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() || Subtarget->hasAVX()) {
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() || Subtarget->hasAVX()) {
926 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
927 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
928 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
929 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
931 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
932 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
933 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
935 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
936 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
939 if (Subtarget->hasSSE42() || Subtarget->hasAVX())
940 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
942 if (!UseSoftFloat && Subtarget->hasAVX()) {
943 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
944 addRegisterClass(MVT::v16i16, X86::VR256RegisterClass);
945 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
946 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
947 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
948 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
950 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
951 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
952 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
954 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
955 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
956 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
957 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
958 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
959 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
961 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
962 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
963 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
964 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
965 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
966 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
968 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
969 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
970 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
972 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f64, Custom);
973 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i64, Custom);
974 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
975 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
976 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i8, Custom);
977 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i16, Custom);
979 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
980 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
981 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
982 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
984 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
985 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
986 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
987 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
989 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
990 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
992 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
993 setOperationAction(ISD::VSETCC, MVT::v4i64, Custom);
995 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
996 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
997 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
999 // Custom lower several nodes for 256-bit types.
1000 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1001 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1002 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
1005 // Extract subvector is special because the value type
1006 // (result) is 128-bit but the source is 256-bit wide.
1007 if (VT.is128BitVector())
1008 setOperationAction(ISD::EXTRACT_SUBVECTOR, SVT, Custom);
1010 // Do not attempt to custom lower other non-256-bit vectors
1011 if (!VT.is256BitVector())
1014 setOperationAction(ISD::BUILD_VECTOR, SVT, Custom);
1015 setOperationAction(ISD::VECTOR_SHUFFLE, SVT, Custom);
1016 setOperationAction(ISD::INSERT_VECTOR_ELT, SVT, Custom);
1017 setOperationAction(ISD::EXTRACT_VECTOR_ELT, SVT, Custom);
1018 setOperationAction(ISD::SCALAR_TO_VECTOR, SVT, Custom);
1019 setOperationAction(ISD::INSERT_SUBVECTOR, SVT, Custom);
1022 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1023 for (unsigned i = (unsigned)MVT::v32i8; i != (unsigned)MVT::v4i64; ++i) {
1024 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
1027 // Do not attempt to promote non-256-bit vectors
1028 if (!VT.is256BitVector())
1031 setOperationAction(ISD::AND, SVT, Promote);
1032 AddPromotedToType (ISD::AND, SVT, MVT::v4i64);
1033 setOperationAction(ISD::OR, SVT, Promote);
1034 AddPromotedToType (ISD::OR, SVT, MVT::v4i64);
1035 setOperationAction(ISD::XOR, SVT, Promote);
1036 AddPromotedToType (ISD::XOR, SVT, MVT::v4i64);
1037 setOperationAction(ISD::LOAD, SVT, Promote);
1038 AddPromotedToType (ISD::LOAD, SVT, MVT::v4i64);
1039 setOperationAction(ISD::SELECT, SVT, Promote);
1040 AddPromotedToType (ISD::SELECT, SVT, MVT::v4i64);
1044 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1045 // of this type with custom code.
1046 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1047 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE; VT++) {
1048 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT, Custom);
1051 // We want to custom lower some of our intrinsics.
1052 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1055 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1056 // handle type legalization for these operations here.
1058 // FIXME: We really should do custom legalization for addition and
1059 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1060 // than generic legalization for 64-bit multiplication-with-overflow, though.
1061 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1062 // Add/Sub/Mul with overflow operations are custom lowered.
1064 setOperationAction(ISD::SADDO, VT, Custom);
1065 setOperationAction(ISD::UADDO, VT, Custom);
1066 setOperationAction(ISD::SSUBO, VT, Custom);
1067 setOperationAction(ISD::USUBO, VT, Custom);
1068 setOperationAction(ISD::SMULO, VT, Custom);
1069 setOperationAction(ISD::UMULO, VT, Custom);
1072 // There are no 8-bit 3-address imul/mul instructions
1073 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1074 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1076 if (!Subtarget->is64Bit()) {
1077 // These libcalls are not available in 32-bit.
1078 setLibcallName(RTLIB::SHL_I128, 0);
1079 setLibcallName(RTLIB::SRL_I128, 0);
1080 setLibcallName(RTLIB::SRA_I128, 0);
1083 // We have target-specific dag combine patterns for the following nodes:
1084 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1085 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1086 setTargetDAGCombine(ISD::BUILD_VECTOR);
1087 setTargetDAGCombine(ISD::SELECT);
1088 setTargetDAGCombine(ISD::SHL);
1089 setTargetDAGCombine(ISD::SRA);
1090 setTargetDAGCombine(ISD::SRL);
1091 setTargetDAGCombine(ISD::OR);
1092 setTargetDAGCombine(ISD::AND);
1093 setTargetDAGCombine(ISD::ADD);
1094 setTargetDAGCombine(ISD::SUB);
1095 setTargetDAGCombine(ISD::STORE);
1096 setTargetDAGCombine(ISD::ZERO_EXTEND);
1097 setTargetDAGCombine(ISD::SINT_TO_FP);
1098 if (Subtarget->is64Bit())
1099 setTargetDAGCombine(ISD::MUL);
1101 computeRegisterProperties();
1103 // On Darwin, -Os means optimize for size without hurting performance,
1104 // do not reduce the limit.
1105 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1106 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1107 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1108 maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1109 maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1110 maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1111 setPrefLoopAlignment(16);
1112 benefitFromCodePlacementOpt = true;
1114 setPrefFunctionAlignment(4);
1118 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1123 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1124 /// the desired ByVal argument alignment.
1125 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1128 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1129 if (VTy->getBitWidth() == 128)
1131 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1132 unsigned EltAlign = 0;
1133 getMaxByValAlign(ATy->getElementType(), EltAlign);
1134 if (EltAlign > MaxAlign)
1135 MaxAlign = EltAlign;
1136 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1137 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1138 unsigned EltAlign = 0;
1139 getMaxByValAlign(STy->getElementType(i), EltAlign);
1140 if (EltAlign > MaxAlign)
1141 MaxAlign = EltAlign;
1149 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1150 /// function arguments in the caller parameter area. For X86, aggregates
1151 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1152 /// are at 4-byte boundaries.
