT::v2i32, Expand);
744 setOperationAction(ISD::OR, MVT::v1i64, Expand);
745 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
746 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
747 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
748 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
749 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
750 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
751 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
752 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
753 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
754 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
755 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
756 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
757 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
758 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
759 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
760 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
761 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
763 if (!UseSoftFloat && Subtarget->hasXMM()) {
764 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
766 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
767 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
768 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
769 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
770 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
771 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
772 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
773 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
774 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
775 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
776 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
777 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
780 if (!UseSoftFloat && Subtarget->hasXMMInt()) {
781 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
783 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
784 // registers cannot be used even for integer operations.
785 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
786 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
787 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
788 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
790 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
791 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
792 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
793 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
794 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
795 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
796 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
797 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
798 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
799 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
800 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
801 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
802 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
803 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
804 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
805 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
807 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
808 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
809 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
810 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
812 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
813 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
814 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
815 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
816 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
818 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
819 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
820 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
821 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
822 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
824 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
825 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
826 EVT VT = (MVT::SimpleValueType)i;
827 // Do not attempt to custom lower non-power-of-2 vectors
828 if (!isPowerOf2_32(VT.getVectorNumElements()))
830 // Do not attempt to custom lower non-128-bit vectors
831 if (!VT.is128BitVector())
833 setOperationAction(ISD::BUILD_VECTOR,
834 VT.getSimpleVT().SimpleTy, Custom);
835 setOperationAction(ISD::VECTOR_SHUFFLE,
836 VT.getSimpleVT().SimpleTy, Custom);
837 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
838 VT.getSimpleVT().SimpleTy, Custom);
841 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
842 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
843 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
844 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
846 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
848 if (Subtarget->is64Bit()) {
849 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
853 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
854 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
855 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
858 // Do not attempt to promote non-128-bit vectors
859 if (!VT.is128BitVector())
862 setOperationAction(ISD::AND, SVT, Promote);
863 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
864 setOperationAction(ISD::OR, SVT, Promote);
865 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
866 setOperationAction(ISD::XOR, SVT, Promote);
867 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
868 setOperationAction(ISD::LOAD, SVT, Promote);
869 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
870 setOperationAction(ISD::SELECT, SVT, Promote);
871 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
874 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
876 // Custom lower v2i64 and v2f64 selects.
877 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
878 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
879 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
880 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
882 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
883 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
886 if (Subtarget->hasSSE41()) {
887 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
888 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
889 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
890 setOperationAction(ISD::FRINT, MVT::f32, Legal);
891 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
892 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
893 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
894 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
895 setOperationAction(ISD::FRINT, MVT::f64, Legal);
896 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
898 // FIXME: Do we need to handle scalar-to-vector here?
899 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
901 // Can turn SHL into an integer multiply.
902 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
903 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
905 // i8 and i16 vectors are custom , because the source register and source
906 // source memory operand types are not the same width. f32 vectors are
907 // custom since the immediate controlling the insert encodes additional
909 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
910 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
912 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
914 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
915 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
916 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
917 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
919 if (Subtarget->is64Bit()) {
920 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
921 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
925 if (Subtarget->hasSSE2()) {
926 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
927 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
928 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
930 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
931 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
932 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
934 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
935 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
938 if (Subtarget->hasSSE42())
939 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
941 if (!UseSoftFloat && Subtarget->hasAVX()) {
942 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
943 addRegisterClass(MVT::v16i16, X86::VR256RegisterClass);
944 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
945 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
946 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
947 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
949 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
950 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
951 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
953 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
954 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
955 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
956 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
957 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
958 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
960 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
961 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
962 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
963 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
964 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
965 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
967 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
968 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
969 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
971 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f64, Custom);
972 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i64, Custom);
973 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
974 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
975 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i8, Custom);
976 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i16, Custom);
978 // Custom lower several nodes for 256-bit types.
979 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
980 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
981 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
984 // Extract subvector is special because the value type
985 // (result) is 128-bit but the source is 256-bit wide.
986 if (VT.is128BitVector())
987 setOperationAction(ISD::EXTRACT_SUBVECTOR, SVT, Custom);
989 // Do not attempt to custom lower other non-256-bit vectors
990 if (!VT.is256BitVector())
993 setOperationAction(ISD::BUILD_VECTOR, SVT, Custom);
994 setOperationAction(ISD::VECTOR_SHUFFLE, SVT, Custom);
995 setOperationAction(ISD::INSERT_VECTOR_ELT, SVT, Custom);
996 setOperationAction(ISD::EXTRACT_VECTOR_ELT, SVT, Custom);
997 setOperationAction(ISD::SCALAR_TO_VECTOR, SVT, Custom);
998 setOperationAction(ISD::INSERT_SUBVECTOR, SVT, Custom);
1001 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1002 for (unsigned i = (unsigned)MVT::v32i8; i != (unsigned)MVT::v4i64; ++i) {
1003 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
1006 // Do not attempt to promote non-256-bit vectors
1007 if (!VT.is256BitVector())
1010 setOperationAction(ISD::AND, SVT, Promote);
1011 AddPromotedToType (ISD::AND, SVT, MVT::v4i64);
1012 setOperationAction(ISD::OR, SVT, Promote);
1013 AddPromotedToType (ISD::OR, SVT, MVT::v4i64);
1014 setOperationAction(ISD::XOR, SVT, Promote);
1015 AddPromotedToType (ISD::XOR, SVT, MVT::v4i64);
1016 setOperationAction(ISD::LOAD, SVT, Promote);
1017 AddPromotedToType (ISD::LOAD, SVT, MVT::v4i64);
1018 setOperationAction(ISD::SELECT, SVT, Promote);
1019 AddPromotedToType (ISD::SELECT, SVT, MVT::v4i64);
1023 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1024 // of this type with custom code.
1025 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1026 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE; VT++) {
1027 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT, Custom);
1030 // We want to custom lower some of our intrinsics.
1031 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1034 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1035 // handle type legalization for these operations here.
1037 // FIXME: We really should do custom legalization for addition and
1038 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1039 // than generic legalization for 64-bit multiplication-with-overflow, though.
1040 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1041 // Add/Sub/Mul with overflow operations are custom lowered.
1043 setOperationAction(ISD::SADDO, VT, Custom);
1044 setOperationAction(ISD::UADDO, VT, Custom);
1045 setOperationAction(ISD::SSUBO, VT, Custom);
1046 setOperationAction(ISD::USUBO, VT, Custom);
1047 setOperationAction(ISD::SMULO, VT, Custom);
1048 setOperationAction(ISD::UMULO, VT, Custom);
1051 // There are no 8-bit 3-address imul/mul instructions
1052 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1053 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1055 if (!Subtarget->is64Bit()) {
1056 // These libcalls are not available in 32-bit.
1057 setLibcallName(RTLIB::SHL_I128, 0);
1058 setLibcallName(RTLIB::SRL_I128, 0);
1059 setLibcallName(RTLIB::SRA_I128, 0);
1062 // We have target-specific dag combine patterns for the following nodes:
1063 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1064 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1065 setTargetDAGCombine(ISD::BUILD_VECTOR);
1066 setTargetDAGCombine(ISD::SELECT);
1067 setTargetDAGCombine(ISD::SHL);
1068 setTargetDAGCombine(ISD::SRA);
1069 setTargetDAGCombine(ISD::SRL);
1070 setTargetDAGCombine(ISD::OR);
1071 setTargetDAGCombine(ISD::AND);
1072 setTargetDAGCombine(ISD::ADD);
1073 setTargetDAGCombine(ISD::SUB);
1074 setTargetDAGCombine(ISD::STORE);
1075 setTargetDAGCombine(ISD::ZERO_EXTEND);
1076 setTargetDAGCombine(ISD::SINT_TO_FP);
1077 if (Subtarget->is64Bit())
1078 setTargetDAGCombine(ISD::MUL);
1080 computeRegisterProperties();
1082 // On Darwin, -Os means optimize for size without hurting performance,
1083 // do not reduce the limit.
1084 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1085 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1086 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1087 maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1088 maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1089 maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1090 setPrefLoopAlignment(16);
1091 benefitFromCodePlacementOpt = true;
1093 setPrefFunctionAlignment(4);
1097 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1102 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1103 /// the desired ByVal argument alignment.
1104 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1107 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1108 if (VTy->getBitWidth() == 128)
1110 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1111 unsigned EltAlign = 0;
1112 getMaxByValAlign(ATy->getElementType(), EltAlign);
1113 if (EltAlign > MaxAlign)
1114 MaxAlign = EltAlign;
1115 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1116 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1117 unsigned EltAlign = 0;
1118 getMaxByValAlign(STy->getElementType(i), EltAlign);
1119 if (EltAlign > MaxAlign)
1120 MaxAlign = EltAlign;
1128 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1129 /// function arguments in the caller parameter area. For X86, aggregates
1130 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1131 /// are at 4-byte boundaries.
1132 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1133 if (Subtarget->is64Bit()) {
1134 // Max of 8 and alignment of type.
1135 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1142 if (Subtarget->hasXMM())
1143 getMaxByValAlign(Ty, Align);
1147 /// getOptimalMemOpType - Returns the target specific optimal type for load
1148 /// and store operations as a result of memset, memcpy, and memmove
1149 /// lowering. If DstAlign is zero that means it's safe to destination
1150 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1151 /// means there isn't a need to check it against alignment requirement,
1152 /// probably because the source does not need to be loaded. If
1153 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1154 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1155 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1156 /// constant so it does not need to be loaded.
1157 /// It returns EVT::Other if the type should be determined using generic
1158 /// target-independent logic.
1160 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1161 unsigned DstAlign, unsigned SrcAlign,
1162 bool NonScalarIntSafe,
1164 MachineFunction &MF) const {
1165 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1166 // linux. This is because the stack realignment code can't handle certain
1167 // cases like PR2962. This should be removed when PR2962 is fixed.
1168 const Function *F = MF.getFunction();
1169 if (NonScalarIntSafe &&
1170 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1172 (Subtarget->isUnalignedMemAccessFast() ||
1173 ((DstAlign == 0 || DstAlign >= 16) &&
1174 (SrcAlign == 0 || SrcAlign >= 16))) &&
1175 Subtarget->getStackAlignment() >= 16) {
1176 if (Subtarget->hasSSE2())
1178 if (Subtarget->hasSSE1())
1180 } else if (!MemcpyStrSrc && Size >= 8 &&
1181 !Subtarget->is64Bit() &&
1182 Subtarget->getStackAlignment() >= 8 &&
1183 Subtarget->hasXMMInt()) {
1184 // Do not use f64 to lower memcpy if source is string constant. It's
1185 // better to use i32 to avoid the loads.
1189 if (Subtarget->is64Bit() && Size >= 8)
1194 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1195 /// current function. The returned value is a member of the
1196 /// MachineJumpTableInfo::JTEntryKind enum.
1197 unsigned X86TargetLowering::getJumpTableEncoding() const {
1198 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1200 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1201 Subtarget->isPICStyleGOT())
1202 return MachineJumpTableInfo::EK_Custom32;
1204 // Otherwise, use the normal jump table encoding heuristics.
1205 return TargetLowering::getJumpTableEncoding();
1209 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1210 const MachineBasicBlock *MBB,
1211 unsigned uid,MCContext &Ctx) const{
1212 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1213 Subtarget->isPICStyleGOT());
1214 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1216 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1217 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1220 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1222 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1223 SelectionDAG &DAG) const {
1224 if (!Subtarget->is64Bit())
1225 // This doesn't have DebugLoc associated with it, but is not really the
1226 // same as a Register.
1227 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1231 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1232 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1234 const MCExpr *X86TargetLowering::
1235 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1236 MCContext &Ctx) const {
1237 // X86-64 uses RIP relative addressing based on the jump table label.
1238 if (Subtarget->isPICStyleRIPRel())
1239 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1241 // Otherwise, the reference is relative to the PIC base.
1242 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1245 // FIXME: Why this routine is here? Move to RegInfo!
1246 std::pair<const TargetRegisterClass*, uint8_t>
1247 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1248 const TargetRegisterClass *RRC = 0;
1250 switch (VT.getSimpleVT().SimpleTy) {
1252 return TargetLowering::findRepresentativeClass(VT);
1253 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1254 RRC = (Subtarget->is64Bit()
1255 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1258 RRC = X86::VR64RegisterClass;
1260 case MVT::f32: case MVT::f64:
1261 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1262 case MVT::v4f32: case MVT::v2f64:
1263 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1265 RRC = X86::VR128RegisterClass;
1268 return std::make_pair(RRC, Cost);
1271 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1272 unsigned &Offset) const {
1273 if (!Subtarget->isTargetLinux())
1276 if (Subtarget->is64Bit()) {
1277 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1279 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1292 //===----------------------------------------------------------------------===//
1293 // Return Value Calling Convention Implementation
1294 //===----------------------------------------------------------------------===//
1296 #include "X86GenCallingConv.inc"
1299 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1300 MachineFunction &MF, bool isVarArg,
1301 const SmallVectorImpl<ISD::OutputArg> &Outs,
1302 LLVMContext &Context) const {
1303 SmallVector<CCValAssign, 16> RVLocs;
1304 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1306 return CCInfo.CheckReturn(Outs, RetCC_X86);
1310 X86TargetLowering::LowerReturn(SDValue Chain,
1311 CallingConv::ID CallConv, bool isVarArg,
1312 const SmallVectorImpl<ISD::OutputArg> &Outs,
1313 const SmallVectorImpl<SDValue> &OutVals,
1314 DebugLoc dl, SelectionDAG &DAG) const {
1315 MachineFunction &MF = DAG.getMachineFunction();
1316 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1318 SmallVector<CCValAssign, 16> RVLocs;
1319 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1320 RVLocs, *DAG.getContext());
1321 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1323 // Add the regs to the liveout set for the function.
1324 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1325 for (unsigned i = 0; i != RVLocs.size(); ++i)
1326 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1327 MRI.addLiveOut(RVLocs[i].getLocReg());
1331 SmallVector<SDValue, 6> RetOps;
1332 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1333 // Operand #1 = Bytes To Pop
1334 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1337 // Copy the result values into the output registers.
1338 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1339 CCValAssign &VA = RVLocs[i];
1340 assert(VA.isRegLoc() && "Can only return in registers!");
1341 SDValue ValToCopy = OutVals[i];
1342 EVT ValVT = ValToCopy.getValueType();
1344 // If this is x86-64, and we disabled SSE, we can't return FP values,
1345 // or SSE or MMX vectors.
1346 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1347 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1348 (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
1349 report_fatal_error("SSE register return with SSE disabled");
1351 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1352 // llvm-gcc has never done it right and no one has noticed, so this
1353 // should be OK for now.
1354 if (ValVT == MVT::f64 &&
1355 (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
1356 report_fatal_error("SSE2 register return with SSE2 disabled");
1358 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1359 // the RET instruction and handled by the FP Stackifier.
1360 if (VA.getLocReg() == X86::ST0 ||
1361 VA.getLocReg() == X86::ST1) {
1362 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1363 // change the value to the FP stack register class.
1364 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1365 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1366 RetOps.push_back(ValToCopy);
1367 // Don't emit a copytoreg.
1371 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1372 // which is returned in RAX / RDX.
1373 if (Subtarget->is64Bit()) {
1374 if (ValVT == MVT::x86mmx) {
1375 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1376 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1377 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1379 // If we don't have SSE2 available, convert to v4f32 so the generated
1380 // register is legal.
1381 if (!Subtarget->hasSSE2())
1382 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1387 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1388 Flag = Chain.getValue(1);
1391 // The x86-64 ABI for returning structs by value requires that we copy
1392 // the sret argument into %rax for the return. We saved the argument into
1393 // a virtual register in the entry block, so now we copy the value out
1395 if (Subtarget->is64Bit() &&
1396 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1397 MachineFunction &MF = DAG.getMachineFunction();
1398 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1399 unsigned Reg = FuncInfo->getSRetReturnReg();
1401 "SRetReturnReg should have been set in LowerFormalArguments().");
1402 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1404 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1405 Flag = Chain.getValue(1);
1407 // RAX now acts like a return value.
1408 MRI.addLiveOut(X86::RAX);
1411 RetOps[0] = Chain; // Update chain.
1413 // Add the flag if we have it.
1415 RetOps.push_back(Flag);
1417 return DAG.getNode(X86ISD::RET_FLAG, dl,
1418 MVT::Other, &RetOps[0], RetOps.size());
1421 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
1422 if (N->getNumValues() != 1)
1424 if (!N->hasNUsesOfValue(1, 0))
1427 SDNode *Copy = *N->use_begin();
1428 if (Copy->getOpcode() != ISD::CopyToReg &&
1429 Copy->getOpcode() != ISD::FP_EXTEND)
1432 bool HasRet = false;
1433 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1435 if (UI->getOpcode() != X86ISD::RET_FLAG)
1444 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1445 ISD::NodeType ExtendKind) const {
1447 // TODO: Is this also valid on 32-bit?
1448 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1449 ReturnMVT = MVT::i8;
1451 ReturnMVT = MVT::i32;
1453 EVT MinVT = getRegisterType(Context, ReturnMVT);
1454 return VT.bitsLT(MinVT) ? MinVT : VT;
1457 /// LowerCallResult - Lower the result values of a call into the
1458 /// appropriate copies out of appropriate physical registers.
1461 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1462 CallingConv::ID CallConv, bool isVarArg,
1463 const SmallVectorImpl<ISD::InputArg> &Ins,
1464 DebugLoc dl, SelectionDAG &DAG,
1465 SmallVectorImpl<SDValue> &InVals) const {
1467 // Assign locations to each value returned by this call.
1468 SmallVector<CCValAssign, 16> RVLocs;
1469 bool Is64Bit = Subtarget->is64Bit();
1470 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1471 getTargetMachine(), RVLocs, *DAG.getContext());
1472 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1474 // Copy all of the result registers out of their specified physreg.
1475 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1476 CCValAssign &VA = RVLocs[i];
1477 EVT CopyVT = VA.getValVT();
1479 // If this is x86-64, and we disabled SSE, we can't return FP values
1480 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1481 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
1482 report_fatal_error("SSE register return with SSE disabled");
1487 // If this is a call to a function that returns an fp value on the floating
1488 // point stack, we must guarantee the the value is popped from the stack, so
1489 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1490 // if the return value is not used. We use the FpPOP_RETVAL instruction
1492 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1493 // If we prefer to use the value in xmm registers, copy it out as f80 and
1494 // use a truncate to move it from fp stack reg to xmm reg.
1495 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1496 SDValue Ops[] = { Chain, InFlag };
1497 Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
1498 MVT::Other, MVT::Glue, Ops, 2), 1);
1499 Val = Chain.getValue(0);
1501 // Round the f80 to the right size, which also moves it to the appropriate
1503 if (CopyVT != VA.getValVT())
1504 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1505 // This truncation won't change the value.
1506 DAG.getIntPtrConstant(1));
1508 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1509 CopyVT, InFlag).getValue(1);
1510 Val = Chain.getValue(0);
1512 InFlag = Chain.getValue(2);
1513 InVals.push_back(Val);
1520 //===----------------------------------------------------------------------===//
1521 // C & StdCall & Fast Calling Convention implementation
1522 //===----------------------------------------------------------------------===//
1523 // StdCall calling convention seems to be standard for many Windows' API
1524 // routines and around. It differs from C calling convention just a little:
1525 // callee should clean up the stack, not caller. Symbols should be also
1526 // decorated in some fancy way :) It doesn't support any vector arguments.
1527 // For info on fast calling convention see Fast Calling Convention (tail call)
1528 // implementation LowerX86_32FastCCCallTo.
1530 /// CallIsStructReturn - Determines whether a call uses struct return
1532 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1536 return Outs[0].Flags.isSRet();
1539 /// ArgsAreStructReturn - Determines whether a function uses struct
1540 /// return semantics.
1542 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1546 return Ins[0].Flags.isSRet();
1549 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1550 /// by "Src" to address "Dst" with size and alignment information specified by
1551 /// the specific parameter attribute. The copy will be passed as a byval
1552 /// function parameter.
1554 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1555 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1557 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1559 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1560 /*isVolatile*/false, /*AlwaysInline=*/true,
1561 MachinePointerInfo(), MachinePointerInfo());
1564 /// IsTailCallConvention - Return true if the calling convention is one that
1565 /// supports tail call optimization.
1566 static bool IsTailCallConvention(CallingConv::ID CC) {
1567 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1570 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1571 if (!CI->isTailCall())
1575 CallingConv::ID CalleeCC = CS.getCallingConv();
1576 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1582 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1583 /// a tailcall target by changing its ABI.
1584 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1585 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1589 X86TargetLowering::LowerMemArgument(SDValue Chain,
1590 CallingConv::ID CallConv,
1591 const SmallVectorImpl<ISD::InputArg> &Ins,
1592 DebugLoc dl, SelectionDAG &DAG,
1593 const CCValAssign &VA,
1594 MachineFrameInfo *MFI,
1596 // Create the nodes corresponding to a load from this parameter slot.
1597 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1598 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1599 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1602 // If value is passed by pointer we have address passed instead of the value
1604 if (VA.getLocInfo() == CCValAssign::Indirect)
1605 ValVT = VA.getLocVT();
1607 ValVT = VA.getValVT();
1609 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1610 // changed with more analysis.
1611 // In case of tail call optimization mark all arguments mutable. Since they
1612 // could be overwritten by lowering of arguments in case of a tail call.
1613 if (Flags.isByVal()) {
1614 unsigned Bytes = Flags.getByValSize();
1615 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
1616 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
1617 return DAG.getFrameIndex(FI, getPointerTy());
1619 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1620 VA.getLocMemOffset(), isImmutable);
1621 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1622 return DAG.getLoad(ValVT, dl, Chain, FIN,
1623 MachinePointerInfo::getFixedStack(FI),
1629 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1630 CallingConv::ID CallConv,
1632 const SmallVectorImpl<ISD::InputArg> &Ins,
1635 SmallVectorImpl<SDValue> &InVals)
1637 MachineFunction &MF = DAG.getMachineFunction();
1638 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1640 const Function* Fn = MF.getFunction();
1641 if (Fn->hasExternalLinkage() &&
1642 Subtarget->isTargetCygMing() &&
1643 Fn->getName() == "main")
1644 FuncInfo->setForceFramePointer(true);
1646 MachineFrameInfo *MFI = MF.getFrameInfo();
1647 bool Is64Bit = Subtarget->is64Bit();
1648 bool IsWin64 = Subtarget->isTargetWin64();
1650 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1651 "Var args not supported with calling convention fastcc or ghc");
1653 // Assign locations to all of the incoming arguments.
1654 SmallVector<CCValAssign, 16> ArgLocs;
1655 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1656 ArgLocs, *DAG.getContext());
1658 // Allocate shadow area for Win64
1660 CCInfo.AllocateStack(32, 8);
1663 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1665 unsigned LastVal = ~0U;
1667 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1668 CCValAssign &VA = ArgLocs[i];
1669 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1671 assert(VA.getValNo() != LastVal &&
1672 "Don't support value assigned to multiple locs yet");
1673 LastVal = VA.getValNo();
1675 if (VA.isRegLoc()) {
1676 EVT RegVT = VA.getLocVT();
1677 TargetRegisterClass *RC = NULL;
1678 if (RegVT == MVT::i32)
1679 RC = X86::GR32RegisterClass;
1680 else if (Is64Bit && RegVT == MVT::i64)
1681 RC = X86::GR64RegisterClass;
1682 else if (RegVT == MVT::f32)
1683 RC = X86::FR32RegisterClass;
1684 else if (RegVT == MVT::f64)
1685 RC = X86::FR64RegisterClass;
1686 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1687 RC = X86::VR256RegisterClass;
1688 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1689 RC = X86::VR128RegisterClass;
1690 else if (RegVT == MVT::x86mmx)
1691 RC = X86::VR64RegisterClass;
1693 llvm_unreachable("Unknown argument type!");
1695 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1696 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1698 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1699 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1701 if (VA.getLocInfo() == CCValAssign::SExt)
1702 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1703 DAG.getValueType(VA.getValVT()));
1704 else if (VA.getLocInfo() == CCValAssign::ZExt)
1705 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1706 DAG.getValueType(VA.getValVT()));
1707 else if (VA.getLocInfo() == CCValAssign::BCvt)
1708 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1710 if (VA.isExtInLoc()) {
1711 // Handle MMX values passed in XMM regs.
1712 if (RegVT.isVector()) {
1713 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1716 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1719 assert(VA.isMemLoc());
1720 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1723 // If value is passed via pointer - do a load.
1724 if (VA.getLocInfo() == CCValAssign::Indirect)
1725 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1726 MachinePointerInfo(), false, false, 0);
1728 InVals.push_back(ArgValue);
1731 // The x86-64 ABI for returning structs by value requires that we copy
1732 // the sret argument into %rax for the return. Save the argument into
1733 // a virtual register so that we can access it from the return points.
1734 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1735 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1736 unsigned Reg = FuncInfo->getSRetReturnReg();
1738 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1739 FuncInfo->setSRetReturnReg(Reg);
1741 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1742 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1745 unsigned StackSize = CCInfo.getNextStackOffset();
1746 // Align stack specially for tail calls.
1747 if (FuncIsMadeTailCallSafe(CallConv))
1748 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1750 // If the function takes variable number of arguments, make a frame index for
1751 // the start of the first vararg value... for expansion of llvm.va_start.
1753 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1754 CallConv != CallingConv::X86_ThisCall)) {
1755 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1758 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1760 // FIXME: We should really autogenerate these arrays
1761 static const unsigned GPR64ArgRegsWin64[] = {
1762 X86::RCX, X86::RDX, X86::R8, X86::R9
1764 static const unsigned GPR64ArgRegs64Bit[] = {
1765 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1767 static const unsigned XMMArgRegs64Bit[] = {
1768 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1769 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1771 const unsigned *GPR64ArgRegs;
1772 unsigned NumXMMRegs = 0;
1775 // The XMM registers which might contain var arg parameters are shadowed
1776 // in their paired GPR. So we only need to save the GPR to their home
1778 TotalNumIntRegs = 4;
1779 GPR64ArgRegs = GPR64ArgRegsWin64;
1781 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1782 GPR64ArgRegs = GPR64ArgRegs64Bit;
1784 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1786 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1789 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1790 assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
1791 "SSE register cannot be used when SSE is disabled!");
1792 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1793 "SSE register cannot be used when SSE is disabled!");
1794 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
1795 // Kernel mode asks for SSE to be disabled, so don't push them
1797 TotalNumXMMRegs = 0;
1800 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
1801 // Get to the caller-allocated home save location. Add 8 to account
1802 // for the return address.
1803 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1804 FuncInfo->setRegSaveFrameIndex(
1805 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1806 // Fixup to set vararg frame on shadow area (4 x i64).
1808 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1810 // For X86-64, if there are vararg parameters that are passed via
1811 // registers, then we must store them to their spots on the stack so they
1812 // may be loaded by deferencing the result of va_next.
1813 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1814 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1815 FuncInfo->setRegSaveFrameIndex(
1816 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1820 // Store the integer parameter registers.
1821 SmallVector<SDValue, 8> MemOps;
1822 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1824 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1825 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1826 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1827 DAG.getIntPtrConstant(Offset));
1828 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1829 X86::GR64RegisterClass);
1830 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1832 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1833 MachinePointerInfo::getFixedStack(
1834 FuncInfo->getRegSaveFrameIndex(), Offset),
1836 MemOps.push_back(Store);
1840 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1841 // Now store the XMM (fp + vector) parameter registers.
1842 SmallVector<SDValue, 11> SaveXMMOps;
1843 SaveXMMOps.push_back(Chain);
1845 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1846 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1847 SaveXMMOps.push_back(ALVal);
1849 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1850 FuncInfo->getRegSaveFrameIndex()));
1851 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1852 FuncInfo->getVarArgsFPOffset()));
1854 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1855 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1856 X86::VR128RegisterClass);
1857 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1858 SaveXMMOps.push_back(Val);
1860 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1862 &SaveXMMOps[0], SaveXMMOps.size()));
1865 if (!MemOps.empty())
1866 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1867 &MemOps[0], MemOps.size());
1871 // Some CCs need callee pop.
1872 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt)) {
1873 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1875 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1876 // If this is an sret function, the return should pop the hidden pointer.
1877 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1878 FuncInfo->setBytesToPopOnReturn(4);
1882 // RegSaveFrameIndex is X86-64 only.
1883 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1884 if (CallConv == CallingConv::X86_FastCall ||
1885 CallConv == CallingConv::X86_ThisCall)
1886 // fastcc functions can't have varargs.
1887 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1894 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1895 SDValue StackPtr, SDValue Arg,
1896 DebugLoc dl, SelectionDAG &DAG,
1897 const CCValAssign &VA,
1898 ISD::ArgFlagsTy Flags) const {
1899 unsigned LocMemOffset = VA.getLocMemOffset();
1900 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1901 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1902 if (Flags.isByVal())
1903 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1905 return DAG.getStore(Chain, dl, Arg, PtrOff,
1906 MachinePointerInfo::getStack(LocMemOffset),
1910 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1911 /// optimization is performed and it is required.
1913 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1914 SDValue &OutRetAddr, SDValue Chain,
1915 bool IsTailCall, bool Is64Bit,
1916 int FPDiff, DebugLoc dl) const {
1917 // Adjust the Return address stack slot.
1918 EVT VT = getPointerTy();
1919 OutRetAddr = getReturnAddressFrameIndex(DAG);
1921 // Load the "old" Return address.
1922 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1924 return SDValue(OutRetAddr.getNode(), 1);
1927 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
1928 /// optimization is performed and it is required (FPDiff!=0).
1930 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1931 SDValue Chain, SDValue RetAddrFrIdx,
1932 bool Is64Bit, int FPDiff, DebugLoc dl) {
1933 // Store the return address to the appropriate stack slot.
1934 if (!FPDiff) return Chain;
1935 // Calculate the new stack slot for the return address.
1936 int SlotSize = Is64Bit ? 8 : 4;
1937 int NewReturnAddrFI =
1938 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1939 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1940 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1941 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1942 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1948 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1949 CallingConv::ID CallConv, bool isVarArg,
1951 const SmallVectorImpl<ISD::OutputArg> &Outs,
1952 const SmallVectorImpl<SDValue> &OutVals,
1953 const SmallVectorImpl<ISD::InputArg> &Ins,
1954 DebugLoc dl, SelectionDAG &DAG,
1955 SmallVectorImpl<SDValue> &InVals) const {
1956 MachineFunction &MF = DAG.getMachineFunction();
1957 bool Is64Bit = Subtarget->is64Bit();
1958 bool IsWin64 = Subtarget->isTargetWin64();
1959 bool IsStructRet = CallIsStructReturn(Outs);
1960 bool IsSibcall = false;
1963 // Check if it's really possible to do a tail call.
1964 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1965 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1966 Outs, OutVals, Ins, DAG);
1968 // Sibcalls are automatically detected tailcalls which do not require
1970 if (!GuaranteedTailCallOpt && isTailCall)
1977 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1978 "Var args not supported with calling convention fastcc or ghc");
1980 // Analyze operands of the call, assigning locations to each operand.
1981 SmallVector<CCValAssign, 16> ArgLocs;
1982 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1983 ArgLocs, *DAG.getContext());
1985 // Allocate shadow area for Win64
1987 CCInfo.AllocateStack(32, 8);
1990 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
1992 // Get a count of how many bytes are to be pushed on the stack.
1993 unsigned NumBytes = CCInfo.getNextStackOffset();
1995 // This is a sibcall. The memory operands are available in caller's
1996 // own caller's stack.
1998 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
1999 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2002 if (isTailCall && !IsSibcall) {
2003 // Lower arguments at fp - stackoffset + fpdiff.
2004 unsigned NumBytesCallerPushed =
2005 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
2006 FPDiff = NumBytesCallerPushed - NumBytes;
2008 // Set the delta of movement of the returnaddr stackslot.
2009 // But only set if delta is greater than previous delta.
2010 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
2011 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
2015 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2017 SDValue RetAddrFrIdx;
2018 // Load return address for tail calls.
2019 if (isTailCall && FPDiff)
2020 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2021 Is64Bit, FPDiff, dl);
2023 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2024 SmallVector<SDValue, 8> MemOpChains;
2027 // Walk the register/memloc assignments, inserting copies/loads. In the case
2028 // of tail call optimization arguments are handle later.
2029 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2030 CCValAssign &VA = ArgLocs[i];
2031 EVT RegVT = VA.getLocVT();
2032 SDValue Arg = OutVals[i];
2033 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2034 bool isByVal = Flags.isByVal();
2036 // Promote the value if needed.
2037 switch (VA.getLocInfo()) {
2038 default: llvm_unreachable("Unknown loc info!");
2039 case CCValAssign::Full: break;
2040 case CCValAssign::SExt:
2041 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2043 case CCValAssign::ZExt:
2044 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2046 case CCValAssign::AExt:
2047 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
2048 // Special case: passing MMX values in XMM registers.
2049 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2050 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2051 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2053 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2055 case CCValAssign::BCvt:
2056 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2058 case CCValAssign::Indirect: {
2059 // Store the argument.
2060 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2061 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2062 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2063 MachinePointerInfo::getFixedStack(FI),
2070 if (VA.isRegLoc()) {
2071 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2072 if (isVarArg && IsWin64) {
2073 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2074 // shadow reg if callee is a varargs function.
2075 unsigned ShadowReg = 0;
2076 switch (VA.getLocReg()) {
2077 case X86::XMM0: ShadowReg = X86::RCX; break;
2078 case X86::XMM1: ShadowReg = X86::RDX; break;
2079 case X86::XMM2: ShadowReg = X86::R8; break;
2080 case X86::XMM3: ShadowReg = X86::R9; break;
2083 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2085 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2086 assert(VA.isMemLoc());
2087 if (StackPtr.getNode() == 0)
2088 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2089 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2090 dl, DAG, VA, Flags));
2094 if (!MemOpChains.empty())
2095 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2096 &MemOpChains[0], MemOpChains.size());
2098 // Build a sequence of copy-to-reg nodes chained together with token chain
2099 // and flag operands which copy the outgoing args into registers.
2101 // Tail call byval lowering might overwrite argument registers so in case of
2102 // tail call optimization the copies to registers are lowered later.
2104 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2105 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2106 RegsToPass[i].second, InFlag);
2107 InFlag = Chain.getValue(1);
2110 if (Subtarget->isPICStyleGOT()) {
2111 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2114 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2115 DAG.getNode(X86ISD::GlobalBaseReg,
2116 DebugLoc(), getPointerTy()),
2118 InFlag = Chain.getValue(1);
2120 // If we are tail calling and generating PIC/GOT style code load the
2121 // address of the callee into ECX. The value in ecx is used as target of
2122 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2123 // for tail calls on PIC/GOT architectures. Normally we would just put the
2124 // address of GOT into ebx and then call target@PLT. But for tail calls
2125 // ebx would be restored (since ebx is callee saved) before jumping to the
2128 // Note: The actual moving to ECX is done further down.
2129 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2130 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2131 !G->getGlobal()->hasProtectedVisibility())
2132 Callee = LowerGlobalAddress(Callee, DAG);
2133 else if (isa<ExternalSymbolSDNode>(Callee))
2134 Callee = LowerExternalSymbol(Callee, DAG);
2138 if (Is64Bit && isVarArg && !IsWin64) {
2139 // From AMD64 ABI document:
2140 // For calls that may call functions that use varargs or stdargs
2141 // (prototype-less calls or calls to functions containing ellipsis (...) in
2142 // the declaration) %al is used as hidden argument to specify the number
2143 // of SSE registers used. The contents of %al do not need to match exactly
2144 // the number of registers, but must be an ubound on the number of SSE
2145 // registers used and is in the range 0 - 8 inclusive.
2147 // Count the number of XMM registers allocated.
2148 static const unsigned XMMArgRegs[] = {
2149 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2150 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2152 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2153 assert((Subtarget->hasXMM() || !NumXMMRegs)
2154 && "SSE registers cannot be used when SSE is disabled");
2156 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2157 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2158 InFlag = Chain.getValue(1);
2162 // For tail calls lower the arguments to the 'real' stack slot.
2164 // Force all the incoming stack arguments to be loaded from the stack
2165 // before any new outgoing arguments are stored to the stack, because the
2166 // outgoing stack slots may alias the incoming argument stack slots, and
2167 // the alias isn't otherwise explicit. This is slightly more conservative
2168 // than necessary, because it means that each store effectively depends
2169 // on every argument instead of just those arguments it would clobber.
2170 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2172 SmallVector<SDValue, 8> MemOpChains2;
2175 // Do not flag preceding copytoreg stuff together with the following stuff.
2177 if (GuaranteedTailCallOpt) {
2178 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2179 CCValAssign &VA = ArgLocs[i];
2182 assert(VA.isMemLoc());
2183 SDValue Arg = OutVals[i];
2184 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2185 // Create frame index.
2186 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2187 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2188 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2189 FIN = DAG.getFrameIndex(FI, getPointerTy());
2191 if (Flags.isByVal()) {
2192 // Copy relative to framepointer.
2193 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2194 if (StackPtr.getNode() == 0)
2195 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2197 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2199 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2203 // Store relative to framepointer.
2204 MemOpChains2.push_back(
2205 DAG.getStore(ArgChain, dl, Arg, FIN,
2206 MachinePointerInfo::getFixedStack(FI),
2212 if (!MemOpChains2.empty())
2213 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2214 &MemOpChains2[0], MemOpChains2.size());
2216 // Copy arguments to their registers.
2217 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2218 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2219 RegsToPass[i].second, InFlag);
2220 InFlag = Chain.getValue(1);
2224 // Store the return address to the appropriate stack slot.
2225 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2229 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2230 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2231 // In the 64-bit large code model, we have to make all calls
2232 // through a register, since the call instruction's 32-bit
2233 // pc-relative offset may not be large enough to hold the whole
2235 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2236 // If the callee is a GlobalAddress node (quite common, every direct call
2237 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2240 // We should use extra load for direct calls to dllimported functions in
2242 const GlobalValue *GV = G->getGlobal();
2243 if (!GV->hasDLLImportLinkage()) {
2244 unsigned char OpFlags = 0;
2245 bool ExtraLoad = false;
2246 unsigned WrapperKind = ISD::DELETED_NODE;
2248 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2249 // external symbols most go through the PLT in PIC mode. If the symbol
2250 // has hidden or protected visibility, or if it is static or local, then
2251 // we don't need to use the PLT - we can directly call it.
2252 if (Subtarget->isTargetELF() &&
2253 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2254 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2255 OpFlags = X86II::MO_PLT;
2256 } else if (Subtarget->isPICStyleStubAny() &&
2257 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2258 (!Subtarget->getTargetTriple().isMacOSX() ||
2259 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2260 // PC-relative references to external symbols should go through $stub,
2261 // unless we're building with the leopard linker or later, which
2262 // automatically synthesizes these stubs.
2263 OpFlags = X86II::MO_DARWIN_STUB;
2264 } else if (Subtarget->isPICStyleRIPRel() &&
2265 isa<Function>(GV) &&
2266 cast<Function>(GV)->hasFnAttr(Attribute::NonLazyBind)) {
2267 // If the function is marked as non-lazy, generate an indirect call
2268 // which loads from the GOT directly. This avoids runtime overhead
2269 // at the cost of eager binding (and one extra byte of encoding).
2270 OpFlags = X86II::MO_GOTPCREL;
2271 WrapperKind = X86ISD::WrapperRIP;
2275 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2276 G->getOffset(), OpFlags);
2278 // Add a wrapper if needed.
2279 if (WrapperKind != ISD::DELETED_NODE)
2280 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2281 // Add extra indirection if needed.
2283 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2284 MachinePointerInfo::getGOT(),
2287 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2288 unsigned char OpFlags = 0;
2290 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2291 // external symbols should go through the PLT.
2292 if (Subtarget->isTargetELF() &&
2293 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2294 OpFlags = X86II::MO_PLT;
2295 } else if (Subtarget->isPICStyleStubAny() &&
2296 (!Subtarget->getTargetTriple().isMacOSX() ||
2297 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2298 // PC-relative references to external symbols should go through $stub,
2299 // unless we're building with the leopard linker or later, which
2300 // automatically synthesizes these stubs.
2301 OpFlags = X86II::MO_DARWIN_STUB;
2304 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2308 // Returns a chain & a flag for retval copy to use.
2309 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2310 SmallVector<SDValue, 8> Ops;
2312 if (!IsSibcall && isTailCall) {
2313 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2314 DAG.getIntPtrConstant(0, true), InFlag);
2315 InFlag = Chain.getValue(1);
2318 Ops.push_back(Chain);
2319 Ops.push_back(Callee);
2322 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2324 // Add argument registers to the end of the list so that they are known live
2326 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2327 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2328 RegsToPass[i].second.getValueType()));
2330 // Add an implicit use GOT pointer in EBX.
2331 if (!isTailCall && Subtarget->isPICStyleGOT())
2332 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2334 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2335 if (Is64Bit && isVarArg && !IsWin64)
2336 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2338 if (InFlag.getNode())
2339 Ops.push_back(InFlag);
2343 //// If this is the first return lowered for this function, add the regs
2344 //// to the liveout set for the function.
2345 // This isn't right, although it's probably harmless on x86; liveouts
2346 // should be computed from returns not tail calls. Consider a void
2347 // function making a tail call to a function returning int.
2348 return DAG.getNode(X86ISD::TC_RETURN, dl,
2349 NodeTys, &Ops[0], Ops.size());
2352 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2353 InFlag = Chain.getValue(1);
2355 // Create the CALLSEQ_END node.
2356 unsigned NumBytesForCalleeToPush;
2357 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt))
2358 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2359 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2360 // If this is a call to a struct-return function, the callee
2361 // pops the hidden struct pointer, so we have to push it back.
2362 // This is common for Darwin/X86, Linux & Mingw32 targets.
2363 NumBytesForCalleeToPush = 4;
2365 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2367 // Returns a flag for retval copy to use.
2369 Chain = DAG.getCALLSEQ_END(Chain,
2370 DAG.getIntPtrConstant(NumBytes, true),
2371 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2374 InFlag = Chain.getValue(1);
2377 // Handle result values, copying them out of physregs into vregs that we
2379 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2380 Ins, dl, DAG, InVals);
2384 //===----------------------------------------------------------------------===//
2385 // Fast Calling Convention (tail call) implementation
2386 //===----------------------------------------------------------------------===//
2388 // Like std call, callee cleans arguments, convention except that ECX is
2389 // reserved for storing the tail called function address. Only 2 registers are
2390 // free for argument passing (inreg). Tail call optimization is performed
2392 // * tailcallopt is enabled
2393 // * caller/callee are fastcc
2394 // On X86_64 architecture with GOT-style position independent code only local
2395 // (within module) calls are supported at the moment.