1153 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1154 if (Subtarget->is64Bit()) {
1155 // Max of 8 and alignment of type.
1156 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1163 if (Subtarget->hasXMM())
1164 getMaxByValAlign(Ty, Align);
1168 /// getOptimalMemOpType - Returns the target specific optimal type for load
1169 /// and store operations as a result of memset, memcpy, and memmove
1170 /// lowering. If DstAlign is zero that means it's safe to destination
1171 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1172 /// means there isn't a need to check it against alignment requirement,
1173 /// probably because the source does not need to be loaded. If
1174 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1175 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1176 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1177 /// constant so it does not need to be loaded.
1178 /// It returns EVT::Other if the type should be determined using generic
1179 /// target-independent logic.
1181 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1182 unsigned DstAlign, unsigned SrcAlign,
1183 bool NonScalarIntSafe,
1185 MachineFunction &MF) const {
1186 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1187 // linux. This is because the stack realignment code can't handle certain
1188 // cases like PR2962. This should be removed when PR2962 is fixed.
1189 const Function *F = MF.getFunction();
1190 if (NonScalarIntSafe &&
1191 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1193 (Subtarget->isUnalignedMemAccessFast() ||
1194 ((DstAlign == 0 || DstAlign >= 16) &&
1195 (SrcAlign == 0 || SrcAlign >= 16))) &&
1196 Subtarget->getStackAlignment() >= 16) {
1197 if (Subtarget->hasSSE2())
1199 if (Subtarget->hasSSE1())
1201 } else if (!MemcpyStrSrc && Size >= 8 &&
1202 !Subtarget->is64Bit() &&
1203 Subtarget->getStackAlignment() >= 8 &&
1204 Subtarget->hasXMMInt()) {
1205 // Do not use f64 to lower memcpy if source is string constant. It's
1206 // better to use i32 to avoid the loads.
1210 if (Subtarget->is64Bit() && Size >= 8)
1215 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1216 /// current function. The returned value is a member of the
1217 /// MachineJumpTableInfo::JTEntryKind enum.
1218 unsigned X86TargetLowering::getJumpTableEncoding() const {
1219 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1221 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1222 Subtarget->isPICStyleGOT())
1223 return MachineJumpTableInfo::EK_Custom32;
1225 // Otherwise, use the normal jump table encoding heuristics.
1226 return TargetLowering::getJumpTableEncoding();
1230 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1231 const MachineBasicBlock *MBB,
1232 unsigned uid,MCContext &Ctx) const{
1233 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1234 Subtarget->isPICStyleGOT());
1235 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1237 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1238 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1241 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1243 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1244 SelectionDAG &DAG) const {
1245 if (!Subtarget->is64Bit())
1246 // This doesn't have DebugLoc associated with it, but is not really the
1247 // same as a Register.
1248 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1252 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1253 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1255 const MCExpr *X86TargetLowering::
1256 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1257 MCContext &Ctx) const {
1258 // X86-64 uses RIP relative addressing based on the jump table label.
1259 if (Subtarget->isPICStyleRIPRel())
1260 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1262 // Otherwise, the reference is relative to the PIC base.
1263 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1266 // FIXME: Why this routine is here? Move to RegInfo!
1267 std::pair<const TargetRegisterClass*, uint8_t>
1268 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1269 const TargetRegisterClass *RRC = 0;
1271 switch (VT.getSimpleVT().SimpleTy) {
1273 return TargetLowering::findRepresentativeClass(VT);
1274 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1275 RRC = (Subtarget->is64Bit()
1276 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1279 RRC = X86::VR64RegisterClass;
1281 case MVT::f32: case MVT::f64:
1282 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1283 case MVT::v4f32: case MVT::v2f64:
1284 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1286 RRC = X86::VR128RegisterClass;
1289 return std::make_pair(RRC, Cost);
1292 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1293 unsigned &Offset) const {
1294 if (!Subtarget->isTargetLinux())
1297 if (Subtarget->is64Bit()) {
1298 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1300 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1313 //===----------------------------------------------------------------------===//
1314 // Return Value Calling Convention Implementation
1315 //===----------------------------------------------------------------------===//
1317 #include "X86GenCallingConv.inc"
1320 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1321 MachineFunction &MF, bool isVarArg,
1322 const SmallVectorImpl<ISD::OutputArg> &Outs,
1323 LLVMContext &Context) const {
1324 SmallVector<CCValAssign, 16> RVLocs;
1325 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1327 return CCInfo.CheckReturn(Outs, RetCC_X86);
1331 X86TargetLowering::LowerReturn(SDValue Chain,
1332 CallingConv::ID CallConv, bool isVarArg,
1333 const SmallVectorImpl<ISD::OutputArg> &Outs,
1334 const SmallVectorImpl<SDValue> &OutVals,
1335 DebugLoc dl, SelectionDAG &DAG) const {
1336 MachineFunction &MF = DAG.getMachineFunction();
1337 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1339 SmallVector<CCValAssign, 16> RVLocs;
1340 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1341 RVLocs, *DAG.getContext());
1342 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1344 // Add the regs to the liveout set for the function.
1345 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1346 for (unsigned i = 0; i != RVLocs.size(); ++i)
1347 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1348 MRI.addLiveOut(RVLocs[i].getLocReg());
1352 SmallVector<SDValue, 6> RetOps;
1353 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1354 // Operand #1 = Bytes To Pop
1355 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1358 // Copy the result values into the output registers.
1359 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1360 CCValAssign &VA = RVLocs[i];
1361 assert(VA.isRegLoc() && "Can only return in registers!");
1362 SDValue ValToCopy = OutVals[i];
1363 EVT ValVT = ValToCopy.getValueType();
1365 // If this is x86-64, and we disabled SSE, we can't return FP values,
1366 // or SSE or MMX vectors.
1367 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1368 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1369 (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
1370 report_fatal_error("SSE register return with SSE disabled");
1372 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1373 // llvm-gcc has never done it right and no one has noticed, so this
1374 // should be OK for now.
1375 if (ValVT == MVT::f64 &&
1376 (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
1377 report_fatal_error("SSE2 register return with SSE2 disabled");
1379 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1380 // the RET instruction and handled by the FP Stackifier.
1381 if (VA.getLocReg() == X86::ST0 ||
1382 VA.getLocReg() == X86::ST1) {
1383 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1384 // change the value to the FP stack register class.