2396 // To keep the stack aligned according to platform abi the function
2397 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2398 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2399 // If a tail called function callee has more arguments than the caller the
2400 // caller needs to make sure that there is room to move the RETADDR to. This is
2401 // achieved by reserving an area the size of the argument delta right after the
2402 // original REtADDR, but before the saved framepointer or the spilled registers
2403 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2415 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2416 /// for a 16 byte align requirement.
2418 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2419 SelectionDAG& DAG) const {
2420 MachineFunction &MF = DAG.getMachineFunction();
2421 const TargetMachine &TM = MF.getTarget();
2422 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2423 unsigned StackAlignment = TFI.getStackAlignment();
2424 uint64_t AlignMask = StackAlignment - 1;
2425 int64_t Offset = StackSize;
2426 uint64_t SlotSize = TD->getPointerSize();
2427 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2428 // Number smaller than 12 so just add the difference.
2429 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2431 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2432 Offset = ((~AlignMask) & Offset) + StackAlignment +
2433 (StackAlignment-SlotSize);
2438 /// MatchingStackOffset - Return true if the given stack call argument is
2439 /// already available in the same position (relatively) of the caller's
2440 /// incoming argument stack.
2442 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2443 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2444 const X86InstrInfo *TII) {
2445 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2447 if (Arg.getOpcode() == ISD::CopyFromReg) {
2448 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2449 if (!TargetRegisterInfo::isVirtualRegister(VR))
2451 MachineInstr *Def = MRI->getVRegDef(VR);
2454 if (!Flags.isByVal()) {
2455 if (!TII->isLoadFromStackSlot(Def, FI))
2458 unsigned Opcode = Def->getOpcode();
2459 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2460 Def->getOperand(1).isFI()) {
2461 FI = Def->getOperand(1).getIndex();
2462 Bytes = Flags.getByValSize();
2466 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2467 if (Flags.isByVal())
2468 // ByVal argument is passed in as a pointer but it's now being
2469 // dereferenced. e.g.
2470 // define @foo(%struct.X* %A) {
2471 // tail call @bar(%struct.X* byval %A)
2474 SDValue Ptr = Ld->getBasePtr();
2475 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2478 FI = FINode->getIndex();
2479 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
2480 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
2481 FI = FINode->getIndex();
2482 Bytes = Flags.getByValSize();
2486 assert(FI != INT_MAX);
2487 if (!MFI->isFixedObjectIndex(FI))
2489 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2492 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2493 /// for tail call optimization. Targets which want to do tail call
2494 /// optimization should implement this function.
2496 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2497 CallingConv::ID CalleeCC,
2499 bool isCalleeStructRet,
2500 bool isCallerStructRet,
2501 const SmallVectorImpl<ISD::OutputArg> &Outs,
2502 const SmallVectorImpl<SDValue> &OutVals,
2503 const SmallVectorImpl<ISD::InputArg> &Ins,
2504 SelectionDAG& DAG) const {
2505 if (!IsTailCallConvention(CalleeCC) &&
2506 CalleeCC != CallingConv::C)
2509 // If -tailcallopt is specified, make fastcc functions tail-callable.
2510 const MachineFunction &MF = DAG.getMachineFunction();
2511 const Function *CallerF = DAG.getMachineFunction().getFunction();
2512 CallingConv::ID CallerCC = CallerF->getCallingConv();
2513 bool CCMatch = CallerCC == CalleeCC;
2515 if (GuaranteedTailCallOpt) {
2516 if (IsTailCallConvention(CalleeCC) && CCMatch)
2521 // Look for obvious safe cases to perform tail call optimization that do not
2522 // require ABI changes. This is what gcc calls sibcall.
2524 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2525 // emit a special epilogue.
2526 if (RegInfo->needsStackRealignment(MF))
2529 // Also avoid sibcall optimization if either caller or callee uses struct
2530 // return semantics.
2531 if (isCalleeStructRet || isCallerStructRet)
2534 // An stdcall caller is expected to clean up its arguments; the callee
2535 // isn't going to do that.
2536 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2539 // Do not sibcall optimize vararg calls unless all arguments are passed via
2541 if (isVarArg && !Outs.empty()) {
2543 // Optimizing for varargs on Win64 is unlikely to be safe without
2544 // additional testing.
2545 if (Subtarget->isTargetWin64())
2548 SmallVector<CCValAssign, 16> ArgLocs;
2549 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2550 getTargetMachine(), ArgLocs, *DAG.getContext());
2552 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2553 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2554 if (!ArgLocs[i].isRegLoc())
2558 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2559 // Therefore if it's not used by the call it is not safe to optimize this into
2561 bool Unused = false;
2562 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2569 SmallVector<CCValAssign, 16> RVLocs;
2570 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
2571 getTargetMachine(), RVLocs, *DAG.getContext());
2572 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2573 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2574 CCValAssign &VA = RVLocs[i];
2575 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2580 // If the calling conventions do not match, then we'd better make sure the
2581 // results are returned in the same way as what the caller expects.
2583 SmallVector<CCValAssign, 16> RVLocs1;
2584 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2585 getTargetMachine(), RVLocs1, *DAG.getContext());
2586 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2588 SmallVector<CCValAssign, 16> RVLocs2;
2589 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2590 getTargetMachine(), RVLocs2, *DAG.getContext());
2591 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2593 if (RVLocs1.size() != RVLocs2.size())
2595 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2596 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2598 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2600 if (RVLocs1[i].isRegLoc()) {
2601 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2604 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2610 // If the callee takes no arguments then go on to check the results of the
2612 if (!Outs.empty()) {
2613 // Check if stack adjustment is needed. For now, do not do this if any
2614 // argument is passed on the stack.
2615 SmallVector<CCValAssign, 16> ArgLocs;
2616 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2617 getTargetMachine(), ArgLocs, *DAG.getContext());
2619 // Allocate shadow area for Win64
2620 if (Subtarget->isTargetWin64()) {
2621 CCInfo.AllocateStack(32, 8);
2624 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2625 if (CCInfo.getNextStackOffset()) {
2626 MachineFunction &MF = DAG.getMachineFunction();
2627 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2630 // Check if the arguments are already laid out in the right way as
2631 // the caller's fixed stack objects.
2632 MachineFrameInfo *MFI = MF.getFrameInfo();
2633 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2634 const X86InstrInfo *TII =
2635 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2636 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2637 CCValAssign &VA = ArgLocs[i];
2638 SDValue Arg = OutVals[i];
2639 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2640 if (VA.getLocInfo() == CCValAssign::Indirect)
2642 if (!VA.isRegLoc()) {
2643 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2650 // If the tailcall address may be in a register, then make sure it's
2651 // possible to register allocate for it. In 32-bit, the call address can
2652 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2653 // callee-saved registers are restored. These happen to be the same
2654 // registers used to pass 'inreg' arguments so watch out for those.
2655 if (!Subtarget->is64Bit() &&
2656 !isa<GlobalAddressSDNode>(Callee) &&
2657 !isa<ExternalSymbolSDNode>(Callee)) {
2658 unsigned NumInRegs = 0;
2659 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2660 CCValAssign &VA = ArgLocs[i];
2663 unsigned Reg = VA.getLocReg();
2666 case X86::EAX: case X86::EDX: case X86::ECX:
2667 if (++NumInRegs == 3)
2679 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2680 return X86::createFastISel(funcInfo);
2684 //===----------------------------------------------------------------------===//
2685 // Other Lowering Hooks
2686 //===----------------------------------------------------------------------===//
2688 static bool MayFoldLoad(SDValue Op) {
2689 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2692 static bool MayFoldIntoStore(SDValue Op) {
2693 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2696 static bool isTargetShuffle(unsigned Opcode) {
2698 default: return false;
2699 case X86ISD::PSHUFD:
2700 case X86ISD::PSHUFHW:
2701 case X86ISD::PSHUFLW:
2702 case X86ISD::SHUFPD:
2703 case X86ISD::PALIGN:
2704 case X86ISD::SHUFPS:
2705 case X86ISD::MOVLHPS:
2706 case X86ISD::MOVLHPD:
2707 case X86ISD::MOVHLPS:
2708 case X86ISD::MOVLPS:
2709 case X86ISD::MOVLPD:
2710 case X86ISD::MOVSHDUP:
2711 case X86ISD::MOVSLDUP:
2712 case X86ISD::MOVDDUP:
2715 case X86ISD::UNPCKLPS:
2716 case X86ISD::UNPCKLPD:
2717 case X86ISD::VUNPCKLPSY:
2718 case X86ISD::VUNPCKLPDY:
2719 case X86ISD::PUNPCKLWD:
2720 case X86ISD::PUNPCKLBW:
2721 case X86ISD::PUNPCKLDQ:
2722 case X86ISD::PUNPCKLQDQ:
2723 case X86ISD::UNPCKHPS:
2724 case X86ISD::UNPCKHPD:
2725 case X86ISD::VUNPCKHPSY:
2726 case X86ISD::VUNPCKHPDY:
2727 case X86ISD::PUNPCKHWD:
2728 case X86ISD::PUNPCKHBW:
2729 case X86ISD::PUNPCKHDQ:
2730 case X86ISD::PUNPCKHQDQ:
2731 case X86ISD::VPERMILPS:
2732 case X86ISD::VPERMILPSY:
2733 case X86ISD::VPERMILPD:
2734 case X86ISD::VPERMILPDY:
2740 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2741 SDValue V1, SelectionDAG &DAG) {
2743 default: llvm_unreachable("Unknown x86 shuffle node");
2744 case X86ISD::MOVSHDUP:
2745 case X86ISD::MOVSLDUP:
2746 case X86ISD::MOVDDUP:
2747 return DAG.getNode(Opc, dl, VT, V1);
2753 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2754 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2756 default: llvm_unreachable("Unknown x86 shuffle node");
2757 case X86ISD::PSHUFD:
2758 case X86ISD::PSHUFHW:
2759 case X86ISD::PSHUFLW:
2760 case X86ISD::VPERMILPS:
2761 case X86ISD::VPERMILPSY:
2762 case X86ISD::VPERMILPD:
2763 case X86ISD::VPERMILPDY:
2764 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2770 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2771 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2773 default: llvm_unreachable("Unknown x86 shuffle node");
2774 case X86ISD::PALIGN:
2775 case X86ISD::SHUFPD:
2776 case X86ISD::SHUFPS:
2777 return DAG.getNode(Opc, dl, VT, V1, V2,
2778 DAG.getConstant(TargetMask, MVT::i8));
2783 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2784 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2786 default: llvm_unreachable("Unknown x86 shuffle node");
2787 case X86ISD::MOVLHPS:
2788 case X86ISD::MOVLHPD:
2789 case X86ISD::MOVHLPS:
2790 case X86ISD::MOVLPS:
2791 case X86ISD::MOVLPD:
2794 case X86ISD::UNPCKLPS:
2795 case X86ISD::UNPCKLPD:
2796 case X86ISD::VUNPCKLPSY:
2797 case X86ISD::VUNPCKLPDY:
2798 case X86ISD::PUNPCKLWD:
2799 case X86ISD::PUNPCKLBW:
2800 case X86ISD::PUNPCKLDQ:
2801 case X86ISD::PUNPCKLQDQ:
2802 case X86ISD::UNPCKHPS:
2803 case X86ISD::UNPCKHPD:
2804 case X86ISD::VUNPCKHPSY:
2805 case X86ISD::VUNPCKHPDY:
2806 case X86ISD::PUNPCKHWD:
2807 case X86ISD::PUNPCKHBW:
2808 case X86ISD::PUNPCKHDQ:
2809 case X86ISD::PUNPCKHQDQ:
2810 return DAG.getNode(Opc, dl, VT, V1, V2);
2815 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2816 MachineFunction &MF = DAG.getMachineFunction();
2817 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2818 int ReturnAddrIndex = FuncInfo->getRAIndex();
2820 if (ReturnAddrIndex == 0) {
2821 // Set up a frame object for the return address.
2822 uint64_t SlotSize = TD->getPointerSize();
2823 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2825 FuncInfo->setRAIndex(ReturnAddrIndex);
2828 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2832 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2833 bool hasSymbolicDisplacement) {
2834 // Offset should fit into 32 bit immediate field.
2835 if (!isInt<32>(Offset))
2838 // If we don't have a symbolic displacement - we don't have any extra
2840 if (!hasSymbolicDisplacement)
2843 // FIXME: Some tweaks might be needed for medium code model.
2844 if (M != CodeModel::Small && M != CodeModel::Kernel)
2847 // For small code model we assume that latest object is 16MB before end of 31
2848 // bits boundary. We may also accept pretty large negative constants knowing
2849 // that all objects are in the positive half of address space.
2850 if (M == CodeModel::Small && Offset < 16*1024*1024)
2853 // For kernel code model we know that all object resist in the negative half
2854 // of 32bits address space. We may not accept negative offsets, since they may
2855 // be just off and we may accept pretty large positive ones.
2856 if (M == CodeModel::Kernel && Offset > 0)
2862 /// isCalleePop - Determines whether the callee is required to pop its
2863 /// own arguments. Callee pop is necessary to support tail calls.
2864 bool X86::isCalleePop(CallingConv::ID CallingConv,
2865 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
2869 switch (CallingConv) {
2872 case CallingConv::X86_StdCall:
2874 case CallingConv::X86_FastCall:
2876 case CallingConv::X86_ThisCall:
2878 case CallingConv::Fast:
2880 case CallingConv::GHC:
2885 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2886 /// specific condition code, returning the condition code and the LHS/RHS of the
2887 /// comparison to make.
2888 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2889 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2891 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2892 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2893 // X > -1 -> X == 0, jump !sign.
2894 RHS = DAG.getConstant(0, RHS.getValueType());
2895 return X86::COND_NS;
2896 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2897 // X < 0 -> X == 0, jump on sign.
2899 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2901 RHS = DAG.getConstant(0, RHS.getValueType());
2902 return X86::COND_LE;
2906 switch (SetCCOpcode) {
2907 default: llvm_unreachable("Invalid integer condition!");
2908 case ISD::SETEQ: return X86::COND_E;
2909 case ISD::SETGT: return X86::COND_G;
2910 case ISD::SETGE: return X86::COND_GE;
2911 case ISD::SETLT: return X86::COND_L;
2912 case ISD::SETLE: return X86::COND_LE;
2913 case ISD::SETNE: return X86::COND_NE;
2914 case ISD::SETULT: return X86::COND_B;
2915 case ISD::SETUGT: return X86::COND_A;
2916 case ISD::SETULE: return X86::COND_BE;
2917 case ISD::SETUGE: return X86::COND_AE;
2921 // First determine if it is required or is profitable to flip the operands.
2923 // If LHS is a foldable load, but RHS is not, flip the condition.
2924 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
2925 !ISD::isNON_EXTLoad(RHS.getNode())) {
2926 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2927 std::swap(LHS, RHS);
2930 switch (SetCCOpcode) {
2936 std::swap(LHS, RHS);
2940 // On a floating point condition, the flags are set as follows:
2942 // 0 | 0 | 0 | X > Y
2943 // 0 | 0 | 1 | X < Y
2944 // 1 | 0 | 0 | X == Y
2945 // 1 | 1 | 1 | unordered
2946 switch (SetCCOpcode) {
2947 default: llvm_unreachable("Condcode should be pre-legalized away");
2949 case ISD::SETEQ: return X86::COND_E;
2950 case ISD::SETOLT: // flipped
2952 case ISD::SETGT: return X86::COND_A;
2953 case ISD::SETOLE: // flipped
2955 case ISD::SETGE: return X86::COND_AE;
2956 case ISD::SETUGT: // flipped
2958 case ISD::SETLT: return X86::COND_B;
2959 case ISD::SETUGE: // flipped
2961 case ISD::SETLE: return X86::COND_BE;
2963 case ISD::SETNE: return X86::COND_NE;
2964 case ISD::SETUO: return X86::COND_P;
2965 case ISD::SETO: return X86::COND_NP;
2967 case ISD::SETUNE: return X86::COND_INVALID;
2971 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2972 /// code. Current x86 isa includes the following FP cmov instructions:
2973 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2974 static bool hasFPCMov(unsigned X86CC) {
2990 /// isFPImmLegal - Returns true if the target can instruction select the
2991 /// specified FP immediate natively. If false, the legalizer will
2992 /// materialize the FP immediate as a load from a constant pool.
2993 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2994 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2995 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3001 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3002 /// the specified range (L, H].
3003 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3004 return (Val < 0) || (Val >= Low && Val < Hi);
3007 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3008 /// specified value.
3009 static bool isUndefOrEqual(int Val, int CmpVal) {
3010 if (Val < 0 || Val == CmpVal)
3015 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3016 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
3017 /// the second operand.
3018 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3019 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
3020 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3021 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3022 return (Mask[0] < 2 && Mask[1] < 2);
3026 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
3027 SmallVector<int, 8> M;
3029 return ::isPSHUFDMask(M, N->getValueType(0));
3032 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3033 /// is suitable for input to PSHUFHW.
3034 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3035 if (VT != MVT::v8i16)
3038 // Lower quadword copied in order or undef.
3039 for (int i = 0; i != 4; ++i)
3040 if (Mask[i] >= 0 && Mask[i] != i)
3043 // Upper quadword shuffled.
3044 for (int i = 4; i != 8; ++i)
3045 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
3051 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
3052 SmallVector<int, 8> M;
3054 return ::isPSHUFHWMask(M, N->getValueType(0));
3057 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3058 /// is suitable for input to PSHUFLW.
3059 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3060 if (VT != MVT::v8i16)
3063 // Upper quadword copied in order.
3064 for (int i = 4; i != 8; ++i)
3065 if (Mask[i] >= 0 && Mask[i] != i)
3068 // Lower quadword shuffled.
3069 for (int i = 0; i != 4; ++i)
3076 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
3077 SmallVector<int, 8> M;
3079 return ::isPSHUFLWMask(M, N->getValueType(0));
3082 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3083 /// is suitable for input to PALIGNR.
3084 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
3086 int i, e = VT.getVectorNumElements();
3087 if (VT.getSizeInBits() != 128 && VT.getSizeInBits() != 64)
3090 // Do not handle v2i64 / v2f64 shuffles with palignr.
3091 if (e < 4 || !hasSSSE3)
3094 for (i = 0; i != e; ++i)
3098 // All undef, not a palignr.
3102 // Make sure we're shifting in the right direction.
3106 int s = Mask[i] - i;
3108 // Check the rest of the elements to see if they are consecutive.
3109 for (++i; i != e; ++i) {
3111 if (m >= 0 && m != s+i)
3117 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3118 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
3119 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3120 int NumElems = VT.getVectorNumElements();
3121 if (NumElems != 2 && NumElems != 4)
3124 int Half = NumElems / 2;
3125 for (int i = 0; i < Half; ++i)
3126 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3128 for (int i = Half; i < NumElems; ++i)
3129 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3135 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3136 SmallVector<int, 8> M;
3138 return ::isSHUFPMask(M, N->getValueType(0));
3141 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3142 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3143 /// half elements to come from vector 1 (which would equal the dest.) and
3144 /// the upper half to come from vector 2.
3145 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3146 int NumElems = VT.getVectorNumElements();
3148 if (NumElems != 2 && NumElems != 4)
3151 int Half = NumElems / 2;
3152 for (int i = 0; i < Half; ++i)
3153 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3155 for (int i = Half; i < NumElems; ++i)
3156 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3161 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3162 SmallVector<int, 8> M;
3164 return isCommutedSHUFPMask(M, N->getValueType(0));
3167 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3168 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3169 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3170 EVT VT = N->getValueType(0);
3171 unsigned NumElems = VT.getVectorNumElements();
3173 if (VT.getSizeInBits() != 128)
3179 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3180 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3181 isUndefOrEqual(N->getMaskElt(1), 7) &&
3182 isUndefOrEqual(N->getMaskElt(2), 2) &&
3183 isUndefOrEqual(N->getMaskElt(3), 3);
3186 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3187 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3189 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3190 EVT VT = N->getValueType(0);
3191 unsigned NumElems = VT.getVectorNumElements();
3193 if (VT.getSizeInBits() != 128)
3199 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3200 isUndefOrEqual(N->getMaskElt(1), 3) &&
3201 isUndefOrEqual(N->getMaskElt(2), 2) &&
3202 isUndefOrEqual(N->getMaskElt(3), 3);
3205 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3206 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3207 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3208 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3210 if (NumElems != 2 && NumElems != 4)
3213 for (unsigned i = 0; i < NumElems/2; ++i)
3214 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3217 for (unsigned i = NumElems/2; i < NumElems; ++i)
3218 if (!isUndefOrEqual(N->getMaskElt(i), i))
3224 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3225 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3226 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3227 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3229 if ((NumElems != 2 && NumElems != 4)
3230 || N->getValueType(0).getSizeInBits() > 128)
3233 for (unsigned i = 0; i < NumElems/2; ++i)
3234 if (!isUndefOrEqual(N->getMaskElt(i), i))
3237 for (unsigned i = 0; i < NumElems/2; ++i)
3238 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3244 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3245 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3246 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3247 bool V2IsSplat = false) {
3248 int NumElts = VT.getVectorNumElements();
3250 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3251 "Unsupported vector type for unpckh");
3253 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8)
3256 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3257 // independently on 128-bit lanes.
3258 unsigned NumLanes = VT.getSizeInBits()/128;
3259 unsigned NumLaneElts = NumElts/NumLanes;
3262 unsigned End = NumLaneElts;
3263 for (unsigned s = 0; s < NumLanes; ++s) {
3264 for (unsigned i = Start, j = s * NumLaneElts;
3268 int BitI1 = Mask[i+1];
3269 if (!isUndefOrEqual(BitI, j))
3272 if (!isUndefOrEqual(BitI1, NumElts))
3275 if (!isUndefOrEqual(BitI1, j + NumElts))
3279 // Process the next 128 bits.
3280 Start += NumLaneElts;
3287 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3288 SmallVector<int, 8> M;
3290 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3293 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3294 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3295 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3296 bool V2IsSplat = false) {
3297 int NumElts = VT.getVectorNumElements();
3299 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3300 "Unsupported vector type for unpckh");
3302 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8)
3305 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3306 // independently on 128-bit lanes.
3307 unsigned NumLanes = VT.getSizeInBits()/128;
3308 unsigned NumLaneElts = NumElts/NumLanes;
3311 unsigned End = NumLaneElts;
3312 for (unsigned l = 0; l != NumLanes; ++l) {
3313 for (unsigned i = Start, j = (l*NumLaneElts)+NumLaneElts/2;
3314 i != End; i += 2, ++j) {
3316 int BitI1 = Mask[i+1];
3317 if (!isUndefOrEqual(BitI, j))
3320 if (isUndefOrEqual(BitI1, NumElts))
3323 if (!isUndefOrEqual(BitI1, j+NumElts))
3327 // Process the next 128 bits.
3328 Start += NumLaneElts;
3334 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3335 SmallVector<int, 8> M;
3337 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3340 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3341 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3343 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3344 int NumElems = VT.getVectorNumElements();
3345 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3348 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3349 // independently on 128-bit lanes.
3350 unsigned NumLanes = VT.getSizeInBits() / 128;
3351 unsigned NumLaneElts = NumElems / NumLanes;
3353 for (unsigned s = 0; s < NumLanes; ++s) {
3354 for (unsigned i = s * NumLaneElts, j = s * NumLaneElts;
3355 i != NumLaneElts * (s + 1);
3358 int BitI1 = Mask[i+1];
3360 if (!isUndefOrEqual(BitI, j))
3362 if (!isUndefOrEqual(BitI1, j))
3370 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3371 SmallVector<int, 8> M;
3373 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3376 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3377 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3379 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3380 int NumElems = VT.getVectorNumElements();
3381 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3384 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3386 int BitI1 = Mask[i+1];
3387 if (!isUndefOrEqual(BitI, j))
3389 if (!isUndefOrEqual(BitI1, j))
3395 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3396 SmallVector<int, 8> M;
3398 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3401 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3402 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3403 /// MOVSD, and MOVD, i.e. setting the lowest element.
3404 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3405 if (VT.getVectorElementType().getSizeInBits() < 32)
3408 int NumElts = VT.getVectorNumElements();
3410 if (!isUndefOrEqual(Mask[0], NumElts))
3413 for (int i = 1; i < NumElts; ++i)
3414 if (!isUndefOrEqual(Mask[i], i))
3420 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3421 SmallVector<int, 8> M;
3423 return ::isMOVLMask(M, N->getValueType(0));
3426 /// isVPERMILPDMask - Return true if the specified VECTOR_SHUFFLE operand
3427 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
3428 /// Note that VPERMIL mask matching is different depending whether theunderlying
3429 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
3430 /// to the same elements of the low, but to the higher half of the source.
3431 /// In VPERMILPD the two lanes could be shuffled independently of each other
3432 /// with the same restriction that lanes can't be crossed.
3433 static bool isVPERMILPDMask(const SmallVectorImpl<int> &Mask, EVT VT,
3434 const X86Subtarget *Subtarget) {
3435 int NumElts = VT.getVectorNumElements();
3436 int NumLanes = VT.getSizeInBits()/128;
3438 if (!Subtarget->hasAVX())
3441 // Match any permutation of 128-bit vector with 64-bit types
3442 if (NumLanes == 1 && NumElts != 2)
3445 // Only match 256-bit with 32 types
3446 if (VT.getSizeInBits() == 256 && NumElts != 4)
3449 // The mask on the high lane is independent of the low. Both can match
3450 // any element in inside its own lane, but can't cross.
3451 int LaneSize = NumElts/NumLanes;
3452 for (int l = 0; l < NumLanes; ++l)
3453 for (int i = l*LaneSize; i < LaneSize*(l+1); ++i) {
3454 int LaneStart = l*LaneSize;
3455 if (!isUndefOrInRange(Mask[i], LaneStart, LaneStart+LaneSize))
3462 /// isVPERMILPSMask - Return true if the specified VECTOR_SHUFFLE operand
3463 /// specifies a shuffle of elements that is suitable for input to VPERMILPS*.
3464 /// Note that VPERMIL mask matching is different depending whether theunderlying
3465 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
3466 /// to the same elements of the low, but to the higher half of the source.
3467 /// In VPERMILPD the two lanes could be shuffled independently of each other
3468 /// with the same restriction that lanes can't be crossed.
3469 static bool isVPERMILPSMask(const SmallVectorImpl<int> &Mask, EVT VT,
3470 const X86Subtarget *Subtarget) {
3471 unsigned NumElts = VT.getVectorNumElements();
3472 unsigned NumLanes = VT.getSizeInBits()/128;
3474 if (!Subtarget->hasAVX())
3477 // Match any permutation of 128-bit vector with 32-bit types
3478 if (NumLanes == 1 && NumElts != 4)
3481 // Only match 256-bit with 32 types
3482 if (VT.getSizeInBits() == 256 && NumElts != 8)
3485 // The mask on the high lane should be the same as the low. Actually,
3486 // they can differ if any of the corresponding index in a lane is undef
3487 // and the other stays in range.
3488 int LaneSize = NumElts/NumLanes;
3489 for (int i = 0; i < LaneSize; ++i) {
3490 int HighElt = i+LaneSize;
3491 if (Mask[i] < 0 && (isUndefOrInRange(Mask[HighElt], LaneSize, NumElts)))
3493 if (Mask[HighElt] < 0 && (isUndefOrInRange(Mask[i], 0, LaneSize)))
3495 if (Mask[HighElt]-Mask[i] != LaneSize)
3502 /// getShuffleVPERMILPSImmediate - Return the appropriate immediate to shuffle
3503 /// the specified VECTOR_MASK mask with VPERMILPS* instructions.
3504 static unsigned getShuffleVPERMILPSImmediate(SDNode *N) {
3505 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3506 EVT VT = SVOp->getValueType(0);
3508 int NumElts = VT.getVectorNumElements();
3509 int NumLanes = VT.getSizeInBits()/128;
3510 int LaneSize = NumElts/NumLanes;
3512 // Although the mask is equal for both lanes do it twice to get the cases
3513 // where a mask will match because the same mask element is undef on the
3514 // first half but valid on the second. This would get pathological cases
3515 // such as: shuffle <u, 0, 1, 2, 4, 4, 5, 6>, which is completely valid.
3517 for (int l = 0; l < NumLanes; ++l) {
3518 for (int i = 0; i < LaneSize; ++i) {
3519 int MaskElt = SVOp->getMaskElt(i+(l*LaneSize));
3522 if (MaskElt >= LaneSize)
3523 MaskElt -= LaneSize;
3524 Mask |= MaskElt << (i*2);
3531 /// getShuffleVPERMILPDImmediate - Return the appropriate immediate to shuffle
3532 /// the specified VECTOR_MASK mask with VPERMILPD* instructions.
3533 static unsigned getShuffleVPERMILPDImmediate(SDNode *N) {
3534 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3535 EVT VT = SVOp->getValueType(0);
3537 int NumElts = VT.getVectorNumElements();
3538 int NumLanes = VT.getSizeInBits()/128;
3541 int LaneSize = NumElts/NumLanes;
3542 for (int l = 0; l < NumLanes; ++l)
3543 for (int i = l*LaneSize; i < LaneSize*(l+1); ++i) {
3544 int MaskElt = SVOp->getMaskElt(i);
3547 Mask |= (MaskElt-l*LaneSize) << i;
3553 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3554 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3555 /// element of vector 2 and the other elements to come from vector 1 in order.
3556 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3557 bool V2IsSplat = false, bool V2IsUndef = false) {
3558 int NumOps = VT.getVectorNumElements();
3559 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3562 if (!isUndefOrEqual(Mask[0], 0))
3565 for (int i = 1; i < NumOps; ++i)
3566 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3567 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3568 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3574 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3575 bool V2IsUndef = false) {
3576 SmallVector<int, 8> M;
3578 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3581 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3582 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3583 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
3584 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N,
3585 const X86Subtarget *Subtarget) {
3586 if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
3589 // The second vector must be undef
3590 if (N->getOperand(1).getOpcode() != ISD::UNDEF)
3593 EVT VT = N->getValueType(0);
3594 unsigned NumElems = VT.getVectorNumElements();
3596 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3597 (VT.getSizeInBits() == 256 && NumElems != 8))
3600 // "i+1" is the value the indexed mask element must have
3601 for (unsigned i = 0; i < NumElems; i += 2)
3602 if (!isUndefOrEqual(N->getMaskElt(i), i+1) ||
3603 !isUndefOrEqual(N->getMaskElt(i+1), i+1))
3609 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3610 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3611 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
3612 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N,
3613 const X86Subtarget *Subtarget) {
3614 if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
3617 // The second vector must be undef
3618 if (N->getOperand(1).getOpcode() != ISD::UNDEF)
3621 EVT VT = N->getValueType(0);
3622 unsigned NumElems = VT.getVectorNumElements();
3624 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3625 (VT.getSizeInBits() == 256 && NumElems != 8))
3628 // "i" is the value the indexed mask element must have
3629 for (unsigned i = 0; i < NumElems; i += 2)
3630 if (!isUndefOrEqual(N->getMaskElt(i), i) ||
3631 !isUndefOrEqual(N->getMaskElt(i+1), i))
3637 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3638 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3639 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3640 int e = N->getValueType(0).getVectorNumElements() / 2;
3642 for (int i = 0; i < e; ++i)
3643 if (!isUndefOrEqual(N->getMaskElt(i), i))
3645 for (int i = 0; i < e; ++i)
3646 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3651 /// isVEXTRACTF128Index - Return true if the specified
3652 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3653 /// suitable for input to VEXTRACTF128.
3654 bool X86::isVEXTRACTF128Index(SDNode *N) {
3655 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3658 // The index should be aligned on a 128-bit boundary.
3660 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3662 unsigned VL = N->getValueType(0).getVectorNumElements();
3663 unsigned VBits = N->getValueType(0).getSizeInBits();
3664 unsigned ElSize = VBits / VL;
3665 bool Result = (Index * ElSize) % 128 == 0;
3670 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
3671 /// operand specifies a subvector insert that is suitable for input to
3673 bool X86::isVINSERTF128Index(SDNode *N) {
3674 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3677 // The index should be aligned on a 128-bit boundary.
3679 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3681 unsigned VL = N->getValueType(0).getVectorNumElements();
3682 unsigned VBits = N->getValueType(0).getSizeInBits();
3683 unsigned ElSize = VBits / VL;
3684 bool Result = (Index * ElSize) % 128 == 0;
3689 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3690 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3691 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3692 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3693 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3695 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3697 for (int i = 0; i < NumOperands; ++i) {
3698 int Val = SVOp->getMaskElt(NumOperands-i-1);
3699 if (Val < 0) Val = 0;
3700 if (Val >= NumOperands) Val -= NumOperands;
3702 if (i != NumOperands - 1)
3708 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3709 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3710 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3711 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3713 // 8 nodes, but we only care about the last 4.
3714 for (unsigned i = 7; i >= 4; --i) {
3715 int Val = SVOp->getMaskElt(i);
3724 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3725 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3726 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3727 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3729 // 8 nodes, but we only care about the first 4.
3730 for (int i = 3; i >= 0; --i) {
3731 int Val = SVOp->getMaskElt(i);
3740 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3741 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3742 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3743 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3744 EVT VVT = N->getValueType(0);
3745 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3749 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3750 Val = SVOp->getMaskElt(i);
3754 assert(Val - i > 0 && "PALIGNR imm should be positive");
3755 return (Val - i) * EltSize;
3758 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
3759 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3761 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
3762 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3763 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
3766 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3768 EVT VecVT = N->getOperand(0).getValueType();
3769 EVT ElVT = VecVT.getVectorElementType();
3771 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3772 return Index / NumElemsPerChunk;
3775 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
3776 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3778 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
3779 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3780 llvm_unreachable("Illegal insert subvector for VINSERTF128");
3783 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3785 EVT VecVT = N->getValueType(0);
3786 EVT ElVT = VecVT.getVectorElementType();
3788 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3789 return Index / NumElemsPerChunk;
3792 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3794 bool X86::isZeroNode(SDValue Elt) {
3795 return ((isa<ConstantSDNode>(Elt) &&
3796 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3797 (isa<ConstantFPSDNode>(Elt) &&
3798 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3801 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3802 /// their permute mask.
3803 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3804 SelectionDAG &DAG) {
3805 EVT VT = SVOp->getValueType(0);
3806 unsigned NumElems = VT.getVectorNumElements();
3807 SmallVector<int, 8> MaskVec;
3809 for (unsigned i = 0; i != NumElems; ++i) {
3810 int idx = SVOp->getMaskElt(i);
3812 MaskVec.push_back(idx);
3813 else if (idx < (int)NumElems)
3814 MaskVec.push_back(idx + NumElems);
3816 MaskVec.push_back(idx - NumElems);
3818 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3819 SVOp->getOperand(0), &MaskVec[0]);
3822 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3823 /// the two vector operands have swapped position.
3824 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3825 unsigned NumElems = VT.getVectorNumElements();
3826 for (unsigned i = 0; i != NumElems; ++i) {
3830 else if (idx < (int)NumElems)
3831 Mask[i] = idx + NumElems;
3833 Mask[i] = idx - NumElems;
3837 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3838 /// match movhlps. The lower half elements should come from upper half of
3839 /// V1 (and in order), and the upper half elements should come from the upper
3840 /// half of V2 (and in order).
3841 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3842 if (Op->getValueType(0).getVectorNumElements() != 4)
3844 for (unsigned i = 0, e = 2; i != e; ++i)
3845 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3847 for (unsigned i = 2; i != 4; ++i)
3848 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3853 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3854 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3856 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3857 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3859 N = N->getOperand(0).getNode();
3860 if (!ISD::isNON_EXTLoad(N))
3863 *LD = cast<LoadSDNode>(N);
3867 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3868 /// match movlp{s|d}. The lower half elements should come from lower half of
3869 /// V1 (and in order), and the upper half elements should come from the upper
3870 /// half of V2 (and in order). And since V1 will become the source of the
3871 /// MOVLP, it must be either a vector load or a scalar load to vector.
3872 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3873 ShuffleVectorSDNode *Op) {
3874 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3876 // Is V2 is a vector load, don't do this transformation. We will try to use
3877 // load folding shufps op.
3878 if (ISD::isNON_EXTLoad(V2))
3881 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3883 if (NumElems != 2 && NumElems != 4)
3885 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3886 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3888 for (unsigned i = NumElems/2; i != NumElems; ++i)
3889 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3894 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3896 static bool isSplatVector(SDNode *N) {
3897 if (N->getOpcode() != ISD::BUILD_VECTOR)
3900 SDValue SplatValue = N->getOperand(0);
3901 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3902 if (N->getOperand(i) != SplatValue)
3907 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3908 /// to an zero vector.
3909 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3910 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3911 SDValue V1 = N->getOperand(0);
3912 SDValue V2 = N->getOperand(1);
3913 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3914 for (unsigned i = 0; i != NumElems; ++i) {
3915 int Idx = N->getMaskElt(i);
3916 if (Idx >= (int)NumElems) {
3917 unsigned Opc = V2.getOpcode();
3918 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3920 if (Opc != ISD::BUILD_VECTOR ||
3921 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3923 } else if (Idx >= 0) {
3924 unsigned Opc = V1.getOpcode();
3925 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3927 if (Opc != ISD::BUILD_VECTOR ||
3928 !X86::isZeroNode(V1.getOperand(Idx)))
3935 /// getZeroVector - Returns a vector of specified type with all zero elements.
3937 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3939 assert(VT.isVector() && "Expected a vector type");
3941 // Always build SSE zero vectors as <4 x i32> bitcasted
3942 // to their dest type. This ensures they get CSE'd.
3944 if (VT.getSizeInBits() == 128) { // SSE
3945 if (HasSSE2) { // SSE2
3946 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3947 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3949 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3950 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3952 } else if (VT.getSizeInBits() == 256) { // AVX
3953 // 256-bit logic and arithmetic instructions in AVX are
3954 // all floating-point, no support for integer ops. Default
3955 // to emitting fp zeroed vectors then.
3956 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3957 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
3958 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
3960 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3963 /// getOnesVector - Returns a vector of specified type with all bits set.
3964 /// Always build ones vectors as <4 x i32>. For 256-bit types, use two
3965 /// <4 x i32> inserted in a <8 x i32> appropriately. Then bitcast to their
3966 /// original type, ensuring they get CSE'd.
3967 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3968 assert(VT.isVector() && "Expected a vector type");
3969 assert((VT.is128BitVector() || VT.is256BitVector())
3970 && "Expected a 128-bit or 256-bit vector type");
3972 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3973 SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
3974 Cst, Cst, Cst, Cst);
3976 if (VT.is256BitVector()) {
3977 SDValue InsV = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, MVT::v8i32),
3978 Vec, DAG.getConstant(0, MVT::i32), DAG, dl);
3979 Vec = Insert128BitVector(InsV, Vec,
3980 DAG.getConstant(4 /* NumElems/2 */, MVT::i32), DAG, dl);
3983 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
3986 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3987 /// that point to V2 points to its first element.
3988 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3989 EVT VT = SVOp->getValueType(0);
3990 unsigned NumElems = VT.getVectorNumElements();
3992 bool Changed = false;
3993 SmallVector<int, 8> MaskVec;
3994 SVOp->getMask(MaskVec);
3996 for (unsigned i = 0; i != NumElems; ++i) {
3997 if (MaskVec[i] > (int)NumElems) {
3998 MaskVec[i] = NumElems;
4003 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
4004 SVOp->getOperand(1), &MaskVec[0]);
4005 return SDValue(SVOp, 0);
4008 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4009 /// operation of specified width.
4010 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4012 unsigned NumElems = VT.getVectorNumElements();
4013 SmallVector<int, 8> Mask;
4014 Mask.push_back(NumElems);
4015 for (unsigned i = 1; i != NumElems; ++i)
4017 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4020 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4021 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4023 unsigned NumElems = VT.getVectorNumElements();
4024 SmallVector<int, 8> Mask;
4025 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4027 Mask.push_back(i + NumElems);
4029 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4032 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4033 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4035 unsigned NumElems = VT.getVectorNumElements();
4036 unsigned Half = NumElems/2;
4037 SmallVector<int, 8> Mask;
4038 for (unsigned i = 0; i != Half; ++i) {
4039 Mask.push_back(i + Half);
4040 Mask.push_back(i + NumElems + Half);
4042 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4045 // PromoteSplatv8v16 - All i16 and i8 vector types can't be used directly by
4046 // a generic shuffle instruction because the target has no such instructions.
4047 // Generate shuffles which repeat i16 and i8 several times until they can be
4048 // represented by v4f32 and then be manipulated by target suported shuffles.
4049 static SDValue PromoteSplatv8v16(SDValue V, SelectionDAG &DAG, int &EltNo) {
4050 EVT VT = V.getValueType();
4051 int NumElems = VT.getVectorNumElements();
4052 DebugLoc dl = V.getDebugLoc();
4054 while (NumElems > 4) {
4055 if (EltNo < NumElems/2) {
4056 V = getUnpackl(DAG, dl, VT, V, V);
4058 V = getUnpackh(DAG, dl, VT, V, V);
4059 EltNo -= NumElems/2;
4066 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
4067 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
4068 EVT VT = V.getValueType();
4069 DebugLoc dl = V.getDebugLoc();
4070 assert((VT.getSizeInBits() == 128 || VT.getSizeInBits() == 256)
4071 && "Vector size not supported");
4073 bool Is128 = VT.getSizeInBits() == 128;
4074 EVT NVT = Is128 ? MVT::v4f32 : MVT::v8f32;
4075 V = DAG.getNode(ISD::BITCAST, dl, NVT, V);
4078 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
4079 V = DAG.getVectorShuffle(NVT, dl, V, DAG.getUNDEF(NVT), &SplatMask[0]);
4081 // The second half of indicies refer to the higher part, which is a
4082 // duplication of the lower one. This makes this shuffle a perfect match
4083 // for the VPERM instruction.
4084 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
4085 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
4086 V = DAG.getVectorShuffle(NVT, dl, V, DAG.getUNDEF(NVT), &SplatMask[0]);
4089 return DAG.getNode(ISD::BITCAST, dl, VT, V);
4092 /// PromoteVectorToScalarSplat - Since there's no native support for
4093 /// scalar_to_vector for 256-bit AVX, a 128-bit scalar_to_vector +
4094 /// INSERT_SUBVECTOR is generated. Recognize this idiom and do the
4095 /// shuffle before the insertion, this yields less instructions in the end.
4096 static SDValue PromoteVectorToScalarSplat(ShuffleVectorSDNode *SV,
4097 SelectionDAG &DAG) {
4098 EVT SrcVT = SV->getValueType(0);
4099 SDValue V1 = SV->getOperand(0);
4100 DebugLoc dl = SV->getDebugLoc();
4101 int NumElems = SrcVT.getVectorNumElements();
4103 assert(SrcVT.is256BitVector() && "unknown howto handle vector type");
4105 SmallVector<int, 4> Mask;
4106 for (int i = 0; i < NumElems/2; ++i)
4107 Mask.push_back(SV->getMaskElt(i));
4109 EVT SVT = EVT::getVectorVT(*DAG.getContext(), SrcVT.getVectorElementType(),
4111 SDValue SV1 = DAG.getVectorShuffle(SVT, dl, V1.getOperand(1),
4112 DAG.getUNDEF(SVT), &Mask[0]);
4113 SDValue InsV = Insert128BitVector(DAG.getUNDEF(SrcVT), SV1,
4114 DAG.getConstant(0, MVT::i32), DAG, dl);
4116 return Insert128BitVector(InsV, SV1,
4117 DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
4120 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32 and
4121 /// v8i32, v16i16 or v32i8 to v8f32.
4122 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
4123 EVT SrcVT = SV->getValueType(0);
4124 SDValue V1 = SV->getOperand(0);
4125 DebugLoc dl = SV->getDebugLoc();
4127 int EltNo = SV->getSplatIndex();
4128 int NumElems = SrcVT.getVectorNumElements();
4129 unsigned Size = SrcVT.getSizeInBits();
4131 // Extract the 128-bit part containing the splat element and update
4132 // the splat element index when it refers to the higher register.
4134 unsigned Idx = (EltNo > NumElems/2) ? NumElems/2 : 0;
4135 V1 = Extract128BitVector(V1, DAG.getConstant(Idx, MVT::i32), DAG, dl);
4137 EltNo -= NumElems/2;
4140 // Make this 128-bit vector duplicate i8 and i16 elements
4142 V1 = PromoteSplatv8v16(V1, DAG, EltNo);
4144 // Recreate the 256-bit vector and place the same 128-bit vector
4145 // into the low and high part. This is necessary because we want
4146 // to use VPERM to shuffle the v8f32 vector, and VPERM only shuffles
4147 // inside each separate v4f32 lane.
4149 SDValue InsV = Insert128BitVector(DAG.getUNDEF(SrcVT), V1,
4150 DAG.getConstant(0, MVT::i32), DAG, dl);
4151 V1 = Insert128BitVector(InsV, V1,
4152 DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
4155 return getLegalSplat(DAG, V1, EltNo);
4158 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4159 /// vector of zero or undef vector. This produces a shuffle where the low
4160 /// element of V2 is swizzled into the zero/undef vector, landing at element
4161 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4162 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4163 bool isZero, bool HasSSE2,
4164 SelectionDAG &DAG) {
4165 EVT VT = V2.getValueType();
4167 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
4168 unsigned NumElems = VT.getVectorNumElements();
4169 SmallVector<int, 16> MaskVec;
4170 for (unsigned i = 0; i != NumElems; ++i)
4171 // If this is the insertion idx, put the low elt of V2 here.
4172 MaskVec.push_back(i == Idx ? NumElems : i);
4173 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
4176 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4177 /// element of the result of the vector shuffle.
4178 static SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
4181 return SDValue(); // Limit search depth.
4183 SDValue V = SDValue(N, 0);
4184 EVT VT = V.getValueType();
4185 unsigned Opcode = V.getOpcode();
4187 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4188 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4189 Index = SV->getMaskElt(Index);
4192 return DAG.getUNDEF(VT.getVectorElementType());
4194 int NumElems = VT.getVectorNumElements();
4195 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
4196 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
4199 // Recurse into target specific vector shuffles to find scalars.
4200 if (isTargetShuffle(Opcode)) {
4201 int NumElems = VT.getVectorNumElements();
4202 SmallVector<unsigned, 16> ShuffleMask;
4206 case X86ISD::SHUFPS:
4207 case X86ISD::SHUFPD:
4208 ImmN = N->getOperand(N->getNumOperands()-1);
4209 DecodeSHUFPSMask(NumElems,
4210 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4213 case X86ISD::PUNPCKHBW:
4214 case X86ISD::PUNPCKHWD:
4215 case X86ISD::PUNPCKHDQ:
4216 case X86ISD::PUNPCKHQDQ:
4217 DecodePUNPCKHMask(NumElems, ShuffleMask);
4219 case X86ISD::UNPCKHPS:
4220 case X86ISD::UNPCKHPD:
4221 case X86ISD::VUNPCKHPSY:
4222 case X86ISD::VUNPCKHPDY:
4223 DecodeUNPCKHPMask(NumElems, ShuffleMask);
4225 case X86ISD::PUNPCKLBW:
4226 case X86ISD::PUNPCKLWD:
4227 case X86ISD::PUNPCKLDQ:
4228 case X86ISD::PUNPCKLQDQ:
4229 DecodePUNPCKLMask(VT, ShuffleMask);
4231 case X86ISD::UNPCKLPS:
4232 case X86ISD::UNPCKLPD:
4233 case X86ISD::VUNPCKLPSY:
4234 case X86ISD::VUNPCKLPDY:
4235 DecodeUNPCKLPMask(VT, ShuffleMask);
4237 case X86ISD::MOVHLPS:
4238 DecodeMOVHLPSMask(NumElems, ShuffleMask);
4240 case X86ISD::MOVLHPS:
4241 DecodeMOVLHPSMask(NumElems, ShuffleMask);
4243 case X86ISD::PSHUFD:
4244 ImmN = N->getOperand(N->getNumOperands()-1);
4245 DecodePSHUFMask(NumElems,
4246 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4249 case X86ISD::PSHUFHW:
4250 ImmN = N->getOperand(N->getNumOperands()-1);
4251 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4254 case X86ISD::PSHUFLW:
4255 ImmN = N->getOperand(N->getNumOperands()-1);
4256 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4260 case X86ISD::MOVSD: {
4261 // The index 0 always comes from the first element of the second source,
4262 // this is why MOVSS and MOVSD are used in the first place. The other
4263 // elements come from the other positions of the first source vector.
4264 unsigned OpNum = (Index == 0) ? 1 : 0;
4265 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
4268 case X86ISD::VPERMILPS:
4269 ImmN = N->getOperand(N->getNumOperands()-1);
4270 DecodeVPERMILPSMask(4, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4273 case X86ISD::VPERMILPSY:
4274 ImmN = N->getOperand(N->getNumOperands()-1);
4275 DecodeVPERMILPSMask(8, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4278 case X86ISD::VPERMILPD:
4279 ImmN = N->getOperand(N->getNumOperands()-1);
4280 DecodeVPERMILPDMask(2, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4283 case X86ISD::VPERMILPDY:
4284 ImmN = N->getOperand(N->getNumOperands()-1);
4285 DecodeVPERMILPDMask(4, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4289 assert("not implemented for target shuffle node");
4293 Index = ShuffleMask[Index];
4295 return DAG.getUNDEF(VT.getVectorElementType());
4297 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
4298 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
4302 // Actual nodes that may contain scalar elements
4303 if (Opcode == ISD::BITCAST) {
4304 V = V.getOperand(0);
4305 EVT SrcVT = V.getValueType();
4306 unsigned NumElems = VT.getVectorNumElements();
4308 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4312 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4313 return (Index == 0) ? V.getOperand(0)
4314 : DAG.getUNDEF(VT.getVectorElementType());
4316 if (V.getOpcode() == ISD::BUILD_VECTOR)
4317 return V.getOperand(Index);
4322 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
4323 /// shuffle operation which come from a consecutively from a zero. The
4324 /// search can start in two different directions, from left or right.
4326 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
4327 bool ZerosFromLeft, SelectionDAG &DAG) {
4330 while (i < NumElems) {
4331 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
4332 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
4333 if (!(Elt.getNode() &&
4334 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
4342 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
4343 /// MaskE correspond consecutively to elements from one of the vector operands,
4344 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
4346 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
4347 int OpIdx, int NumElems, unsigned &OpNum) {
4348 bool SeenV1 = false;
4349 bool SeenV2 = false;
4351 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
4352 int Idx = SVOp->getMaskElt(i);
4353 // Ignore undef indicies
4362 // Only accept consecutive elements from the same vector
4363 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
4367 OpNum = SeenV1 ? 0 : 1;
4371 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
4372 /// logical left shift of a vector.
4373 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4374 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4375 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4376 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4377 false /* check zeros from right */, DAG);
4383 // Considering the elements in the mask that are not consecutive zeros,
4384 // check if they consecutively come from only one of the source vectors.
4386 // V1 = {X, A, B, C} 0
4388 // vector_shuffle V1, V2 <1, 2, 3, X>
4390 if (!isShuffleMaskConsecutive(SVOp,
4391 0, // Mask Start Index
4392 NumElems-NumZeros-1, // Mask End Index
4393 NumZeros, // Where to start looking in the src vector
4394 NumElems, // Number of elements in vector
4395 OpSrc)) // Which source operand ?
4400 ShVal = SVOp->getOperand(OpSrc);
4404 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4405 /// logical left shift of a vector.
4406 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4407 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4408 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4409 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4410 true /* check zeros from left */, DAG);
4416 // Considering the elements in the mask that are not consecutive zeros,
4417 // check if they consecutively come from only one of the source vectors.
4419 // 0 { A, B, X, X } = V2
4421 // vector_shuffle V1, V2 <X, X, 4, 5>
4423 if (!isShuffleMaskConsecutive(SVOp,
4424 NumZeros, // Mask Start Index
4425 NumElems-1, // Mask End Index
4426 0, // Where to start looking in the src vector
4427 NumElems, // Number of elements in vector
4428 OpSrc)) // Which source operand ?
4433 ShVal = SVOp->getOperand(OpSrc);
4437 /// isVectorShift - Returns true if the shuffle can be implemented as a
4438 /// logical left or right shift of a vector.
4439 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4440 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4441 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4442 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4448 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4450 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4451 unsigned NumNonZero, unsigned NumZero,
4453 const TargetLowering &TLI) {
4457 DebugLoc dl = Op.getDebugLoc();
4460 for (unsigned i = 0; i < 16; ++i) {
4461 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4462 if (ThisIsNonZero && First) {
4464 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4466 V = DAG.getUNDEF(MVT::v8i16);
4471 SDValue ThisElt(0, 0), LastElt(0, 0);
4472 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4473 if (LastIsNonZero) {
4474 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4475 MVT::i16, Op.getOperand(i-1));
4477 if (ThisIsNonZero) {
4478 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4479 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4480 ThisElt, DAG.getConstant(8, MVT::i8));
4482 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4486 if (ThisElt.getNode())
4487 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4488 DAG.getIntPtrConstant(i/2));
4492 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4495 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4497 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4498 unsigned NumNonZero, unsigned NumZero,
4500 const TargetLowering &TLI) {
4504 DebugLoc dl = Op.getDebugLoc();
4507 for (unsigned i = 0; i < 8; ++i) {
4508 bool isNonZero = (NonZeros & (1 << i)) != 0;
4512 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4514 V = DAG.getUNDEF(MVT::v8i16);
4517 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4518 MVT::v8i16, V, Op.getOperand(i),
4519 DAG.getIntPtrConstant(i));
4526 /// getVShift - Return a vector logical shift node.
4528 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4529 unsigned NumBits, SelectionDAG &DAG,
4530 const TargetLowering &TLI, DebugLoc dl) {
4531 EVT ShVT = MVT::v2i64;
4532 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4533 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4534 return DAG.getNode(ISD::BITCAST, dl, VT,
4535 DAG.getNode(Opc, dl, ShVT, SrcOp,
4536 DAG.getConstant(NumBits,
4537 TLI.getShiftAmountTy(SrcOp.getValueType()))));
4541 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4542 SelectionDAG &DAG) const {
4544 // Check if the scalar load can be widened into a vector load. And if
4545 // the address is "base + cst" see if the cst can be "absorbed" into
4546 // the shuffle mask.
4547 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4548 SDValue Ptr = LD->getBasePtr();
4549 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4551 EVT PVT = LD->getValueType(0);
4552 if (PVT != MVT::i32 && PVT != MVT::f32)
4557 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4558 FI = FINode->getIndex();
4560 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4561 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4562 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4563 Offset = Ptr.getConstantOperandVal(1);
4564 Ptr = Ptr.getOperand(0);
4569 // FIXME: 256-bit vector instructions don't require a strict alignment,
4570 // improve this code to support it better.
4571 unsigned RequiredAlign = VT.getSizeInBits()/8;
4572 SDValue Chain = LD->getChain();
4573 // Make sure the stack object alignment is at least 16 or 32.
4574 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4575 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4576 if (MFI->isFixedObjectIndex(FI)) {
4577 // Can't change the alignment. FIXME: It's possible to compute
4578 // the exact stack offset and reference FI + adjust offset instead.
4579 // If someone *really* cares about this. That's the way to implement it.
4582 MFI->setObjectAlignment(FI, RequiredAlign);
4586 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4587 // Ptr + (Offset & ~15).
4590 if ((Offset % RequiredAlign) & 3)
4592 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4594 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4595 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4597 int EltNo = (Offset - StartOffset) >> 2;
4598 int NumElems = VT.getVectorNumElements();
4600 EVT CanonVT = VT.getSizeInBits() == 128 ? MVT::v4i32 : MVT::v8i32;
4601 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4602 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4603 LD->getPointerInfo().getWithOffset(StartOffset),
4606 // Canonicalize it to a v4i32 or v8i32 shuffle.
4607 SmallVector<int, 8> Mask;
4608 for (int i = 0; i < NumElems; ++i)
4609 Mask.push_back(EltNo);
4611 V1 = DAG.getNode(ISD::BITCAST, dl, CanonVT, V1);
4612 return DAG.getNode(ISD::BITCAST, dl, NVT,
4613 DAG.getVectorShuffle(CanonVT, dl, V1,
4614 DAG.getUNDEF(CanonVT),&Mask[0]));
4620 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4621 /// vector of type 'VT', see if the elements can be replaced by a single large
4622 /// load which has the same value as a build_vector whose operands are 'elts'.
4624 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4626 /// FIXME: we'd also like to handle the case where the last elements are zero
4627 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4628 /// There's even a handy isZeroNode for that purpose.
4629 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4630 DebugLoc &DL, SelectionDAG &DAG) {
4631 EVT EltVT = VT.getVectorElementType();
4632 unsigned NumElems = Elts.size();
4634 LoadSDNode *LDBase = NULL;
4635 unsigned LastLoadedElt = -1U;
4637 // For each element in the initializer, see if we've found a load or an undef.
4638 // If we don't find an initial load element, or later load elements are
4639 // non-consecutive, bail out.
4640 for (unsigned i = 0; i < NumElems; ++i) {
4641 SDValue Elt = Elts[i];
4643 if (!Elt.getNode() ||
4644 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4647 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4649 LDBase = cast<LoadSDNode>(Elt.getNode());
4653 if (Elt.getOpcode() == ISD::UNDEF)
4656 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4657 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4662 // If we have found an entire vector of loads and undefs, then return a large
4663 // load of the entire vector width starting at the base pointer. If we found
4664 // consecutive loads for the low half, generate a vzext_load node.
4665 if (LastLoadedElt == NumElems - 1) {
4666 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4667 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4668 LDBase->getPointerInfo(),
4669 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4670 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4671 LDBase->getPointerInfo(),
4672 LDBase->isVolatile(), LDBase->isNonTemporal(),
4673 LDBase->getAlignment());
4674 } else if (NumElems == 4 && LastLoadedElt == 1 &&
4675 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4676 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4677 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4678 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4680 LDBase->getMemOperand());
4681 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4687 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4688 DebugLoc dl = Op.getDebugLoc();
4690 EVT VT = Op.getValueType();
4691 EVT ExtVT = VT.getVectorElementType();
4692 unsigned NumElems = Op.getNumOperands();
4694 // Vectors containing all zeros can be matched by pxor and xorps later
4695 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
4696 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
4697 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
4698 if (Op.getValueType() == MVT::v4i32 ||
4699 Op.getValueType() == MVT::v8i32)
4702 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4705 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
4706 // vectors or broken into v4i32 operations on 256-bit vectors.
4707 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
4708 if (Op.getValueType() == MVT::v4i32)
4711 return getOnesVector(Op.getValueType(), DAG, dl);
4714 unsigned EVTBits = ExtVT.getSizeInBits();
4716 unsigned NumZero = 0;
4717 unsigned NumNonZero = 0;
4718 unsigned NonZeros = 0;
4719 bool IsAllConstants = true;
4720 SmallSet<SDValue, 8> Values;
4721 for (unsigned i = 0; i < NumElems; ++i) {
4722 SDValue Elt = Op.getOperand(i);
4723 if (Elt.getOpcode() == ISD::UNDEF)
4726 if (Elt.getOpcode() != ISD::Constant &&
4727 Elt.getOpcode() != ISD::ConstantFP)
4728 IsAllConstants = false;
4729 if (X86::isZeroNode(Elt))
4732 NonZeros |= (1 << i);
4737 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4738 if (NumNonZero == 0)
4739 return DAG.getUNDEF(VT);
4741 // Special case for single non-zero, non-undef, element.
4742 if (NumNonZero == 1) {
4743 unsigned Idx = CountTrailingZeros_32(NonZeros);
4744 SDValue Item = Op.getOperand(Idx);
4746 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4747 // the value are obviously zero, truncate the value to i32 and do the
4748 // insertion that way. Only do this if the value is non-constant or if the
4749 // value is a constant being inserted into element 0. It is cheaper to do
4750 // a constant pool load than it is to do a movd + shuffle.
4751 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4752 (!IsAllConstants || Idx == 0)) {
4753 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4755 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
4756 EVT VecVT = MVT::v4i32;
4757 unsigned VecElts = 4;
4759 // Truncate the value (which may itself be a constant) to i32, and
4760 // convert it to a vector with movd (S2V+shuffle to zero extend).
4761 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4762 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4763 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4764 Subtarget->hasSSE2(), DAG);
4766 // Now we have our 32-bit value zero extended in the low element of
4767 // a vector. If Idx != 0, swizzle it into place.
4769 SmallVector<int, 4> Mask;
4770 Mask.push_back(Idx);
4771 for (unsigned i = 1; i != VecElts; ++i)
4773 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4774 DAG.getUNDEF(Item.getValueType()),
4777 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
4781 // If we have a constant or non-constant insertion into the low element of
4782 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4783 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4784 // depending on what the source datatype is.
4787 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4788 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4789 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4790 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4791 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4792 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4794 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4795 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4796 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
4797 EVT MiddleVT = MVT::v4i32;
4798 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4799 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4800 Subtarget->hasSSE2(), DAG);
4801 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
4805 // Is it a vector logical left shift?
4806 if (NumElems == 2 && Idx == 1 &&
4807 X86::isZeroNode(Op.getOperand(0)) &&
4808 !X86::isZeroNode(Op.getOperand(1))) {
4809 unsigned NumBits = VT.getSizeInBits();
4810 return getVShift(true, VT,
4811 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4812 VT, Op.getOperand(1)),
4813 NumBits/2, DAG, *this, dl);
4816 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4819 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4820 // is a non-constant being inserted into an element other than the low one,
4821 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4822 // movd/movss) to move this into the low element, then shuffle it into
4824 if (EVTBits == 32) {
4825 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4827 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4828 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4829 Subtarget->hasSSE2(), DAG);
4830 SmallVector<int, 8> MaskVec;
4831 for (unsigned i = 0; i < NumElems; i++)
4832 MaskVec.push_back(i == Idx ? 0 : 1);
4833 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4837 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4838 if (Values.size() == 1) {
4839 if (EVTBits == 32) {
4840 // Instead of a shuffle like this:
4841 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4842 // Check if it's possible to issue this instead.
4843 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4844 unsigned Idx = CountTrailingZeros_32(NonZeros);
4845 SDValue Item = Op.getOperand(Idx);
4846 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4847 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4852 // A vector full of immediates; various special cases are already
4853 // handled, so this is best done with a single constant-pool load.
4857 // For AVX-length vectors, build the individual 128-bit pieces and use
4858 // shuffles to put them in place.
4859 if (VT.getSizeInBits() == 256 && !ISD::isBuildVectorAllZeros(Op.getNode())) {
4860 SmallVector<SDValue, 32> V;
4861 for (unsigned i = 0; i < NumElems; ++i)
4862 V.push_back(Op.getOperand(i));
4864 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
4866 // Build both the lower and upper subvector.
4867 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
4868 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
4871 // Recreate the wider vector with the lower and upper part.
4872 SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Lower,
4873 DAG.getConstant(0, MVT::i32), DAG, dl);
4874 return Insert128BitVector(Vec, Upper, DAG.getConstant(NumElems/2, MVT::i32),
4878 // Let legalizer expand 2-wide build_vectors.
4879 if (EVTBits == 64) {
4880 if (NumNonZero == 1) {
4881 // One half is zero or undef.
4882 unsigned Idx = CountTrailingZeros_32(NonZeros);
4883 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4884 Op.getOperand(Idx));
4885 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4886 Subtarget->hasSSE2(), DAG);
4891 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4892 if (EVTBits == 8 && NumElems == 16) {
4893 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4895 if (V.getNode()) return V;
4898 if (EVTBits == 16 && NumElems == 8) {
4899 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4901 if (V.getNode()) return V;
4904 // If element VT is == 32 bits, turn it into a number of shuffles.
4905 SmallVector<SDValue, 8> V;
4907 if (NumElems == 4 && NumZero > 0) {
4908 for (unsigned i = 0; i < 4; ++i) {
4909 bool isZero = !(NonZeros & (1 << i));
4911 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4913 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4916 for (unsigned i = 0; i < 2; ++i) {
4917 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
4920 V[i] = V[i*2]; // Must be a zero vector.
4923 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
4926 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
4929 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
4934 SmallVector<int, 8> MaskVec;
4935 bool Reverse = (NonZeros & 0x3) == 2;
4936 for (unsigned i = 0; i < 2; ++i)
4937 MaskVec.push_back(Reverse ? 1-i : i);
4938 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
4939 for (unsigned i = 0; i < 2; ++i)
4940 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
4941 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
4944 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
4945 // Check for a build vector of consecutive loads.
4946 for (unsigned i = 0; i < NumElems; ++i)
4947 V[i] = Op.getOperand(i);
4949 // Check for elements which are consecutive loads.
4950 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
4954 // For SSE 4.1, use insertps to put the high elements into the low element.
4955 if (getSubtarget()->hasSSE41()) {
4957 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
4958 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
4960 Result = DAG.getUNDEF(VT);
4962 for (unsigned i = 1; i < NumElems; ++i) {
4963 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
4964 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
4965 Op.getOperand(i), DAG.getIntPtrConstant(i));
4970 // Otherwise, expand into a number of unpckl*, start by extending each of
4971 // our (non-undef) elements to the full vector width with the element in the
4972 // bottom slot of the vector (which generates no code for SSE).
4973 for (unsigned i = 0; i < NumElems; ++i) {
4974 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
4975 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
4977 V[i] = DAG.getUNDEF(VT);
4980 // Next, we iteratively mix elements, e.g. for v4f32:
4981 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
4982 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
4983 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
4984 unsigned EltStride = NumElems >> 1;
4985 while (EltStride != 0) {
4986 for (unsigned i = 0; i < EltStride; ++i) {
4987 // If V[i+EltStride] is undef and this is the first round of mixing,
4988 // then it is safe to just drop this shuffle: V[i] is already in the
4989 // right place, the one element (since it's the first round) being
4990 // inserted as undef can be dropped. This isn't safe for successive
4991 // rounds because they will permute elements within both vectors.
4992 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
4993 EltStride == NumElems/2)
4996 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
5005 // LowerMMXCONCAT_VECTORS - We support concatenate two MMX registers and place
5006 // them in a MMX register. This is better than doing a stack convert.
5007 static SDValue LowerMMXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5008 DebugLoc dl = Op.getDebugLoc();
5009 EVT ResVT = Op.getValueType();
5011 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
5012 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
5014 SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
5015 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
5016 InVec = Op.getOperand(1);
5017 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5018 unsigned NumElts = ResVT.getVectorNumElements();
5019 VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
5020 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
5021 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
5023 InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
5024 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
5025 Mask[0] = 0; Mask[1] = 2;
5026 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
5028 return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
5031 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
5032 // to create 256-bit vectors from two other 128-bit ones.
5033 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5034 DebugLoc dl = Op.getDebugLoc();
5035 EVT ResVT = Op.getValueType();
5037 assert(ResVT.getSizeInBits() == 256 && "Value type must be 256-bit wide");
5039 SDValue V1 = Op.getOperand(0);
5040 SDValue V2 = Op.getOperand(1);
5041 unsigned NumElems = ResVT.getVectorNumElements();
5043 SDValue V = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, ResVT), V1,
5044 DAG.getConstant(0, MVT::i32), DAG, dl);
5045 return Insert128BitVector(V, V2, DAG.getConstant(NumElems/2, MVT::i32),
5050 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
5051 DebugLoc dl = Op.getDebugLoc();
5052 EVT ResVT = Op.getValueType();
5054 assert(Op.getNumOperands() == 2);
5055 assert((ResVT.getSizeInBits() == 128 || ResVT.getSizeInBits() == 256) &&
5056 "Unsupported CONCAT_VECTORS for value type");
5058 // We support concatenate two MMX registers and place them in a MMX register.
5059 // This is better than doing a stack convert.
5060 if (ResVT.is128BitVector())
5061 return LowerMMXCONCAT_VECTORS(Op, DAG);
5063 // 256-bit AVX can use the vinsertf128 instruction to create 256-bit vectors
5064 // from two other 128-bit ones.
5065 return LowerAVXCONCAT_VECTORS(Op, DAG);
5068 // v8i16 shuffles - Prefer shuffles in the following order:
5069 // 1. [all] pshuflw, pshufhw, optional move
5070 // 2. [ssse3] 1 x pshufb
5071 // 3. [ssse3] 2 x pshufb + 1 x por
5072 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
5074 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
5075 SelectionDAG &DAG) const {
5076 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5077 SDValue V1 = SVOp->getOperand(0);
5078 SDValue V2 = SVOp->getOperand(1);
5079 DebugLoc dl = SVOp->getDebugLoc();
5080 SmallVector<int, 8> MaskVals;
5082 // Determine if more than 1 of the words in each of the low and high quadwords
5083 // of the result come from the same quadword of one of the two inputs. Undef
5084 // mask values count as coming from any quadword, for better codegen.
5085 SmallVector<unsigned, 4> LoQuad(4);
5086 SmallVector<unsigned, 4> HiQuad(4);
5087 BitVector InputQuads(4);
5088 for (unsigned i = 0; i < 8; ++i) {
5089 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
5090 int EltIdx = SVOp->getMaskElt(i);
5091 MaskVals.push_back(EltIdx);
5100 InputQuads.set(EltIdx / 4);
5103 int BestLoQuad = -1;
5104 unsigned MaxQuad = 1;
5105 for (unsigned i = 0; i < 4; ++i) {
5106 if (LoQuad[i] > MaxQuad) {
5108 MaxQuad = LoQuad[i];
5112 int BestHiQuad = -1;
5114 for (unsigned i = 0; i < 4; ++i) {
5115 if (HiQuad[i] > MaxQuad) {
5117 MaxQuad = HiQuad[i];
5121 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
5122 // of the two input vectors, shuffle them into one input vector so only a
5123 // single pshufb instruction is necessary. If There are more than 2 input
5124 // quads, disable the next transformation since it does not help SSSE3.
5125 bool V1Used = InputQuads[0] || InputQuads[1];
5126 bool V2Used = InputQuads[2] || InputQuads[3];
5127 if (Subtarget->hasSSSE3()) {
5128 if (InputQuads.count() == 2 && V1Used && V2Used) {
5129 BestLoQuad = InputQuads.find_first();
5130 BestHiQuad = InputQuads.find_next(BestLoQuad);
5132 if (InputQuads.count() > 2) {
5138 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
5139 // the shuffle mask. If a quad is scored as -1, that means that it contains
5140 // words from all 4 input quadwords.
5142 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
5143 SmallVector<int, 8> MaskV;
5144 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
5145 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
5146 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
5147 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
5148 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
5149 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
5151 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
5152 // source words for the shuffle, to aid later transformations.
5153 bool AllWordsInNewV = true;
5154 bool InOrder[2] = { true, true };
5155 for (unsigned i = 0; i != 8; ++i) {
5156 int idx = MaskVals[i];
5158 InOrder[i/4] = false;
5159 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
5161 AllWordsInNewV = false;
5165 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
5166 if (AllWordsInNewV) {
5167 for (int i = 0; i != 8; ++i) {
5168 int idx = MaskVals[i];
5171 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
5172 if ((idx != i) && idx < 4)
5174 if ((idx != i) && idx > 3)
5183 // If we've eliminated the use of V2, and the new mask is a pshuflw or
5184 // pshufhw, that's as cheap as it gets. Return the new shuffle.
5185 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
5186 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
5187 unsigned TargetMask = 0;
5188 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
5189 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
5190 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
5191 X86::getShufflePSHUFLWImmediate(NewV.getNode());
5192 V1 = NewV.getOperand(0);
5193 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
5197 // If we have SSSE3, and all words of the result are from 1 input vector,
5198 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
5199 // is present, fall back to case 4.
5200 if (Subtarget->hasSSSE3()) {
5201 SmallVector<SDValue,16> pshufbMask;
5203 // If we have elements from both input vectors, set the high bit of the
5204 // shuffle mask element to zero out elements that come from V2 in the V1
5205 // mask, and elements that come from V1 in the V2 mask, so that the two
5206 // results can be OR'd together.
5207 bool TwoInputs = V1Used && V2Used;
5208 for (unsigned i = 0; i != 8; ++i) {
5209 int EltIdx = MaskVals[i] * 2;
5210 if (TwoInputs && (EltIdx >= 16)) {
5211 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5212 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5215 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5216 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
5218 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
5219 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5220 DAG.getNode(ISD::BUILD_VECTOR, dl,
5221 MVT::v16i8, &pshufbMask[0], 16));
5223 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5225 // Calculate the shuffle mask for the second input, shuffle it, and
5226 // OR it with the first shuffled input.
5228 for (unsigned i = 0; i != 8; ++i) {
5229 int EltIdx = MaskVals[i] * 2;
5231 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5232 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5235 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5236 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
5238 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
5239 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5240 DAG.getNode(ISD::BUILD_VECTOR, dl,
5241 MVT::v16i8, &pshufbMask[0], 16));
5242 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5243 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5246 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
5247 // and update MaskVals with new element order.
5248 BitVector InOrder(8);
5249 if (BestLoQuad >= 0) {
5250 SmallVector<int, 8> MaskV;
5251 for (int i = 0; i != 4; ++i) {
5252 int idx = MaskVals[i];
5254 MaskV.push_back(-1);
5256 } else if ((idx / 4) == BestLoQuad) {
5257 MaskV.push_back(idx & 3);
5260 MaskV.push_back(-1);
5263 for (unsigned i = 4; i != 8; ++i)
5265 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5268 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5269 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
5271 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
5275 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
5276 // and update MaskVals with the new element order.
5277 if (BestHiQuad >= 0) {
5278 SmallVector<int, 8> MaskV;
5279 for (unsigned i = 0; i != 4; ++i)
5281 for (unsigned i = 4; i != 8; ++i) {
5282 int idx = MaskVals[i];
5284 MaskV.push_back(-1);
5286 } else if ((idx / 4) == BestHiQuad) {
5287 MaskV.push_back((idx & 3) + 4);
5290 MaskV.push_back(-1);
5293 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5296 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5297 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
5299 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
5303 // In case BestHi & BestLo were both -1, which means each quadword has a word
5304 // from each of the four input quadwords, calculate the InOrder bitvector now
5305 // before falling through to the insert/extract cleanup.
5306 if (BestLoQuad == -1 && BestHiQuad == -1) {
5308 for (int i = 0; i != 8; ++i)
5309 if (MaskVals[i] < 0 || MaskVals[i] == i)
5313 // The other elements are put in the right place using pextrw and pinsrw.
5314 for (unsigned i = 0; i != 8; ++i) {
5317 int EltIdx = MaskVals[i];
5320 SDValue ExtOp = (EltIdx < 8)
5321 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
5322 DAG.getIntPtrConstant(EltIdx))
5323 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
5324 DAG.getIntPtrConstant(EltIdx - 8));
5325 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
5326 DAG.getIntPtrConstant(i));
5331 // v16i8 shuffles - Prefer shuffles in the following order:
5332 // 1. [ssse3] 1 x pshufb
5333 // 2. [ssse3] 2 x pshufb + 1 x por
5334 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
5336 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
5338 const X86TargetLowering &TLI) {
5339 SDValue V1 = SVOp->getOperand(0);
5340 SDValue V2 = SVOp->getOperand(1);
5341 DebugLoc dl = SVOp->getDebugLoc();
5342 SmallVector<int, 16> MaskVals;
5343 SVOp->getMask(MaskVals);
5345 // If we have SSSE3, case 1 is generated when all result bytes come from
5346 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
5347 // present, fall back to case 3.
5348 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
5351 for (unsigned i = 0; i < 16; ++i) {
5352 int EltIdx = MaskVals[i];
5361 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
5362 if (TLI.getSubtarget()->hasSSSE3()) {
5363 SmallVector<SDValue,16> pshufbMask;
5365 // If all result elements are from one input vector, then only translate
5366 // undef mask values to 0x80 (zero out result) in the pshufb mask.
5368 // Otherwise, we have elements from both input vectors, and must zero out
5369 // elements that come from V2 in the first mask, and V1 in the second mask
5370 // so that we can OR them together.
5371 bool TwoInputs = !(V1Only || V2Only);
5372 for (unsigned i = 0; i != 16; ++i) {
5373 int EltIdx = MaskVals[i];
5374 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
5375 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5378 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5380 // If all the elements are from V2, assign it to V1 and return after
5381 // building the first pshufb.
5384 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5385 DAG.getNode(ISD::BUILD_VECTOR, dl,
5386 MVT::v16i8, &pshufbMask[0], 16));
5390 // Calculate the shuffle mask for the second input, shuffle it, and
5391 // OR it with the first shuffled input.
5393 for (unsigned i = 0; i != 16; ++i) {
5394 int EltIdx = MaskVals[i];
5396 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5399 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5401 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5402 DAG.getNode(ISD::BUILD_VECTOR, dl,
5403 MVT::v16i8, &pshufbMask[0], 16));
5404 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5407 // No SSSE3 - Calculate in place words and then fix all out of place words
5408 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
5409 // the 16 different words that comprise the two doublequadword input vectors.
5410 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5411 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
5412 SDValue NewV = V2Only ? V2 : V1;
5413 for (int i = 0; i != 8; ++i) {
5414 int Elt0 = MaskVals[i*2];
5415 int Elt1 = MaskVals[i*2+1];
5417 // This word of the result is all undef, skip it.
5418 if (Elt0 < 0 && Elt1 < 0)
5421 // This word of the result is already in the correct place, skip it.
5422 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
5424 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
5427 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
5428 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
5431 // If Elt0 and Elt1 are defined, are consecutive, and can be load
5432 // using a single extract together, load it and store it.
5433 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
5434 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5435 DAG.getIntPtrConstant(Elt1 / 2));
5436 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5437 DAG.getIntPtrConstant(i));
5441 // If Elt1 is defined, extract it from the appropriate source. If the
5442 // source byte is not also odd, shift the extracted word left 8 bits
5443 // otherwise clear the bottom 8 bits if we need to do an or.
5445 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5446 DAG.getIntPtrConstant(Elt1 / 2));
5447 if ((Elt1 & 1) == 0)
5448 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
5450 TLI.getShiftAmountTy(InsElt.getValueType())));
5452 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
5453 DAG.getConstant(0xFF00, MVT::i16));
5455 // If Elt0 is defined, extract it from the appropriate source. If the
5456 // source byte is not also even, shift the extracted word right 8 bits. If
5457 // Elt1 was also defined, OR the extracted values together before
5458 // inserting them in the result.
5460 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
5461 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
5462 if ((Elt0 & 1) != 0)
5463 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
5465 TLI.getShiftAmountTy(InsElt0.getValueType())));
5467 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
5468 DAG.getConstant(0x00FF, MVT::i16));
5469 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
5472 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5473 DAG.getIntPtrConstant(i));
5475 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
5478 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
5479 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
5480 /// done when every pair / quad of shuffle mask elements point to elements in
5481 /// the right sequence. e.g.
5482 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
5484 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
5485 SelectionDAG &DAG, DebugLoc dl) {
5486 EVT VT = SVOp->getValueType(0);
5487 SDValue V1 = SVOp->getOperand(0);
5488 SDValue V2 = SVOp->getOperand(1);
5489 unsigned NumElems = VT.getVectorNumElements();
5490 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
5492 switch (VT.getSimpleVT().SimpleTy) {
5493 default: assert(false && "Unexpected!");
5494 case MVT::v4f32: NewVT = MVT::v2f64; break;
5495 case MVT::v4i32: NewVT = MVT::v2i64; break;
5496 case MVT::v8i16: NewVT = MVT::v4i32; break;
5497 case MVT::v16i8: NewVT = MVT::v4i32; break;
5500 int Scale = NumElems / NewWidth;
5501 SmallVector<int, 8> MaskVec;
5502 for (unsigned i = 0; i < NumElems; i += Scale) {
5504 for (int j = 0; j < Scale; ++j) {
5505 int EltIdx = SVOp->getMaskElt(i+j);
5509 StartIdx = EltIdx - (EltIdx % Scale);
5510 if (EltIdx != StartIdx + j)
5514 MaskVec.push_back(-1);
5516 MaskVec.push_back(StartIdx / Scale);
5519 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
5520 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
5521 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
5524 /// getVZextMovL - Return a zero-extending vector move low node.
5526 static SDValue getVZextMovL(EVT VT, EVT OpVT,
5527 SDValue SrcOp, SelectionDAG &DAG,
5528 const X86Subtarget *Subtarget, DebugLoc dl) {
5529 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
5530 LoadSDNode *LD = NULL;
5531 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
5532 LD = dyn_cast<LoadSDNode>(SrcOp);
5534 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
5536 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
5537 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
5538 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
5539 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
5540 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
5542 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
5543 return DAG.getNode(ISD::BITCAST, dl, VT,
5544 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5545 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5553 return DAG.getNode(ISD::BITCAST, dl, VT,
5554 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5555 DAG.getNode(ISD::BITCAST, dl,
5559 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
5560 /// which could not be matched by any known target speficic shuffle
5562 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5566 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
5567 /// 4 elements, and match them with several different shuffle types.
5569 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5570 SDValue V1 = SVOp->getOperand(0);
5571 SDValue V2 = SVOp->getOperand(1);
5572 DebugLoc dl = SVOp->getDebugLoc();
5573 EVT VT = SVOp->getValueType(0);
5575 assert(VT.getSizeInBits() == 128 && "Unsupported vector size");
5577 SmallVector<std::pair<int, int>, 8> Locs;
5579 SmallVector<int, 8> Mask1(4U, -1);
5580 SmallVector<int, 8> PermMask;
5581 SVOp->getMask(PermMask);
5585 for (unsigned i = 0; i != 4; ++i) {
5586 int Idx = PermMask[i];
5588 Locs[i] = std::make_pair(-1, -1);
5590 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5592 Locs[i] = std::make_pair(0, NumLo);
5596 Locs[i] = std::make_pair(1, NumHi);
5598 Mask1[2+NumHi] = Idx;
5604 if (NumLo <= 2 && NumHi <= 2) {
5605 // If no more than two elements come from either vector. This can be
5606 // implemented with two shuffles. First shuffle gather the elements.
5607 // The second shuffle, which takes the first shuffle as both of its
5608 // vector operands, put the elements into the right order.
5609 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5611 SmallVector<int, 8> Mask2(4U, -1);
5613 for (unsigned i = 0; i != 4; ++i) {
5614 if (Locs[i].first == -1)
5617 unsigned Idx = (i < 2) ? 0 : 4;
5618 Idx += Locs[i].first * 2 + Locs[i].second;
5623 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
5624 } else if (NumLo == 3 || NumHi == 3) {
5625 // Otherwise, we must have three elements from one vector, call it X, and
5626 // one element from the other, call it Y. First, use a shufps to build an
5627 // intermediate vector with the one element from Y and the element from X
5628 // that will be in the same half in the final destination (the indexes don't
5629 // matter). Then, use a shufps to build the final vector, taking the half
5630 // containing the element from Y from the intermediate, and the other half
5633 // Normalize it so the 3 elements come from V1.
5634 CommuteVectorShuffleMask(PermMask, VT);
5638 // Find the element from V2.
5640 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5641 int Val = PermMask[HiIndex];
5648 Mask1[0] = PermMask[HiIndex];
5650 Mask1[2] = PermMask[HiIndex^1];
5652 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5655 Mask1[0] = PermMask[0];
5656 Mask1[1] = PermMask[1];
5657 Mask1[2] = HiIndex & 1 ? 6 : 4;
5658 Mask1[3] = HiIndex & 1 ? 4 : 6;
5659 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5661 Mask1[0] = HiIndex & 1 ? 2 : 0;
5662 Mask1[1] = HiIndex & 1 ? 0 : 2;
5663 Mask1[2] = PermMask[2];
5664 Mask1[3] = PermMask[3];
5669 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5673 // Break it into (shuffle shuffle_hi, shuffle_lo).