1385 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1386 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1387 RetOps.push_back(ValToCopy);
1388 // Don't emit a copytoreg.
1392 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1393 // which is returned in RAX / RDX.
1394 if (Subtarget->is64Bit()) {
1395 if (ValVT == MVT::x86mmx) {
1396 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1397 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1398 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1400 // If we don't have SSE2 available, convert to v4f32 so the generated
1401 // register is legal.
1402 if (!Subtarget->hasSSE2())
1403 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1408 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1409 Flag = Chain.getValue(1);
1412 // The x86-64 ABI for returning structs by value requires that we copy
1413 // the sret argument into %rax for the return. We saved the argument into
1414 // a virtual register in the entry block, so now we copy the value out
1416 if (Subtarget->is64Bit() &&
1417 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1418 MachineFunction &MF = DAG.getMachineFunction();
1419 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1420 unsigned Reg = FuncInfo->getSRetReturnReg();
1422 "SRetReturnReg should have been set in LowerFormalArguments().");
1423 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1425 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1426 Flag = Chain.getValue(1);
1428 // RAX now acts like a return value.
1429 MRI.addLiveOut(X86::RAX);
1432 RetOps[0] = Chain; // Update chain.
1434 // Add the flag if we have it.
1436 RetOps.push_back(Flag);
1438 return DAG.getNode(X86ISD::RET_FLAG, dl,
1439 MVT::Other, &RetOps[0], RetOps.size());
1442 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
1443 if (N->getNumValues() != 1)
1445 if (!N->hasNUsesOfValue(1, 0))
1448 SDNode *Copy = *N->use_begin();
1449 if (Copy->getOpcode() != ISD::CopyToReg &&
1450 Copy->getOpcode() != ISD::FP_EXTEND)
1453 bool HasRet = false;
1454 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1456 if (UI->getOpcode() != X86ISD::RET_FLAG)
1465 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1466 ISD::NodeType ExtendKind) const {
1468 // TODO: Is this also valid on 32-bit?
1469 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1470 ReturnMVT = MVT::i8;
1472 ReturnMVT = MVT::i32;
1474 EVT MinVT = getRegisterType(Context, ReturnMVT);
1475 return VT.bitsLT(MinVT) ? MinVT : VT;
1478 /// LowerCallResult - Lower the result values of a call into the
1479 /// appropriate copies out of appropriate physical registers.
1482 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1483 CallingConv::ID CallConv, bool isVarArg,
1484 const SmallVectorImpl<ISD::InputArg> &Ins,
1485 DebugLoc dl, SelectionDAG &DAG,
1486 SmallVectorImpl<SDValue> &InVals) const {
1488 // Assign locations to each value returned by this call.
1489 SmallVector<CCValAssign, 16> RVLocs;
1490 bool Is64Bit = Subtarget->is64Bit();
1491 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1492 getTargetMachine(), RVLocs, *DAG.getContext());
1493 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1495 // Copy all of the result registers out of their specified physreg.
1496 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1497 CCValAssign &VA = RVLocs[i];
1498 EVT CopyVT = VA.getValVT();
1500 // If this is x86-64, and we disabled SSE, we can't return FP values
1501 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1502 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
1503 report_fatal_error("SSE register return with SSE disabled");
1508 // If this is a call to a function that returns an fp value on the floating
1509 // point stack, we must guarantee the the value is popped from the stack, so
1510 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1511 // if the return value is not used. We use the FpPOP_RETVAL instruction
1513 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1514 // If we prefer to use the value in xmm registers, copy it out as f80 and
1515 // use a truncate to move it from fp stack reg to xmm reg.
1516 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1517 SDValue Ops[] = { Chain, InFlag };
1518 Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
1519 MVT::Other, MVT::Glue, Ops, 2), 1);
1520 Val = Chain.getValue(0);
1522 // Round the f80 to the right size, which also moves it to the appropriate
1524 if (CopyVT != VA.getValVT())
1525 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1526 // This truncation won't change the value.
1527 DAG.getIntPtrConstant(1));
1529 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1530 CopyVT, InFlag).getValue(1);
1531 Val = Chain.getValue(0);
1533 InFlag = Chain.getValue(2);
1534 InVals.push_back(Val);
1541 //===----------------------------------------------------------------------===//
1542 // C & StdCall & Fast Calling Convention implementation
1543 //===----------------------------------------------------------------------===//
1544 // StdCall calling convention seems to be standard for many Windows' API
1545 // routines and around. It differs from C calling convention just a little:
1546 // callee should clean up the stack, not caller. Symbols should be also
1547 // decorated in some fancy way :) It doesn't support any vector arguments.
1548 // For info on fast calling convention see Fast Calling Convention (tail call)
1549 // implementation LowerX86_32FastCCCallTo.
1551 /// CallIsStructReturn - Determines whether a call uses struct return
1553 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1557 return Outs[0].Flags.isSRet();
1560 /// ArgsAreStructReturn - Determines whether a function uses struct
1561 /// return semantics.
1563 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1567 return Ins[0].Flags.isSRet();
1570 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1571 /// by "Src" to address "Dst" with size and alignment information specified by
1572 /// the specific parameter attribute. The copy will be passed as a byval
1573 /// function parameter.
1575 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1576 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1578 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1580 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1581 /*isVolatile*/false, /*AlwaysInline=*/true,
1582 MachinePointerInfo(), MachinePointerInfo());
1585 /// IsTailCallConvention - Return true if the calling convention is one that
1586 /// supports tail call optimization.
1587 static bool IsTailCallConvention(CallingConv::ID CC) {
1588 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1591 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1592 if (!CI->isTailCall())
1596 CallingConv::ID CalleeCC = CS.getCallingConv();
1597 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1603 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1604 /// a tailcall target by changing its ABI.
1605 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1606 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1610 X86TargetLowering::LowerMemArgument(SDValue Chain,
1611 CallingConv::ID CallConv,
1612 const SmallVectorImpl<ISD::InputArg> &Ins,
1613 DebugLoc dl, SelectionDAG &DAG,
1614 const CCValAssign &VA,
1615 MachineFrameInfo *MFI,
1617 // Create the nodes corresponding to a load from this parameter slot.