5676 SmallVector<int,8> LoMask(4U, -1);
5677 SmallVector<int,8> HiMask(4U, -1);
5679 SmallVector<int,8> *MaskPtr = &LoMask;
5680 unsigned MaskIdx = 0;
5683 for (unsigned i = 0; i != 4; ++i) {
5690 int Idx = PermMask[i];
5692 Locs[i] = std::make_pair(-1, -1);
5693 } else if (Idx < 4) {
5694 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5695 (*MaskPtr)[LoIdx] = Idx;
5698 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5699 (*MaskPtr)[HiIdx] = Idx;
5704 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5705 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5706 SmallVector<int, 8> MaskOps;
5707 for (unsigned i = 0; i != 4; ++i) {
5708 if (Locs[i].first == -1) {
5709 MaskOps.push_back(-1);
5711 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5712 MaskOps.push_back(Idx);
5715 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5718 static bool MayFoldVectorLoad(SDValue V) {
5719 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5720 V = V.getOperand(0);
5721 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5722 V = V.getOperand(0);
5728 // FIXME: the version above should always be used. Since there's
5729 // a bug where several vector shuffles can't be folded because the
5730 // DAG is not updated during lowering and a node claims to have two
5731 // uses while it only has one, use this version, and let isel match
5732 // another instruction if the load really happens to have more than
5733 // one use. Remove this version after this bug get fixed.
5734 // rdar://8434668, PR8156
5735 static bool RelaxedMayFoldVectorLoad(SDValue V) {
5736 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5737 V = V.getOperand(0);
5738 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5739 V = V.getOperand(0);
5740 if (ISD::isNormalLoad(V.getNode()))
5745 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
5746 /// a vector extract, and if both can be later optimized into a single load.
5747 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
5748 /// here because otherwise a target specific shuffle node is going to be
5749 /// emitted for this shuffle, and the optimization not done.
5750 /// FIXME: This is probably not the best approach, but fix the problem
5751 /// until the right path is decided.
5753 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
5754 const TargetLowering &TLI) {
5755 EVT VT = V.getValueType();
5756 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
5758 // Be sure that the vector shuffle is present in a pattern like this:
5759 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
5763 SDNode *N = *V.getNode()->use_begin();
5764 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5767 SDValue EltNo = N->getOperand(1);
5768 if (!isa<ConstantSDNode>(EltNo))
5771 // If the bit convert changed the number of elements, it is unsafe
5772 // to examine the mask.
5773 bool HasShuffleIntoBitcast = false;
5774 if (V.getOpcode() == ISD::BITCAST) {
5775 EVT SrcVT = V.getOperand(0).getValueType();
5776 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
5778 V = V.getOperand(0);
5779 HasShuffleIntoBitcast = true;
5782 // Select the input vector, guarding against out of range extract vector.
5783 unsigned NumElems = VT.getVectorNumElements();
5784 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5785 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
5786 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
5788 // Skip one more bit_convert if necessary
5789 if (V.getOpcode() == ISD::BITCAST)
5790 V = V.getOperand(0);
5792 if (ISD::isNormalLoad(V.getNode())) {
5793 // Is the original load suitable?
5794 LoadSDNode *LN0 = cast<LoadSDNode>(V);
5796 // FIXME: avoid the multi-use bug that is preventing lots of
5797 // of foldings to be detected, this is still wrong of course, but
5798 // give the temporary desired behavior, and if it happens that
5799 // the load has real more uses, during isel it will not fold, and
5800 // will generate poor code.
5801 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
5804 if (!HasShuffleIntoBitcast)
5807 // If there's a bitcast before the shuffle, check if the load type and
5808 // alignment is valid.
5809 unsigned Align = LN0->getAlignment();
5811 TLI.getTargetData()->getABITypeAlignment(
5812 VT.getTypeForEVT(*DAG.getContext()));
5814 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
5822 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
5823 EVT VT = Op.getValueType();
5825 // Canonizalize to v2f64.
5826 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
5827 return DAG.getNode(ISD::BITCAST, dl, VT,
5828 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
5833 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
5835 SDValue V1 = Op.getOperand(0);
5836 SDValue V2 = Op.getOperand(1);
5837 EVT VT = Op.getValueType();
5839 assert(VT != MVT::v2i64 && "unsupported shuffle type");
5841 if (HasSSE2 && VT == MVT::v2f64)
5842 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
5845 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
5849 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
5850 SDValue V1 = Op.getOperand(0);
5851 SDValue V2 = Op.getOperand(1);
5852 EVT VT = Op.getValueType();
5854 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
5855 "unsupported shuffle type");
5857 if (V2.getOpcode() == ISD::UNDEF)
5861 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
5865 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
5866 SDValue V1 = Op.getOperand(0);
5867 SDValue V2 = Op.getOperand(1);
5868 EVT VT = Op.getValueType();
5869 unsigned NumElems = VT.getVectorNumElements();
5871 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
5872 // operand of these instructions is only memory, so check if there's a
5873 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
5875 bool CanFoldLoad = false;
5877 // Trivial case, when V2 comes from a load.
5878 if (MayFoldVectorLoad(V2))
5881 // When V1 is a load, it can be folded later into a store in isel, example:
5882 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
5884 // (MOVLPSmr addr:$src1, VR128:$src2)
5885 // So, recognize this potential and also use MOVLPS or MOVLPD
5886 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
5889 // Both of them can't be memory operations though.
5890 if (MayFoldVectorLoad(V1) && MayFoldVectorLoad(V2))
5891 CanFoldLoad = false;
5894 if (HasSSE2 && NumElems == 2)
5895 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
5898 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
5901 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5902 // movl and movlp will both match v2i64, but v2i64 is never matched by
5903 // movl earlier because we make it strict to avoid messing with the movlp load
5904 // folding logic (see the code above getMOVLP call). Match it here then,
5905 // this is horrible, but will stay like this until we move all shuffle
5906 // matching to x86 specific nodes. Note that for the 1st condition all
5907 // types are matched with movsd.
5908 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
5909 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
5911 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
5914 assert(VT != MVT::v4i32 && "unsupported shuffle type");
5916 // Invert the operand order and use SHUFPS to match it.
5917 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
5918 X86::getShuffleSHUFImmediate(SVOp), DAG);
5921 static inline unsigned getUNPCKLOpcode(EVT VT) {
5922 switch(VT.getSimpleVT().SimpleTy) {
5923 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
5924 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
5925 case MVT::v4f32: return X86ISD::UNPCKLPS;
5926 case MVT::v2f64: return X86ISD::UNPCKLPD;
5927 case MVT::v8f32: return X86ISD::VUNPCKLPSY;
5928 case MVT::v4f64: return X86ISD::VUNPCKLPDY;
5929 case MVT::v16i8: return X86ISD::PUNPCKLBW;
5930 case MVT::v8i16: return X86ISD::PUNPCKLWD;
5932 llvm_unreachable("Unknown type for unpckl");
5937 static inline unsigned getUNPCKHOpcode(EVT VT) {
5938 switch(VT.getSimpleVT().SimpleTy) {
5939 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
5940 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
5941 case MVT::v4f32: return X86ISD::UNPCKHPS;
5942 case MVT::v2f64: return X86ISD::UNPCKHPD;
5943 case MVT::v8f32: return X86ISD::VUNPCKHPSY;
5944 case MVT::v4f64: return X86ISD::VUNPCKHPDY;
5945 case MVT::v16i8: return X86ISD::PUNPCKHBW;
5946 case MVT::v8i16: return X86ISD::PUNPCKHWD;
5948 llvm_unreachable("Unknown type for unpckh");
5953 static inline unsigned getVPERMILOpcode(EVT VT) {
5954 switch(VT.getSimpleVT().SimpleTy) {
5956 case MVT::v4f32: return X86ISD::VPERMILPS;
5958 case MVT::v2f64: return X86ISD::VPERMILPD;
5960 case MVT::v8f32: return X86ISD::VPERMILPSY;
5962 case MVT::v4f64: return X86ISD::VPERMILPDY;
5964 llvm_unreachable("Unknown type for vpermil");
5970 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
5971 const TargetLowering &TLI,
5972 const X86Subtarget *Subtarget) {
5973 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5974 EVT VT = Op.getValueType();
5975 DebugLoc dl = Op.getDebugLoc();
5976 SDValue V1 = Op.getOperand(0);
5977 SDValue V2 = Op.getOperand(1);
5979 if (isZeroShuffle(SVOp))
5980 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5982 // Handle splat operations
5983 if (SVOp->isSplat()) {
5984 unsigned NumElem = VT.getVectorNumElements();
5985 // Special case, this is the only place now where it's allowed to return
5986 // a vector_shuffle operation without using a target specific node, because
5987 // *hopefully* it will be optimized away by the dag combiner. FIXME: should
5988 // this be moved to DAGCombine instead?
5989 if (NumElem <= 4 && CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
5992 // Since there's no native support for scalar_to_vector for 256-bit AVX, a
5993 // 128-bit scalar_to_vector + INSERT_SUBVECTOR is generated. Recognize this
5994 // idiom and do the shuffle before the insertion, this yields less
5995 // instructions in the end.
5996 if (VT.is256BitVector() &&
5997 V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
5998 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
5999 V1.getOperand(1).getOpcode() == ISD::SCALAR_TO_VECTOR)
6000 return PromoteVectorToScalarSplat(SVOp, DAG);
6002 // Handle splats by matching through known shuffle masks
6003 if ((VT.is128BitVector() && NumElem <= 4) ||
6004 (VT.is256BitVector() && NumElem <= 8))
6007 // All i16 and i8 vector types can't be used directly by a generic shuffle
6008 // instruction because the target has no such instruction. Generate shuffles
6009 // which repeat i16 and i8 several times until they fit in i32, and then can
6010 // be manipulated by target suported shuffles. After the insertion of the
6011 // necessary shuffles, the result is bitcasted back to v4f32 or v8f32.
6012 return PromoteSplat(SVOp, DAG);
6015 // If the shuffle can be profitably rewritten as a narrower shuffle, then
6017 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
6018 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6019 if (NewOp.getNode())
6020 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
6021 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
6022 // FIXME: Figure out a cleaner way to do this.
6023 // Try to make use of movq to zero out the top part.
6024 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
6025 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6026 if (NewOp.getNode()) {
6027 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
6028 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
6029 DAG, Subtarget, dl);
6031 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
6032 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6033 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
6034 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
6035 DAG, Subtarget, dl);
6042 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
6043 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6044 SDValue V1 = Op.getOperand(0);
6045 SDValue V2 = Op.getOperand(1);
6046 EVT VT = Op.getValueType();
6047 DebugLoc dl = Op.getDebugLoc();
6048 unsigned NumElems = VT.getVectorNumElements();
6049 bool isMMX = VT.getSizeInBits() == 64;
6050 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
6051 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
6052 bool V1IsSplat = false;
6053 bool V2IsSplat = false;
6054 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
6055 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
6056 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
6057 MachineFunction &MF = DAG.getMachineFunction();
6058 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
6060 // Shuffle operations on MMX not supported.
6064 // Vector shuffle lowering takes 3 steps:
6066 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
6067 // narrowing and commutation of operands should be handled.
6068 // 2) Matching of shuffles with known shuffle masks to x86 target specific
6070 // 3) Rewriting of unmatched masks into new generic shuffle operations,
6071 // so the shuffle can be broken into other shuffles and the legalizer can
6072 // try the lowering again.
6074 // The general ideia is that no vector_shuffle operation should be left to
6075 // be matched during isel, all of them must be converted to a target specific
6078 // Normalize the input vectors. Here splats, zeroed vectors, profitable
6079 // narrowing and commutation of operands should be handled. The actual code
6080 // doesn't include all of those, work in progress...
6081 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
6082 if (NewOp.getNode())
6085 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
6086 // unpckh_undef). Only use pshufd if speed is more important than size.
6087 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
6088 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
6089 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
6090 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6092 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
6093 RelaxedMayFoldVectorLoad(V1))
6094 return getMOVDDup(Op, dl, V1, DAG);
6096 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
6097 return getMOVHighToLow(Op, dl, DAG);
6099 // Use to match splats
6100 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
6101 (VT == MVT::v2f64 || VT == MVT::v2i64))
6102 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6104 if (X86::isPSHUFDMask(SVOp)) {
6105 // The actual implementation will match the mask in the if above and then
6106 // during isel it can match several different instructions, not only pshufd
6107 // as its name says, sad but true, emulate the behavior for now...
6108 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
6109 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
6111 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
6113 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
6114 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
6116 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
6117 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
6120 if (VT == MVT::v4f32)
6121 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
6125 // Check if this can be converted into a logical shift.
6126 bool isLeft = false;
6129 bool isShift = getSubtarget()->hasSSE2() &&
6130 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
6131 if (isShift && ShVal.hasOneUse()) {
6132 // If the shifted value has multiple uses, it may be cheaper to use
6133 // v_set0 + movlhps or movhlps, etc.
6134 EVT EltVT = VT.getVectorElementType();
6135 ShAmt *= EltVT.getSizeInBits();
6136 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6139 if (X86::isMOVLMask(SVOp)) {
6142 if (ISD::isBuildVectorAllZeros(V1.getNode()))
6143 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
6144 if (!X86::isMOVLPMask(SVOp)) {
6145 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
6146 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
6148 if (VT == MVT::v4i32 || VT == MVT::v4f32)
6149 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
6153 // FIXME: fold these into legal mask.
6154 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
6155 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
6157 if (X86::isMOVHLPSMask(SVOp))
6158 return getMOVHighToLow(Op, dl, DAG);
6160 if (X86::isMOVSHDUPMask(SVOp, Subtarget))
6161 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
6163 if (X86::isMOVSLDUPMask(SVOp, Subtarget))
6164 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
6166 if (X86::isMOVLPMask(SVOp))
6167 return getMOVLP(Op, dl, DAG, HasSSE2);
6169 if (ShouldXformToMOVHLPS(SVOp) ||
6170 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
6171 return CommuteVectorShuffle(SVOp, DAG);
6174 // No better options. Use a vshl / vsrl.
6175 EVT EltVT = VT.getVectorElementType();
6176 ShAmt *= EltVT.getSizeInBits();
6177 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6180 bool Commuted = false;
6181 // FIXME: This should also accept a bitcast of a splat? Be careful, not
6182 // 1,1,1,1 -> v8i16 though.
6183 V1IsSplat = isSplatVector(V1.getNode());
6184 V2IsSplat = isSplatVector(V2.getNode());
6186 // Canonicalize the splat or undef, if present, to be on the RHS.
6187 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
6188 Op = CommuteVectorShuffle(SVOp, DAG);
6189 SVOp = cast<ShuffleVectorSDNode>(Op);
6190 V1 = SVOp->getOperand(0);
6191 V2 = SVOp->getOperand(1);
6192 std::swap(V1IsSplat, V2IsSplat);
6193 std::swap(V1IsUndef, V2IsUndef);
6197 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
6198 // Shuffling low element of v1 into undef, just return v1.
6201 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
6202 // the instruction selector will not match, so get a canonical MOVL with
6203 // swapped operands to undo the commute.
6204 return getMOVL(DAG, dl, VT, V2, V1);
6207 if (X86::isUNPCKLMask(SVOp))
6208 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
6210 if (X86::isUNPCKHMask(SVOp))
6211 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
6214 // Normalize mask so all entries that point to V2 points to its first
6215 // element then try to match unpck{h|l} again. If match, return a
6216 // new vector_shuffle with the corrected mask.
6217 SDValue NewMask = NormalizeMask(SVOp, DAG);
6218 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
6219 if (NSVOp != SVOp) {
6220 if (X86::isUNPCKLMask(NSVOp, true)) {
6222 } else if (X86::isUNPCKHMask(NSVOp, true)) {
6229 // Commute is back and try unpck* again.
6230 // FIXME: this seems wrong.
6231 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
6232 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
6234 if (X86::isUNPCKLMask(NewSVOp))
6235 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
6237 if (X86::isUNPCKHMask(NewSVOp))
6238 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
6241 // Normalize the node to match x86 shuffle ops if needed
6242 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
6243 return CommuteVectorShuffle(SVOp, DAG);
6245 // The checks below are all present in isShuffleMaskLegal, but they are
6246 // inlined here right now to enable us to directly emit target specific
6247 // nodes, and remove one by one until they don't return Op anymore.
6248 SmallVector<int, 16> M;
6251 if (isPALIGNRMask(M, VT, HasSSSE3))
6252 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
6253 X86::getShufflePALIGNRImmediate(SVOp),
6256 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
6257 SVOp->getSplatIndex() == 0 && V2IsUndef) {
6258 if (VT == MVT::v2f64)
6259 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
6260 if (VT == MVT::v2i64)
6261 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
6264 if (isPSHUFHWMask(M, VT))
6265 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
6266 X86::getShufflePSHUFHWImmediate(SVOp),
6269 if (isPSHUFLWMask(M, VT))
6270 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
6271 X86::getShufflePSHUFLWImmediate(SVOp),
6274 if (isSHUFPMask(M, VT)) {
6275 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
6276 if (VT == MVT::v4f32 || VT == MVT::v4i32)
6277 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
6279 if (VT == MVT::v2f64 || VT == MVT::v2i64)
6280 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
6284 if (X86::isUNPCKL_v_undef_Mask(SVOp))
6285 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
6286 if (X86::isUNPCKH_v_undef_Mask(SVOp))
6287 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6289 //===--------------------------------------------------------------------===//
6290 // Generate target specific nodes for 128 or 256-bit shuffles only
6291 // supported in the AVX instruction set.
6294 // Handle VPERMILPS* permutations
6295 if (isVPERMILPSMask(M, VT, Subtarget))
6296 return getTargetShuffleNode(getVPERMILOpcode(VT), dl, VT, V1,
6297 getShuffleVPERMILPSImmediate(SVOp), DAG);
6299 // Handle VPERMILPD* permutations
6300 if (isVPERMILPDMask(M, VT, Subtarget))
6301 return getTargetShuffleNode(getVPERMILOpcode(VT), dl, VT, V1,
6302 getShuffleVPERMILPDImmediate(SVOp), DAG);
6304 //===--------------------------------------------------------------------===//
6305 // Since no target specific shuffle was selected for this generic one,
6306 // lower it into other known shuffles. FIXME: this isn't true yet, but
6307 // this is the plan.
6310 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
6311 if (VT == MVT::v8i16) {
6312 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
6313 if (NewOp.getNode())
6317 if (VT == MVT::v16i8) {
6318 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
6319 if (NewOp.getNode())
6323 // Handle all 128-bit wide vectors with 4 elements, and match them with
6324 // several different shuffle types.
6325 if (NumElems == 4 && VT.getSizeInBits() == 128)
6326 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
6328 // Handle general 256-bit shuffles
6329 if (VT.is256BitVector())
6330 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
6336 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
6337 SelectionDAG &DAG) const {
6338 EVT VT = Op.getValueType();
6339 DebugLoc dl = Op.getDebugLoc();
6341 if (Op.getOperand(0).getValueType().getSizeInBits() != 128)
6344 if (VT.getSizeInBits() == 8) {
6345 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
6346 Op.getOperand(0), Op.getOperand(1));
6347 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6348 DAG.getValueType(VT));
6349 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6350 } else if (VT.getSizeInBits() == 16) {
6351 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6352 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
6354 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6355 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6356 DAG.getNode(ISD::BITCAST, dl,
6360 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
6361 Op.getOperand(0), Op.getOperand(1));
6362 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6363 DAG.getValueType(VT));
6364 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6365 } else if (VT == MVT::f32) {
6366 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
6367 // the result back to FR32 register. It's only worth matching if the
6368 // result has a single use which is a store or a bitcast to i32. And in
6369 // the case of a store, it's not worth it if the index is a constant 0,
6370 // because a MOVSSmr can be used instead, which is smaller and faster.
6371 if (!Op.hasOneUse())
6373 SDNode *User = *Op.getNode()->use_begin();
6374 if ((User->getOpcode() != ISD::STORE ||
6375 (isa<ConstantSDNode>(Op.getOperand(1)) &&
6376 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
6377 (User->getOpcode() != ISD::BITCAST ||
6378 User->getValueType(0) != MVT::i32))
6380 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6381 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
6384 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
6385 } else if (VT == MVT::i32) {
6386 // ExtractPS works with constant index.
6387 if (isa<ConstantSDNode>(Op.getOperand(1)))
6395 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6396 SelectionDAG &DAG) const {
6397 if (!isa<ConstantSDNode>(Op.getOperand(1)))
6400 SDValue Vec = Op.getOperand(0);
6401 EVT VecVT = Vec.getValueType();
6403 // If this is a 256-bit vector result, first extract the 128-bit vector and
6404 // then extract the element from the 128-bit vector.
6405 if (VecVT.getSizeInBits() == 256) {
6406 DebugLoc dl = Op.getNode()->getDebugLoc();
6407 unsigned NumElems = VecVT.getVectorNumElements();
6408 SDValue Idx = Op.getOperand(1);
6409 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
6411 // Get the 128-bit vector.
6412 bool Upper = IdxVal >= NumElems/2;
6413 Vec = Extract128BitVector(Vec,
6414 DAG.getConstant(Upper ? NumElems/2 : 0, MVT::i32), DAG, dl);
6416 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
6417 Upper ? DAG.getConstant(IdxVal-NumElems/2, MVT::i32) : Idx);
6420 assert(Vec.getValueSizeInBits() <= 128 && "Unexpected vector length");
6422 if (Subtarget->hasSSE41() || Subtarget->hasAVX()) {
6423 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
6428 EVT VT = Op.getValueType();
6429 DebugLoc dl = Op.getDebugLoc();
6430 // TODO: handle v16i8.
6431 if (VT.getSizeInBits() == 16) {
6432 SDValue Vec = Op.getOperand(0);
6433 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6435 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6436 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6437 DAG.getNode(ISD::BITCAST, dl,
6440 // Transform it so it match pextrw which produces a 32-bit result.
6441 EVT EltVT = MVT::i32;
6442 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
6443 Op.getOperand(0), Op.getOperand(1));
6444 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
6445 DAG.getValueType(VT));
6446 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6447 } else if (VT.getSizeInBits() == 32) {
6448 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6452 // SHUFPS the element to the lowest double word, then movss.
6453 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
6454 EVT VVT = Op.getOperand(0).getValueType();
6455 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6456 DAG.getUNDEF(VVT), Mask);
6457 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6458 DAG.getIntPtrConstant(0));
6459 } else if (VT.getSizeInBits() == 64) {
6460 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
6461 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
6462 // to match extract_elt for f64.
6463 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6467 // UNPCKHPD the element to the lowest double word, then movsd.
6468 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
6469 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
6470 int Mask[2] = { 1, -1 };
6471 EVT VVT = Op.getOperand(0).getValueType();
6472 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6473 DAG.getUNDEF(VVT), Mask);
6474 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6475 DAG.getIntPtrConstant(0));
6482 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
6483 SelectionDAG &DAG) const {
6484 EVT VT = Op.getValueType();
6485 EVT EltVT = VT.getVectorElementType();
6486 DebugLoc dl = Op.getDebugLoc();
6488 SDValue N0 = Op.getOperand(0);
6489 SDValue N1 = Op.getOperand(1);
6490 SDValue N2 = Op.getOperand(2);
6492 if (VT.getSizeInBits() == 256)
6495 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
6496 isa<ConstantSDNode>(N2)) {
6498 if (VT == MVT::v8i16)
6499 Opc = X86ISD::PINSRW;
6500 else if (VT == MVT::v16i8)
6501 Opc = X86ISD::PINSRB;
6503 Opc = X86ISD::PINSRB;
6505 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
6507 if (N1.getValueType() != MVT::i32)
6508 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6509 if (N2.getValueType() != MVT::i32)
6510 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6511 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
6512 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
6513 // Bits [7:6] of the constant are the source select. This will always be
6514 // zero here. The DAG Combiner may combine an extract_elt index into these
6515 // bits. For example (insert (extract, 3), 2) could be matched by putting
6516 // the '3' into bits [7:6] of X86ISD::INSERTPS.
6517 // Bits [5:4] of the constant are the destination select. This is the
6518 // value of the incoming immediate.
6519 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
6520 // combine either bitwise AND or insert of float 0.0 to set these bits.
6521 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
6522 // Create this as a scalar to vector..
6523 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
6524 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
6525 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
6526 // PINSR* works with constant index.
6533 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
6534 EVT VT = Op.getValueType();
6535 EVT EltVT = VT.getVectorElementType();
6537 DebugLoc dl = Op.getDebugLoc();
6538 SDValue N0 = Op.getOperand(0);
6539 SDValue N1 = Op.getOperand(1);
6540 SDValue N2 = Op.getOperand(2);
6542 // If this is a 256-bit vector result, first extract the 128-bit vector,
6543 // insert the element into the extracted half and then place it back.
6544 if (VT.getSizeInBits() == 256) {
6545 if (!isa<ConstantSDNode>(N2))
6548 // Get the desired 128-bit vector half.
6549 unsigned NumElems = VT.getVectorNumElements();
6550 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
6551 bool Upper = IdxVal >= NumElems/2;
6552 SDValue Ins128Idx = DAG.getConstant(Upper ? NumElems/2 : 0, MVT::i32);
6553 SDValue V = Extract128BitVector(N0, Ins128Idx, DAG, dl);
6555 // Insert the element into the desired half.
6556 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V,
6557 N1, Upper ? DAG.getConstant(IdxVal-NumElems/2, MVT::i32) : N2);
6559 // Insert the changed part back to the 256-bit vector
6560 return Insert128BitVector(N0, V, Ins128Idx, DAG, dl);
6563 if (Subtarget->hasSSE41() || Subtarget->hasAVX())
6564 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
6566 if (EltVT == MVT::i8)
6569 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
6570 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
6571 // as its second argument.
6572 if (N1.getValueType() != MVT::i32)
6573 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6574 if (N2.getValueType() != MVT::i32)
6575 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6576 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
6582 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6583 LLVMContext *Context = DAG.getContext();
6584 DebugLoc dl = Op.getDebugLoc();
6585 EVT OpVT = Op.getValueType();
6587 // If this is a 256-bit vector result, first insert into a 128-bit
6588 // vector and then insert into the 256-bit vector.
6589 if (OpVT.getSizeInBits() > 128) {
6590 // Insert into a 128-bit vector.
6591 EVT VT128 = EVT::getVectorVT(*Context,
6592 OpVT.getVectorElementType(),
6593 OpVT.getVectorNumElements() / 2);
6595 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
6597 // Insert the 128-bit vector.
6598 return Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, OpVT), Op,
6599 DAG.getConstant(0, MVT::i32),
6603 if (Op.getValueType() == MVT::v1i64 &&
6604 Op.getOperand(0).getValueType() == MVT::i64)
6605 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
6607 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
6608 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
6609 "Expected an SSE type!");
6610 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
6611 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
6614 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
6615 // a simple subregister reference or explicit instructions to grab
6616 // upper bits of a vector.
6618 X86TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6619 if (Subtarget->hasAVX()) {
6620 DebugLoc dl = Op.getNode()->getDebugLoc();
6621 SDValue Vec = Op.getNode()->getOperand(0);
6622 SDValue Idx = Op.getNode()->getOperand(1);
6624 if (Op.getNode()->getValueType(0).getSizeInBits() == 128
6625 && Vec.getNode()->getValueType(0).getSizeInBits() == 256) {
6626 return Extract128BitVector(Vec, Idx, DAG, dl);
6632 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
6633 // simple superregister reference or explicit instructions to insert
6634 // the upper bits of a vector.
6636 X86TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6637 if (Subtarget->hasAVX()) {
6638 DebugLoc dl = Op.getNode()->getDebugLoc();
6639 SDValue Vec = Op.getNode()->getOperand(0);
6640 SDValue SubVec = Op.getNode()->getOperand(1);
6641 SDValue Idx = Op.getNode()->getOperand(2);
6643 if (Op.getNode()->getValueType(0).getSizeInBits() == 256
6644 && SubVec.getNode()->getValueType(0).getSizeInBits() == 128) {
6645 return Insert128BitVector(Vec, SubVec, Idx, DAG, dl);
6651 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
6652 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
6653 // one of the above mentioned nodes. It has to be wrapped because otherwise
6654 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
6655 // be used to form addressing mode. These wrapped nodes will be selected
6658 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
6659 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
6661 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6663 unsigned char OpFlag = 0;
6664 unsigned WrapperKind = X86ISD::Wrapper;
6665 CodeModel::Model M = getTargetMachine().getCodeModel();
6667 if (Subtarget->isPICStyleRIPRel() &&
6668 (M == CodeModel::Small || M == CodeModel::Kernel))
6669 WrapperKind = X86ISD::WrapperRIP;
6670 else if (Subtarget->isPICStyleGOT())
6671 OpFlag = X86II::MO_GOTOFF;
6672 else if (Subtarget->isPICStyleStubPIC())
6673 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6675 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
6677 CP->getOffset(), OpFlag);
6678 DebugLoc DL = CP->getDebugLoc();
6679 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6680 // With PIC, the address is actually $g + Offset.
6682 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6683 DAG.getNode(X86ISD::GlobalBaseReg,
6684 DebugLoc(), getPointerTy()),
6691 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
6692 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
6694 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6696 unsigned char OpFlag = 0;
6697 unsigned WrapperKind = X86ISD::Wrapper;
6698 CodeModel::Model M = getTargetMachine().getCodeModel();
6700 if (Subtarget->isPICStyleRIPRel() &&
6701 (M == CodeModel::Small || M == CodeModel::Kernel))
6702 WrapperKind = X86ISD::WrapperRIP;
6703 else if (Subtarget->isPICStyleGOT())
6704 OpFlag = X86II::MO_GOTOFF;
6705 else if (Subtarget->isPICStyleStubPIC())
6706 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6708 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
6710 DebugLoc DL = JT->getDebugLoc();
6711 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6713 // With PIC, the address is actually $g + Offset.
6715 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6716 DAG.getNode(X86ISD::GlobalBaseReg,
6717 DebugLoc(), getPointerTy()),
6724 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
6725 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
6727 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6729 unsigned char OpFlag = 0;
6730 unsigned WrapperKind = X86ISD::Wrapper;
6731 CodeModel::Model M = getTargetMachine().getCodeModel();
6733 if (Subtarget->isPICStyleRIPRel() &&
6734 (M == CodeModel::Small || M == CodeModel::Kernel))
6735 WrapperKind = X86ISD::WrapperRIP;
6736 else if (Subtarget->isPICStyleGOT())
6737 OpFlag = X86II::MO_GOTOFF;
6738 else if (Subtarget->isPICStyleStubPIC())
6739 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6741 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
6743 DebugLoc DL = Op.getDebugLoc();
6744 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6747 // With PIC, the address is actually $g + Offset.
6748 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
6749 !Subtarget->is64Bit()) {
6750 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6751 DAG.getNode(X86ISD::GlobalBaseReg,
6752 DebugLoc(), getPointerTy()),
6760 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
6761 // Create the TargetBlockAddressAddress node.
6762 unsigned char OpFlags =
6763 Subtarget->ClassifyBlockAddressReference();
6764 CodeModel::Model M = getTargetMachine().getCodeModel();
6765 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
6766 DebugLoc dl = Op.getDebugLoc();
6767 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
6768 /*isTarget=*/true, OpFlags);
6770 if (Subtarget->isPICStyleRIPRel() &&
6771 (M == CodeModel::Small || M == CodeModel::Kernel))
6772 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6774 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6776 // With PIC, the address is actually $g + Offset.
6777 if (isGlobalRelativeToPICBase(OpFlags)) {
6778 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6779 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6787 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
6789 SelectionDAG &DAG) const {
6790 // Create the TargetGlobalAddress node, folding in the constant
6791 // offset if it is legal.
6792 unsigned char OpFlags =
6793 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
6794 CodeModel::Model M = getTargetMachine().getCodeModel();
6796 if (OpFlags == X86II::MO_NO_FLAG &&
6797 X86::isOffsetSuitableForCodeModel(Offset, M)) {
6798 // A direct static reference to a global.
6799 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
6802 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
6805 if (Subtarget->isPICStyleRIPRel() &&
6806 (M == CodeModel::Small || M == CodeModel::Kernel))
6807 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6809 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6811 // With PIC, the address is actually $g + Offset.
6812 if (isGlobalRelativeToPICBase(OpFlags)) {
6813 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6814 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6818 // For globals that require a load from a stub to get the address, emit the
6820 if (isGlobalStubReference(OpFlags))
6821 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
6822 MachinePointerInfo::getGOT(), false, false, 0);
6824 // If there was a non-zero offset that we didn't fold, create an explicit
6827 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
6828 DAG.getConstant(Offset, getPointerTy()));
6834 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
6835 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
6836 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
6837 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
6841 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
6842 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
6843 unsigned char OperandFlags) {
6844 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6845 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6846 DebugLoc dl = GA->getDebugLoc();
6847 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6848 GA->getValueType(0),
6852 SDValue Ops[] = { Chain, TGA, *InFlag };
6853 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
6855 SDValue Ops[] = { Chain, TGA };
6856 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
6859 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
6860 MFI->setAdjustsStack(true);
6862 SDValue Flag = Chain.getValue(1);
6863 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
6866 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
6868 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6871 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
6872 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
6873 DAG.getNode(X86ISD::GlobalBaseReg,
6874 DebugLoc(), PtrVT), InFlag);
6875 InFlag = Chain.getValue(1);
6877 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
6880 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
6882 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6884 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
6885 X86::RAX, X86II::MO_TLSGD);
6888 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
6889 // "local exec" model.
6890 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
6891 const EVT PtrVT, TLSModel::Model model,
6893 DebugLoc dl = GA->getDebugLoc();
6895 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
6896 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
6897 is64Bit ? 257 : 256));
6899 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
6900 DAG.getIntPtrConstant(0),
6901 MachinePointerInfo(Ptr), false, false, 0);
6903 unsigned char OperandFlags = 0;
6904 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
6906 unsigned WrapperKind = X86ISD::Wrapper;
6907 if (model == TLSModel::LocalExec) {
6908 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
6909 } else if (is64Bit) {
6910 assert(model == TLSModel::InitialExec);
6911 OperandFlags = X86II::MO_GOTTPOFF;
6912 WrapperKind = X86ISD::WrapperRIP;
6914 assert(model == TLSModel::InitialExec);
6915 OperandFlags = X86II::MO_INDNTPOFF;
6918 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
6920 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
6921 GA->getValueType(0),
6922 GA->getOffset(), OperandFlags);
6923 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
6925 if (model == TLSModel::InitialExec)
6926 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
6927 MachinePointerInfo::getGOT(), false, false, 0);
6929 // The address of the thread local variable is the add of the thread
6930 // pointer with the offset of the variable.
6931 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
6935 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
6937 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
6938 const GlobalValue *GV = GA->getGlobal();
6940 if (Subtarget->isTargetELF()) {
6941 // TODO: implement the "local dynamic" model
6942 // TODO: implement the "initial exec"model for pic executables
6944 // If GV is an alias then use the aliasee for determining
6945 // thread-localness.
6946 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
6947 GV = GA->resolveAliasedGlobal(false);
6949 TLSModel::Model model
6950 = getTLSModel(GV, getTargetMachine().getRelocationModel());
6953 case TLSModel::GeneralDynamic:
6954 case TLSModel::LocalDynamic: // not implemented
6955 if (Subtarget->is64Bit())
6956 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
6957 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
6959 case TLSModel::InitialExec:
6960 case TLSModel::LocalExec:
6961 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
6962 Subtarget->is64Bit());
6964 } else if (Subtarget->isTargetDarwin()) {
6965 // Darwin only has one model of TLS. Lower to that.
6966 unsigned char OpFlag = 0;
6967 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
6968 X86ISD::WrapperRIP : X86ISD::Wrapper;
6970 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6972 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
6973 !Subtarget->is64Bit();
6975 OpFlag = X86II::MO_TLVP_PIC_BASE;
6977 OpFlag = X86II::MO_TLVP;
6978 DebugLoc DL = Op.getDebugLoc();
6979 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
6980 GA->getValueType(0),
6981 GA->getOffset(), OpFlag);
6982 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6984 // With PIC32, the address is actually $g + Offset.
6986 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6987 DAG.getNode(X86ISD::GlobalBaseReg,
6988 DebugLoc(), getPointerTy()),
6991 // Lowering the machine isd will make sure everything is in the right
6993 SDValue Chain = DAG.getEntryNode();
6994 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
6995 SDValue Args[] = { Chain, Offset };
6996 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
6998 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
6999 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7000 MFI->setAdjustsStack(true);
7002 // And our return value (tls address) is in the standard call return value
7004 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
7005 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
7009 "TLS not implemented for this target.");
7011 llvm_unreachable("Unreachable");
7016 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values and
7017 /// take a 2 x i32 value to shift plus a shift amount.
7018 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const {
7019 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
7020 EVT VT = Op.getValueType();
7021 unsigned VTBits = VT.getSizeInBits();
7022 DebugLoc dl = Op.getDebugLoc();
7023 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
7024 SDValue ShOpLo = Op.getOperand(0);
7025 SDValue ShOpHi = Op.getOperand(1);
7026 SDValue ShAmt = Op.getOperand(2);
7027 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
7028 DAG.getConstant(VTBits - 1, MVT::i8))
7029 : DAG.getConstant(0, VT);
7032 if (Op.getOpcode() == ISD::SHL_PARTS) {
7033 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
7034 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
7036 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
7037 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
7040 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
7041 DAG.getConstant(VTBits, MVT::i8));
7042 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
7043 AndNode, DAG.getConstant(0, MVT::i8));
7046 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7047 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
7048 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
7050 if (Op.getOpcode() == ISD::SHL_PARTS) {
7051 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7052 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7054 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7055 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7058 SDValue Ops[2] = { Lo, Hi };
7059 return DAG.getMergeValues(Ops, 2, dl);
7062 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
7063 SelectionDAG &DAG) const {
7064 EVT SrcVT = Op.getOperand(0).getValueType();
7066 if (SrcVT.isVector())
7069 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
7070 "Unknown SINT_TO_FP to lower!");
7072 // These are really Legal; return the operand so the caller accepts it as
7074 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
7076 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
7077 Subtarget->is64Bit()) {
7081 DebugLoc dl = Op.getDebugLoc();
7082 unsigned Size = SrcVT.getSizeInBits()/8;
7083 MachineFunction &MF = DAG.getMachineFunction();
7084 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
7085 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7086 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7088 MachinePointerInfo::getFixedStack(SSFI),
7090 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
7093 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
7095 SelectionDAG &DAG) const {
7097 DebugLoc DL = Op.getDebugLoc();
7099 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
7101 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
7103 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
7105 unsigned ByteSize = SrcVT.getSizeInBits()/8;
7107 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
7108 MachineMemOperand *MMO;
7110 int SSFI = FI->getIndex();
7112 DAG.getMachineFunction()
7113 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7114 MachineMemOperand::MOLoad, ByteSize, ByteSize);
7116 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
7117 StackSlot = StackSlot.getOperand(1);
7119 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
7120 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
7122 Tys, Ops, array_lengthof(Ops),
7126 Chain = Result.getValue(1);
7127 SDValue InFlag = Result.getValue(2);
7129 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
7130 // shouldn't be necessary except that RFP cannot be live across
7131 // multiple blocks. When stackifier is fixed, they can be uncoupled.
7132 MachineFunction &MF = DAG.getMachineFunction();
7133 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
7134 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
7135 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7136 Tys = DAG.getVTList(MVT::Other);
7138 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
7140 MachineMemOperand *MMO =
7141 DAG.getMachineFunction()
7142 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7143 MachineMemOperand::MOStore, SSFISize, SSFISize);
7145 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
7146 Ops, array_lengthof(Ops),
7147 Op.getValueType(), MMO);
7148 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
7149 MachinePointerInfo::getFixedStack(SSFI),
7156 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
7157 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
7158 SelectionDAG &DAG) const {
7159 // This algorithm is not obvious. Here it is in C code, more or less:
7161 double uint64_to_double( uint32_t hi, uint32_t lo ) {
7162 static const __m128i exp = { 0x4330000045300000ULL, 0 };
7163 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
7165 // Copy ints to xmm registers.
7166 __m128i xh = _mm_cvtsi32_si128( hi );
7167 __m128i xl = _mm_cvtsi32_si128( lo );
7169 // Combine into low half of a single xmm register.
7170 __m128i x = _mm_unpacklo_epi32( xh, xl );
7174 // Merge in appropriate exponents to give the integer bits the right
7176 x = _mm_unpacklo_epi32( x, exp );
7178 // Subtract away the biases to deal with the IEEE-754 double precision
7180 d = _mm_sub_pd( (__m128d) x, bias );
7182 // All conversions up to here are exact. The correctly rounded result is
7183 // calculated using the current rounding mode using the following
7185 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
7186 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
7187 // store doesn't really need to be here (except
7188 // maybe to zero the other double)
7193 DebugLoc dl = Op.getDebugLoc();
7194 LLVMContext *Context = DAG.getContext();
7196 // Build some magic constants.
7197 std::vector<Constant*> CV0;
7198 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
7199 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
7200 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
7201 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
7202 Constant *C0 = ConstantVector::get(CV0);
7203 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
7205 std::vector<Constant*> CV1;
7207 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
7209 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
7210 Constant *C1 = ConstantVector::get(CV1);
7211 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
7213 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7214 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7216 DAG.getIntPtrConstant(1)));
7217 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7218 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7220 DAG.getIntPtrConstant(0)));
7221 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
7222 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
7223 MachinePointerInfo::getConstantPool(),
7225 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
7226 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
7227 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
7228 MachinePointerInfo::getConstantPool(),
7230 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
7232 // Add the halves; easiest way is to swap them into another reg first.
7233 int ShufMask[2] = { 1, -1 };
7234 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
7235 DAG.getUNDEF(MVT::v2f64), ShufMask);
7236 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
7237 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
7238 DAG.getIntPtrConstant(0));
7241 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
7242 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
7243 SelectionDAG &DAG) const {
7244 DebugLoc dl = Op.getDebugLoc();
7245 // FP constant to bias correct the final result.
7246 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
7249 // Load the 32-bit value into an XMM register.
7250 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7251 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7253 DAG.getIntPtrConstant(0)));
7255 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7256 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
7257 DAG.getIntPtrConstant(0));
7259 // Or the load with the bias.
7260 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
7261 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7262 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7264 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7265 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7266 MVT::v2f64, Bias)));
7267 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7268 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
7269 DAG.getIntPtrConstant(0));
7271 // Subtract the bias.
7272 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
7274 // Handle final rounding.
7275 EVT DestVT = Op.getValueType();
7277 if (DestVT.bitsLT(MVT::f64)) {
7278 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
7279 DAG.getIntPtrConstant(0));
7280 } else if (DestVT.bitsGT(MVT::f64)) {
7281 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
7284 // Handle final rounding.
7288 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
7289 SelectionDAG &DAG) const {
7290 SDValue N0 = Op.getOperand(0);
7291 DebugLoc dl = Op.getDebugLoc();
7293 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
7294 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
7295 // the optimization here.
7296 if (DAG.SignBitIsZero(N0))
7297 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
7299 EVT SrcVT = N0.getValueType();
7300 EVT DstVT = Op.getValueType();
7301 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
7302 return LowerUINT_TO_FP_i64(Op, DAG);
7303 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
7304 return LowerUINT_TO_FP_i32(Op, DAG);
7306 // Make a 64-bit buffer, and use it to build an FILD.
7307 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
7308 if (SrcVT == MVT::i32) {
7309 SDValue WordOff = DAG.getConstant(4, getPointerTy());
7310 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
7311 getPointerTy(), StackSlot, WordOff);
7312 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7313 StackSlot, MachinePointerInfo(),
7315 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
7316 OffsetSlot, MachinePointerInfo(),
7318 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
7322 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
7323 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7324 StackSlot, MachinePointerInfo(),
7326 // For i64 source, we need to add the appropriate power of 2 if the input
7327 // was negative. This is the same as the optimization in
7328 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
7329 // we must be careful to do the computation in x87 extended precision, not
7330 // in SSE. (The generic code can't know it's OK to do this, or how to.)
7331 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
7332 MachineMemOperand *MMO =
7333 DAG.getMachineFunction()
7334 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7335 MachineMemOperand::MOLoad, 8, 8);
7337 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
7338 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
7339 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
7342 APInt FF(32, 0x5F800000ULL);
7344 // Check whether the sign bit is set.
7345 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
7346 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
7349 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
7350 SDValue FudgePtr = DAG.getConstantPool(
7351 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
7354 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
7355 SDValue Zero = DAG.getIntPtrConstant(0);
7356 SDValue Four = DAG.getIntPtrConstant(4);
7357 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
7359 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
7361 // Load the value out, extending it from f32 to f80.
7362 // FIXME: Avoid the extend by constructing the right constant pool?
7363 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
7364 FudgePtr, MachinePointerInfo::getConstantPool(),
7365 MVT::f32, false, false, 4);
7366 // Extend everything to 80 bits to force it to be done on x87.
7367 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
7368 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
7371 std::pair<SDValue,SDValue> X86TargetLowering::
7372 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
7373 DebugLoc DL = Op.getDebugLoc();
7375 EVT DstTy = Op.getValueType();
7378 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
7382 assert(DstTy.getSimpleVT() <= MVT::i64 &&
7383 DstTy.getSimpleVT() >= MVT::i16 &&
7384 "Unknown FP_TO_SINT to lower!");
7386 // These are really Legal.
7387 if (DstTy == MVT::i32 &&
7388 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7389 return std::make_pair(SDValue(), SDValue());
7390 if (Subtarget->is64Bit() &&
7391 DstTy == MVT::i64 &&
7392 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7393 return std::make_pair(SDValue(), SDValue());
7395 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
7397 MachineFunction &MF = DAG.getMachineFunction();
7398 unsigned MemSize = DstTy.getSizeInBits()/8;
7399 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7400 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7405 switch (DstTy.getSimpleVT().SimpleTy) {
7406 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
7407 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
7408 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
7409 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
7412 SDValue Chain = DAG.getEntryNode();
7413 SDValue Value = Op.getOperand(0);
7414 EVT TheVT = Op.getOperand(0).getValueType();
7415 if (isScalarFPTypeInSSEReg(TheVT)) {
7416 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
7417 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
7418 MachinePointerInfo::getFixedStack(SSFI),
7420 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
7422 Chain, StackSlot, DAG.getValueType(TheVT)
7425 MachineMemOperand *MMO =
7426 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7427 MachineMemOperand::MOLoad, MemSize, MemSize);
7428 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
7430 Chain = Value.getValue(1);
7431 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7432 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7435 MachineMemOperand *MMO =
7436 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7437 MachineMemOperand::MOStore, MemSize, MemSize);
7439 // Build the FP_TO_INT*_IN_MEM
7440 SDValue Ops[] = { Chain, Value, StackSlot };
7441 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
7442 Ops, 3, DstTy, MMO);
7444 return std::make_pair(FIST, StackSlot);
7447 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
7448 SelectionDAG &DAG) const {
7449 if (Op.getValueType().isVector())
7452 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
7453 SDValue FIST = Vals.first, StackSlot = Vals.second;
7454 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
7455 if (FIST.getNode() == 0) return Op;
7458 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7459 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7462 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
7463 SelectionDAG &DAG) const {
7464 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
7465 SDValue FIST = Vals.first, StackSlot = Vals.second;
7466 assert(FIST.getNode() && "Unexpected failure");
7469 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7470 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7473 SDValue X86TargetLowering::LowerFABS(SDValue Op,
7474 SelectionDAG &DAG) const {
7475 LLVMContext *Context = DAG.getContext();
7476 DebugLoc dl = Op.getDebugLoc();
7477 EVT VT = Op.getValueType();
7480 EltVT = VT.getVectorElementType();
7481 std::vector<Constant*> CV;
7482 if (EltVT == MVT::f64) {
7483 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
7487 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
7493 Constant *C = ConstantVector::get(CV);
7494 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7495 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7496 MachinePointerInfo::getConstantPool(),
7498 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
7501 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
7502 LLVMContext *Context = DAG.getContext();
7503 DebugLoc dl = Op.getDebugLoc();
7504 EVT VT = Op.getValueType();
7507 EltVT = VT.getVectorElementType();
7508 std::vector<Constant*> CV;
7509 if (EltVT == MVT::f64) {
7510 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
7514 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
7520 Constant *C = ConstantVector::get(CV);
7521 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7522 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7523 MachinePointerInfo::getConstantPool(),
7525 if (VT.isVector()) {
7526 return DAG.getNode(ISD::BITCAST, dl, VT,
7527 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
7528 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7530 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
7532 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
7536 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
7537 LLVMContext *Context = DAG.getContext();
7538 SDValue Op0 = Op.getOperand(0);
7539 SDValue Op1 = Op.getOperand(1);
7540 DebugLoc dl = Op.getDebugLoc();
7541 EVT VT = Op.getValueType();
7542 EVT SrcVT = Op1.getValueType();
7544 // If second operand is smaller, extend it first.
7545 if (SrcVT.bitsLT(VT)) {
7546 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
7549 // And if it is bigger, shrink it first.
7550 if (SrcVT.bitsGT(VT)) {
7551 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
7555 // At this point the operands and the result should have the same
7556 // type, and that won't be f80 since that is not custom lowered.
7558 // First get the sign bit of second operand.
7559 std::vector<Constant*> CV;
7560 if (SrcVT == MVT::f64) {
7561 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
7562 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7564 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
7565 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7566 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7567 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7569 Constant *C = ConstantVector::get(CV);
7570 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7571 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
7572 MachinePointerInfo::getConstantPool(),
7574 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
7576 // Shift sign bit right or left if the two operands have different types.
7577 if (SrcVT.bitsGT(VT)) {
7578 // Op0 is MVT::f32, Op1 is MVT::f64.
7579 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
7580 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
7581 DAG.getConstant(32, MVT::i32));
7582 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
7583 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
7584 DAG.getIntPtrConstant(0));
7587 // Clear first operand sign bit.
7589 if (VT == MVT::f64) {
7590 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
7591 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7593 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
7594 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7595 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7596 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7598 C = ConstantVector::get(CV);
7599 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7600 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7601 MachinePointerInfo::getConstantPool(),
7603 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
7605 // Or the value with the sign bit.
7606 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
7609 SDValue X86TargetLowering::LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) const {
7610 SDValue N0 = Op.getOperand(0);
7611 DebugLoc dl = Op.getDebugLoc();
7612 EVT VT = Op.getValueType();
7614 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
7615 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
7616 DAG.getConstant(1, VT));
7617 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
7620 /// Emit nodes that will be selected as "test Op0,Op0", or something
7622 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
7623 SelectionDAG &DAG) const {
7624 DebugLoc dl = Op.getDebugLoc();
7626 // CF and OF aren't always set the way we want. Determine which
7627 // of these we need.
7628 bool NeedCF = false;
7629 bool NeedOF = false;
7632 case X86::COND_A: case X86::COND_AE:
7633 case X86::COND_B: case X86::COND_BE:
7636 case X86::COND_G: case X86::COND_GE:
7637 case X86::COND_L: case X86::COND_LE:
7638 case X86::COND_O: case X86::COND_NO:
7643 // See if we can use the EFLAGS value from the operand instead of
7644 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
7645 // we prove that the arithmetic won't overflow, we can't use OF or CF.
7646 if (Op.getResNo() != 0 || NeedOF || NeedCF)
7647 // Emit a CMP with 0, which is the TEST pattern.
7648 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7649 DAG.getConstant(0, Op.getValueType()));
7651 unsigned Opcode = 0;
7652 unsigned NumOperands = 0;
7653 switch (Op.getNode()->getOpcode()) {
7655 // Due to an isel shortcoming, be conservative if this add is likely to be
7656 // selected as part of a load-modify-store instruction. When the root node
7657 // in a match is a store, isel doesn't know how to remap non-chain non-flag
7658 // uses of other nodes in the match, such as the ADD in this case. This
7659 // leads to the ADD being left around and reselected, with the result being
7660 // two adds in the output. Alas, even if none our users are stores, that
7661 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
7662 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
7663 // climbing the DAG back to the root, and it doesn't seem to be worth the
7665 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7666 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7667 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
7670 if (ConstantSDNode *C =
7671 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
7672 // An add of one will be selected as an INC.
7673 if (C->getAPIntValue() == 1) {
7674 Opcode = X86ISD::INC;
7679 // An add of negative one (subtract of one) will be selected as a DEC.
7680 if (C->getAPIntValue().isAllOnesValue()) {
7681 Opcode = X86ISD::DEC;
7687 // Otherwise use a regular EFLAGS-setting add.
7688 Opcode = X86ISD::ADD;
7692 // If the primary and result isn't used, don't bother using X86ISD::AND,
7693 // because a TEST instruction will be better.
7694 bool NonFlagUse = false;
7695 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7696 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
7698 unsigned UOpNo = UI.getOperandNo();
7699 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
7700 // Look pass truncate.
7701 UOpNo = User->use_begin().getOperandNo();
7702 User = *User->use_begin();
7705 if (User->getOpcode() != ISD::BRCOND &&
7706 User->getOpcode() != ISD::SETCC &&
7707 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
7720 // Due to the ISEL shortcoming noted above, be conservative if this op is
7721 // likely to be selected as part of a load-modify-store instruction.
7722 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7723 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7724 if (UI->getOpcode() == ISD::STORE)
7727 // Otherwise use a regular EFLAGS-setting instruction.
7728 switch (Op.getNode()->getOpcode()) {
7729 default: llvm_unreachable("unexpected operator!");
7730 case ISD::SUB: Opcode = X86ISD::SUB; break;
7731 case ISD::OR: Opcode = X86ISD::OR; break;
7732 case ISD::XOR: Opcode = X86ISD::XOR; break;
7733 case ISD::AND: Opcode = X86ISD::AND; break;
7745 return SDValue(Op.getNode(), 1);
7752 // Emit a CMP with 0, which is the TEST pattern.
7753 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7754 DAG.getConstant(0, Op.getValueType()));
7756 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
7757 SmallVector<SDValue, 4> Ops;
7758 for (unsigned i = 0; i != NumOperands; ++i)
7759 Ops.push_back(Op.getOperand(i));
7761 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
7762 DAG.ReplaceAllUsesWith(Op, New);
7763 return SDValue(New.getNode(), 1);
7766 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
7768 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
7769 SelectionDAG &DAG) const {
7770 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
7771 if (C->getAPIntValue() == 0)
7772 return EmitTest(Op0, X86CC, DAG);
7774 DebugLoc dl = Op0.getDebugLoc();
7775 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
7778 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
7779 /// if it's possible.
7780 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
7781 DebugLoc dl, SelectionDAG &DAG) const {
7782 SDValue Op0 = And.getOperand(0);
7783 SDValue Op1 = And.getOperand(1);
7784 if (Op0.getOpcode() == ISD::TRUNCATE)
7785 Op0 = Op0.getOperand(0);
7786 if (Op1.getOpcode() == ISD::TRUNCATE)
7787 Op1 = Op1.getOperand(0);
7790 if (Op1.getOpcode() == ISD::SHL)
7791 std::swap(Op0, Op1);
7792 if (Op0.getOpcode() == ISD::SHL) {
7793 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
7794 if (And00C->getZExtValue() == 1) {
7795 // If we looked past a truncate, check that it's only truncating away
7797 unsigned BitWidth = Op0.getValueSizeInBits();
7798 unsigned AndBitWidth = And.getValueSizeInBits();
7799 if (BitWidth > AndBitWidth) {
7800 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
7801 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
7802 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
7806 RHS = Op0.getOperand(1);
7808 } else if (Op1.getOpcode() == ISD::Constant) {
7809 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
7810 SDValue AndLHS = Op0;
7811 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
7812 LHS = AndLHS.getOperand(0);
7813 RHS = AndLHS.getOperand(1);
7817 if (LHS.getNode()) {
7818 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
7819 // instruction. Since the shift amount is in-range-or-undefined, we know
7820 // that doing a bittest on the i32 value is ok. We extend to i32 because
7821 // the encoding for the i16 version is larger than the i32 version.
7822 // Also promote i16 to i32 for performance / code size reason.
7823 if (LHS.getValueType() == MVT::i8 ||
7824 LHS.getValueType() == MVT::i16)
7825 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
7827 // If the operand types disagree, extend the shift amount to match. Since
7828 // BT ignores high bits (like shifts) we can use anyextend.
7829 if (LHS.getValueType() != RHS.getValueType())
7830 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
7832 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
7833 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
7834 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7835 DAG.getConstant(Cond, MVT::i8), BT);
7841 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
7842 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
7843 SDValue Op0 = Op.getOperand(0);
7844 SDValue Op1 = Op.getOperand(1);
7845 DebugLoc dl = Op.getDebugLoc();
7846 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
7848 // Optimize to BT if possible.
7849 // Lower (X & (1 << N)) == 0 to BT(X, N).
7850 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
7851 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
7852 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
7853 Op1.getOpcode() == ISD::Constant &&
7854 cast<ConstantSDNode>(Op1)->isNullValue() &&
7855 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7856 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
7857 if (NewSetCC.getNode())
7861 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
7863 if (Op1.getOpcode() == ISD::Constant &&
7864 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
7865 cast<ConstantSDNode>(Op1)->isNullValue()) &&
7866 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
7868 // If the input is a setcc, then reuse the input setcc or use a new one with
7869 // the inverted condition.
7870 if (Op0.getOpcode() == X86ISD::SETCC) {
7871 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
7872 bool Invert = (CC == ISD::SETNE) ^
7873 cast<ConstantSDNode>(Op1)->isNullValue();
7874 if (!Invert) return Op0;
7876 CCode = X86::GetOppositeBranchCondition(CCode);
7877 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7878 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
7882 bool isFP = Op1.getValueType().isFloatingPoint();
7883 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
7884 if (X86CC == X86::COND_INVALID)
7887 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
7888 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
7889 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
7892 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
7894 SDValue Op0 = Op.getOperand(0);
7895 SDValue Op1 = Op.getOperand(1);
7896 SDValue CC = Op.getOperand(2);
7897 EVT VT = Op.getValueType();
7898 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
7899 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
7900 DebugLoc dl = Op.getDebugLoc();
7904 EVT VT0 = Op0.getValueType();
7905 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
7906 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
7909 switch (SetCCOpcode) {
7912 case ISD::SETEQ: SSECC = 0; break;
7914 case ISD::SETGT: Swap = true; // Fallthrough
7916 case ISD::SETOLT: SSECC = 1; break;
7918 case ISD::SETGE: Swap = true; // Fallthrough
7920 case ISD::SETOLE: SSECC = 2; break;
7921 case ISD::SETUO: SSECC = 3; break;
7923 case ISD::SETNE: SSECC = 4; break;
7924 case ISD::SETULE: Swap = true;
7925 case ISD::SETUGE: SSECC = 5; break;
7926 case ISD::SETULT: Swap = true;
7927 case ISD::SETUGT: SSECC = 6; break;
7928 case ISD::SETO: SSECC = 7; break;
7931 std::swap(Op0, Op1);
7933 // In the two special cases we can't handle, emit two comparisons.
7935 if (SetCCOpcode == ISD::SETUEQ) {
7937 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
7938 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
7939 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
7941 else if (SetCCOpcode == ISD::SETONE) {
7943 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
7944 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
7945 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
7947 llvm_unreachable("Illegal FP comparison");
7949 // Handle all other FP comparisons here.
7950 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
7953 // We are handling one of the integer comparisons here. Since SSE only has
7954 // GT and EQ comparisons for integer, swapping operands and multiple
7955 // operations may be required for some comparisons.
7956 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
7957 bool Swap = false, Invert = false, FlipSigns = false;
7959 switch (VT.getSimpleVT().SimpleTy) {
7961 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
7962 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
7963 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
7964 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
7967 switch (SetCCOpcode) {
7969 case ISD::SETNE: Invert = true;
7970 case ISD::SETEQ: Opc = EQOpc; break;
7971 case ISD::SETLT: Swap = true;
7972 case ISD::SETGT: Opc = GTOpc; break;
7973 case ISD::SETGE: Swap = true;
7974 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
7975 case ISD::SETULT: Swap = true;
7976 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
7977 case ISD::SETUGE: Swap = true;
7978 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
7981 std::swap(Op0, Op1);
7983 // Since SSE has no unsigned integer comparisons, we need to flip the sign
7984 // bits of the inputs before performing those operations.
7986 EVT EltVT = VT.getVectorElementType();
7987 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
7989 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
7990 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
7992 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
7993 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
7996 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
7998 // If the logical-not of the result is required, perform that now.
8000 Result = DAG.getNOT(dl, Result, VT);
8005 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
8006 static bool isX86LogicalCmp(SDValue Op) {
8007 unsigned Opc = Op.getNode()->getOpcode();
8008 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
8010 if (Op.getResNo() == 1 &&
8011 (Opc == X86ISD::ADD ||
8012 Opc == X86ISD::SUB ||
8013 Opc == X86ISD::ADC ||
8014 Opc == X86ISD::SBB ||
8015 Opc == X86ISD::SMUL ||
8016 Opc == X86ISD::UMUL ||
8017 Opc == X86ISD::INC ||
8018 Opc == X86ISD::DEC ||
8019 Opc == X86ISD::OR ||
8020 Opc == X86ISD::XOR ||
8021 Opc == X86ISD::AND))
8024 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
8030 static bool isZero(SDValue V) {
8031 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
8032 return C && C->isNullValue();
8035 static bool isAllOnes(SDValue V) {
8036 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
8037 return C && C->isAllOnesValue();
8040 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
8041 bool addTest = true;
8042 SDValue Cond = Op.getOperand(0);
8043 SDValue Op1 = Op.getOperand(1);
8044 SDValue Op2 = Op.getOperand(2);
8045 DebugLoc DL = Op.getDebugLoc();
8048 if (Cond.getOpcode() == ISD::SETCC) {
8049 SDValue NewCond = LowerSETCC(Cond, DAG);
8050 if (NewCond.getNode())
8054 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
8055 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
8056 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
8057 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
8058 if (Cond.getOpcode() == X86ISD::SETCC &&
8059 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
8060 isZero(Cond.getOperand(1).getOperand(1))) {
8061 SDValue Cmp = Cond.getOperand(1);
8063 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
8065 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
8066 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
8067 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
8069 SDValue CmpOp0 = Cmp.getOperand(0);
8070 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
8071 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
8073 SDValue Res = // Res = 0 or -1.
8074 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
8075 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
8077 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
8078 Res = DAG.getNOT(DL, Res, Res.getValueType());
8080 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
8081 if (N2C == 0 || !N2C->isNullValue())
8082 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
8087 // Look past (and (setcc_carry (cmp ...)), 1).
8088 if (Cond.getOpcode() == ISD::AND &&
8089 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
8090 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
8091 if (C && C->getAPIntValue() == 1)
8092 Cond = Cond.getOperand(0);
8095 // If condition flag is set by a X86ISD::CMP, then use it as the condition
8096 // setting operand in place of the X86ISD::SETCC.
8097 if (Cond.getOpcode() == X86ISD::SETCC ||
8098 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
8099 CC = Cond.getOperand(0);
8101 SDValue Cmp = Cond.getOperand(1);
8102 unsigned Opc = Cmp.getOpcode();
8103 EVT VT = Op.getValueType();
8105 bool IllegalFPCMov = false;
8106 if (VT.isFloatingPoint() && !VT.isVector() &&
8107 !isScalarFPTypeInSSEReg(VT)) // FPStack?
8108 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
8110 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
8111 Opc == X86ISD::BT) { // FIXME
8118 // Look pass the truncate.
8119 if (Cond.getOpcode() == ISD::TRUNCATE)
8120 Cond = Cond.getOperand(0);
8122 // We know the result of AND is compared against zero. Try to match
8124 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
8125 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
8126 if (NewSetCC.getNode()) {
8127 CC = NewSetCC.getOperand(0);
8128 Cond = NewSetCC.getOperand(1);
8135 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8136 Cond = EmitTest(Cond, X86::COND_NE, DAG);
8139 // a < b ? -1 : 0 -> RES = ~setcc_carry
8140 // a < b ? 0 : -1 -> RES = setcc_carry
8141 // a >= b ? -1 : 0 -> RES = setcc_carry
8142 // a >= b ? 0 : -1 -> RES = ~setcc_carry
8143 if (Cond.getOpcode() == X86ISD::CMP) {
8144 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
8146 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
8147 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
8148 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
8149 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
8150 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
8151 return DAG.getNOT(DL, Res, Res.getValueType());
8156 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
8157 // condition is true.
8158 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
8159 SDValue Ops[] = { Op2, Op1, CC, Cond };
8160 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
8163 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
8164 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
8165 // from the AND / OR.
8166 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
8167 Opc = Op.getOpcode();
8168 if (Opc != ISD::OR && Opc != ISD::AND)
8170 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
8171 Op.getOperand(0).hasOneUse() &&
8172 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
8173 Op.getOperand(1).hasOneUse());
8176 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
8177 // 1 and that the SETCC node has a single use.
8178 static bool isXor1OfSetCC(SDValue Op) {
8179 if (Op.getOpcode() != ISD::XOR)
8181 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8182 if (N1C && N1C->getAPIntValue() == 1) {
8183 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
8184 Op.getOperand(0).hasOneUse();
8189 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
8190 bool addTest = true;
8191 SDValue Chain = Op.getOperand(0);
8192 SDValue Cond = Op.getOperand(1);
8193 SDValue Dest = Op.getOperand(2);
8194 DebugLoc dl = Op.getDebugLoc();
8197 if (Cond.getOpcode() == ISD::SETCC) {
8198 SDValue NewCond = LowerSETCC(Cond, DAG);
8199 if (NewCond.getNode())
8203 // FIXME: LowerXALUO doesn't handle these!!
8204 else if (Cond.getOpcode() == X86ISD::ADD ||
8205 Cond.getOpcode() == X86ISD::SUB ||
8206 Cond.getOpcode() == X86ISD::SMUL ||
8207 Cond.getOpcode() == X86ISD::UMUL)
8208 Cond = LowerXALUO(Cond, DAG);
8211 // Look pass (and (setcc_carry (cmp ...)), 1).
8212 if (Cond.getOpcode() == ISD::AND &&
8213 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
8214 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
8215 if (C && C->getAPIntValue() == 1)
8216 Cond = Cond.getOperand(0);
8219 // If condition flag is set by a X86ISD::CMP, then use it as the condition
8220 // setting operand in place of the X86ISD::SETCC.
8221 if (Cond.getOpcode() == X86ISD::SETCC ||
8222 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
8223 CC = Cond.getOperand(0);
8225 SDValue Cmp = Cond.getOperand(1);
8226 unsigned Opc = Cmp.getOpcode();
8227 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
8228 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
8232 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
8236 // These can only come from an arithmetic instruction with overflow,
8237 // e.g. SADDO, UADDO.
8238 Cond = Cond.getNode()->getOperand(1);
8245 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
8246 SDValue Cmp = Cond.getOperand(0).getOperand(1);
8247 if (CondOpc == ISD::OR) {
8248 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
8249 // two branches instead of an explicit OR instruction with a
8251 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8252 isX86LogicalCmp(Cmp)) {
8253 CC = Cond.getOperand(0).getOperand(0);
8254 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8255 Chain, Dest, CC, Cmp);
8256 CC = Cond.getOperand(1).getOperand(0);
8260 } else { // ISD::AND
8261 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
8262 // two branches instead of an explicit AND instruction with a
8263 // separate test. However, we only do this if this block doesn't
8264 // have a fall-through edge, because this requires an explicit
8265 // jmp when the condition is false.
8266 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8267 isX86LogicalCmp(Cmp) &&
8268 Op.getNode()->hasOneUse()) {
8269 X86::CondCode CCode =
8270 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8271 CCode = X86::GetOppositeBranchCondition(CCode);
8272 CC = DAG.getConstant(CCode, MVT::i8);
8273 SDNode *User = *Op.getNode()->use_begin();
8274 // Look for an unconditional branch following this conditional branch.
8275 // We need this because we need to reverse the successors in order
8276 // to implement FCMP_OEQ.
8277 if (User->getOpcode() == ISD::BR) {
8278 SDValue FalseBB = User->getOperand(1);
8280 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
8281 assert(NewBR == User);
8285 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8286 Chain, Dest, CC, Cmp);
8287 X86::CondCode CCode =
8288 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
8289 CCode = X86::GetOppositeBranchCondition(CCode);
8290 CC = DAG.getConstant(CCode, MVT::i8);
8296 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
8297 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
8298 // It should be transformed during dag combiner except when the condition
8299 // is set by a arithmetics with overflow node.
8300 X86::CondCode CCode =
8301 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8302 CCode = X86::GetOppositeBranchCondition(CCode);
8303 CC = DAG.getConstant(CCode, MVT::i8);
8304 Cond = Cond.getOperand(0).getOperand(1);
8310 // Look pass the truncate.
8311 if (Cond.getOpcode() == ISD::TRUNCATE)
8312 Cond = Cond.getOperand(0);
8314 // We know the result of AND is compared against zero. Try to match
8316 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
8317 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
8318 if (NewSetCC.getNode()) {
8319 CC = NewSetCC.getOperand(0);
8320 Cond = NewSetCC.getOperand(1);
8327 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8328 Cond = EmitTest(Cond, X86::COND_NE, DAG);
8330 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8331 Chain, Dest, CC, Cond);
8335 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
8336 // Calls to _alloca is needed to probe the stack when allocating more than 4k
8337 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
8338 // that the guard pages used by the OS virtual memory manager are allocated in
8339 // correct sequence.
8341 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
8342 SelectionDAG &DAG) const {
8343 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows()) &&
8344 "This should be used only on Windows targets");
8345 assert(!Subtarget->isTargetEnvMacho());
8346 DebugLoc dl = Op.getDebugLoc();
8349 SDValue Chain = Op.getOperand(0);
8350 SDValue Size = Op.getOperand(1);
8351 // FIXME: Ensure alignment here
8355 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
8356 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
8358 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
8359 Flag = Chain.getValue(1);
8361 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
8363 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
8364 Flag = Chain.getValue(1);
8366 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
8368 SDValue Ops1[2] = { Chain.getValue(0), Chain };
8369 return DAG.getMergeValues(Ops1, 2, dl);
8372 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
8373 MachineFunction &MF = DAG.getMachineFunction();
8374 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
8376 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8377 DebugLoc DL = Op.getDebugLoc();
8379 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
8380 // vastart just stores the address of the VarArgsFrameIndex slot into the
8381 // memory location argument.
8382 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8384 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
8385 MachinePointerInfo(SV), false, false, 0);
8389 // gp_offset (0 - 6 * 8)
8390 // fp_offset (48 - 48 + 8 * 16)
8391 // overflow_arg_area (point to parameters coming in memory).
8393 SmallVector<SDValue, 8> MemOps;
8394 SDValue FIN = Op.getOperand(1);
8396 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
8397 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
8399 FIN, MachinePointerInfo(SV), false, false, 0);
8400 MemOps.push_back(Store);
8403 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8404 FIN, DAG.getIntPtrConstant(4));
8405 Store = DAG.getStore(Op.getOperand(0), DL,
8406 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
8408 FIN, MachinePointerInfo(SV, 4), false, false, 0);
8409 MemOps.push_back(Store);
8411 // Store ptr to overflow_arg_area
8412 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8413 FIN, DAG.getIntPtrConstant(4));
8414 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8416 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
8417 MachinePointerInfo(SV, 8),
8419 MemOps.push_back(Store);
8421 // Store ptr to reg_save_area.
8422 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8423 FIN, DAG.getIntPtrConstant(8));
8424 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
8426 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
8427 MachinePointerInfo(SV, 16), false, false, 0);
8428 MemOps.push_back(Store);
8429 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
8430 &MemOps[0], MemOps.size());
8433 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
8434 assert(Subtarget->is64Bit() &&
8435 "LowerVAARG only handles 64-bit va_arg!");
8436 assert((Subtarget->isTargetLinux() ||
8437 Subtarget->isTargetDarwin()) &&
8438 "Unhandled target in LowerVAARG");
8439 assert(Op.getNode()->getNumOperands() == 4);
8440 SDValue Chain = Op.getOperand(0);
8441 SDValue SrcPtr = Op.getOperand(1);
8442 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8443 unsigned Align = Op.getConstantOperandVal(3);
8444 DebugLoc dl = Op.getDebugLoc();
8446 EVT ArgVT = Op.getNode()->getValueType(0);
8447 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
8448 uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
8451 // Decide which area this value should be read from.
8452 // TODO: Implement the AMD64 ABI in its entirety. This simple
8453 // selection mechanism works only for the basic types.
8454 if (ArgVT == MVT::f80) {
8455 llvm_unreachable("va_arg for f80 not yet implemented");
8456 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
8457 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
8458 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
8459 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
8461 llvm_unreachable("Unhandled argument type in LowerVAARG");
8465 // Sanity Check: Make sure using fp_offset makes sense.
8466 assert(!UseSoftFloat &&
8467 !(DAG.getMachineFunction()
8468 .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
8469 Subtarget->hasXMM());
8472 // Insert VAARG_64 node into the DAG
8473 // VAARG_64 returns two values: Variable Argument Address, Chain
8474 SmallVector<SDValue, 11> InstOps;
8475 InstOps.push_back(Chain);
8476 InstOps.push_back(SrcPtr);
8477 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
8478 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
8479 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
8480 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
8481 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
8482 VTs, &InstOps[0], InstOps.size(),
8484 MachinePointerInfo(SV),
8489 Chain = VAARG.getValue(1);
8491 // Load the next argument and return it
8492 return DAG.getLoad(ArgVT, dl,
8495 MachinePointerInfo(),
8499 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
8500 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
8501 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
8502 SDValue Chain = Op.getOperand(0);
8503 SDValue DstPtr = Op.getOperand(1);
8504 SDValue SrcPtr = Op.getOperand(2);
8505 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
8506 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8507 DebugLoc DL = Op.getDebugLoc();
8509 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
8510 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
8512 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
8516 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
8517 DebugLoc dl = Op.getDebugLoc();
8518 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8520 default: return SDValue(); // Don't custom lower most intrinsics.
8521 // Comparison intrinsics.
8522 case Intrinsic::x86_sse_comieq_ss:
8523 case Intrinsic::x86_sse_comilt_ss:
8524 case Intrinsic::x86_sse_comile_ss:
8525 case Intrinsic::x86_sse_comigt_ss:
8526 case Intrinsic::x86_sse_comige_ss:
8527 case Intrinsic::x86_sse_comineq_ss:
8528 case Intrinsic::x86_sse_ucomieq_ss:
8529 case Intrinsic::x86_sse_ucomilt_ss:
8530 case Intrinsic::x86_sse_ucomile_ss:
8531 case Intrinsic::x86_sse_ucomigt_ss:
8532 case Intrinsic::x86_sse_ucomige_ss:
8533 case Intrinsic::x86_sse_ucomineq_ss:
8534 case Intrinsic::x86_sse2_comieq_sd:
8535 case Intrinsic::x86_sse2_comilt_sd:
8536 case Intrinsic::x86_sse2_comile_sd:
8537 case Intrinsic::x86_sse2_comigt_sd:
8538 case Intrinsic::x86_sse2_comige_sd:
8539 case Intrinsic::x86_sse2_comineq_sd:
8540 case Intrinsic::x86_sse2_ucomieq_sd:
8541 case Intrinsic::x86_sse2_ucomilt_sd:
8542 case Intrinsic::x86_sse2_ucomile_sd:
8543 case Intrinsic::x86_sse2_ucomigt_sd:
8544 case Intrinsic::x86_sse2_ucomige_sd:
8545 case Intrinsic::x86_sse2_ucomineq_sd: {
8547 ISD::CondCode CC = ISD::SETCC_INVALID;
8550 case Intrinsic::x86_sse_comieq_ss:
8551 case Intrinsic::x86_sse2_comieq_sd:
8555 case Intrinsic::x86_sse_comilt_ss:
8556 case Intrinsic::x86_sse2_comilt_sd:
8560 case Intrinsic::x86_sse_comile_ss:
8561 case Intrinsic::x86_sse2_comile_sd:
8565 case Intrinsic::x86_sse_comigt_ss:
8566 case Intrinsic::x86_sse2_comigt_sd:
8570 case Intrinsic::x86_sse_comige_ss:
8571 case Intrinsic::x86_sse2_comige_sd:
8575 case Intrinsic::x86_sse_comineq_ss:
8576 case Intrinsic::x86_sse2_comineq_sd:
8580 case Intrinsic::x86_sse_ucomieq_ss:
8581 case Intrinsic::x86_sse2_ucomieq_sd:
8582 Opc = X86ISD::UCOMI;
8585 case Intrinsic::x86_sse_ucomilt_ss:
8586 case Intrinsic::x86_sse2_ucomilt_sd:
8587 Opc = X86ISD::UCOMI;
8590 case Intrinsic::x86_sse_ucomile_ss:
8591 case Intrinsic::x86_sse2_ucomile_sd:
8592 Opc = X86ISD::UCOMI;
8595 case Intrinsic::x86_sse_ucomigt_ss:
8596 case Intrinsic::x86_sse2_ucomigt_sd:
8597 Opc = X86ISD::UCOMI;
8600 case Intrinsic::x86_sse_ucomige_ss:
8601 case Intrinsic::x86_sse2_ucomige_sd:
8602 Opc = X86ISD::UCOMI;
8605 case Intrinsic::x86_sse_ucomineq_ss:
8606 case Intrinsic::x86_sse2_ucomineq_sd:
8607 Opc = X86ISD::UCOMI;
8612 SDValue LHS = Op.getOperand(1);
8613 SDValue RHS = Op.getOperand(2);
8614 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
8615 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
8616 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
8617 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8618 DAG.getConstant(X86CC, MVT::i8), Cond);
8619 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8621 // ptest and testp intrinsics. The intrinsic these come from are designed to
8622 // return an integer value, not just an instruction so lower it to the ptest
8623 // or testp pattern and a setcc for the result.
8624 case Intrinsic::x86_sse41_ptestz:
8625 case Intrinsic::x86_sse41_ptestc:
8626 case Intrinsic::x86_sse41_ptestnzc:
8627 case Intrinsic::x86_avx_ptestz_256:
8628 case Intrinsic::x86_avx_ptestc_256:
8629 case Intrinsic::x86_avx_ptestnzc_256:
8630 case Intrinsic::x86_avx_vtestz_ps:
8631 case Intrinsic::x86_avx_vtestc_ps:
8632 case Intrinsic::x86_avx_vtestnzc_ps:
8633 case Intrinsic::x86_avx_vtestz_pd:
8634 case Intrinsic::x86_avx_vtestc_pd:
8635 case Intrinsic::x86_avx_vtestnzc_pd:
8636 case Intrinsic::x86_avx_vtestz_ps_256:
8637 case Intrinsic::x86_avx_vtestc_ps_256:
8638 case Intrinsic::x86_avx_vtestnzc_ps_256:
8639 case Intrinsic::x86_avx_vtestz_pd_256:
8640 case Intrinsic::x86_avx_vtestc_pd_256:
8641 case Intrinsic::x86_avx_vtestnzc_pd_256: {
8642 bool IsTestPacked = false;
8645 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
8646 case Intrinsic::x86_avx_vtestz_ps:
8647 case Intrinsic::x86_avx_vtestz_pd:
8648 case Intrinsic::x86_avx_vtestz_ps_256:
8649 case Intrinsic::x86_avx_vtestz_pd_256:
8650 IsTestPacked = true; // Fallthrough
8651 case Intrinsic::x86_sse41_ptestz:
8652 case Intrinsic::x86_avx_ptestz_256:
8654 X86CC = X86::COND_E;
8656 case Intrinsic::x86_avx_vtestc_ps:
8657 case Intrinsic::x86_avx_vtestc_pd:
8658 case Intrinsic::x86_avx_vtestc_ps_256:
8659 case Intrinsic::x86_avx_vtestc_pd_256:
8660 IsTestPacked = true; // Fallthrough
8661 case Intrinsic::x86_sse41_ptestc:
8662 case Intrinsic::x86_avx_ptestc_256:
8664 X86CC = X86::COND_B;
8666 case Intrinsic::x86_avx_vtestnzc_ps:
8667 case Intrinsic::x86_avx_vtestnzc_pd:
8668 case Intrinsic::x86_avx_vtestnzc_ps_256:
8669 case Intrinsic::x86_avx_vtestnzc_pd_256:
8670 IsTestPacked = true; // Fallthrough
8671 case Intrinsic::x86_sse41_ptestnzc:
8672 case Intrinsic::x86_avx_ptestnzc_256:
8674 X86CC = X86::COND_A;
8678 SDValue LHS = Op.getOperand(1);
8679 SDValue RHS = Op.getOperand(2);
8680 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
8681 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
8682 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
8683 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
8684 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8687 // Fix vector shift instructions where the last operand is a non-immediate
8689 case Intrinsic::x86_sse2_pslli_w:
8690 case Intrinsic::x86_sse2_pslli_d:
8691 case Intrinsic::x86_sse2_pslli_q:
8692 case Intrinsic::x86_sse2_psrli_w:
8693 case Intrinsic::x86_sse2_psrli_d:
8694 case Intrinsic::x86_sse2_psrli_q:
8695 case Intrinsic::x86_sse2_psrai_w:
8696 case Intrinsic::x86_sse2_psrai_d:
8697 case Intrinsic::x86_mmx_pslli_w:
8698 case Intrinsic::x86_mmx_pslli_d:
8699 case Intrinsic::x86_mmx_pslli_q:
8700 case Intrinsic::x86_mmx_psrli_w:
8701 case Intrinsic::x86_mmx_psrli_d:
8702 case Intrinsic::x86_mmx_psrli_q:
8703 case Intrinsic::x86_mmx_psrai_w:
8704 case Intrinsic::x86_mmx_psrai_d: {
8705 SDValue ShAmt = Op.getOperand(2);
8706 if (isa<ConstantSDNode>(ShAmt))
8709 unsigned NewIntNo = 0;
8710 EVT ShAmtVT = MVT::v4i32;
8712 case Intrinsic::x86_sse2_pslli_w:
8713 NewIntNo = Intrinsic::x86_sse2_psll_w;
8715 case Intrinsic::x86_sse2_pslli_d:
8716 NewIntNo = Intrinsic::x86_sse2_psll_d;
8718 case Intrinsic::x86_sse2_pslli_q:
8719 NewIntNo = Intrinsic::x86_sse2_psll_q;
8721 case Intrinsic::x86_sse2_psrli_w:
8722 NewIntNo = Intrinsic::x86_sse2_psrl_w;
8724 case Intrinsic::x86_sse2_psrli_d:
8725 NewIntNo = Intrinsic::x86_sse2_psrl_d;
8727 case Intrinsic::x86_sse2_psrli_q:
8728 NewIntNo = Intrinsic::x86_sse2_psrl_q;
8730 case Intrinsic::x86_sse2_psrai_w:
8731 NewIntNo = Intrinsic::x86_sse2_psra_w;
8733 case Intrinsic::x86_sse2_psrai_d:
8734 NewIntNo = Intrinsic::x86_sse2_psra_d;
8737 ShAmtVT = MVT::v2i32;
8739 case Intrinsic::x86_mmx_pslli_w:
8740 NewIntNo = Intrinsic::x86_mmx_psll_w;
8742 case Intrinsic::x86_mmx_pslli_d:
8743 NewIntNo = Intrinsic::x86_mmx_psll_d;
8745 case Intrinsic::x86_mmx_pslli_q:
8746 NewIntNo = Intrinsic::x86_mmx_psll_q;
8748 case Intrinsic::x86_mmx_psrli_w:
8749 NewIntNo = Intrinsic::x86_mmx_psrl_w;
8751 case Intrinsic::x86_mmx_psrli_d:
8752 NewIntNo = Intrinsic::x86_mmx_psrl_d;
8754 case Intrinsic::x86_mmx_psrli_q:
8755 NewIntNo = Intrinsic::x86_mmx_psrl_q;
8757 case Intrinsic::x86_mmx_psrai_w:
8758 NewIntNo = Intrinsic::x86_mmx_psra_w;
8760 case Intrinsic::x86_mmx_psrai_d:
8761 NewIntNo = Intrinsic::x86_mmx_psra_d;
8763 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
8769 // The vector shift intrinsics with scalars uses 32b shift amounts but
8770 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
8774 ShOps[1] = DAG.getConstant(0, MVT::i32);
8775 if (ShAmtVT == MVT::v4i32) {
8776 ShOps[2] = DAG.getUNDEF(MVT::i32);
8777 ShOps[3] = DAG.getUNDEF(MVT::i32);
8778 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
8780 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
8781 // FIXME this must be lowered to get rid of the invalid type.
8784 EVT VT = Op.getValueType();
8785 ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
8786 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8787 DAG.getConstant(NewIntNo, MVT::i32),
8788 Op.getOperand(1), ShAmt);
8793 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
8794 SelectionDAG &DAG) const {
8795 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8796 MFI->setReturnAddressIsTaken(true);
8798 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8799 DebugLoc dl = Op.getDebugLoc();
8802 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
8804 DAG.getConstant(TD->getPointerSize(),
8805 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8806 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8807 DAG.getNode(ISD::ADD, dl, getPointerTy(),
8809 MachinePointerInfo(), false, false, 0);
8812 // Just load the return address.