1618 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1619 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1620 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1623 // If value is passed by pointer we have address passed instead of the value
1625 if (VA.getLocInfo() == CCValAssign::Indirect)
1626 ValVT = VA.getLocVT();
1628 ValVT = VA.getValVT();
1630 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1631 // changed with more analysis.
1632 // In case of tail call optimization mark all arguments mutable. Since they
1633 // could be overwritten by lowering of arguments in case of a tail call.
1634 if (Flags.isByVal()) {
1635 unsigned Bytes = Flags.getByValSize();
1636 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
1637 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
1638 return DAG.getFrameIndex(FI, getPointerTy());
1640 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1641 VA.getLocMemOffset(), isImmutable);
1642 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1643 return DAG.getLoad(ValVT, dl, Chain, FIN,
1644 MachinePointerInfo::getFixedStack(FI),
1650 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1651 CallingConv::ID CallConv,
1653 const SmallVectorImpl<ISD::InputArg> &Ins,
1656 SmallVectorImpl<SDValue> &InVals)
1658 MachineFunction &MF = DAG.getMachineFunction();
1659 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1661 const Function* Fn = MF.getFunction();
1662 if (Fn->hasExternalLinkage() &&
1663 Subtarget->isTargetCygMing() &&
1664 Fn->getName() == "main")
1665 FuncInfo->setForceFramePointer(true);
1667 MachineFrameInfo *MFI = MF.getFrameInfo();
1668 bool Is64Bit = Subtarget->is64Bit();
1669 bool IsWin64 = Subtarget->isTargetWin64();
1671 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1672 "Var args not supported with calling convention fastcc or ghc");
1674 // Assign locations to all of the incoming arguments.
1675 SmallVector<CCValAssign, 16> ArgLocs;
1676 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1677 ArgLocs, *DAG.getContext());
1679 // Allocate shadow area for Win64
1681 CCInfo.AllocateStack(32, 8);
1684 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1686 unsigned LastVal = ~0U;
1688 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1689 CCValAssign &VA = ArgLocs[i];
1690 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1692 assert(VA.getValNo() != LastVal &&
1693 "Don't support value assigned to multiple locs yet");
1694 LastVal = VA.getValNo();
1696 if (VA.isRegLoc()) {
1697 EVT RegVT = VA.getLocVT();
1698 TargetRegisterClass *RC = NULL;
1699 if (RegVT == MVT::i32)
1700 RC = X86::GR32RegisterClass;
1701 else if (Is64Bit && RegVT == MVT::i64)
1702 RC = X86::GR64RegisterClass;
1703 else if (RegVT == MVT::f32)
1704 RC = X86::FR32RegisterClass;
1705 else if (RegVT == MVT::f64)
1706 RC = X86::FR64RegisterClass;
1707 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1708 RC = X86::VR256RegisterClass;
1709 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1710 RC = X86::VR128RegisterClass;
1711 else if (RegVT == MVT::x86mmx)
1712 RC = X86::VR64RegisterClass;
1714 llvm_unreachable("Unknown argument type!");
1716 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1717 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1719 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1720 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1722 if (VA.getLocInfo() == CCValAssign::SExt)
1723 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1724 DAG.getValueType(VA.getValVT()));
1725 else if (VA.getLocInfo() == CCValAssign::ZExt)
1726 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1727 DAG.getValueType(VA.getValVT()));
1728 else if (VA.getLocInfo() == CCValAssign::BCvt)
1729 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1731 if (VA.isExtInLoc()) {
1732 // Handle MMX values passed in XMM regs.
1733 if (RegVT.isVector()) {
1734 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1737 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1740 assert(VA.isMemLoc());
1741 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1744 // If value is passed via pointer - do a load.
1745 if (VA.getLocInfo() == CCValAssign::Indirect)
1746 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1747 MachinePointerInfo(), false, false, 0);
1749 InVals.push_back(ArgValue);
1752 // The x86-64 ABI for returning structs by value requires that we copy
1753 // the sret argument into %rax for the return. Save the argument into
1754 // a virtual register so that we can access it from the return points.
1755 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1756 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1757 unsigned Reg = FuncInfo->getSRetReturnReg();
1759 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1760 FuncInfo->setSRetReturnReg(Reg);
1762 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1763 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1766 unsigned StackSize = CCInfo.getNextStackOffset();
1767 // Align stack specially for tail calls.
1768 if (FuncIsMadeTailCallSafe(CallConv))
1769 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1771 // If the function takes variable number of arguments, make a frame index for
1772 // the start of the first vararg value... for expansion of llvm.va_start.
1774 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1775 CallConv != CallingConv::X86_ThisCall)) {
1776 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1779 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1781 // FIXME: We should really autogenerate these arrays
1782 static const unsigned GPR64ArgRegsWin64[] = {
1783 X86::RCX, X86::RDX, X86::R8, X86::R9
1785 static const unsigned GPR64ArgRegs64Bit[] = {
1786 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1788 static const unsigned XMMArgRegs64Bit[] = {
1789 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1790 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1792 const unsigned *GPR64ArgRegs;
1793 unsigned NumXMMRegs = 0;
1796 // The XMM registers which might contain var arg parameters are shadowed
1797 // in their paired GPR. So we only need to save the GPR to their home
1799 TotalNumIntRegs = 4;
1800 GPR64ArgRegs = GPR64ArgRegsWin64;
1802 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1803 GPR64ArgRegs = GPR64ArgRegs64Bit;
1805 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1807 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1810 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1811 assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
1812 "SSE register cannot be used when SSE is disabled!");
1813 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1814 "SSE register cannot be used when SSE is disabled!");
1815 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
1816 // Kernel mode asks for SSE to be disabled, so don't push them
1818 TotalNumXMMRegs = 0;
1821 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
1822 // Get to the caller-allocated home save location. Add 8 to account
1823 // for the return address.
1824 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1825 FuncInfo->setRegSaveFrameIndex(
1826 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1827 // Fixup to set vararg frame on shadow area (4 x i64).
1829 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1831 // For X86-64, if there are vararg parameters that are passed via
1832 // registers, then we must store them to their spots on the stack so they
1833 // may be loaded by deferencing the result of va_next.
1834 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1835 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1836 FuncInfo->setRegSaveFrameIndex(
1837 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1841 // Store the integer parameter registers.