8813 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
8814 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8815 RetAddrFI, MachinePointerInfo(), false, false, 0);
8818 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
8819 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8820 MFI->setFrameAddressIsTaken(true);
8822 EVT VT = Op.getValueType();
8823 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
8824 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8825 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
8826 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
8828 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
8829 MachinePointerInfo(),
8834 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
8835 SelectionDAG &DAG) const {
8836 return DAG.getIntPtrConstant(2*TD->getPointerSize());
8839 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
8840 MachineFunction &MF = DAG.getMachineFunction();
8841 SDValue Chain = Op.getOperand(0);
8842 SDValue Offset = Op.getOperand(1);
8843 SDValue Handler = Op.getOperand(2);
8844 DebugLoc dl = Op.getDebugLoc();
8846 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
8847 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
8849 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
8851 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
8852 DAG.getIntPtrConstant(TD->getPointerSize()));
8853 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
8854 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
8856 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
8857 MF.getRegInfo().addLiveOut(StoreAddrReg);
8859 return DAG.getNode(X86ISD::EH_RETURN, dl,
8861 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
8864 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
8865 SelectionDAG &DAG) const {
8866 SDValue Root = Op.getOperand(0);
8867 SDValue Trmp = Op.getOperand(1); // trampoline
8868 SDValue FPtr = Op.getOperand(2); // nested function
8869 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
8870 DebugLoc dl = Op.getDebugLoc();
8872 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8874 if (Subtarget->is64Bit()) {
8875 SDValue OutChains[6];
8877 // Large code-model.
8878 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
8879 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
8881 const unsigned char N86R10 = X86_MC::getX86RegNum(X86::R10);
8882 const unsigned char N86R11 = X86_MC::getX86RegNum(X86::R11);
8884 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
8886 // Load the pointer to the nested function into R11.
8887 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
8888 SDValue Addr = Trmp;
8889 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8890 Addr, MachinePointerInfo(TrmpAddr),
8893 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8894 DAG.getConstant(2, MVT::i64));
8895 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
8896 MachinePointerInfo(TrmpAddr, 2),
8899 // Load the 'nest' parameter value into R10.
8900 // R10 is specified in X86CallingConv.td
8901 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
8902 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8903 DAG.getConstant(10, MVT::i64));
8904 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8905 Addr, MachinePointerInfo(TrmpAddr, 10),
8908 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8909 DAG.getConstant(12, MVT::i64));
8910 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
8911 MachinePointerInfo(TrmpAddr, 12),
8914 // Jump to the nested function.
8915 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
8916 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8917 DAG.getConstant(20, MVT::i64));
8918 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
8919 Addr, MachinePointerInfo(TrmpAddr, 20),
8922 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
8923 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
8924 DAG.getConstant(22, MVT::i64));
8925 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
8926 MachinePointerInfo(TrmpAddr, 22),
8930 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
8931 return DAG.getMergeValues(Ops, 2, dl);
8933 const Function *Func =
8934 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
8935 CallingConv::ID CC = Func->getCallingConv();
8940 llvm_unreachable("Unsupported calling convention");
8941 case CallingConv::C:
8942 case CallingConv::X86_StdCall: {
8943 // Pass 'nest' parameter in ECX.
8944 // Must be kept in sync with X86CallingConv.td
8947 // Check that ECX wasn't needed by an 'inreg' parameter.
8948 FunctionType *FTy = Func->getFunctionType();
8949 const AttrListPtr &Attrs = Func->getAttributes();
8951 if (!Attrs.isEmpty() && !Func->isVarArg()) {
8952 unsigned InRegCount = 0;
8955 for (FunctionType::param_iterator I = FTy->param_begin(),
8956 E = FTy->param_end(); I != E; ++I, ++Idx)
8957 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
8958 // FIXME: should only count parameters that are lowered to integers.
8959 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
8961 if (InRegCount > 2) {
8962 report_fatal_error("Nest register in use - reduce number of inreg"
8968 case CallingConv::X86_FastCall:
8969 case CallingConv::X86_ThisCall:
8970 case CallingConv::Fast:
8971 // Pass 'nest' parameter in EAX.
8972 // Must be kept in sync with X86CallingConv.td
8977 SDValue OutChains[4];
8980 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8981 DAG.getConstant(10, MVT::i32));
8982 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
8984 // This is storing the opcode for MOV32ri.
8985 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
8986 const unsigned char N86Reg = X86_MC::getX86RegNum(NestReg);
8987 OutChains[0] = DAG.getStore(Root, dl,
8988 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
8989 Trmp, MachinePointerInfo(TrmpAddr),
8992 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
8993 DAG.getConstant(1, MVT::i32));
8994 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
8995 MachinePointerInfo(TrmpAddr, 1),
8998 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
8999 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9000 DAG.getConstant(5, MVT::i32));
9001 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
9002 MachinePointerInfo(TrmpAddr, 5),
9005 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9006 DAG.getConstant(6, MVT::i32));
9007 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
9008 MachinePointerInfo(TrmpAddr, 6),
9012 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
9013 return DAG.getMergeValues(Ops, 2, dl);
9017 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
9018 SelectionDAG &DAG) const {
9020 The rounding mode is in bits 11:10 of FPSR, and has the following
9027 FLT_ROUNDS, on the other hand, expects the following:
9034 To perform the conversion, we do:
9035 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
9038 MachineFunction &MF = DAG.getMachineFunction();
9039 const TargetMachine &TM = MF.getTarget();
9040 const TargetFrameLowering &TFI = *TM.getFrameLowering();
9041 unsigned StackAlignment = TFI.getStackAlignment();
9042 EVT VT = Op.getValueType();
9043 DebugLoc DL = Op.getDebugLoc();
9045 // Save FP Control Word to stack slot
9046 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
9047 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
9050 MachineMemOperand *MMO =
9051 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
9052 MachineMemOperand::MOStore, 2, 2);
9054 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
9055 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
9056 DAG.getVTList(MVT::Other),
9057 Ops, 2, MVT::i16, MMO);
9059 // Load FP Control Word from stack slot
9060 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
9061 MachinePointerInfo(), false, false, 0);
9063 // Transform as necessary
9065 DAG.getNode(ISD::SRL, DL, MVT::i16,
9066 DAG.getNode(ISD::AND, DL, MVT::i16,
9067 CWD, DAG.getConstant(0x800, MVT::i16)),
9068 DAG.getConstant(11, MVT::i8));
9070 DAG.getNode(ISD::SRL, DL, MVT::i16,
9071 DAG.getNode(ISD::AND, DL, MVT::i16,
9072 CWD, DAG.getConstant(0x400, MVT::i16)),
9073 DAG.getConstant(9, MVT::i8));
9076 DAG.getNode(ISD::AND, DL, MVT::i16,
9077 DAG.getNode(ISD::ADD, DL, MVT::i16,
9078 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
9079 DAG.getConstant(1, MVT::i16)),
9080 DAG.getConstant(3, MVT::i16));
9083 return DAG.getNode((VT.getSizeInBits() < 16 ?
9084 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
9087 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
9088 EVT VT = Op.getValueType();
9090 unsigned NumBits = VT.getSizeInBits();
9091 DebugLoc dl = Op.getDebugLoc();
9093 Op = Op.getOperand(0);
9094 if (VT == MVT::i8) {
9095 // Zero extend to i32 since there is not an i8 bsr.
9097 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
9100 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
9101 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
9102 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
9104 // If src is zero (i.e. bsr sets ZF), returns NumBits.
9107 DAG.getConstant(NumBits+NumBits-1, OpVT),
9108 DAG.getConstant(X86::COND_E, MVT::i8),
9111 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
9113 // Finally xor with NumBits-1.
9114 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
9117 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
9121 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
9122 EVT VT = Op.getValueType();
9124 unsigned NumBits = VT.getSizeInBits();
9125 DebugLoc dl = Op.getDebugLoc();
9127 Op = Op.getOperand(0);
9128 if (VT == MVT::i8) {
9130 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
9133 // Issue a bsf (scan bits forward) which also sets EFLAGS.
9134 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
9135 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
9137 // If src is zero (i.e. bsf sets ZF), returns NumBits.
9140 DAG.getConstant(NumBits, OpVT),
9141 DAG.getConstant(X86::COND_E, MVT::i8),
9144 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
9147 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
9151 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
9152 EVT VT = Op.getValueType();
9153 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
9154 DebugLoc dl = Op.getDebugLoc();
9156 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
9157 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
9158 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
9159 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
9160 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
9162 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
9163 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
9164 // return AloBlo + AloBhi + AhiBlo;
9166 SDValue A = Op.getOperand(0);
9167 SDValue B = Op.getOperand(1);
9169 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9170 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9171 A, DAG.getConstant(32, MVT::i32));
9172 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9173 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9174 B, DAG.getConstant(32, MVT::i32));
9175 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9176 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9178 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9179 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9181 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9182 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9184 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9185 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9186 AloBhi, DAG.getConstant(32, MVT::i32));
9187 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9188 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9189 AhiBlo, DAG.getConstant(32, MVT::i32));
9190 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
9191 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
9195 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
9197 EVT VT = Op.getValueType();
9198 DebugLoc dl = Op.getDebugLoc();
9199 SDValue R = Op.getOperand(0);
9200 SDValue Amt = Op.getOperand(1);
9202 LLVMContext *Context = DAG.getContext();
9205 if (!Subtarget->hasSSE2()) return SDValue();
9207 // Optimize shl/srl/sra with constant shift amount.
9208 if (isSplatVector(Amt.getNode())) {
9209 SDValue SclrAmt = Amt->getOperand(0);
9210 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
9211 uint64_t ShiftAmt = C->getZExtValue();
9213 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SHL)
9214 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9215 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9216 R, DAG.getConstant(ShiftAmt, MVT::i32));
9218 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SHL)
9219 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9220 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9221 R, DAG.getConstant(ShiftAmt, MVT::i32));
9223 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SHL)
9224 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9225 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9226 R, DAG.getConstant(ShiftAmt, MVT::i32));
9228 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SRL)
9229 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9230 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9231 R, DAG.getConstant(ShiftAmt, MVT::i32));
9233 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRL)
9234 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9235 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9236 R, DAG.getConstant(ShiftAmt, MVT::i32));
9238 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRL)
9239 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9240 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9241 R, DAG.getConstant(ShiftAmt, MVT::i32));
9243 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRA)
9244 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9245 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9246 R, DAG.getConstant(ShiftAmt, MVT::i32));
9248 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRA)
9249 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9250 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9251 R, DAG.getConstant(ShiftAmt, MVT::i32));
9255 // Lower SHL with variable shift amount.
9256 // Cannot lower SHL without SSE2 or later.
9257 if (!Subtarget->hasSSE2()) return SDValue();
9259 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
9260 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9261 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9262 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
9264 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
9266 std::vector<Constant*> CV(4, CI);
9267 Constant *C = ConstantVector::get(CV);
9268 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9269 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9270 MachinePointerInfo::getConstantPool(),
9273 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
9274 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
9275 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
9276 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
9278 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
9280 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9281 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9282 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
9284 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
9285 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
9287 std::vector<Constant*> CVM1(16, CM1);
9288 std::vector<Constant*> CVM2(16, CM2);
9289 Constant *C = ConstantVector::get(CVM1);
9290 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9291 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9292 MachinePointerInfo::getConstantPool(),
9295 // r = pblendv(r, psllw(r & (char16)15, 4), a);
9296 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9297 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9298 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9299 DAG.getConstant(4, MVT::i32));
9300 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9302 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9304 C = ConstantVector::get(CVM2);
9305 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9306 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9307 MachinePointerInfo::getConstantPool(),
9310 // r = pblendv(r, psllw(r & (char16)63, 2), a);
9311 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9312 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9313 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9314 DAG.getConstant(2, MVT::i32));
9315 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9317 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9319 // return pblendv(r, r+r, a);
9320 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
9321 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
9327 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
9328 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
9329 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
9330 // looks for this combo and may remove the "setcc" instruction if the "setcc"
9331 // has only one use.
9332 SDNode *N = Op.getNode();
9333 SDValue LHS = N->getOperand(0);
9334 SDValue RHS = N->getOperand(1);
9335 unsigned BaseOp = 0;
9337 DebugLoc DL = Op.getDebugLoc();
9338 switch (Op.getOpcode()) {
9339 default: llvm_unreachable("Unknown ovf instruction!");
9341 // A subtract of one will be selected as a INC. Note that INC doesn't
9342 // set CF, so we can't do this for UADDO.
9343 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9345 BaseOp = X86ISD::INC;
9349 BaseOp = X86ISD::ADD;
9353 BaseOp = X86ISD::ADD;
9357 // A subtract of one will be selected as a DEC. Note that DEC doesn't
9358 // set CF, so we can't do this for USUBO.
9359 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9361 BaseOp = X86ISD::DEC;
9365 BaseOp = X86ISD::SUB;
9369 BaseOp = X86ISD::SUB;
9373 BaseOp = X86ISD::SMUL;
9376 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
9377 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
9379 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
9382 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9383 DAG.getConstant(X86::COND_O, MVT::i32),
9384 SDValue(Sum.getNode(), 2));
9386 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
9390 // Also sets EFLAGS.
9391 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
9392 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
9395 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
9396 DAG.getConstant(Cond, MVT::i32),
9397 SDValue(Sum.getNode(), 1));
9399 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
9402 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const{
9403 DebugLoc dl = Op.getDebugLoc();
9404 SDNode* Node = Op.getNode();
9405 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
9406 EVT VT = Node->getValueType(0);
9408 if (Subtarget->hasSSE2() && VT.isVector()) {
9409 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
9410 ExtraVT.getScalarType().getSizeInBits();
9411 SDValue ShAmt = DAG.getConstant(BitsDiff, MVT::i32);
9413 unsigned SHLIntrinsicsID = 0;
9414 unsigned SRAIntrinsicsID = 0;
9415 switch (VT.getSimpleVT().SimpleTy) {
9419 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_q;
9420 SRAIntrinsicsID = 0;
9424 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_d;
9425 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_d;
9429 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_w;
9430 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_w;
9435 SDValue Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9436 DAG.getConstant(SHLIntrinsicsID, MVT::i32),
9437 Node->getOperand(0), ShAmt);
9439 // In case of 1 bit sext, no need to shr
9440 if (ExtraVT.getScalarType().getSizeInBits() == 1) return Tmp1;
9442 if (SRAIntrinsicsID) {
9443 Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9444 DAG.getConstant(SRAIntrinsicsID, MVT::i32),
9454 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
9455 DebugLoc dl = Op.getDebugLoc();
9457 // Go ahead and emit the fence on x86-64 even if we asked for no-sse2.
9458 // There isn't any reason to disable it if the target processor supports it.
9459 if (!Subtarget->hasSSE2() && !Subtarget->is64Bit()) {
9460 SDValue Chain = Op.getOperand(0);
9461 SDValue Zero = DAG.getConstant(0, MVT::i32);
9463 DAG.getRegister(X86::ESP, MVT::i32), // Base
9464 DAG.getTargetConstant(1, MVT::i8), // Scale
9465 DAG.getRegister(0, MVT::i32), // Index
9466 DAG.getTargetConstant(0, MVT::i32), // Disp
9467 DAG.getRegister(0, MVT::i32), // Segment.
9472 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
9473 array_lengthof(Ops));
9474 return SDValue(Res, 0);
9477 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
9479 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
9481 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
9482 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
9483 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
9484 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
9486 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
9487 if (!Op1 && !Op2 && !Op3 && Op4)
9488 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
9490 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
9491 if (Op1 && !Op2 && !Op3 && !Op4)
9492 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
9494 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
9496 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
9499 SDValue X86TargetLowering::LowerATOMIC_FENCE(SDValue Op,
9500 SelectionDAG &DAG) const {
9501 DebugLoc dl = Op.getDebugLoc();
9502 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
9503 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
9504 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
9505 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
9507 // The only fence that needs an instruction is a sequentially-consistent
9508 // cross-thread fence.
9509 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
9510 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
9511 // no-sse2). There isn't any reason to disable it if the target processor
9513 if (Subtarget->hasSSE2() || Subtarget->is64Bit())
9514 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
9516 SDValue Chain = Op.getOperand(0);
9517 SDValue Zero = DAG.getConstant(0, MVT::i32);
9519 DAG.getRegister(X86::ESP, MVT::i32), // Base
9520 DAG.getTargetConstant(1, MVT::i8), // Scale
9521 DAG.getRegister(0, MVT::i32), // Index
9522 DAG.getTargetConstant(0, MVT::i32), // Disp
9523 DAG.getRegister(0, MVT::i32), // Segment.
9528 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
9529 array_lengthof(Ops));
9530 return SDValue(Res, 0);
9533 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
9534 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
9538 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
9539 EVT T = Op.getValueType();
9540 DebugLoc DL = Op.getDebugLoc();
9543 switch(T.getSimpleVT().SimpleTy) {
9545 assert(false && "Invalid value type!");
9546 case MVT::i8: Reg = X86::AL; size = 1; break;
9547 case MVT::i16: Reg = X86::AX; size = 2; break;
9548 case MVT::i32: Reg = X86::EAX; size = 4; break;
9550 assert(Subtarget->is64Bit() && "Node not type legal!");
9551 Reg = X86::RAX; size = 8;
9554 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
9555 Op.getOperand(2), SDValue());
9556 SDValue Ops[] = { cpIn.getValue(0),
9559 DAG.getTargetConstant(size, MVT::i8),
9561 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9562 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
9563 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
9566 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
9570 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
9571 SelectionDAG &DAG) const {
9572 assert(Subtarget->is64Bit() && "Result not type legalized?");
9573 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9574 SDValue TheChain = Op.getOperand(0);
9575 DebugLoc dl = Op.getDebugLoc();
9576 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9577 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
9578 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
9580 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
9581 DAG.getConstant(32, MVT::i8));
9583 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
9586 return DAG.getMergeValues(Ops, 2, dl);
9589 SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
9590 SelectionDAG &DAG) const {
9591 EVT SrcVT = Op.getOperand(0).getValueType();
9592 EVT DstVT = Op.getValueType();
9593 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
9594 Subtarget->hasMMX() && "Unexpected custom BITCAST");
9595 assert((DstVT == MVT::i64 ||
9596 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
9597 "Unexpected custom BITCAST");
9598 // i64 <=> MMX conversions are Legal.
9599 if (SrcVT==MVT::i64 && DstVT.isVector())
9601 if (DstVT==MVT::i64 && SrcVT.isVector())
9603 // MMX <=> MMX conversions are Legal.
9604 if (SrcVT.isVector() && DstVT.isVector())
9606 // All other conversions need to be expanded.
9610 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
9611 SDNode *Node = Op.getNode();
9612 DebugLoc dl = Node->getDebugLoc();
9613 EVT T = Node->getValueType(0);
9614 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
9615 DAG.getConstant(0, T), Node->getOperand(2));
9616 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
9617 cast<AtomicSDNode>(Node)->getMemoryVT(),
9618 Node->getOperand(0),
9619 Node->getOperand(1), negOp,
9620 cast<AtomicSDNode>(Node)->getSrcValue(),
9621 cast<AtomicSDNode>(Node)->getAlignment(),
9622 cast<AtomicSDNode>(Node)->getOrdering(),
9623 cast<AtomicSDNode>(Node)->getSynchScope());
9626 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
9627 EVT VT = Op.getNode()->getValueType(0);
9629 // Let legalize expand this if it isn't a legal type yet.
9630 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
9633 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
9636 bool ExtraOp = false;
9637 switch (Op.getOpcode()) {
9638 default: assert(0 && "Invalid code");
9639 case ISD::ADDC: Opc = X86ISD::ADD; break;
9640 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
9641 case ISD::SUBC: Opc = X86ISD::SUB; break;
9642 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
9646 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9648 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9649 Op.getOperand(1), Op.getOperand(2));
9652 /// LowerOperation - Provide custom lowering hooks for some operations.
9654 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
9655 switch (Op.getOpcode()) {
9656 default: llvm_unreachable("Should not custom lower this!");
9657 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
9658 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
9659 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op,DAG);
9660 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
9661 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
9662 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
9663 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
9664 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
9665 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
9666 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
9667 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
9668 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, DAG);
9669 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
9670 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
9671 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
9672 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
9673 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
9674 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
9675 case ISD::SHL_PARTS:
9676 case ISD::SRA_PARTS:
9677 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
9678 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
9679 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
9680 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
9681 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
9682 case ISD::FABS: return LowerFABS(Op, DAG);
9683 case ISD::FNEG: return LowerFNEG(Op, DAG);
9684 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
9685 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
9686 case ISD::SETCC: return LowerSETCC(Op, DAG);
9687 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
9688 case ISD::SELECT: return LowerSELECT(Op, DAG);
9689 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
9690 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
9691 case ISD::VASTART: return LowerVASTART(Op, DAG);
9692 case ISD::VAARG: return LowerVAARG(Op, DAG);
9693 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
9694 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
9695 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
9696 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
9697 case ISD::FRAME_TO_ARGS_OFFSET:
9698 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
9699 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
9700 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
9701 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
9702 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
9703 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
9704 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
9705 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
9708 case ISD::SHL: return LowerShift(Op, DAG);
9714 case ISD::UMULO: return LowerXALUO(Op, DAG);
9715 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
9716 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
9720 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
9724 void X86TargetLowering::
9725 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
9726 SelectionDAG &DAG, unsigned NewOp) const {
9727 EVT T = Node->getValueType(0);
9728 DebugLoc dl = Node->getDebugLoc();
9729 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
9731 SDValue Chain = Node->getOperand(0);
9732 SDValue In1 = Node->getOperand(1);
9733 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9734 Node->getOperand(2), DAG.getIntPtrConstant(0));
9735 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9736 Node->getOperand(2), DAG.getIntPtrConstant(1));
9737 SDValue Ops[] = { Chain, In1, In2L, In2H };
9738 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
9740 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
9741 cast<MemSDNode>(Node)->getMemOperand());
9742 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
9743 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9744 Results.push_back(Result.getValue(2));
9747 /// ReplaceNodeResults - Replace a node with an illegal result type
9748 /// with a new node built out of custom code.
9749 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
9750 SmallVectorImpl<SDValue>&Results,
9751 SelectionDAG &DAG) const {
9752 DebugLoc dl = N->getDebugLoc();
9753 switch (N->getOpcode()) {
9755 assert(false && "Do not know how to custom type legalize this operation!");
9757 case ISD::SIGN_EXTEND_INREG:
9762 // We don't want to expand or promote these.
9764 case ISD::FP_TO_SINT: {
9765 std::pair<SDValue,SDValue> Vals =
9766 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
9767 SDValue FIST = Vals.first, StackSlot = Vals.second;
9768 if (FIST.getNode() != 0) {
9769 EVT VT = N->getValueType(0);
9770 // Return a load from the stack slot.
9771 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
9772 MachinePointerInfo(), false, false, 0));
9776 case ISD::READCYCLECOUNTER: {
9777 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9778 SDValue TheChain = N->getOperand(0);
9779 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9780 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
9782 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
9784 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
9785 SDValue Ops[] = { eax, edx };
9786 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
9787 Results.push_back(edx.getValue(1));
9790 case ISD::ATOMIC_CMP_SWAP: {
9791 EVT T = N->getValueType(0);
9792 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
9793 SDValue cpInL, cpInH;
9794 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9795 DAG.getConstant(0, MVT::i32));
9796 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
9797 DAG.getConstant(1, MVT::i32));
9798 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
9799 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
9801 SDValue swapInL, swapInH;
9802 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9803 DAG.getConstant(0, MVT::i32));
9804 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
9805 DAG.getConstant(1, MVT::i32));
9806 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
9808 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
9809 swapInL.getValue(1));
9810 SDValue Ops[] = { swapInH.getValue(0),
9812 swapInH.getValue(1) };
9813 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9814 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
9815 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
9817 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
9818 MVT::i32, Result.getValue(1));
9819 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
9820 MVT::i32, cpOutL.getValue(2));
9821 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
9822 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9823 Results.push_back(cpOutH.getValue(1));
9826 case ISD::ATOMIC_LOAD_ADD:
9827 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
9829 case ISD::ATOMIC_LOAD_AND:
9830 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
9832 case ISD::ATOMIC_LOAD_NAND:
9833 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
9835 case ISD::ATOMIC_LOAD_OR:
9836 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
9838 case ISD::ATOMIC_LOAD_SUB:
9839 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
9841 case ISD::ATOMIC_LOAD_XOR:
9842 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
9844 case ISD::ATOMIC_SWAP:
9845 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
9850 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
9852 default: return NULL;
9853 case X86ISD::BSF: return "X86ISD::BSF";
9854 case X86ISD::BSR: return "X86ISD::BSR";
9855 case X86ISD::SHLD: return "X86ISD::SHLD";
9856 case X86ISD::SHRD: return "X86ISD::SHRD";
9857 case X86ISD::FAND: return "X86ISD::FAND";
9858 case X86ISD::FOR: return "X86ISD::FOR";
9859 case X86ISD::FXOR: return "X86ISD::FXOR";
9860 case X86ISD::FSRL: return "X86ISD::FSRL";
9861 case X86ISD::FILD: return "X86ISD::FILD";
9862 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
9863 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
9864 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
9865 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
9866 case X86ISD::FLD: return "X86ISD::FLD";
9867 case X86ISD::FST: return "X86ISD::FST";
9868 case X86ISD::CALL: return "X86ISD::CALL";
9869 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
9870 case X86ISD::BT: return "X86ISD::BT";
9871 case X86ISD::CMP: return "X86ISD::CMP";
9872 case X86ISD::COMI: return "X86ISD::COMI";
9873 case X86ISD::UCOMI: return "X86ISD::UCOMI";
9874 case X86ISD::SETCC: return "X86ISD::SETCC";
9875 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
9876 case X86ISD::FSETCCsd: return "X86ISD::FSETCCsd";
9877 case X86ISD::FSETCCss: return "X86ISD::FSETCCss";
9878 case X86ISD::CMOV: return "X86ISD::CMOV";
9879 case X86ISD::BRCOND: return "X86ISD::BRCOND";
9880 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
9881 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
9882 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
9883 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
9884 case X86ISD::Wrapper: return "X86ISD::Wrapper";
9885 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
9886 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
9887 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
9888 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
9889 case X86ISD::PINSRB: return "X86ISD::PINSRB";
9890 case X86ISD::PINSRW: return "X86ISD::PINSRW";
9891 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
9892 case X86ISD::ANDNP: return "X86ISD::ANDNP";
9893 case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
9894 case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
9895 case X86ISD::PSIGND: return "X86ISD::PSIGND";
9896 case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
9897 case X86ISD::FMAX: return "X86ISD::FMAX";
9898 case X86ISD::FMIN: return "X86ISD::FMIN";
9899 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
9900 case X86ISD::FRCP: return "X86ISD::FRCP";
9901 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
9902 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
9903 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
9904 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
9905 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
9906 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
9907 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
9908 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
9909 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
9910 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
9911 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
9912 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
9913 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
9914 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
9915 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
9916 case X86ISD::VSHL: return "X86ISD::VSHL";
9917 case X86ISD::VSRL: return "X86ISD::VSRL";
9918 case X86ISD::CMPPD: return "X86ISD::CMPPD";
9919 case X86ISD::CMPPS: return "X86ISD::CMPPS";
9920 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
9921 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
9922 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
9923 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
9924 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
9925 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
9926 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
9927 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
9928 case X86ISD::ADD: return "X86ISD::ADD";
9929 case X86ISD::SUB: return "X86ISD::SUB";
9930 case X86ISD::ADC: return "X86ISD::ADC";
9931 case X86ISD::SBB: return "X86ISD::SBB";
9932 case X86ISD::SMUL: return "X86ISD::SMUL";
9933 case X86ISD::UMUL: return "X86ISD::UMUL";
9934 case X86ISD::INC: return "X86ISD::INC";
9935 case X86ISD::DEC: return "X86ISD::DEC";
9936 case X86ISD::OR: return "X86ISD::OR";
9937 case X86ISD::XOR: return "X86ISD::XOR";
9938 case X86ISD::AND: return "X86ISD::AND";
9939 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
9940 case X86ISD::PTEST: return "X86ISD::PTEST";
9941 case X86ISD::TESTP: return "X86ISD::TESTP";
9942 case X86ISD::PALIGN: return "X86ISD::PALIGN";
9943 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
9944 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
9945 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
9946 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
9947 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
9948 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
9949 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
9950 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
9951 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
9952 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
9953 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
9954 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
9955 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
9956 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
9957 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
9958 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
9959 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
9960 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
9961 case X86ISD::MOVSD: return "X86ISD::MOVSD";
9962 case X86ISD::MOVSS: return "X86ISD::MOVSS";
9963 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
9964 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
9965 case X86ISD::VUNPCKLPDY: return "X86ISD::VUNPCKLPDY";
9966 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
9967 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
9968 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
9969 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
9970 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
9971 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
9972 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
9973 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
9974 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
9975 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
9976 case X86ISD::VPERMILPS: return "X86ISD::VPERMILPS";
9977 case X86ISD::VPERMILPSY: return "X86ISD::VPERMILPSY";
9978 case X86ISD::VPERMILPD: return "X86ISD::VPERMILPD";
9979 case X86ISD::VPERMILPDY: return "X86ISD::VPERMILPDY";
9980 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
9981 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
9982 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
9983 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
9987 // isLegalAddressingMode - Return true if the addressing mode represented
9988 // by AM is legal for this target, for a load/store of the specified type.
9989 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
9991 // X86 supports extremely general addressing modes.
9992 CodeModel::Model M = getTargetMachine().getCodeModel();
9993 Reloc::Model R = getTargetMachine().getRelocationModel();
9995 // X86 allows a sign-extended 32-bit immediate field as a displacement.
9996 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
10001 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
10003 // If a reference to this global requires an extra load, we can't fold it.
10004 if (isGlobalStubReference(GVFlags))
10007 // If BaseGV requires a register for the PIC base, we cannot also have a
10008 // BaseReg specified.
10009 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
10012 // If lower 4G is not available, then we must use rip-relative addressing.
10013 if ((M != CodeModel::Small || R != Reloc::Static) &&
10014 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
10018 switch (AM.Scale) {
10024 // These scales always work.
10029 // These scales are formed with basereg+scalereg. Only accept if there is
10034 default: // Other stuff never works.
10042 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
10043 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10045 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
10046 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
10047 if (NumBits1 <= NumBits2)
10052 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
10053 if (!VT1.isInteger() || !VT2.isInteger())
10055 unsigned NumBits1 = VT1.getSizeInBits();
10056 unsigned NumBits2 = VT2.getSizeInBits();
10057 if (NumBits1 <= NumBits2)
10062 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
10063 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
10064 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
10067 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
10068 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
10069 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
10072 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
10073 // i16 instructions are longer (0x66 prefix) and potentially slower.
10074 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
10077 /// isShuffleMaskLegal - Targets can use this to indicate that they only
10078 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
10079 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
10080 /// are assumed to be legal.
10082 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
10084 // Very little shuffling can be done for 64-bit vectors right now.
10085 if (VT.getSizeInBits() == 64)
10086 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
10088 // FIXME: pshufb, blends, shifts.
10089 return (VT.getVectorNumElements() == 2 ||
10090 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
10091 isMOVLMask(M, VT) ||
10092 isSHUFPMask(M, VT) ||
10093 isPSHUFDMask(M, VT) ||
10094 isPSHUFHWMask(M, VT) ||
10095 isPSHUFLWMask(M, VT) ||
10096 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
10097 isUNPCKLMask(M, VT) ||
10098 isUNPCKHMask(M, VT) ||
10099 isUNPCKL_v_undef_Mask(M, VT) ||
10100 isUNPCKH_v_undef_Mask(M, VT));
10104 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
10106 unsigned NumElts = VT.getVectorNumElements();
10107 // FIXME: This collection of masks seems suspect.
10110 if (NumElts == 4 && VT.getSizeInBits() == 128) {
10111 return (isMOVLMask(Mask, VT) ||
10112 isCommutedMOVLMask(Mask, VT, true) ||
10113 isSHUFPMask(Mask, VT) ||
10114 isCommutedSHUFPMask(Mask, VT));
10119 //===----------------------------------------------------------------------===//
10120 // X86 Scheduler Hooks
10121 //===----------------------------------------------------------------------===//
10123 // private utility function
10124 MachineBasicBlock *
10125 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
10126 MachineBasicBlock *MBB,
10133 TargetRegisterClass *RC,
10134 bool invSrc) const {
10135 // For the atomic bitwise operator, we generate
10138 // ld t1 = [bitinstr.addr]
10139 // op t2 = t1, [bitinstr.val]
10141 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
10143 // fallthrough -->nextMBB
10144 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10145 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10146 MachineFunction::iterator MBBIter = MBB;
10149 /// First build the CFG
10150 MachineFunction *F = MBB->getParent();
10151 MachineBasicBlock *thisMBB = MBB;
10152 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10153 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10154 F->insert(MBBIter, newMBB);
10155 F->insert(MBBIter, nextMBB);
10157 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10158 nextMBB->splice(nextMBB->begin(), thisMBB,
10159 llvm::next(MachineBasicBlock::iterator(bInstr)),
10161 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10163 // Update thisMBB to fall through to newMBB
10164 thisMBB->addSuccessor(newMBB);
10166 // newMBB jumps to itself and fall through to nextMBB
10167 newMBB->addSuccessor(nextMBB);
10168 newMBB->addSuccessor(newMBB);
10170 // Insert instructions into newMBB based on incoming instruction
10171 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
10172 "unexpected number of operands");
10173 DebugLoc dl = bInstr->getDebugLoc();
10174 MachineOperand& destOper = bInstr->getOperand(0);
10175 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10176 int numArgs = bInstr->getNumOperands() - 1;
10177 for (int i=0; i < numArgs; ++i)
10178 argOpers[i] = &bInstr->getOperand(i+1);
10180 // x86 address has 4 operands: base, index, scale, and displacement
10181 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10182 int valArgIndx = lastAddrIndx + 1;
10184 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
10185 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
10186 for (int i=0; i <= lastAddrIndx; ++i)
10187 (*MIB).addOperand(*argOpers[i]);
10189 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
10191 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
10196 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
10197 assert((argOpers[valArgIndx]->isReg() ||
10198 argOpers[valArgIndx]->isImm()) &&
10199 "invalid operand");
10200 if (argOpers[valArgIndx]->isReg())
10201 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
10203 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
10205 (*MIB).addOperand(*argOpers[valArgIndx]);
10207 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
10210 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
10211 for (int i=0; i <= lastAddrIndx; ++i)
10212 (*MIB).addOperand(*argOpers[i]);
10214 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10215 (*MIB).setMemRefs(bInstr->memoperands_begin(),
10216 bInstr->memoperands_end());
10218 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
10219 MIB.addReg(EAXreg);
10222 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10224 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
10228 // private utility function: 64 bit atomics on 32 bit host.
10229 MachineBasicBlock *
10230 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
10231 MachineBasicBlock *MBB,
10236 bool invSrc) const {
10237 // For the atomic bitwise operator, we generate
10238 // thisMBB (instructions are in pairs, except cmpxchg8b)
10239 // ld t1,t2 = [bitinstr.addr]
10241 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
10242 // op t5, t6 <- out1, out2, [bitinstr.val]
10243 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
10244 // mov ECX, EBX <- t5, t6
10245 // mov EAX, EDX <- t1, t2
10246 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
10247 // mov t3, t4 <- EAX, EDX
10249 // result in out1, out2
10250 // fallthrough -->nextMBB
10252 const TargetRegisterClass *RC = X86::GR32RegisterClass;
10253 const unsigned LoadOpc = X86::MOV32rm;
10254 const unsigned NotOpc = X86::NOT32r;
10255 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10256 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10257 MachineFunction::iterator MBBIter = MBB;
10260 /// First build the CFG
10261 MachineFunction *F = MBB->getParent();
10262 MachineBasicBlock *thisMBB = MBB;
10263 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10264 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10265 F->insert(MBBIter, newMBB);
10266 F->insert(MBBIter, nextMBB);
10268 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10269 nextMBB->splice(nextMBB->begin(), thisMBB,
10270 llvm::next(MachineBasicBlock::iterator(bInstr)),
10272 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10274 // Update thisMBB to fall through to newMBB
10275 thisMBB->addSuccessor(newMBB);
10277 // newMBB jumps to itself and fall through to nextMBB
10278 newMBB->addSuccessor(nextMBB);
10279 newMBB->addSuccessor(newMBB);
10281 DebugLoc dl = bInstr->getDebugLoc();
10282 // Insert instructions into newMBB based on incoming instruction
10283 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
10284 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
10285 "unexpected number of operands");
10286 MachineOperand& dest1Oper = bInstr->getOperand(0);
10287 MachineOperand& dest2Oper = bInstr->getOperand(1);
10288 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10289 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
10290 argOpers[i] = &bInstr->getOperand(i+2);
10292 // We use some of the operands multiple times, so conservatively just
10293 // clear any kill flags that might be present.
10294 if (argOpers[i]->isReg() && argOpers[i]->isUse())
10295 argOpers[i]->setIsKill(false);
10298 // x86 address has 5 operands: base, index, scale, displacement, and segment.
10299 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10301 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
10302 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
10303 for (int i=0; i <= lastAddrIndx; ++i)
10304 (*MIB).addOperand(*argOpers[i]);
10305 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
10306 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
10307 // add 4 to displacement.
10308 for (int i=0; i <= lastAddrIndx-2; ++i)
10309 (*MIB).addOperand(*argOpers[i]);
10310 MachineOperand newOp3 = *(argOpers[3]);
10311 if (newOp3.isImm())
10312 newOp3.setImm(newOp3.getImm()+4);
10314 newOp3.setOffset(newOp3.getOffset()+4);
10315 (*MIB).addOperand(newOp3);
10316 (*MIB).addOperand(*argOpers[lastAddrIndx]);
10318 // t3/4 are defined later, at the bottom of the loop
10319 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
10320 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
10321 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
10322 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
10323 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
10324 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
10326 // The subsequent operations should be using the destination registers of
10327 //the PHI instructions.
10329 t1 = F->getRegInfo().createVirtualRegister(RC);
10330 t2 = F->getRegInfo().createVirtualRegister(RC);
10331 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
10332 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
10334 t1 = dest1Oper.getReg();
10335 t2 = dest2Oper.getReg();
10338 int valArgIndx = lastAddrIndx + 1;
10339 assert((argOpers[valArgIndx]->isReg() ||
10340 argOpers[valArgIndx]->isImm()) &&
10341 "invalid operand");
10342 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
10343 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
10344 if (argOpers[valArgIndx]->isReg())
10345 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
10347 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
10348 if (regOpcL != X86::MOV32rr)
10350 (*MIB).addOperand(*argOpers[valArgIndx]);
10351 assert(argOpers[valArgIndx + 1]->isReg() ==
10352 argOpers[valArgIndx]->isReg());
10353 assert(argOpers[valArgIndx + 1]->isImm() ==
10354 argOpers[valArgIndx]->isImm());
10355 if (argOpers[valArgIndx + 1]->isReg())
10356 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
10358 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
10359 if (regOpcH != X86::MOV32rr)
10361 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
10363 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
10365 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
10368 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
10370 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
10373 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
10374 for (int i=0; i <= lastAddrIndx; ++i)
10375 (*MIB).addOperand(*argOpers[i]);
10377 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10378 (*MIB).setMemRefs(bInstr->memoperands_begin(),
10379 bInstr->memoperands_end());
10381 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
10382 MIB.addReg(X86::EAX);
10383 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
10384 MIB.addReg(X86::EDX);
10387 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10389 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
10393 // private utility function
10394 MachineBasicBlock *
10395 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
10396 MachineBasicBlock *MBB,
10397 unsigned cmovOpc) const {
10398 // For the atomic min/max operator, we generate
10401 // ld t1 = [min/max.addr]
10402 // mov t2 = [min/max.val]
10404 // cmov[cond] t2 = t1
10406 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
10408 // fallthrough -->nextMBB
10410 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10411 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10412 MachineFunction::iterator MBBIter = MBB;
10415 /// First build the CFG
10416 MachineFunction *F = MBB->getParent();
10417 MachineBasicBlock *thisMBB = MBB;
10418 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10419 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10420 F->insert(MBBIter, newMBB);
10421 F->insert(MBBIter, nextMBB);
10423 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10424 nextMBB->splice(nextMBB->begin(), thisMBB,
10425 llvm::next(MachineBasicBlock::iterator(mInstr)),
10427 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10429 // Update thisMBB to fall through to newMBB
10430 thisMBB->addSuccessor(newMBB);
10432 // newMBB jumps to newMBB and fall through to nextMBB
10433 newMBB->addSuccessor(nextMBB);
10434 newMBB->addSuccessor(newMBB);
10436 DebugLoc dl = mInstr->getDebugLoc();
10437 // Insert instructions into newMBB based on incoming instruction
10438 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
10439 "unexpected number of operands");
10440 MachineOperand& destOper = mInstr->getOperand(0);
10441 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10442 int numArgs = mInstr->getNumOperands() - 1;
10443 for (int i=0; i < numArgs; ++i)
10444 argOpers[i] = &mInstr->getOperand(i+1);
10446 // x86 address has 4 operands: base, index, scale, and displacement
10447 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10448 int valArgIndx = lastAddrIndx + 1;
10450 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10451 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
10452 for (int i=0; i <= lastAddrIndx; ++i)
10453 (*MIB).addOperand(*argOpers[i]);
10455 // We only support register and immediate values
10456 assert((argOpers[valArgIndx]->isReg() ||
10457 argOpers[valArgIndx]->isImm()) &&
10458 "invalid operand");
10460 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10461 if (argOpers[valArgIndx]->isReg())
10462 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
10464 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
10465 (*MIB).addOperand(*argOpers[valArgIndx]);
10467 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
10470 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
10475 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10476 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
10480 // Cmp and exchange if none has modified the memory location
10481 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
10482 for (int i=0; i <= lastAddrIndx; ++i)
10483 (*MIB).addOperand(*argOpers[i]);
10485 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10486 (*MIB).setMemRefs(mInstr->memoperands_begin(),
10487 mInstr->memoperands_end());
10489 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
10490 MIB.addReg(X86::EAX);
10493 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10495 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
10499 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
10500 // or XMM0_V32I8 in AVX all of this code can be replaced with that
10501 // in the .td file.
10502 MachineBasicBlock *
10503 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
10504 unsigned numArgs, bool memArg) const {
10505 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
10506 "Target must have SSE4.2 or AVX features enabled");
10508 DebugLoc dl = MI->getDebugLoc();
10509 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10511 if (!Subtarget->hasAVX()) {
10513 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
10515 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
10518 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
10520 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
10523 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
10524 for (unsigned i = 0; i < numArgs; ++i) {
10525 MachineOperand &Op = MI->getOperand(i+1);
10526 if (!(Op.isReg() && Op.isImplicit()))
10527 MIB.addOperand(Op);
10529 BuildMI(*BB, MI, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
10530 .addReg(X86::XMM0);
10532 MI->eraseFromParent();
10536 MachineBasicBlock *
10537 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
10538 DebugLoc dl = MI->getDebugLoc();
10539 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10541 // Address into RAX/EAX, other two args into ECX, EDX.
10542 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
10543 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
10544 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
10545 for (int i = 0; i < X86::AddrNumOperands; ++i)
10546 MIB.addOperand(MI->getOperand(i));
10548 unsigned ValOps = X86::AddrNumOperands;
10549 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10550 .addReg(MI->getOperand(ValOps).getReg());
10551 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
10552 .addReg(MI->getOperand(ValOps+1).getReg());
10554 // The instruction doesn't actually take any operands though.
10555 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
10557 MI->eraseFromParent(); // The pseudo is gone now.
10561 MachineBasicBlock *
10562 X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
10563 DebugLoc dl = MI->getDebugLoc();
10564 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10566 // First arg in ECX, the second in EAX.
10567 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10568 .addReg(MI->getOperand(0).getReg());
10569 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
10570 .addReg(MI->getOperand(1).getReg());
10572 // The instruction doesn't actually take any operands though.
10573 BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
10575 MI->eraseFromParent(); // The pseudo is gone now.
10579 MachineBasicBlock *
10580 X86TargetLowering::EmitVAARG64WithCustomInserter(
10582 MachineBasicBlock *MBB) const {
10583 // Emit va_arg instruction on X86-64.
10585 // Operands to this pseudo-instruction:
10586 // 0 ) Output : destination address (reg)
10587 // 1-5) Input : va_list address (addr, i64mem)
10588 // 6 ) ArgSize : Size (in bytes) of vararg type
10589 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
10590 // 8 ) Align : Alignment of type
10591 // 9 ) EFLAGS (implicit-def)
10593 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
10594 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
10596 unsigned DestReg = MI->getOperand(0).getReg();
10597 MachineOperand &Base = MI->getOperand(1);
10598 MachineOperand &Scale = MI->getOperand(2);
10599 MachineOperand &Index = MI->getOperand(3);
10600 MachineOperand &Disp = MI->getOperand(4);
10601 MachineOperand &Segment = MI->getOperand(5);
10602 unsigned ArgSize = MI->getOperand(6).getImm();
10603 unsigned ArgMode = MI->getOperand(7).getImm();
10604 unsigned Align = MI->getOperand(8).getImm();
10606 // Memory Reference
10607 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
10608 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
10609 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
10611 // Machine Information
10612 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10613 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
10614 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
10615 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
10616 DebugLoc DL = MI->getDebugLoc();
10618 // struct va_list {
10621 // i64 overflow_area (address)
10622 // i64 reg_save_area (address)
10624 // sizeof(va_list) = 24
10625 // alignment(va_list) = 8
10627 unsigned TotalNumIntRegs = 6;
10628 unsigned TotalNumXMMRegs = 8;
10629 bool UseGPOffset = (ArgMode == 1);
10630 bool UseFPOffset = (ArgMode == 2);
10631 unsigned MaxOffset = TotalNumIntRegs * 8 +
10632 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
10634 /* Align ArgSize to a multiple of 8 */
10635 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
10636 bool NeedsAlign = (Align > 8);
10638 MachineBasicBlock *thisMBB = MBB;
10639 MachineBasicBlock *overflowMBB;
10640 MachineBasicBlock *offsetMBB;
10641 MachineBasicBlock *endMBB;
10643 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
10644 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
10645 unsigned OffsetReg = 0;
10647 if (!UseGPOffset && !UseFPOffset) {
10648 // If we only pull from the overflow region, we don't create a branch.
10649 // We don't need to alter control flow.
10650 OffsetDestReg = 0; // unused
10651 OverflowDestReg = DestReg;
10654 overflowMBB = thisMBB;
10657 // First emit code to check if gp_offset (or fp_offset) is below the bound.
10658 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
10659 // If not, pull from overflow_area. (branch to overflowMBB)
10664 // offsetMBB overflowMBB
10669 // Registers for the PHI in endMBB
10670 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
10671 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
10673 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10674 MachineFunction *MF = MBB->getParent();
10675 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10676 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10677 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10679 MachineFunction::iterator MBBIter = MBB;
10682 // Insert the new basic blocks
10683 MF->insert(MBBIter, offsetMBB);
10684 MF->insert(MBBIter, overflowMBB);
10685 MF->insert(MBBIter, endMBB);
10687 // Transfer the remainder of MBB and its successor edges to endMBB.
10688 endMBB->splice(endMBB->begin(), thisMBB,
10689 llvm::next(MachineBasicBlock::iterator(MI)),
10691 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10693 // Make offsetMBB and overflowMBB successors of thisMBB
10694 thisMBB->addSuccessor(offsetMBB);
10695 thisMBB->addSuccessor(overflowMBB);
10697 // endMBB is a successor of both offsetMBB and overflowMBB
10698 offsetMBB->addSuccessor(endMBB);
10699 overflowMBB->addSuccessor(endMBB);
10701 // Load the offset value into a register
10702 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10703 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
10707 .addDisp(Disp, UseFPOffset ? 4 : 0)
10708 .addOperand(Segment)
10709 .setMemRefs(MMOBegin, MMOEnd);
10711 // Check if there is enough room left to pull this argument.
10712 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
10714 .addImm(MaxOffset + 8 - ArgSizeA8);
10716 // Branch to "overflowMBB" if offset >= max
10717 // Fall through to "offsetMBB" otherwise
10718 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
10719 .addMBB(overflowMBB);
10722 // In offsetMBB, emit code to use the reg_save_area.
10724 assert(OffsetReg != 0);
10726 // Read the reg_save_area address.
10727 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
10728 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
10733 .addOperand(Segment)
10734 .setMemRefs(MMOBegin, MMOEnd);
10736 // Zero-extend the offset
10737 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
10738 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
10741 .addImm(X86::sub_32bit);
10743 // Add the offset to the reg_save_area to get the final address.
10744 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
10745 .addReg(OffsetReg64)
10746 .addReg(RegSaveReg);
10748 // Compute the offset for the next argument
10749 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10750 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
10752 .addImm(UseFPOffset ? 16 : 8);
10754 // Store it back into the va_list.
10755 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
10759 .addDisp(Disp, UseFPOffset ? 4 : 0)
10760 .addOperand(Segment)
10761 .addReg(NextOffsetReg)
10762 .setMemRefs(MMOBegin, MMOEnd);
10765 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
10770 // Emit code to use overflow area
10773 // Load the overflow_area address into a register.
10774 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
10775 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
10780 .addOperand(Segment)
10781 .setMemRefs(MMOBegin, MMOEnd);
10783 // If we need to align it, do so. Otherwise, just copy the address
10784 // to OverflowDestReg.
10786 // Align the overflow address
10787 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
10788 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
10790 // aligned_addr = (addr + (align-1)) & ~(align-1)
10791 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
10792 .addReg(OverflowAddrReg)
10795 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
10797 .addImm(~(uint64_t)(Align-1));
10799 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
10800 .addReg(OverflowAddrReg);
10803 // Compute the next overflow address after this argument.
10804 // (the overflow address should be kept 8-byte aligned)
10805 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
10806 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
10807 .addReg(OverflowDestReg)
10808 .addImm(ArgSizeA8);
10810 // Store the new overflow address.
10811 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
10816 .addOperand(Segment)
10817 .addReg(NextAddrReg)
10818 .setMemRefs(MMOBegin, MMOEnd);
10820 // If we branched, emit the PHI to the front of endMBB.
10822 BuildMI(*endMBB, endMBB->begin(), DL,
10823 TII->get(X86::PHI), DestReg)
10824 .addReg(OffsetDestReg).addMBB(offsetMBB)
10825 .addReg(OverflowDestReg).addMBB(overflowMBB);
10828 // Erase the pseudo instruction
10829 MI->eraseFromParent();
10834 MachineBasicBlock *
10835 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
10837 MachineBasicBlock *MBB) const {
10838 // Emit code to save XMM registers to the stack. The ABI says that the
10839 // number of registers to save is given in %al, so it's theoretically
10840 // possible to do an indirect jump trick to avoid saving all of them,
10841 // however this code takes a simpler approach and just executes all
10842 // of the stores if %al is non-zero. It's less code, and it's probably
10843 // easier on the hardware branch predictor, and stores aren't all that
10844 // expensive anyway.
10846 // Create the new basic blocks. One block contains all the XMM stores,
10847 // and one block is the final destination regardless of whether any
10848 // stores were performed.
10849 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10850 MachineFunction *F = MBB->getParent();
10851 MachineFunction::iterator MBBIter = MBB;
10853 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
10854 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
10855 F->insert(MBBIter, XMMSaveMBB);
10856 F->insert(MBBIter, EndMBB);
10858 // Transfer the remainder of MBB and its successor edges to EndMBB.
10859 EndMBB->splice(EndMBB->begin(), MBB,
10860 llvm::next(MachineBasicBlock::iterator(MI)),
10862 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
10864 // The original block will now fall through to the XMM save block.
10865 MBB->addSuccessor(XMMSaveMBB);
10866 // The XMMSaveMBB will fall through to the end block.
10867 XMMSaveMBB->addSuccessor(EndMBB);
10869 // Now add the instructions.
10870 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10871 DebugLoc DL = MI->getDebugLoc();
10873 unsigned CountReg = MI->getOperand(0).getReg();
10874 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
10875 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
10877 if (!Subtarget->isTargetWin64()) {
10878 // If %al is 0, branch around the XMM save block.
10879 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
10880 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
10881 MBB->addSuccessor(EndMBB);
10884 // In the XMM save block, save all the XMM argument registers.
10885 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
10886 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
10887 MachineMemOperand *MMO =
10888 F->getMachineMemOperand(
10889 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
10890 MachineMemOperand::MOStore,
10891 /*Size=*/16, /*Align=*/16);
10892 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
10893 .addFrameIndex(RegSaveFrameIndex)
10894 .addImm(/*Scale=*/1)
10895 .addReg(/*IndexReg=*/0)
10896 .addImm(/*Disp=*/Offset)
10897 .addReg(/*Segment=*/0)
10898 .addReg(MI->getOperand(i).getReg())
10899 .addMemOperand(MMO);
10902 MI->eraseFromParent(); // The pseudo instruction is gone now.
10907 MachineBasicBlock *
10908 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
10909 MachineBasicBlock *BB) const {
10910 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10911 DebugLoc DL = MI->getDebugLoc();
10913 // To "insert" a SELECT_CC instruction, we actually have to insert the
10914 // diamond control-flow pattern. The incoming instruction knows the
10915 // destination vreg to set, the condition code register to branch on, the
10916 // true/false values to select between, and a branch opcode to use.
10917 const BasicBlock *LLVM_BB = BB->getBasicBlock();
10918 MachineFunction::iterator It = BB;
10924 // cmpTY ccX, r1, r2
10926 // fallthrough --> copy0MBB
10927 MachineBasicBlock *thisMBB = BB;
10928 MachineFunction *F = BB->getParent();
10929 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
10930 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
10931 F->insert(It, copy0MBB);
10932 F->insert(It, sinkMBB);
10934 // If the EFLAGS register isn't dead in the terminator, then claim that it's
10935 // live into the sink and copy blocks.
10936 const MachineFunction *MF = BB->getParent();
10937 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
10938 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
10940 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
10941 const MachineOperand &MO = MI->getOperand(I);
10942 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
10943 unsigned Reg = MO.getReg();
10944 if (Reg != X86::EFLAGS) continue;
10945 copy0MBB->addLiveIn(Reg);
10946 sinkMBB->addLiveIn(Reg);
10949 // Transfer the remainder of BB and its successor edges to sinkMBB.
10950 sinkMBB->splice(sinkMBB->begin(), BB,
10951 llvm::next(MachineBasicBlock::iterator(MI)),
10953 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
10955 // Add the true and fallthrough blocks as its successors.
10956 BB->addSuccessor(copy0MBB);
10957 BB->addSuccessor(sinkMBB);
10959 // Create the conditional branch instruction.
10961 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
10962 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
10965 // %FalseValue = ...
10966 // # fallthrough to sinkMBB
10967 copy0MBB->addSuccessor(sinkMBB);
10970 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
10972 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
10973 TII->get(X86::PHI), MI->getOperand(0).getReg())
10974 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
10975 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
10977 MI->eraseFromParent(); // The pseudo instruction is gone now.
10981 MachineBasicBlock *
10982 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
10983 MachineBasicBlock *BB) const {
10984 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10985 DebugLoc DL = MI->getDebugLoc();
10987 assert(!Subtarget->isTargetEnvMacho());
10989 // The lowering is pretty easy: we're just emitting the call to _alloca. The
10990 // non-trivial part is impdef of ESP.
10992 if (Subtarget->isTargetWin64()) {
10993 if (Subtarget->isTargetCygMing()) {
10994 // ___chkstk(Mingw64):
10995 // Clobbers R10, R11, RAX and EFLAGS.
10997 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
10998 .addExternalSymbol("___chkstk")
10999 .addReg(X86::RAX, RegState::Implicit)
11000 .addReg(X86::RSP, RegState::Implicit)
11001 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
11002 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
11003 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11005 // __chkstk(MSVCRT): does not update stack pointer.
11006 // Clobbers R10, R11 and EFLAGS.
11007 // FIXME: RAX(allocated size) might be reused and not killed.
11008 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
11009 .addExternalSymbol("__chkstk")
11010 .addReg(X86::RAX, RegState::Implicit)
11011 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11012 // RAX has the offset to subtracted from RSP.
11013 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
11018 const char *StackProbeSymbol =
11019 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
11021 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
11022 .addExternalSymbol(StackProbeSymbol)
11023 .addReg(X86::EAX, RegState::Implicit)
11024 .addReg(X86::ESP, RegState::Implicit)
11025 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
11026 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
11027 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11030 MI->eraseFromParent(); // The pseudo instruction is gone now.
11034 MachineBasicBlock *
11035 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
11036 MachineBasicBlock *BB) const {
11037 // This is pretty easy. We're taking the value that we received from
11038 // our load from the relocation, sticking it in either RDI (x86-64)
11039 // or EAX and doing an indirect call. The return value will then
11040 // be in the normal return register.
11041 const X86InstrInfo *TII
11042 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
11043 DebugLoc DL = MI->getDebugLoc();
11044 MachineFunction *F = BB->getParent();
11046 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
11047 assert(MI->getOperand(3).isGlobal() && "This should be a global");
11049 if (Subtarget->is64Bit()) {
11050 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11051 TII->get(X86::MOV64rm), X86::RDI)
11053 .addImm(0).addReg(0)
11054 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11055 MI->getOperand(3).getTargetFlags())
11057 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
11058 addDirectMem(MIB, X86::RDI);
11059 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
11060 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11061 TII->get(X86::MOV32rm), X86::EAX)
11063 .addImm(0).addReg(0)
11064 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11065 MI->getOperand(3).getTargetFlags())
11067 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
11068 addDirectMem(MIB, X86::EAX);
11070 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11071 TII->get(X86::MOV32rm), X86::EAX)
11072 .addReg(TII->getGlobalBaseReg(F))
11073 .addImm(0).addReg(0)
11074 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11075 MI->getOperand(3).getTargetFlags())
11077 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
11078 addDirectMem(MIB, X86::EAX);
11081 MI->eraseFromParent(); // The pseudo instruction is gone now.
11085 MachineBasicBlock *
11086 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
11087 MachineBasicBlock *BB) const {
11088 switch (MI->getOpcode()) {
11089 default: assert(false && "Unexpected instr type to insert");
11090 case X86::TAILJMPd64:
11091 case X86::TAILJMPr64:
11092 case X86::TAILJMPm64:
11093 assert(!"TAILJMP64 would not be touched here.");
11094 case X86::TCRETURNdi64:
11095 case X86::TCRETURNri64:
11096 case X86::TCRETURNmi64:
11097 // Defs of TCRETURNxx64 has Win64's callee-saved registers, as subset.
11098 // On AMD64, additional defs should be added before register allocation.
11099 if (!Subtarget->isTargetWin64()) {
11100 MI->addRegisterDefined(X86::RSI);
11101 MI->addRegisterDefined(X86::RDI);
11102 MI->addRegisterDefined(X86::XMM6);
11103 MI->addRegisterDefined(X86::XMM7);
11104 MI->addRegisterDefined(X86::XMM8);
11105 MI->addRegisterDefined(X86::XMM9);
11106 MI->addRegisterDefined(X86::XMM10);
11107 MI->addRegisterDefined(X86::XMM11);
11108 MI->addRegisterDefined(X86::XMM12);
11109 MI->addRegisterDefined(X86::XMM13);
11110 MI->addRegisterDefined(X86::XMM14);
11111 MI->addRegisterDefined(X86::XMM15);
11114 case X86::WIN_ALLOCA:
11115 return EmitLoweredWinAlloca(MI, BB);
11116 case X86::TLSCall_32:
11117 case X86::TLSCall_64:
11118 return EmitLoweredTLSCall(MI, BB);
11119 case X86::CMOV_GR8:
11120 case X86::CMOV_FR32:
11121 case X86::CMOV_FR64:
11122 case X86::CMOV_V4F32:
11123 case X86::CMOV_V2F64:
11124 case X86::CMOV_V2I64:
11125 case X86::CMOV_GR16:
11126 case X86::CMOV_GR32:
11127 case X86::CMOV_RFP32:
11128 case X86::CMOV_RFP64:
11129 case X86::CMOV_RFP80:
11130 return EmitLoweredSelect(MI, BB);
11132 case X86::FP32_TO_INT16_IN_MEM:
11133 case X86::FP32_TO_INT32_IN_MEM:
11134 case X86::FP32_TO_INT64_IN_MEM:
11135 case X86::FP64_TO_INT16_IN_MEM:
11136 case X86::FP64_TO_INT32_IN_MEM:
11137 case X86::FP64_TO_INT64_IN_MEM:
11138 case X86::FP80_TO_INT16_IN_MEM:
11139 case X86::FP80_TO_INT32_IN_MEM:
11140 case X86::FP80_TO_INT64_IN_MEM: {
11141 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11142 DebugLoc DL = MI->getDebugLoc();
11144 // Change the floating point control register to use "round towards zero"
11145 // mode when truncating to an integer value.
11146 MachineFunction *F = BB->getParent();
11147 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
11148 addFrameReference(BuildMI(*BB, MI, DL,
11149 TII->get(X86::FNSTCW16m)), CWFrameIdx);
11151 // Load the old value of the high byte of the control word...
11153 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
11154 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
11157 // Set the high part to be round to zero...
11158 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
11161 // Reload the modified control word now...
11162 addFrameReference(BuildMI(*BB, MI, DL,
11163 TII->get(X86::FLDCW16m)), CWFrameIdx);
11165 // Restore the memory image of control word to original value
11166 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
11169 // Get the X86 opcode to use.
11171 switch (MI->getOpcode()) {
11172 default: llvm_unreachable("illegal opcode!");
11173 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
11174 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
11175 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
11176 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
11177 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
11178 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
11179 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
11180 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
11181 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
11185 MachineOperand &Op = MI->getOperand(0);
11187 AM.BaseType = X86AddressMode::RegBase;
11188 AM.Base.Reg = Op.getReg();
11190 AM.BaseType = X86AddressMode::FrameIndexBase;
11191 AM.Base.FrameIndex = Op.getIndex();
11193 Op = MI->getOperand(1);
11195 AM.Scale = Op.getImm();
11196 Op = MI->getOperand(2);
11198 AM.IndexReg = Op.getImm();
11199 Op = MI->getOperand(3);
11200 if (Op.isGlobal()) {
11201 AM.GV = Op.getGlobal();
11203 AM.Disp = Op.getImm();
11205 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
11206 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
11208 // Reload the original control word now.
11209 addFrameReference(BuildMI(*BB, MI, DL,
11210 TII->get(X86::FLDCW16m)), CWFrameIdx);
11212 MI->eraseFromParent(); // The pseudo instruction is gone now.
11215 // String/text processing lowering.
11216 case X86::PCMPISTRM128REG:
11217 case X86::VPCMPISTRM128REG:
11218 return EmitPCMP(MI, BB, 3, false /* in-mem */);
11219 case X86::PCMPISTRM128MEM:
11220 case X86::VPCMPISTRM128MEM:
11221 return EmitPCMP(MI, BB, 3, true /* in-mem */);
11222 case X86::PCMPESTRM128REG:
11223 case X86::VPCMPESTRM128REG:
11224 return EmitPCMP(MI, BB, 5, false /* in mem */);
11225 case X86::PCMPESTRM128MEM:
11226 case X86::VPCMPESTRM128MEM:
11227 return EmitPCMP(MI, BB, 5, true /* in mem */);
11229 // Thread synchronization.
11231 return EmitMonitor(MI, BB);
11233 return EmitMwait(MI, BB);
11235 // Atomic Lowering.
11236 case X86::ATOMAND32:
11237 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
11238 X86::AND32ri, X86::MOV32rm,
11240 X86::NOT32r, X86::EAX,
11241 X86::GR32RegisterClass);
11242 case X86::ATOMOR32:
11243 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
11244 X86::OR32ri, X86::MOV32rm,
11246 X86::NOT32r, X86::EAX,
11247 X86::GR32RegisterClass);
11248 case X86::ATOMXOR32:
11249 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
11250 X86::XOR32ri, X86::MOV32rm,
11252 X86::NOT32r, X86::EAX,
11253 X86::GR32RegisterClass);
11254 case X86::ATOMNAND32:
11255 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
11256 X86::AND32ri, X86::MOV32rm,
11258 X86::NOT32r, X86::EAX,
11259 X86::GR32RegisterClass, true);
11260 case X86::ATOMMIN32:
11261 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
11262 case X86::ATOMMAX32:
11263 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
11264 case X86::ATOMUMIN32:
11265 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
11266 case X86::ATOMUMAX32:
11267 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
11269 case X86::ATOMAND16:
11270 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
11271 X86::AND16ri, X86::MOV16rm,
11273 X86::NOT16r, X86::AX,
11274 X86::GR16RegisterClass);
11275 case X86::ATOMOR16:
11276 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
11277 X86::OR16ri, X86::MOV16rm,
11279 X86::NOT16r, X86::AX,
11280 X86::GR16RegisterClass);
11281 case X86::ATOMXOR16:
11282 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
11283 X86::XOR16ri, X86::MOV16rm,
11285 X86::NOT16r, X86::AX,
11286 X86::GR16RegisterClass);
11287 case X86::ATOMNAND16:
11288 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
11289 X86::AND16ri, X86::MOV16rm,
11291 X86::NOT16r, X86::AX,
11292 X86::GR16RegisterClass, true);
11293 case X86::ATOMMIN16:
11294 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
11295 case X86::ATOMMAX16:
11296 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
11297 case X86::ATOMUMIN16:
11298 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
11299 case X86::ATOMUMAX16:
11300 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
11302 case X86::ATOMAND8:
11303 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
11304 X86::AND8ri, X86::MOV8rm,
11306 X86::NOT8r, X86::AL,
11307 X86::GR8RegisterClass);
11309 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
11310 X86::OR8ri, X86::MOV8rm,
11312 X86::NOT8r, X86::AL,
11313 X86::GR8RegisterClass);
11314 case X86::ATOMXOR8:
11315 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
11316 X86::XOR8ri, X86::MOV8rm,
11318 X86::NOT8r, X86::AL,
11319 X86::GR8RegisterClass);
11320 case X86::ATOMNAND8:
11321 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
11322 X86::AND8ri, X86::MOV8rm,
11324 X86::NOT8r, X86::AL,
11325 X86::GR8RegisterClass, true);
11326 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
11327 // This group is for 64-bit host.
11328 case X86::ATOMAND64:
11329 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
11330 X86::AND64ri32, X86::MOV64rm,
11332 X86::NOT64r, X86::RAX,
11333 X86::GR64RegisterClass);
11334 case X86::ATOMOR64:
11335 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
11336 X86::OR64ri32, X86::MOV64rm,
11338 X86::NOT64r, X86::RAX,
11339 X86::GR64RegisterClass);
11340 case X86::ATOMXOR64:
11341 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
11342 X86::XOR64ri32, X86::MOV64rm,
11344 X86::NOT64r, X86::RAX,
11345 X86::GR64RegisterClass);
11346 case X86::ATOMNAND64:
11347 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
11348 X86::AND64ri32, X86::MOV64rm,
11350 X86::NOT64r, X86::RAX,
11351 X86::GR64RegisterClass, true);
11352 case X86::ATOMMIN64:
11353 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
11354 case X86::ATOMMAX64:
11355 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
11356 case X86::ATOMUMIN64:
11357 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
11358 case X86::ATOMUMAX64:
11359 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
11361 // This group does 64-bit operations on a 32-bit host.
11362 case X86::ATOMAND6432:
11363 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11364 X86::AND32rr, X86::AND32rr,
11365 X86::AND32ri, X86::AND32ri,
11367 case X86::ATOMOR6432:
11368 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11369 X86::OR32rr, X86::OR32rr,
11370 X86::OR32ri, X86::OR32ri,
11372 case X86::ATOMXOR6432:
11373 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11374 X86::XOR32rr, X86::XOR32rr,
11375 X86::XOR32ri, X86::XOR32ri,
11377 case X86::ATOMNAND6432:
11378 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11379 X86::AND32rr, X86::AND32rr,
11380 X86::AND32ri, X86::AND32ri,
11382 case X86::ATOMADD6432:
11383 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11384 X86::ADD32rr, X86::ADC32rr,
11385 X86::ADD32ri, X86::ADC32ri,
11387 case X86::ATOMSUB6432:
11388 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11389 X86::SUB32rr, X86::SBB32rr,
11390 X86::SUB32ri, X86::SBB32ri,
11392 case X86::ATOMSWAP6432:
11393 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11394 X86::MOV32rr, X86::MOV32rr,
11395 X86::MOV32ri, X86::MOV32ri,
11397 case X86::VASTART_SAVE_XMM_REGS:
11398 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
11400 case X86::VAARG_64:
11401 return EmitVAARG64WithCustomInserter(MI, BB);
11405 //===----------------------------------------------------------------------===//
11406 // X86 Optimization Hooks
11407 //===----------------------------------------------------------------------===//
11409 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
11413 const SelectionDAG &DAG,
11414 unsigned Depth) const {
11415 unsigned Opc = Op.getOpcode();
11416 assert((Opc >= ISD::BUILTIN_OP_END ||
11417 Opc == ISD::INTRINSIC_WO_CHAIN ||
11418 Opc == ISD::INTRINSIC_W_CHAIN ||
11419 Opc == ISD::INTRINSIC_VOID) &&
11420 "Should use MaskedValueIsZero if you don't know whether Op"
11421 " is a target node!");
11423 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
11437 // These nodes' second result is a boolean.
11438 if (Op.getResNo() == 0)
11441 case X86ISD::SETCC:
11442 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
11443 Mask.getBitWidth() - 1);
11448 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
11449 unsigned Depth) const {
11450 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
11451 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
11452 return Op.getValueType().getScalarType().getSizeInBits();
11458 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
11459 /// node is a GlobalAddress + offset.
11460 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
11461 const GlobalValue* &GA,
11462 int64_t &Offset) const {
11463 if (N->getOpcode() == X86ISD::Wrapper) {
11464 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
11465 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
11466 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
11470 return TargetLowering::isGAPlusOffset(N, GA, Offset);
11473 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
11474 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
11475 TargetLowering::DAGCombinerInfo &DCI) {
11476 DebugLoc dl = N->getDebugLoc();
11477 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
11478 SDValue V1 = SVOp->getOperand(0);
11479 SDValue V2 = SVOp->getOperand(1);
11480 EVT VT = SVOp->getValueType(0);
11482 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
11483 V2.getOpcode() == ISD::CONCAT_VECTORS) {
11487 // V UNDEF BUILD_VECTOR UNDEF
11489 // CONCAT_VECTOR CONCAT_VECTOR
11492 // RESULT: V + zero extended
11494 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
11495 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
11496 V1.getOperand(1).getOpcode() != ISD::UNDEF)
11499 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
11502 // To match the shuffle mask, the first half of the mask should
11503 // be exactly the first vector, and all the rest a splat with the
11504 // first element of the second one.
11505 int NumElems = VT.getVectorNumElements();
11506 for (int i = 0; i < NumElems/2; ++i)
11507 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
11508 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
11511 // Emit a zeroed vector and insert the desired subvector on its
11513 SDValue Zeros = getZeroVector(VT, true /* HasSSE2 */, DAG, dl);
11514 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0),
11515 DAG.getConstant(0, MVT::i32), DAG, dl);
11516 return DCI.CombineTo(N, InsV);
11522 /// PerformShuffleCombine - Performs several different shuffle combines.
11523 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
11524 TargetLowering::DAGCombinerInfo &DCI) {
11525 DebugLoc dl = N->getDebugLoc();
11526 EVT VT = N->getValueType(0);
11528 // Don't create instructions with illegal types after legalize types has run.
11529 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11530 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
11533 // Only handle pure VECTOR_SHUFFLE nodes.
11534 if (VT.getSizeInBits() == 256 && N->getOpcode() == ISD::VECTOR_SHUFFLE)
11535 return PerformShuffleCombine256(N, DAG, DCI);
11537 // Only handle 128 wide vector from here on.
11538 if (VT.getSizeInBits() != 128)
11541 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
11542 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
11543 // consecutive, non-overlapping, and in the right order.
11544 SmallVector<SDValue, 16> Elts;
11545 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
11546 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
11548 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
11551 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
11552 /// generation and convert it from being a bunch of shuffles and extracts
11553 /// to a simple store and scalar loads to extract the elements.
11554 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
11555 const TargetLowering &TLI) {
11556 SDValue InputVector = N->getOperand(0);
11558 // Only operate on vectors of 4 elements, where the alternative shuffling
11559 // gets to be more expensive.
11560 if (InputVector.getValueType() != MVT::v4i32)
11563 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
11564 // single use which is a sign-extend or zero-extend, and all elements are
11566 SmallVector<SDNode *, 4> Uses;
11567 unsigned ExtractedElements = 0;
11568 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
11569 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
11570 if (UI.getUse().getResNo() != InputVector.getResNo())
11573 SDNode *Extract = *UI;
11574 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
11577 if (Extract->getValueType(0) != MVT::i32)
11579 if (!Extract->hasOneUse())
11581 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
11582 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
11584 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
11587 // Record which element was extracted.
11588 ExtractedElements |=
11589 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
11591 Uses.push_back(Extract);
11594 // If not all the elements were used, this may not be worthwhile.
11595 if (ExtractedElements != 15)
11598 // Ok, we've now decided to do the transformation.
11599 DebugLoc dl = InputVector.getDebugLoc();
11601 // Store the value to a temporary stack slot.
11602 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
11603 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
11604 MachinePointerInfo(), false, false, 0);
11606 // Replace each use (extract) with a load of the appropriate element.
11607 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
11608 UE = Uses.end(); UI != UE; ++UI) {
11609 SDNode *Extract = *UI;
11611 // cOMpute the element's address.
11612 SDValue Idx = Extract->getOperand(1);
11614 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
11615 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
11616 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
11618 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
11619 StackPtr, OffsetVal);
11621 // Load the scalar.
11622 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
11623 ScalarAddr, MachinePointerInfo(),
11626 // Replace the exact with the load.
11627 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
11630 // The replacement was made in place; don't return anything.
11634 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
11635 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
11636 const X86Subtarget *Subtarget) {
11637 DebugLoc DL = N->getDebugLoc();
11638 SDValue Cond = N->getOperand(0);
11639 // Get the LHS/RHS of the select.
11640 SDValue LHS = N->getOperand(1);
11641 SDValue RHS = N->getOperand(2);
11643 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
11644 // instructions match the semantics of the common C idiom x<y?x:y but not
11645 // x<=y?x:y, because of how they handle negative zero (which can be
11646 // ignored in unsafe-math mode).
11647 if (Subtarget->hasSSE2() &&
11648 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
11649 Cond.getOpcode() == ISD::SETCC) {
11650 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
11652 unsigned Opcode = 0;
11653 // Check for x CC y ? x : y.
11654 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
11655 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
11659 // Converting this to a min would handle NaNs incorrectly, and swapping
11660 // the operands would cause it to handle comparisons between positive
11661 // and negative zero incorrectly.
11662 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11663 if (!UnsafeFPMath &&
11664 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11666 std::swap(LHS, RHS);
11668 Opcode = X86ISD::FMIN;
11671 // Converting this to a min would handle comparisons between positive
11672 // and negative zero incorrectly.
11673 if (!UnsafeFPMath &&
11674 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
11676 Opcode = X86ISD::FMIN;
11679 // Converting this to a min would handle both negative zeros and NaNs
11680 // incorrectly, but we can swap the operands to fix both.
11681 std::swap(LHS, RHS);
11685 Opcode = X86ISD::FMIN;
11689 // Converting this to a max would handle comparisons between positive
11690 // and negative zero incorrectly.
11691 if (!UnsafeFPMath &&
11692 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(LHS))
11694 Opcode = X86ISD::FMAX;
11697 // Converting this to a max would handle NaNs incorrectly, and swapping
11698 // the operands would cause it to handle comparisons between positive
11699 // and negative zero incorrectly.
11700 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11701 if (!UnsafeFPMath &&
11702 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11704 std::swap(LHS, RHS);
11706 Opcode = X86ISD::FMAX;
11709 // Converting this to a max would handle both negative zeros and NaNs
11710 // incorrectly, but we can swap the operands to fix both.
11711 std::swap(LHS, RHS);
11715 Opcode = X86ISD::FMAX;
11718 // Check for x CC y ? y : x -- a min/max with reversed arms.
11719 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
11720 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
11724 // Converting this to a min would handle comparisons between positive
11725 // and negative zero incorrectly, and swapping the operands would
11726 // cause it to handle NaNs incorrectly.
11727 if (!UnsafeFPMath &&
11728 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
11729 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11731 std::swap(LHS, RHS);
11733 Opcode = X86ISD::FMIN;
11736 // Converting this to a min would handle NaNs incorrectly.
11737 if (!UnsafeFPMath &&
11738 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
11740 Opcode = X86ISD::FMIN;
11743 // Converting this to a min would handle both negative zeros and NaNs
11744 // incorrectly, but we can swap the operands to fix both.
11745 std::swap(LHS, RHS);
11749 Opcode = X86ISD::FMIN;
11753 // Converting this to a max would handle NaNs incorrectly.
11754 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11756 Opcode = X86ISD::FMAX;
11759 // Converting this to a max would handle comparisons between positive
11760 // and negative zero incorrectly, and swapping the operands would
11761 // cause it to handle NaNs incorrectly.
11762 if (!UnsafeFPMath &&
11763 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
11764 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
11766 std::swap(LHS, RHS);
11768 Opcode = X86ISD::FMAX;
11771 // Converting this to a max would handle both negative zeros and NaNs
11772 // incorrectly, but we can swap the operands to fix both.
11773 std::swap(LHS, RHS);
11777 Opcode = X86ISD::FMAX;
11783 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
11786 // If this is a select between two integer constants, try to do some
11788 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
11789 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
11790 // Don't do this for crazy integer types.
11791 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
11792 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
11793 // so that TrueC (the true value) is larger than FalseC.
11794 bool NeedsCondInvert = false;
11796 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
11797 // Efficiently invertible.
11798 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
11799 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
11800 isa<ConstantSDNode>(Cond.getOperand(1))))) {
11801 NeedsCondInvert = true;
11802 std::swap(TrueC, FalseC);
11805 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
11806 if (FalseC->getAPIntValue() == 0 &&
11807 TrueC->getAPIntValue().isPowerOf2()) {
11808 if (NeedsCondInvert) // Invert the condition if needed.
11809 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11810 DAG.getConstant(1, Cond.getValueType()));
11812 // Zero extend the condition if needed.
11813 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
11815 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11816 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
11817 DAG.getConstant(ShAmt, MVT::i8));
11820 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
11821 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11822 if (NeedsCondInvert) // Invert the condition if needed.
11823 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11824 DAG.getConstant(1, Cond.getValueType()));
11826 // Zero extend the condition if needed.
11827 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11828 FalseC->getValueType(0), Cond);
11829 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11830 SDValue(FalseC, 0));
11833 // Optimize cases that will turn into an LEA instruction. This requires
11834 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11835 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11836 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11837 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11839 bool isFastMultiplier = false;
11841 switch ((unsigned char)Diff) {
11843 case 1: // result = add base, cond
11844 case 2: // result = lea base( , cond*2)
11845 case 3: // result = lea base(cond, cond*2)
11846 case 4: // result = lea base( , cond*4)
11847 case 5: // result = lea base(cond, cond*4)
11848 case 8: // result = lea base( , cond*8)
11849 case 9: // result = lea base(cond, cond*8)
11850 isFastMultiplier = true;
11855 if (isFastMultiplier) {
11856 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11857 if (NeedsCondInvert) // Invert the condition if needed.
11858 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
11859 DAG.getConstant(1, Cond.getValueType()));
11861 // Zero extend the condition if needed.
11862 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11864 // Scale the condition by the difference.
11866 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11867 DAG.getConstant(Diff, Cond.getValueType()));
11869 // Add the base if non-zero.
11870 if (FalseC->getAPIntValue() != 0)
11871 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11872 SDValue(FalseC, 0));
11882 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
11883 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
11884 TargetLowering::DAGCombinerInfo &DCI) {
11885 DebugLoc DL = N->getDebugLoc();
11887 // If the flag operand isn't dead, don't touch this CMOV.
11888 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
11891 SDValue FalseOp = N->getOperand(0);
11892 SDValue TrueOp = N->getOperand(1);
11893 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
11894 SDValue Cond = N->getOperand(3);
11895 if (CC == X86::COND_E || CC == X86::COND_NE) {
11896 switch (Cond.getOpcode()) {
11900 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
11901 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
11902 return (CC == X86::COND_E) ? FalseOp : TrueOp;
11906 // If this is a select between two integer constants, try to do some
11907 // optimizations. Note that the operands are ordered the opposite of SELECT
11909 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
11910 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
11911 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
11912 // larger than FalseC (the false value).
11913 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
11914 CC = X86::GetOppositeBranchCondition(CC);
11915 std::swap(TrueC, FalseC);
11918 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
11919 // This is efficient for any integer data type (including i8/i16) and
11921 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
11922 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11923 DAG.getConstant(CC, MVT::i8), Cond);
11925 // Zero extend the condition if needed.
11926 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
11928 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
11929 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
11930 DAG.getConstant(ShAmt, MVT::i8));
11931 if (N->getNumValues() == 2) // Dead flag value?
11932 return DCI.CombineTo(N, Cond, SDValue());
11936 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
11937 // for any integer data type, including i8/i16.
11938 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
11939 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11940 DAG.getConstant(CC, MVT::i8), Cond);
11942 // Zero extend the condition if needed.
11943 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
11944 FalseC->getValueType(0), Cond);
11945 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11946 SDValue(FalseC, 0));
11948 if (N->getNumValues() == 2) // Dead flag value?
11949 return DCI.CombineTo(N, Cond, SDValue());
11953 // Optimize cases that will turn into an LEA instruction. This requires
11954 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
11955 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
11956 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
11957 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
11959 bool isFastMultiplier = false;
11961 switch ((unsigned char)Diff) {
11963 case 1: // result = add base, cond
11964 case 2: // result = lea base( , cond*2)
11965 case 3: // result = lea base(cond, cond*2)
11966 case 4: // result = lea base( , cond*4)
11967 case 5: // result = lea base(cond, cond*4)
11968 case 8: // result = lea base( , cond*8)
11969 case 9: // result = lea base(cond, cond*8)
11970 isFastMultiplier = true;
11975 if (isFastMultiplier) {
11976 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
11977 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
11978 DAG.getConstant(CC, MVT::i8), Cond);
11979 // Zero extend the condition if needed.
11980 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
11982 // Scale the condition by the difference.
11984 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
11985 DAG.getConstant(Diff, Cond.getValueType()));
11987 // Add the base if non-zero.
11988 if (FalseC->getAPIntValue() != 0)
11989 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
11990 SDValue(FalseC, 0));
11991 if (N->getNumValues() == 2) // Dead flag value?
11992 return DCI.CombineTo(N, Cond, SDValue());
12002 /// PerformMulCombine - Optimize a single multiply with constant into two
12003 /// in order to implement it with two cheaper instructions, e.g.
12004 /// LEA + SHL, LEA + LEA.
12005 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
12006 TargetLowering::DAGCombinerInfo &DCI) {
12007 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
12010 EVT VT = N->getValueType(0);
12011 if (VT != MVT::i64)
12014 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
12017 uint64_t MulAmt = C->getZExtValue();
12018 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
12021 uint64_t MulAmt1 = 0;
12022 uint64_t MulAmt2 = 0;
12023 if ((MulAmt % 9) == 0) {
12025 MulAmt2 = MulAmt / 9;
12026 } else if ((MulAmt % 5) == 0) {
12028 MulAmt2 = MulAmt / 5;
12029 } else if ((MulAmt % 3) == 0) {
12031 MulAmt2 = MulAmt / 3;
12034 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
12035 DebugLoc DL = N->getDebugLoc();
12037 if (isPowerOf2_64(MulAmt2) &&
12038 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
12039 // If second multiplifer is pow2, issue it first. We want the multiply by
12040 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
12042 std::swap(MulAmt1, MulAmt2);
12045 if (isPowerOf2_64(MulAmt1))
12046 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
12047 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
12049 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
12050 DAG.getConstant(MulAmt1, VT));
12052 if (isPowerOf2_64(MulAmt2))
12053 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
12054 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
12056 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
12057 DAG.getConstant(MulAmt2, VT));
12059 // Do not add new nodes to DAG combiner worklist.