1842 SmallVector<SDValue, 8> MemOps;
1843 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1845 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1846 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1847 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1848 DAG.getIntPtrConstant(Offset));
1849 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1850 X86::GR64RegisterClass);
1851 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1853 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1854 MachinePointerInfo::getFixedStack(
1855 FuncInfo->getRegSaveFrameIndex(), Offset),
1857 MemOps.push_back(Store);
1861 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1862 // Now store the XMM (fp + vector) parameter registers.
1863 SmallVector<SDValue, 11> SaveXMMOps;
1864 SaveXMMOps.push_back(Chain);
1866 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1867 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1868 SaveXMMOps.push_back(ALVal);
1870 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1871 FuncInfo->getRegSaveFrameIndex()));
1872 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1873 FuncInfo->getVarArgsFPOffset()));
1875 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1876 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1877 X86::VR128RegisterClass);
1878 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1879 SaveXMMOps.push_back(Val);
1881 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1883 &SaveXMMOps[0], SaveXMMOps.size()));
1886 if (!MemOps.empty())
1887 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1888 &MemOps[0], MemOps.size());
1892 // Some CCs need callee pop.
1893 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt)) {
1894 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1896 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1897 // If this is an sret function, the return should pop the hidden pointer.
1898 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1899 FuncInfo->setBytesToPopOnReturn(4);
1903 // RegSaveFrameIndex is X86-64 only.
1904 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1905 if (CallConv == CallingConv::X86_FastCall ||
1906 CallConv == CallingConv::X86_ThisCall)
1907 // fastcc functions can't have varargs.
1908 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1915 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1916 SDValue StackPtr, SDValue Arg,
1917 DebugLoc dl, SelectionDAG &DAG,
1918 const CCValAssign &VA,
1919 ISD::ArgFlagsTy Flags) const {
1920 unsigned LocMemOffset = VA.getLocMemOffset();
1921 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1922 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1923 if (Flags.isByVal())
1924 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1926 return DAG.getStore(Chain, dl, Arg, PtrOff,
1927 MachinePointerInfo::getStack(LocMemOffset),
1931 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1932 /// optimization is performed and it is required.
1934 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1935 SDValue &OutRetAddr, SDValue Chain,
1936 bool IsTailCall, bool Is64Bit,
1937 int FPDiff, DebugLoc dl) const {
1938 // Adjust the Return address stack slot.
1939 EVT VT = getPointerTy();
1940 OutRetAddr = getReturnAddressFrameIndex(DAG);
1942 // Load the "old" Return address.
1943 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1945 return SDValue(OutRetAddr.getNode(), 1);
1948 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
1949 /// optimization is performed and it is required (FPDiff!=0).
1951 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1952 SDValue Chain, SDValue RetAddrFrIdx,
1953 bool Is64Bit, int FPDiff, DebugLoc dl) {
1954 // Store the return address to the appropriate stack slot.
1955 if (!FPDiff) return Chain;
1956 // Calculate the new stack slot for the return address.
1957 int SlotSize = Is64Bit ? 8 : 4;
1958 int NewReturnAddrFI =
1959 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1960 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1961 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1962 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1963 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
1969 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1970 CallingConv::ID CallConv, bool isVarArg,
1972 const SmallVectorImpl<ISD::OutputArg> &Outs,
1973 const SmallVectorImpl<SDValue> &OutVals,
1974 const SmallVectorImpl<ISD::InputArg> &Ins,
1975 DebugLoc dl, SelectionDAG &DAG,
1976 SmallVectorImpl<SDValue> &InVals) const {
1977 MachineFunction &MF = DAG.getMachineFunction();
1978 bool Is64Bit = Subtarget->is64Bit();
1979 bool IsWin64 = Subtarget->isTargetWin64();
1980 bool IsStructRet = CallIsStructReturn(Outs);
1981 bool IsSibcall = false;
1984 // Check if it's really possible to do a tail call.
1985 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
1986 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
1987 Outs, OutVals, Ins, DAG);
1989 // Sibcalls are automatically detected tailcalls which do not require
1991 if (!GuaranteedTailCallOpt && isTailCall)
1998 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1999 "Var args not supported with calling convention fastcc or ghc");
2001 // Analyze operands of the call, assigning locations to each operand.
2002 SmallVector<CCValAssign, 16> ArgLocs;
2003 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
2004 ArgLocs, *DAG.getContext());
2006 // Allocate shadow area for Win64
2008 CCInfo.AllocateStack(32, 8);
2011 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2013 // Get a count of how many bytes are to be pushed on the stack.
2014 unsigned NumBytes = CCInfo.getNextStackOffset();
2016 // This is a sibcall. The memory operands are available in caller's
2017 // own caller's stack.
2019 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
2020 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2023 if (isTailCall && !IsSibcall) {
2024 // Lower arguments at fp - stackoffset + fpdiff.
2025 unsigned NumBytesCallerPushed =
2026 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
2027 FPDiff = NumBytesCallerPushed - NumBytes;
2029 // Set the delta of movement of the returnaddr stackslot.
2030 // But only set if delta is greater than previous delta.
2031 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
2032 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
2036 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2038 SDValue RetAddrFrIdx;
2039 // Load return address for tail calls.
2040 if (isTailCall && FPDiff)
2041 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2042 Is64Bit, FPDiff, dl);
2044 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2045 SmallVector<SDValue, 8> MemOpChains;
2048 // Walk the register/memloc assignments, inserting copies/loads. In the case
2049 // of tail call optimization arguments are handle later.
2050 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2051 CCValAssign &VA = ArgLocs[i];
2052 EVT RegVT = VA.getLocVT();
2053 SDValue Arg = OutVals[i];
2054 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2055 bool isByVal = Flags.isByVal();
2057 // Promote the value if needed.
2058 switch (VA.getLocInfo()) {
2059 default: llvm_unreachable("Unknown loc info!");
2060 case CCValAssign::Full: break;
2061 case CCValAssign::SExt:
2062 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2064 case CCValAssign::ZExt:
2065 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2067 case CCValAssign::AExt:
2068 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
2069 // Special case: passing MMX values in XMM registers.
2070 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2071 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2072 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2074 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2076 case CCValAssign::BCvt:
2077 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2079 case CCValAssign::Indirect: {
2080 // Store the argument.