12060 DCI.CombineTo(N, NewMul, false);
12065 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
12066 SDValue N0 = N->getOperand(0);
12067 SDValue N1 = N->getOperand(1);
12068 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
12069 EVT VT = N0.getValueType();
12071 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
12072 // since the result of setcc_c is all zero's or all ones.
12073 if (N1C && N0.getOpcode() == ISD::AND &&
12074 N0.getOperand(1).getOpcode() == ISD::Constant) {
12075 SDValue N00 = N0.getOperand(0);
12076 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
12077 ((N00.getOpcode() == ISD::ANY_EXTEND ||
12078 N00.getOpcode() == ISD::ZERO_EXTEND) &&
12079 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
12080 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
12081 APInt ShAmt = N1C->getAPIntValue();
12082 Mask = Mask.shl(ShAmt);
12084 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
12085 N00, DAG.getConstant(Mask, VT));
12092 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
12094 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
12095 const X86Subtarget *Subtarget) {
12096 EVT VT = N->getValueType(0);
12097 if (!VT.isVector() && VT.isInteger() &&
12098 N->getOpcode() == ISD::SHL)
12099 return PerformSHLCombine(N, DAG);
12101 // On X86 with SSE2 support, we can transform this to a vector shift if
12102 // all elements are shifted by the same amount. We can't do this in legalize
12103 // because the a constant vector is typically transformed to a constant pool
12104 // so we have no knowledge of the shift amount.
12105 if (!Subtarget->hasSSE2())
12108 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
12111 SDValue ShAmtOp = N->getOperand(1);
12112 EVT EltVT = VT.getVectorElementType();
12113 DebugLoc DL = N->getDebugLoc();
12114 SDValue BaseShAmt = SDValue();
12115 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
12116 unsigned NumElts = VT.getVectorNumElements();
12118 for (; i != NumElts; ++i) {
12119 SDValue Arg = ShAmtOp.getOperand(i);
12120 if (Arg.getOpcode() == ISD::UNDEF) continue;
12124 for (; i != NumElts; ++i) {
12125 SDValue Arg = ShAmtOp.getOperand(i);
12126 if (Arg.getOpcode() == ISD::UNDEF) continue;
12127 if (Arg != BaseShAmt) {
12131 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
12132 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
12133 SDValue InVec = ShAmtOp.getOperand(0);
12134 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
12135 unsigned NumElts = InVec.getValueType().getVectorNumElements();
12137 for (; i != NumElts; ++i) {
12138 SDValue Arg = InVec.getOperand(i);
12139 if (Arg.getOpcode() == ISD::UNDEF) continue;
12143 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
12144 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
12145 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
12146 if (C->getZExtValue() == SplatIdx)
12147 BaseShAmt = InVec.getOperand(1);
12150 if (BaseShAmt.getNode() == 0)
12151 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
12152 DAG.getIntPtrConstant(0));
12156 // The shift amount is an i32.
12157 if (EltVT.bitsGT(MVT::i32))
12158 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
12159 else if (EltVT.bitsLT(MVT::i32))
12160 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
12162 // The shift amount is identical so we can do a vector shift.
12163 SDValue ValOp = N->getOperand(0);
12164 switch (N->getOpcode()) {
12166 llvm_unreachable("Unknown shift opcode!");
12169 if (VT == MVT::v2i64)
12170 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12171 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
12173 if (VT == MVT::v4i32)
12174 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12175 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
12177 if (VT == MVT::v8i16)
12178 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12179 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
12183 if (VT == MVT::v4i32)
12184 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12185 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
12187 if (VT == MVT::v8i16)
12188 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12189 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
12193 if (VT == MVT::v2i64)
12194 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12195 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
12197 if (VT == MVT::v4i32)
12198 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12199 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
12201 if (VT == MVT::v8i16)
12202 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12203 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
12211 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
12212 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
12213 // and friends. Likewise for OR -> CMPNEQSS.
12214 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
12215 TargetLowering::DAGCombinerInfo &DCI,
12216 const X86Subtarget *Subtarget) {
12219 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
12220 // we're requiring SSE2 for both.
12221 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
12222 SDValue N0 = N->getOperand(0);
12223 SDValue N1 = N->getOperand(1);
12224 SDValue CMP0 = N0->getOperand(1);
12225 SDValue CMP1 = N1->getOperand(1);
12226 DebugLoc DL = N->getDebugLoc();
12228 // The SETCCs should both refer to the same CMP.
12229 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
12232 SDValue CMP00 = CMP0->getOperand(0);
12233 SDValue CMP01 = CMP0->getOperand(1);
12234 EVT VT = CMP00.getValueType();
12236 if (VT == MVT::f32 || VT == MVT::f64) {
12237 bool ExpectingFlags = false;
12238 // Check for any users that want flags:
12239 for (SDNode::use_iterator UI = N->use_begin(),
12241 !ExpectingFlags && UI != UE; ++UI)
12242 switch (UI->getOpcode()) {
12247 ExpectingFlags = true;
12249 case ISD::CopyToReg:
12250 case ISD::SIGN_EXTEND:
12251 case ISD::ZERO_EXTEND:
12252 case ISD::ANY_EXTEND:
12256 if (!ExpectingFlags) {
12257 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
12258 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
12260 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
12261 X86::CondCode tmp = cc0;
12266 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
12267 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
12268 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
12269 X86ISD::NodeType NTOperator = is64BitFP ?
12270 X86ISD::FSETCCsd : X86ISD::FSETCCss;
12271 // FIXME: need symbolic constants for these magic numbers.
12272 // See X86ATTInstPrinter.cpp:printSSECC().
12273 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
12274 SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
12275 DAG.getConstant(x86cc, MVT::i8));
12276 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
12278 SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
12279 DAG.getConstant(1, MVT::i32));
12280 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
12281 return OneBitOfTruth;
12289 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
12290 /// so it can be folded inside ANDNP.
12291 static bool CanFoldXORWithAllOnes(const SDNode *N) {
12292 EVT VT = N->getValueType(0);
12294 // Match direct AllOnes for 128 and 256-bit vectors
12295 if (ISD::isBuildVectorAllOnes(N))
12298 // Look through a bit convert.
12299 if (N->getOpcode() == ISD::BITCAST)
12300 N = N->getOperand(0).getNode();
12302 // Sometimes the operand may come from a insert_subvector building a 256-bit
12304 SDValue V1 = N->getOperand(0);
12305 SDValue V2 = N->getOperand(1);
12307 if (VT.getSizeInBits() == 256 &&
12308 N->getOpcode() == ISD::INSERT_SUBVECTOR &&
12309 V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
12310 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
12311 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
12312 ISD::isBuildVectorAllOnes(V2.getNode()))
12318 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
12319 TargetLowering::DAGCombinerInfo &DCI,
12320 const X86Subtarget *Subtarget) {
12321 if (DCI.isBeforeLegalizeOps())
12324 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
12328 // Want to form ANDNP nodes:
12329 // 1) In the hopes of then easily combining them with OR and AND nodes
12330 // to form PBLEND/PSIGN.
12331 // 2) To match ANDN packed intrinsics
12332 EVT VT = N->getValueType(0);
12333 if (VT != MVT::v2i64 && VT != MVT::v4i64)
12336 SDValue N0 = N->getOperand(0);
12337 SDValue N1 = N->getOperand(1);
12338 DebugLoc DL = N->getDebugLoc();
12340 // Check LHS for vnot
12341 if (N0.getOpcode() == ISD::XOR &&
12342 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
12343 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
12344 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
12346 // Check RHS for vnot
12347 if (N1.getOpcode() == ISD::XOR &&
12348 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
12349 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
12350 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
12355 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
12356 TargetLowering::DAGCombinerInfo &DCI,
12357 const X86Subtarget *Subtarget) {
12358 if (DCI.isBeforeLegalizeOps())
12361 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
12365 EVT VT = N->getValueType(0);
12366 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
12369 SDValue N0 = N->getOperand(0);
12370 SDValue N1 = N->getOperand(1);
12372 // look for psign/blend
12373 if (Subtarget->hasSSSE3()) {
12374 if (VT == MVT::v2i64) {
12375 // Canonicalize pandn to RHS
12376 if (N0.getOpcode() == X86ISD::ANDNP)
12378 // or (and (m, x), (pandn m, y))
12379 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
12380 SDValue Mask = N1.getOperand(0);
12381 SDValue X = N1.getOperand(1);
12383 if (N0.getOperand(0) == Mask)
12384 Y = N0.getOperand(1);
12385 if (N0.getOperand(1) == Mask)
12386 Y = N0.getOperand(0);
12388 // Check to see if the mask appeared in both the AND and ANDNP and
12392 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
12393 if (Mask.getOpcode() != ISD::BITCAST ||
12394 X.getOpcode() != ISD::BITCAST ||
12395 Y.getOpcode() != ISD::BITCAST)
12398 // Look through mask bitcast.
12399 Mask = Mask.getOperand(0);
12400 EVT MaskVT = Mask.getValueType();
12402 // Validate that the Mask operand is a vector sra node. The sra node
12403 // will be an intrinsic.
12404 if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
12407 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
12408 // there is no psrai.b
12409 switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
12410 case Intrinsic::x86_sse2_psrai_w:
12411 case Intrinsic::x86_sse2_psrai_d:
12413 default: return SDValue();
12416 // Check that the SRA is all signbits.
12417 SDValue SraC = Mask.getOperand(2);
12418 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
12419 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
12420 if ((SraAmt + 1) != EltBits)
12423 DebugLoc DL = N->getDebugLoc();
12425 // Now we know we at least have a plendvb with the mask val. See if
12426 // we can form a psignb/w/d.
12427 // psign = x.type == y.type == mask.type && y = sub(0, x);
12428 X = X.getOperand(0);
12429 Y = Y.getOperand(0);
12430 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
12431 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
12432 X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
12435 case 8: Opc = X86ISD::PSIGNB; break;
12436 case 16: Opc = X86ISD::PSIGNW; break;
12437 case 32: Opc = X86ISD::PSIGND; break;
12441 SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
12442 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
12445 // PBLENDVB only available on SSE 4.1
12446 if (!Subtarget->hasSSE41())
12449 X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
12450 Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
12451 Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
12452 Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
12453 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
12458 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
12459 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
12461 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
12463 if (!N0.hasOneUse() || !N1.hasOneUse())
12466 SDValue ShAmt0 = N0.getOperand(1);
12467 if (ShAmt0.getValueType() != MVT::i8)
12469 SDValue ShAmt1 = N1.getOperand(1);
12470 if (ShAmt1.getValueType() != MVT::i8)
12472 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
12473 ShAmt0 = ShAmt0.getOperand(0);
12474 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
12475 ShAmt1 = ShAmt1.getOperand(0);
12477 DebugLoc DL = N->getDebugLoc();
12478 unsigned Opc = X86ISD::SHLD;
12479 SDValue Op0 = N0.getOperand(0);
12480 SDValue Op1 = N1.getOperand(0);
12481 if (ShAmt0.getOpcode() == ISD::SUB) {
12482 Opc = X86ISD::SHRD;
12483 std::swap(Op0, Op1);
12484 std::swap(ShAmt0, ShAmt1);
12487 unsigned Bits = VT.getSizeInBits();
12488 if (ShAmt1.getOpcode() == ISD::SUB) {
12489 SDValue Sum = ShAmt1.getOperand(0);
12490 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
12491 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
12492 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
12493 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
12494 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
12495 return DAG.getNode(Opc, DL, VT,
12497 DAG.getNode(ISD::TRUNCATE, DL,
12500 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
12501 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
12503 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
12504 return DAG.getNode(Opc, DL, VT,
12505 N0.getOperand(0), N1.getOperand(0),
12506 DAG.getNode(ISD::TRUNCATE, DL,
12513 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
12514 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
12515 const X86Subtarget *Subtarget) {
12516 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
12517 // the FP state in cases where an emms may be missing.
12518 // A preferable solution to the general problem is to figure out the right
12519 // places to insert EMMS. This qualifies as a quick hack.
12521 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
12522 StoreSDNode *St = cast<StoreSDNode>(N);
12523 EVT VT = St->getValue().getValueType();
12524 if (VT.getSizeInBits() != 64)
12527 const Function *F = DAG.getMachineFunction().getFunction();
12528 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
12529 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
12530 && Subtarget->hasSSE2();
12531 if ((VT.isVector() ||
12532 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
12533 isa<LoadSDNode>(St->getValue()) &&
12534 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
12535 St->getChain().hasOneUse() && !St->isVolatile()) {
12536 SDNode* LdVal = St->getValue().getNode();
12537 LoadSDNode *Ld = 0;
12538 int TokenFactorIndex = -1;
12539 SmallVector<SDValue, 8> Ops;
12540 SDNode* ChainVal = St->getChain().getNode();
12541 // Must be a store of a load. We currently handle two cases: the load
12542 // is a direct child, and it's under an intervening TokenFactor. It is
12543 // possible to dig deeper under nested TokenFactors.
12544 if (ChainVal == LdVal)
12545 Ld = cast<LoadSDNode>(St->getChain());
12546 else if (St->getValue().hasOneUse() &&
12547 ChainVal->getOpcode() == ISD::TokenFactor) {
12548 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
12549 if (ChainVal->getOperand(i).getNode() == LdVal) {
12550 TokenFactorIndex = i;
12551 Ld = cast<LoadSDNode>(St->getValue());
12553 Ops.push_back(ChainVal->getOperand(i));
12557 if (!Ld || !ISD::isNormalLoad(Ld))
12560 // If this is not the MMX case, i.e. we are just turning i64 load/store
12561 // into f64 load/store, avoid the transformation if there are multiple
12562 // uses of the loaded value.
12563 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
12566 DebugLoc LdDL = Ld->getDebugLoc();
12567 DebugLoc StDL = N->getDebugLoc();
12568 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
12569 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
12571 if (Subtarget->is64Bit() || F64IsLegal) {
12572 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
12573 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
12574 Ld->getPointerInfo(), Ld->isVolatile(),
12575 Ld->isNonTemporal(), Ld->getAlignment());
12576 SDValue NewChain = NewLd.getValue(1);
12577 if (TokenFactorIndex != -1) {
12578 Ops.push_back(NewChain);
12579 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12582 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
12583 St->getPointerInfo(),
12584 St->isVolatile(), St->isNonTemporal(),
12585 St->getAlignment());
12588 // Otherwise, lower to two pairs of 32-bit loads / stores.
12589 SDValue LoAddr = Ld->getBasePtr();
12590 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
12591 DAG.getConstant(4, MVT::i32));
12593 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
12594 Ld->getPointerInfo(),
12595 Ld->isVolatile(), Ld->isNonTemporal(),
12596 Ld->getAlignment());
12597 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
12598 Ld->getPointerInfo().getWithOffset(4),
12599 Ld->isVolatile(), Ld->isNonTemporal(),
12600 MinAlign(Ld->getAlignment(), 4));
12602 SDValue NewChain = LoLd.getValue(1);
12603 if (TokenFactorIndex != -1) {
12604 Ops.push_back(LoLd);
12605 Ops.push_back(HiLd);
12606 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12610 LoAddr = St->getBasePtr();
12611 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
12612 DAG.getConstant(4, MVT::i32));
12614 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
12615 St->getPointerInfo(),
12616 St->isVolatile(), St->isNonTemporal(),
12617 St->getAlignment());
12618 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
12619 St->getPointerInfo().getWithOffset(4),
12621 St->isNonTemporal(),
12622 MinAlign(St->getAlignment(), 4));
12623 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
12628 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
12629 /// X86ISD::FXOR nodes.
12630 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
12631 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
12632 // F[X]OR(0.0, x) -> x
12633 // F[X]OR(x, 0.0) -> x
12634 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
12635 if (C->getValueAPF().isPosZero())
12636 return N->getOperand(1);
12637 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
12638 if (C->getValueAPF().isPosZero())
12639 return N->getOperand(0);
12643 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
12644 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
12645 // FAND(0.0, x) -> 0.0
12646 // FAND(x, 0.0) -> 0.0
12647 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
12648 if (C->getValueAPF().isPosZero())
12649 return N->getOperand(0);
12650 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
12651 if (C->getValueAPF().isPosZero())
12652 return N->getOperand(1);
12656 static SDValue PerformBTCombine(SDNode *N,
12658 TargetLowering::DAGCombinerInfo &DCI) {
12659 // BT ignores high bits in the bit index operand.
12660 SDValue Op1 = N->getOperand(1);
12661 if (Op1.hasOneUse()) {
12662 unsigned BitWidth = Op1.getValueSizeInBits();
12663 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
12664 APInt KnownZero, KnownOne;
12665 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
12666 !DCI.isBeforeLegalizeOps());
12667 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12668 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
12669 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
12670 DCI.CommitTargetLoweringOpt(TLO);
12675 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
12676 SDValue Op = N->getOperand(0);
12677 if (Op.getOpcode() == ISD::BITCAST)
12678 Op = Op.getOperand(0);
12679 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
12680 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
12681 VT.getVectorElementType().getSizeInBits() ==
12682 OpVT.getVectorElementType().getSizeInBits()) {
12683 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
12688 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
12689 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
12690 // (and (i32 x86isd::setcc_carry), 1)
12691 // This eliminates the zext. This transformation is necessary because
12692 // ISD::SETCC is always legalized to i8.
12693 DebugLoc dl = N->getDebugLoc();
12694 SDValue N0 = N->getOperand(0);
12695 EVT VT = N->getValueType(0);
12696 if (N0.getOpcode() == ISD::AND &&
12698 N0.getOperand(0).hasOneUse()) {
12699 SDValue N00 = N0.getOperand(0);
12700 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
12702 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
12703 if (!C || C->getZExtValue() != 1)
12705 return DAG.getNode(ISD::AND, dl, VT,
12706 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
12707 N00.getOperand(0), N00.getOperand(1)),
12708 DAG.getConstant(1, VT));
12714 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
12715 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
12716 unsigned X86CC = N->getConstantOperandVal(0);
12717 SDValue EFLAG = N->getOperand(1);
12718 DebugLoc DL = N->getDebugLoc();
12720 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
12721 // a zext and produces an all-ones bit which is more useful than 0/1 in some
12723 if (X86CC == X86::COND_B)
12724 return DAG.getNode(ISD::AND, DL, MVT::i8,
12725 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
12726 DAG.getConstant(X86CC, MVT::i8), EFLAG),
12727 DAG.getConstant(1, MVT::i8));
12732 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
12733 const X86TargetLowering *XTLI) {
12734 SDValue Op0 = N->getOperand(0);
12735 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
12736 // a 32-bit target where SSE doesn't support i64->FP operations.
12737 if (Op0.getOpcode() == ISD::LOAD) {
12738 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
12739 EVT VT = Ld->getValueType(0);
12740 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
12741 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
12742 !XTLI->getSubtarget()->is64Bit() &&
12743 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
12744 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
12745 Ld->getChain(), Op0, DAG);
12746 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
12753 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
12754 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
12755 X86TargetLowering::DAGCombinerInfo &DCI) {
12756 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
12757 // the result is either zero or one (depending on the input carry bit).
12758 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
12759 if (X86::isZeroNode(N->getOperand(0)) &&
12760 X86::isZeroNode(N->getOperand(1)) &&
12761 // We don't have a good way to replace an EFLAGS use, so only do this when
12763 SDValue(N, 1).use_empty()) {
12764 DebugLoc DL = N->getDebugLoc();
12765 EVT VT = N->getValueType(0);
12766 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
12767 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
12768 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
12769 DAG.getConstant(X86::COND_B,MVT::i8),
12771 DAG.getConstant(1, VT));
12772 return DCI.CombineTo(N, Res1, CarryOut);
12778 // fold (add Y, (sete X, 0)) -> adc 0, Y
12779 // (add Y, (setne X, 0)) -> sbb -1, Y
12780 // (sub (sete X, 0), Y) -> sbb 0, Y
12781 // (sub (setne X, 0), Y) -> adc -1, Y
12782 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
12783 DebugLoc DL = N->getDebugLoc();
12785 // Look through ZExts.
12786 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
12787 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
12790 SDValue SetCC = Ext.getOperand(0);
12791 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
12794 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
12795 if (CC != X86::COND_E && CC != X86::COND_NE)
12798 SDValue Cmp = SetCC.getOperand(1);
12799 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
12800 !X86::isZeroNode(Cmp.getOperand(1)) ||
12801 !Cmp.getOperand(0).getValueType().isInteger())
12804 SDValue CmpOp0 = Cmp.getOperand(0);
12805 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
12806 DAG.getConstant(1, CmpOp0.getValueType()));
12808 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
12809 if (CC == X86::COND_NE)
12810 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
12811 DL, OtherVal.getValueType(), OtherVal,
12812 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
12813 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
12814 DL, OtherVal.getValueType(), OtherVal,
12815 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
12818 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG) {
12819 SDValue Op0 = N->getOperand(0);
12820 SDValue Op1 = N->getOperand(1);
12822 // X86 can't encode an immediate LHS of a sub. See if we can push the
12823 // negation into a preceding instruction.
12824 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
12825 uint64_t Op0C = C->getSExtValue();
12827 // If the RHS of the sub is a XOR with one use and a constant, invert the
12828 // immediate. Then add one to the LHS of the sub so we can turn
12829 // X-Y -> X+~Y+1, saving one register.
12830 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
12831 isa<ConstantSDNode>(Op1.getOperand(1))) {
12832 uint64_t XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getSExtValue();
12833 EVT VT = Op0.getValueType();
12834 SDValue NewXor = DAG.getNode(ISD::XOR, Op1.getDebugLoc(), VT,
12836 DAG.getConstant(~XorC, VT));
12837 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, NewXor,
12838 DAG.getConstant(Op0C+1, VT));
12842 return OptimizeConditionalInDecrement(N, DAG);
12845 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
12846 DAGCombinerInfo &DCI) const {
12847 SelectionDAG &DAG = DCI.DAG;
12848 switch (N->getOpcode()) {
12850 case ISD::EXTRACT_VECTOR_ELT:
12851 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
12852 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
12853 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
12854 case ISD::ADD: return OptimizeConditionalInDecrement(N, DAG);
12855 case ISD::SUB: return PerformSubCombine(N, DAG);
12856 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
12857 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
12860 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
12861 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
12862 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
12863 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
12864 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
12866 case X86ISD::FOR: return PerformFORCombine(N, DAG);
12867 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
12868 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
12869 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
12870 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
12871 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
12872 case X86ISD::SHUFPS: // Handle all target specific shuffles
12873 case X86ISD::SHUFPD:
12874 case X86ISD::PALIGN:
12875 case X86ISD::PUNPCKHBW:
12876 case X86ISD::PUNPCKHWD:
12877 case X86ISD::PUNPCKHDQ:
12878 case X86ISD::PUNPCKHQDQ:
12879 case X86ISD::UNPCKHPS:
12880 case X86ISD::UNPCKHPD:
12881 case X86ISD::VUNPCKHPSY:
12882 case X86ISD::VUNPCKHPDY:
12883 case X86ISD::PUNPCKLBW:
12884 case X86ISD::PUNPCKLWD:
12885 case X86ISD::PUNPCKLDQ:
12886 case X86ISD::PUNPCKLQDQ:
12887 case X86ISD::UNPCKLPS:
12888 case X86ISD::UNPCKLPD:
12889 case X86ISD::VUNPCKLPSY:
12890 case X86ISD::VUNPCKLPDY:
12891 case X86ISD::MOVHLPS:
12892 case X86ISD::MOVLHPS:
12893 case X86ISD::PSHUFD:
12894 case X86ISD::PSHUFHW:
12895 case X86ISD::PSHUFLW:
12896 case X86ISD::MOVSS:
12897 case X86ISD::MOVSD:
12898 case X86ISD::VPERMILPS:
12899 case X86ISD::VPERMILPSY:
12900 case X86ISD::VPERMILPD:
12901 case X86ISD::VPERMILPDY:
12902 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI);
12908 /// isTypeDesirableForOp - Return true if the target has native support for
12909 /// the specified value type and it is 'desirable' to use the type for the
12910 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
12911 /// instruction encodings are longer and some i16 instructions are slow.
12912 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
12913 if (!isTypeLegal(VT))
12915 if (VT != MVT::i16)
12922 case ISD::SIGN_EXTEND:
12923 case ISD::ZERO_EXTEND:
12924 case ISD::ANY_EXTEND:
12937 /// IsDesirableToPromoteOp - This method query the target whether it is
12938 /// beneficial for dag combiner to promote the specified node. If true, it
12939 /// should return the desired promotion type by reference.
12940 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
12941 EVT VT = Op.getValueType();
12942 if (VT != MVT::i16)
12945 bool Promote = false;
12946 bool Commute = false;
12947 switch (Op.getOpcode()) {
12950 LoadSDNode *LD = cast<LoadSDNode>(Op);
12951 // If the non-extending load has a single use and it's not live out, then it
12952 // might be folded.
12953 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
12954 Op.hasOneUse()*/) {
12955 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
12956 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
12957 // The only case where we'd want to promote LOAD (rather then it being
12958 // promoted as an operand is when it's only use is liveout.
12959 if (UI->getOpcode() != ISD::CopyToReg)
12966 case ISD::SIGN_EXTEND:
12967 case ISD::ZERO_EXTEND:
12968 case ISD::ANY_EXTEND:
12973 SDValue N0 = Op.getOperand(0);
12974 // Look out for (store (shl (load), x)).
12975 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
12988 SDValue N0 = Op.getOperand(0);
12989 SDValue N1 = Op.getOperand(1);
12990 if (!Commute && MayFoldLoad(N1))
12992 // Avoid disabling potential load folding opportunities.
12993 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
12995 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
13005 //===----------------------------------------------------------------------===//
13006 // X86 Inline Assembly Support
13007 //===----------------------------------------------------------------------===//
13009 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
13010 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
13012 std::string AsmStr = IA->getAsmString();
13014 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
13015 SmallVector<StringRef, 4> AsmPieces;
13016 SplitString(AsmStr, AsmPieces, ";\n");
13018 switch (AsmPieces.size()) {
13019 default: return false;
13021 AsmStr = AsmPieces[0];
13023 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
13025 // FIXME: this should verify that we are targeting a 486 or better. If not,
13026 // we will turn this bswap into something that will be lowered to logical ops
13027 // instead of emitting the bswap asm. For now, we don't support 486 or lower
13028 // so don't worry about this.
13030 if (AsmPieces.size() == 2 &&
13031 (AsmPieces[0] == "bswap" ||
13032 AsmPieces[0] == "bswapq" ||
13033 AsmPieces[0] == "bswapl") &&
13034 (AsmPieces[1] == "$0" ||
13035 AsmPieces[1] == "${0:q}")) {
13036 // No need to check constraints, nothing other than the equivalent of
13037 // "=r,0" would be valid here.
13038 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
13039 if (!Ty || Ty->getBitWidth() % 16 != 0)
13041 return IntrinsicLowering::LowerToByteSwap(CI);
13043 // rorw $$8, ${0:w} --> llvm.bswap.i16
13044 if (CI->getType()->isIntegerTy(16) &&
13045 AsmPieces.size() == 3 &&
13046 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
13047 AsmPieces[1] == "$$8," &&
13048 AsmPieces[2] == "${0:w}" &&
13049 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
13051 const std::string &ConstraintsStr = IA->getConstraintString();
13052 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
13053 std::sort(AsmPieces.begin(), AsmPieces.end());
13054 if (AsmPieces.size() == 4 &&
13055 AsmPieces[0] == "~{cc}" &&
13056 AsmPieces[1] == "~{dirflag}" &&
13057 AsmPieces[2] == "~{flags}" &&
13058 AsmPieces[3] == "~{fpsr}") {
13059 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
13060 if (!Ty || Ty->getBitWidth() % 16 != 0)
13062 return IntrinsicLowering::LowerToByteSwap(CI);
13067 if (CI->getType()->isIntegerTy(32) &&
13068 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
13069 SmallVector<StringRef, 4> Words;
13070 SplitString(AsmPieces[0], Words, " \t,");
13071 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
13072 Words[2] == "${0:w}") {
13074 SplitString(AsmPieces[1], Words, " \t,");
13075 if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
13076 Words[2] == "$0") {
13078 SplitString(AsmPieces[2], Words, " \t,");
13079 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
13080 Words[2] == "${0:w}") {
13082 const std::string &ConstraintsStr = IA->getConstraintString();
13083 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
13084 std::sort(AsmPieces.begin(), AsmPieces.end());
13085 if (AsmPieces.size() == 4 &&
13086 AsmPieces[0] == "~{cc}" &&
13087 AsmPieces[1] == "~{dirflag}" &&
13088 AsmPieces[2] == "~{flags}" &&
13089 AsmPieces[3] == "~{fpsr}") {
13090 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
13091 if (!Ty || Ty->getBitWidth() % 16 != 0)
13093 return IntrinsicLowering::LowerToByteSwap(CI);
13100 if (CI->getType()->isIntegerTy(64)) {
13101 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
13102 if (Constraints.size() >= 2 &&
13103 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
13104 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
13105 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
13106 SmallVector<StringRef, 4> Words;
13107 SplitString(AsmPieces[0], Words, " \t");
13108 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
13110 SplitString(AsmPieces[1], Words, " \t");
13111 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
13113 SplitString(AsmPieces[2], Words, " \t,");
13114 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
13115 Words[2] == "%edx") {
13116 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
13117 if (!Ty || Ty->getBitWidth() % 16 != 0)
13119 return IntrinsicLowering::LowerToByteSwap(CI);
13132 /// getConstraintType - Given a constraint letter, return the type of
13133 /// constraint it is for this target.
13134 X86TargetLowering::ConstraintType
13135 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
13136 if (Constraint.size() == 1) {
13137 switch (Constraint[0]) {
13148 return C_RegisterClass;
13172 return TargetLowering::getConstraintType(Constraint);
13175 /// Examine constraint type and operand type and determine a weight value.
13176 /// This object must already have been set up with the operand type
13177 /// and the current alternative constraint selected.
13178 TargetLowering::ConstraintWeight
13179 X86TargetLowering::getSingleConstraintMatchWeight(
13180 AsmOperandInfo &info, const char *constraint) const {
13181 ConstraintWeight weight = CW_Invalid;
13182 Value *CallOperandVal = info.CallOperandVal;
13183 // If we don't have a value, we can't do a match,
13184 // but allow it at the lowest weight.
13185 if (CallOperandVal == NULL)
13187 Type *type = CallOperandVal->getType();
13188 // Look at the constraint type.
13189 switch (*constraint) {
13191 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
13202 if (CallOperandVal->getType()->isIntegerTy())
13203 weight = CW_SpecificReg;
13208 if (type->isFloatingPointTy())
13209 weight = CW_SpecificReg;
13212 if (type->isX86_MMXTy() && Subtarget->hasMMX())
13213 weight = CW_SpecificReg;
13217 if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
13218 weight = CW_Register;
13221 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
13222 if (C->getZExtValue() <= 31)
13223 weight = CW_Constant;
13227 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13228 if (C->getZExtValue() <= 63)
13229 weight = CW_Constant;
13233 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13234 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
13235 weight = CW_Constant;
13239 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13240 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
13241 weight = CW_Constant;
13245 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13246 if (C->getZExtValue() <= 3)
13247 weight = CW_Constant;
13251 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13252 if (C->getZExtValue() <= 0xff)
13253 weight = CW_Constant;
13258 if (dyn_cast<ConstantFP>(CallOperandVal)) {
13259 weight = CW_Constant;
13263 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13264 if ((C->getSExtValue() >= -0x80000000LL) &&
13265 (C->getSExtValue() <= 0x7fffffffLL))
13266 weight = CW_Constant;
13270 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13271 if (C->getZExtValue() <= 0xffffffff)
13272 weight = CW_Constant;
13279 /// LowerXConstraint - try to replace an X constraint, which matches anything,
13280 /// with another that has more specific requirements based on the type of the
13281 /// corresponding operand.
13282 const char *X86TargetLowering::
13283 LowerXConstraint(EVT ConstraintVT) const {
13284 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
13285 // 'f' like normal targets.
13286 if (ConstraintVT.isFloatingPoint()) {
13287 if (Subtarget->hasXMMInt())
13289 if (Subtarget->hasXMM())
13293 return TargetLowering::LowerXConstraint(ConstraintVT);
13296 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
13297 /// vector. If it is invalid, don't add anything to Ops.
13298 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
13299 std::string &Constraint,
13300 std::vector<SDValue>&Ops,
13301 SelectionDAG &DAG) const {
13302 SDValue Result(0, 0);
13304 // Only support length 1 constraints for now.
13305 if (Constraint.length() > 1) return;
13307 char ConstraintLetter = Constraint[0];
13308 switch (ConstraintLetter) {
13311 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13312 if (C->getZExtValue() <= 31) {
13313 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13319 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13320 if (C->getZExtValue() <= 63) {
13321 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13327 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13328 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
13329 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13335 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13336 if (C->getZExtValue() <= 255) {
13337 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13343 // 32-bit signed value
13344 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13345 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
13346 C->getSExtValue())) {
13347 // Widen to 64 bits here to get it sign extended.
13348 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
13351 // FIXME gcc accepts some relocatable values here too, but only in certain
13352 // memory models; it's complicated.
13357 // 32-bit unsigned value
13358 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13359 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
13360 C->getZExtValue())) {
13361 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13365 // FIXME gcc accepts some relocatable values here too, but only in certain
13366 // memory models; it's complicated.
13370 // Literal immediates are always ok.
13371 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
13372 // Widen to 64 bits here to get it sign extended.
13373 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
13377 // In any sort of PIC mode addresses need to be computed at runtime by
13378 // adding in a register or some sort of table lookup. These can't
13379 // be used as immediates.
13380 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
13383 // If we are in non-pic codegen mode, we allow the address of a global (with
13384 // an optional displacement) to be used with 'i'.
13385 GlobalAddressSDNode *GA = 0;
13386 int64_t Offset = 0;
13388 // Match either (GA), (GA+C), (GA+C1+C2), etc.
13390 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
13391 Offset += GA->getOffset();
13393 } else if (Op.getOpcode() == ISD::ADD) {
13394 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
13395 Offset += C->getZExtValue();
13396 Op = Op.getOperand(0);
13399 } else if (Op.getOpcode() == ISD::SUB) {
13400 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
13401 Offset += -C->getZExtValue();
13402 Op = Op.getOperand(0);
13407 // Otherwise, this isn't something we can handle, reject it.
13411 const GlobalValue *GV = GA->getGlobal();
13412 // If we require an extra load to get this address, as in PIC mode, we
13413 // can't accept it.
13414 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
13415 getTargetMachine())))
13418 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
13419 GA->getValueType(0), Offset);
13424 if (Result.getNode()) {
13425 Ops.push_back(Result);
13428 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
13431 std::pair<unsigned, const TargetRegisterClass*>
13432 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
13434 // First, see if this is a constraint that directly corresponds to an LLVM
13436 if (Constraint.size() == 1) {
13437 // GCC Constraint Letters
13438 switch (Constraint[0]) {
13440 // TODO: Slight differences here in allocation order and leaving
13441 // RIP in the class. Do they matter any more here than they do
13442 // in the normal allocation?
13443 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
13444 if (Subtarget->is64Bit()) {
13445 if (VT == MVT::i32 || VT == MVT::f32)
13446 return std::make_pair(0U, X86::GR32RegisterClass);
13447 else if (VT == MVT::i16)
13448 return std::make_pair(0U, X86::GR16RegisterClass);
13449 else if (VT == MVT::i8 || VT == MVT::i1)
13450 return std::make_pair(0U, X86::GR8RegisterClass);
13451 else if (VT == MVT::i64 || VT == MVT::f64)
13452 return std::make_pair(0U, X86::GR64RegisterClass);
13455 // 32-bit fallthrough
13456 case 'Q': // Q_REGS
13457 if (VT == MVT::i32 || VT == MVT::f32)
13458 return std::make_pair(0U, X86::GR32_ABCDRegisterClass);
13459 else if (VT == MVT::i16)
13460 return std::make_pair(0U, X86::GR16_ABCDRegisterClass);
13461 else if (VT == MVT::i8 || VT == MVT::i1)
13462 return std::make_pair(0U, X86::GR8_ABCD_LRegisterClass);
13463 else if (VT == MVT::i64)
13464 return std::make_pair(0U, X86::GR64_ABCDRegisterClass);
13466 case 'r': // GENERAL_REGS
13467 case 'l': // INDEX_REGS
13468 if (VT == MVT::i8 || VT == MVT::i1)
13469 return std::make_pair(0U, X86::GR8RegisterClass);
13470 if (VT == MVT::i16)
13471 return std::make_pair(0U, X86::GR16RegisterClass);
13472 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
13473 return std::make_pair(0U, X86::GR32RegisterClass);
13474 return std::make_pair(0U, X86::GR64RegisterClass);
13475 case 'R': // LEGACY_REGS
13476 if (VT == MVT::i8 || VT == MVT::i1)
13477 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
13478 if (VT == MVT::i16)
13479 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
13480 if (VT == MVT::i32 || !Subtarget->is64Bit())
13481 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
13482 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
13483 case 'f': // FP Stack registers.
13484 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
13485 // value to the correct fpstack register class.
13486 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
13487 return std::make_pair(0U, X86::RFP32RegisterClass);
13488 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
13489 return std::make_pair(0U, X86::RFP64RegisterClass);
13490 return std::make_pair(0U, X86::RFP80RegisterClass);
13491 case 'y': // MMX_REGS if MMX allowed.
13492 if (!Subtarget->hasMMX()) break;
13493 return std::make_pair(0U, X86::VR64RegisterClass);
13494 case 'Y': // SSE_REGS if SSE2 allowed
13495 if (!Subtarget->hasXMMInt()) break;
13497 case 'x': // SSE_REGS if SSE1 allowed
13498 if (!Subtarget->hasXMM()) break;
13500 switch (VT.getSimpleVT().SimpleTy) {
13502 // Scalar SSE types.
13505 return std::make_pair(0U, X86::FR32RegisterClass);
13508 return std::make_pair(0U, X86::FR64RegisterClass);
13516 return std::make_pair(0U, X86::VR128RegisterClass);
13522 // Use the default implementation in TargetLowering to convert the register
13523 // constraint into a member of a register class.
13524 std::pair<unsigned, const TargetRegisterClass*> Res;
13525 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
13527 // Not found as a standard register?
13528 if (Res.second == 0) {
13529 // Map st(0) -> st(7) -> ST0
13530 if (Constraint.size() == 7 && Constraint[0] == '{' &&
13531 tolower(Constraint[1]) == 's' &&
13532 tolower(Constraint[2]) == 't' &&
13533 Constraint[3] == '(' &&
13534 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
13535 Constraint[5] == ')' &&
13536 Constraint[6] == '}') {
13538 Res.first = X86::ST0+Constraint[4]-'0';
13539 Res.second = X86::RFP80RegisterClass;
13543 // GCC allows "st(0)" to be called just plain "st".
13544 if (StringRef("{st}").equals_lower(Constraint)) {
13545 Res.first = X86::ST0;
13546 Res.second = X86::RFP80RegisterClass;
13551 if (StringRef("{flags}").equals_lower(Constraint)) {
13552 Res.first = X86::EFLAGS;
13553 Res.second = X86::CCRRegisterClass;
13557 // 'A' means EAX + EDX.
13558 if (Constraint == "A") {
13559 Res.first = X86::EAX;
13560 Res.second = X86::GR32_ADRegisterClass;
13566 // Otherwise, check to see if this is a register class of the wrong value
13567 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
13568 // turn into {ax},{dx}.
13569 if (Res.second->hasType(VT))
13570 return Res; // Correct type already, nothing to do.
13572 // All of the single-register GCC register classes map their values onto
13573 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
13574 // really want an 8-bit or 32-bit register, map to the appropriate register
13575 // class and return the appropriate register.
13576 if (Res.second == X86::GR16RegisterClass) {
13577 if (VT == MVT::i8) {
13578 unsigned DestReg = 0;
13579 switch (Res.first) {
13581 case X86::AX: DestReg = X86::AL; break;
13582 case X86::DX: DestReg = X86::DL; break;
13583 case X86::CX: DestReg = X86::CL; break;
13584 case X86::BX: DestReg = X86::BL; break;
13587 Res.first = DestReg;
13588 Res.second = X86::GR8RegisterClass;
13590 } else if (VT == MVT::i32) {
13591 unsigned DestReg = 0;
13592 switch (Res.first) {
13594 case X86::AX: DestReg = X86::EAX; break;
13595 case X86::DX: DestReg = X86::EDX; break;
13596 case X86::CX: DestReg = X86::ECX; break;
13597 case X86::BX: DestReg = X86::EBX; break;
13598 case X86::SI: DestReg = X86::ESI; break;
13599 case X86::DI: DestReg = X86::EDI; break;
13600 case X86::BP: DestReg = X86::EBP; break;
13601 case X86::SP: DestReg = X86::ESP; break;
13604 Res.first = DestReg;
13605 Res.second = X86::GR32RegisterClass;
13607 } else if (VT == MVT::i64) {
13608 unsigned DestReg = 0;
13609 switch (Res.first) {
13611 case X86::AX: DestReg = X86::RAX; break;
13612 case X86::DX: DestReg = X86::RDX; break;
13613 case X86::CX: DestReg = X86::RCX; break;
13614 case X86::BX: DestReg = X86::RBX; break;
13615 case X86::SI: DestReg = X86::RSI; break;
13616 case X86::DI: DestReg = X86::RDI; break;
13617 case X86::BP: DestReg = X86::RBP; break;
13618 case X86::SP: DestReg = X86::RSP; break;
13621 Res.first = DestReg;
13622 Res.second = X86::GR64RegisterClass;
13625 } else if (Res.second == X86::FR32RegisterClass ||
13626 Res.second == X86::FR64RegisterClass ||
13627 Res.second == X86::VR128RegisterClass) {
13628 // Handle references to XMM physical registers that got mapped into the
13629 // wrong class. This can happen with constraints like {xmm0} where the
13630 // target independent register mapper will just pick the first match it can
13631 // find, ignoring the required type.
13632 if (VT == MVT::f32)
13633 Res.second = X86::FR32RegisterClass;
13634 else if (VT == MVT::f64)
13635 Res.second = X86::FR64RegisterClass;
13636 else if (X86::VR128RegisterClass->hasType(VT))
13637 Res.second = X86::VR128RegisterClass;