2081 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2082 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2083 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2084 MachinePointerInfo::getFixedStack(FI),
2091 if (VA.isRegLoc()) {
2092 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2093 if (isVarArg && IsWin64) {
2094 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2095 // shadow reg if callee is a varargs function.
2096 unsigned ShadowReg = 0;
2097 switch (VA.getLocReg()) {
2098 case X86::XMM0: ShadowReg = X86::RCX; break;
2099 case X86::XMM1: ShadowReg = X86::RDX; break;
2100 case X86::XMM2: ShadowReg = X86::R8; break;
2101 case X86::XMM3: ShadowReg = X86::R9; break;
2104 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2106 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2107 assert(VA.isMemLoc());
2108 if (StackPtr.getNode() == 0)
2109 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2110 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2111 dl, DAG, VA, Flags));
2115 if (!MemOpChains.empty())
2116 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2117 &MemOpChains[0], MemOpChains.size());
2119 // Build a sequence of copy-to-reg nodes chained together with token chain
2120 // and flag operands which copy the outgoing args into registers.
2122 // Tail call byval lowering might overwrite argument registers so in case of
2123 // tail call optimization the copies to registers are lowered later.
2125 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2126 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2127 RegsToPass[i].second, InFlag);
2128 InFlag = Chain.getValue(1);
2131 if (Subtarget->isPICStyleGOT()) {
2132 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2135 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2136 DAG.getNode(X86ISD::GlobalBaseReg,
2137 DebugLoc(), getPointerTy()),
2139 InFlag = Chain.getValue(1);
2141 // If we are tail calling and generating PIC/GOT style code load the
2142 // address of the callee into ECX. The value in ecx is used as target of
2143 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2144 // for tail calls on PIC/GOT architectures. Normally we would just put the
2145 // address of GOT into ebx and then call target@PLT. But for tail calls
2146 // ebx would be restored (since ebx is callee saved) before jumping to the
2149 // Note: The actual moving to ECX is done further down.
2150 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2151 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2152 !G->getGlobal()->hasProtectedVisibility())
2153 Callee = LowerGlobalAddress(Callee, DAG);
2154 else if (isa<ExternalSymbolSDNode>(Callee))
2155 Callee = LowerExternalSymbol(Callee, DAG);
2159 if (Is64Bit && isVarArg && !IsWin64) {
2160 // From AMD64 ABI document:
2161 // For calls that may call functions that use varargs or stdargs
2162 // (prototype-less calls or calls to functions containing ellipsis (...) in
2163 // the declaration) %al is used as hidden argument to specify the number
2164 // of SSE registers used. The contents of %al do not need to match exactly
2165 // the number of registers, but must be an ubound on the number of SSE
2166 // registers used and is in the range 0 - 8 inclusive.
2168 // Count the number of XMM registers allocated.
2169 static const unsigned XMMArgRegs[] = {
2170 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2171 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2173 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2174 assert((Subtarget->hasXMM() || !NumXMMRegs)
2175 && "SSE registers cannot be used when SSE is disabled");
2177 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2178 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2179 InFlag = Chain.getValue(1);
2183 // For tail calls lower the arguments to the 'real' stack slot.
2185 // Force all the incoming stack arguments to be loaded from the stack
2186 // before any new outgoing arguments are stored to the stack, because the
2187 // outgoing stack slots may alias the incoming argument stack slots, and
2188 // the alias isn't otherwise explicit. This is slightly more conservative
2189 // than necessary, because it means that each store effectively depends
2190 // on every argument instead of just those arguments it would clobber.
2191 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2193 SmallVector<SDValue, 8> MemOpChains2;
2196 // Do not flag preceding copytoreg stuff together with the following stuff.
2198 if (GuaranteedTailCallOpt) {
2199 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2200 CCValAssign &VA = ArgLocs[i];
2203 assert(VA.isMemLoc());
2204 SDValue Arg = OutVals[i];
2205 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2206 // Create frame index.
2207 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2208 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2209 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2210 FIN = DAG.getFrameIndex(FI, getPointerTy());
2212 if (Flags.isByVal()) {
2213 // Copy relative to framepointer.
2214 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2215 if (StackPtr.getNode() == 0)
2216 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2218 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2220 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2224 // Store relative to framepointer.
2225 MemOpChains2.push_back(
2226 DAG.getStore(ArgChain, dl, Arg, FIN,
2227 MachinePointerInfo::getFixedStack(FI),
2233 if (!MemOpChains2.empty())
2234 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2235 &MemOpChains2[0], MemOpChains2.size());
2237 // Copy arguments to their registers.
2238 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2239 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2240 RegsToPass[i].second, InFlag);
2241 InFlag = Chain.getValue(1);
2245 // Store the return address to the appropriate stack slot.
2246 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2250 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2251 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2252 // In the 64-bit large code model, we have to make all calls
2253 // through a register, since the call instruction's 32-bit
2254 // pc-relative offset may not be large enough to hold the whole
2256 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2257 // If the callee is a GlobalAddress node (quite common, every direct call
2258 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2261 // We should use extra load for direct calls to dllimported functions in
2263 const GlobalValue *GV = G->getGlobal();
2264 if (!GV->hasDLLImportLinkage()) {
2265 unsigned char OpFlags = 0;
2266 bool ExtraLoad = false;
2267 unsigned WrapperKind = ISD::DELETED_NODE;
2269 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2270 // external symbols most go through the PLT in PIC mode. If the symbol
2271 // has hidden or protected visibility, or if it is static or local, then
2272 // we don't need to use the PLT - we can directly call it.
2273 if (Subtarget->isTargetELF() &&
2274 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2275 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2276 OpFlags = X86II::MO_PLT;
2277 } else if (Subtarget->isPICStyleStubAny() &&
2278 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2279 (!Subtarget->getTargetTriple().isMacOSX() ||
2280 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2281 // PC-relative references to external symbols should go through $stub,
2282 // unless we're building with the leopard linker or later, which
2283 // automatically synthesizes these stubs.
2284 OpFlags = X86II::MO_DARWIN_STUB;
2285 } else if (Subtarget->isPICStyleRIPRel() &&
2286 isa<Function>(GV) &&
2287 cast<Function>(GV)->hasFnAttr(Attribute::NonLazyBind)) {
2288 // If the function is marked as non-lazy, generate an indirect call
2289 // which loads from the GOT directly. This avoids runtime overhead
2290 // at the cost of eager binding (and one extra byte of encoding).
2291 OpFlags = X86II::MO_GOTPCREL;
2292 WrapperKind = X86ISD::WrapperRIP;
2296 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2297 G->getOffset(), OpFlags);
2299 // Add a wrapper if needed.
2300 if (WrapperKind != ISD::DELETED_NODE)
2301 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2302 // Add extra indirection if needed.
2304 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2305 MachinePointerInfo::getGOT(),
2308 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2309 unsigned char OpFlags = 0;
2311 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2312 // external symbols should go through the PLT.
2313 if (Subtarget->isTargetELF() &&
2314 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2315 OpFlags = X86II::MO_PLT;
2316 } else if (Subtarget->isPICStyleStubAny() &&
2317 (!Subtarget->getTargetTriple().isMacOSX() ||
2318 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2319 // PC-relative references to external symbols should go through $stub,
2320 // unless we're building with the leopard linker or later, which
2321 // automatically synthesizes these stubs.
2322 OpFlags = X86II::MO_DARWIN_STUB;
2325 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2329 // Returns a chain & a flag for retval copy to use.
2330 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2331 SmallVector<SDValue, 8> Ops;
2333 if (!IsSibcall && isTailCall) {
2334 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2335 DAG.getIntPtrConstant(0, true), InFlag);
2336 InFlag = Chain.getValue(1);
2339 Ops.push_back(Chain);
2340 Ops.push_back(Callee);
2343 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2345 // Add argument registers to the end of the list so that they are known live
2347 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2348 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2349 RegsToPass[i].second.getValueType()));
2351 // Add an implicit use GOT pointer in EBX.
2352 if (!isTailCall && Subtarget->isPICStyleGOT())
2353 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2355 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2356 if (Is64Bit && isVarArg && !IsWin64)
2357 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2359 if (InFlag.getNode())
2360 Ops.push_back(InFlag);
2364 //// If this is the first return lowered for this function, add the regs
2365 //// to the liveout set for the function.
2366 // This isn't right, although it's probably harmless on x86; liveouts
2367 // should be computed from returns not tail calls. Consider a void
2368 // function making a tail call to a function returning int.
2369 return DAG.getNode(X86ISD::TC_RETURN, dl,
2370 NodeTys, &Ops[0], Ops.size());
2373 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2374 InFlag = Chain.getValue(1);
2376 // Create the CALLSEQ_END node.
2377 unsigned NumBytesForCalleeToPush;
2378 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt))
2379 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2380 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2381 // If this is a call to a struct-return function, the callee
2382 // pops the hidden struct pointer, so we have to push it back.
2383 // This is common for Darwin/X86, Linux & Mingw32 targets.
2384 NumBytesForCalleeToPush = 4;
2386 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2388 // Returns a flag for retval copy to use.
2390 Chain = DAG.getCALLSEQ_END(Chain,
2391 DAG.getIntPtrConstant(NumBytes, true),
2392 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2395 InFlag = Chain.getValue(1);
2398 // Handle result values, copying them out of physregs into vregs that we
2400 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2401 Ins, dl, DAG, InVals);
2405 //===----------------------------------------------------------------------===//
2406 // Fast Calling Convention (tail call) implementation
2407 //===----------------------------------------------------------------------===//
2409 // Like std call, callee cleans arguments, convention except that ECX is
2410 // reserved for storing the tail called function address. Only 2 registers are
2411 // free for argument passing (inreg). Tail call optimization is performed
2413 // * tailcallopt is enabled
2414 // * caller/callee are fastcc
2415 // On X86_64 architecture with GOT-style position independent code only local
2416 // (within module) calls are supported at the moment.
2417 // To keep the stack aligned according to platform abi the function
2418 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2419 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2420 // If a tail called function callee has more arguments than the caller the
2421 // caller needs to make sure that there is room to move the RETADDR to. This is
2422 // achieved by reserving an area the size of the argument delta right after the
2423 // original REtADDR, but before the saved framepointer or the spilled registers
2424 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2436 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2437 /// for a 16 byte align requirement.
2439 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2440 SelectionDAG& DAG) const {
2441 MachineFunction &MF = DAG.getMachineFunction();
2442 const TargetMachine &TM = MF.getTarget();
2443 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2444 unsigned StackAlignment = TFI.getStackAlignment();
2445 uint64_t AlignMask = StackAlignment - 1;
2446 int64_t Offset = StackSize;
2447 uint64_t SlotSize = TD->getPointerSize();
2448 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2449 // Number smaller than 12 so just add the difference.
2450 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2452 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2453 Offset = ((~AlignMask) & Offset) + StackAlignment +
2454 (StackAlignment-SlotSize);
2459 /// MatchingStackOffset - Return true if the given stack call argument is
2460 /// already available in the same position (relatively) of the caller's
2461 /// incoming argument stack.
2463 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2464 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2465 const X86InstrInfo *TII) {
2466 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2468 if (Arg.getOpcode() == ISD::CopyFromReg) {
2469 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2470 if (!TargetRegisterInfo::isVirtualRegister(VR))
2472 MachineInstr *Def = MRI->getVRegDef(VR);
2475 if (!Flags.isByVal()) {
2476 if (!TII->isLoadFromStackSlot(Def, FI))
2479 unsigned Opcode = Def->getOpcode();
2480 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2481 Def->getOperand(1).isFI()) {
2482 FI = Def->getOperand(1).getIndex();
2483 Bytes = Flags.getByValSize();
2487 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2488 if (Flags.isByVal())
2489 // ByVal argument is passed in as a pointer but it's now being
2490 // dereferenced. e.g.
2491 // define @foo(%struct.X* %A) {
2492 // tail call @bar(%struct.X* byval %A)
2495 SDValue Ptr = Ld->getBasePtr();
2496 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2499 FI = FINode->getIndex();
2500 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
2501 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
2502 FI = FINode->getIndex();
2503 Bytes = Flags.getByValSize();
2507 assert(FI != INT_MAX);
2508 if (!MFI->isFixedObjectIndex(FI))
2510 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2513 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2514 /// for tail call optimization. Targets which want to do tail call
2515 /// optimization should implement this function.
2517 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2518 CallingConv::ID CalleeCC,
2520 bool isCalleeStructRet,
2521 bool isCallerStructRet,
2522 const SmallVectorImpl<ISD::OutputArg> &Outs,
2523 const SmallVectorImpl<SDValue> &OutVals,
2524 const SmallVectorImpl<ISD::InputArg> &Ins,
2525 SelectionDAG& DAG) const {
2526 if (!IsTailCallConvention(CalleeCC) &&
2527 CalleeCC != CallingConv::C)
2530 // If -tailcallopt is specified, make fastcc functions tail-callable.
2531 const MachineFunction &MF = DAG.getMachineFunction();
2532 const Function *CallerF = DAG.getMachineFunction().getFunction();
2533 CallingConv::ID CallerCC = CallerF->getCallingConv();
2534 bool CCMatch = CallerCC == CalleeCC;
2536 if (GuaranteedTailCallOpt) {
2537 if (IsTailCallConvention(CalleeCC) && CCMatch)
2542 // Look for obvious safe cases to perform tail call optimization that do not
2543 // require ABI changes. This is what gcc calls sibcall.
2545 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2546 // emit a special epilogue.
2547 if (RegInfo->needsStackRealignment(MF))
2550 // Also avoid sibcall optimization if either caller or callee uses struct
2551 // return semantics.
2552 if (isCalleeStructRet || isCallerStructRet)
2555 // An stdcall caller is expected to clean up its arguments; the callee
2556 // isn't going to do that.
2557 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2560 // Do not sibcall optimize vararg calls unless all arguments are passed via
2562 if (isVarArg && !Outs.empty()) {
2564 // Optimizing for varargs on Win64 is unlikely to be safe without
2565 // additional testing.
2566 if (Subtarget->isTargetWin64())
2569 SmallVector<CCValAssign, 16> ArgLocs;
2570 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2571 getTargetMachine(), ArgLocs, *DAG.getContext());
2573 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2574 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2575 if (!ArgLocs[i].isRegLoc())
2579 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2580 // Therefore if it's not used by the call it is not safe to optimize this into
2582 bool Unused = false;
2583 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2590 SmallVector<CCValAssign, 16> RVLocs;
2591 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
2592 getTargetMachine(), RVLocs, *DAG.getContext());
2593 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2594 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2595 CCValAssign &VA = RVLocs[i];
2596 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2601 // If the calling conventions do not match, then we'd better make sure the
2602 // results are returned in the same way as what the caller expects.
2604 SmallVector<CCValAssign, 16> RVLocs1;
2605 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2606 getTargetMachine(), RVLocs1, *DAG.getContext());
2607 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2609 SmallVector<CCValAssign, 16> RVLocs2;
2610 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2611 getTargetMachine(), RVLocs2, *DAG.getContext());
2612 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2614 if (RVLocs1.size() != RVLocs2.size())
2616 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2617 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2619 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2621 if (RVLocs1[i].isRegLoc()) {
2622 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2625 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2631 // If the callee takes no arguments then go on to check the results of the
2633 if (!Outs.empty()) {
2634 // Check if stack adjustment is needed. For now, do not do this if any
2635 // argument is passed on the stack.
2636 SmallVector<CCValAssign, 16> ArgLocs;
2637 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2638 getTargetMachine(), ArgLocs, *DAG.getContext());
2640 // Allocate shadow area for Win64
2641 if (Subtarget->isTargetWin64()) {
2642 CCInfo.AllocateStack(32, 8);
2645 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2646 if (CCInfo.getNextStackOffset()) {
2647 MachineFunction &MF = DAG.getMachineFunction();
2648 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2651 // Check if the arguments are already laid out in the right way as
2652 // the caller's fixed stack objects.
2653 MachineFrameInfo *MFI = MF.getFrameInfo();
2654 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2655 const X86InstrInfo *TII =
2656 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2657 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2658 CCValAssign &VA = ArgLocs[i];
2659 SDValue Arg = OutVals[i];
2660 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2661 if (VA.getLocInfo() == CCValAssign::Indirect)
2663 if (!VA.isRegLoc()) {
2664 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2671 // If the tailcall address may be in a register, then make sure it's
2672 // possible to register allocate for it. In 32-bit, the call address can
2673 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2674 // callee-saved registers are restored. These happen to be the same
2675 // registers used to pass 'inreg' arguments so watch out for those.
2676 if (!Subtarget->is64Bit() &&
2677 !isa<GlobalAddressSDNode>(Callee) &&
2678 !isa<ExternalSymbolSDNode>(Callee)) {
2679 unsigned NumInRegs = 0;
2680 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2681 CCValAssign &VA = ArgLocs[i];
2684 unsigned Reg = VA.getLocReg();
2687 case X86::EAX: case X86::EDX: case X86::ECX:
2688 if (++NumInRegs == 3)
2700 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2701 return X86::createFastISel(funcInfo);
2705 //===----------------------------------------------------------------------===//
2706 // Other Lowering Hooks
2707 //===----------------------------------------------------------------------===//
2709 static bool MayFoldLoad(SDValue Op) {
2710 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2713 static bool MayFoldIntoStore(SDValue Op) {
2714 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2717 static bool isTargetShuffle(unsigned Opcode) {
2719 default: return false;
2720 case X86ISD::PSHUFD:
2721 case X86ISD::PSHUFHW:
2722 case X86ISD::PSHUFLW:
2723 case X86ISD::SHUFPD:
2724 case X86ISD::PALIGN:
2725 case X86ISD::SHUFPS:
2726 case X86ISD::MOVLHPS:
2727 case X86ISD::MOVLHPD:
2728 case X86ISD::MOVHLPS:
2729 case X86ISD::MOVLPS:
2730 case X86ISD::MOVLPD:
2731 case X86ISD::MOVSHDUP:
2732 case X86ISD::MOVSLDUP:
2733 case X86ISD::MOVDDUP:
2736 case X86ISD::UNPCKLPS:
2737 case X86ISD::UNPCKLPD:
2738 case X86ISD::VUNPCKLPSY:
2739 case X86ISD::VUNPCKLPDY:
2740 case X86ISD::PUNPCKLWD:
2741 case X86ISD::PUNPCKLBW:
2742 case X86ISD::PUNPCKLDQ:
2743 case X86ISD::PUNPCKLQDQ:
2744 case X86ISD::UNPCKHPS:
2745 case X86ISD::UNPCKHPD:
2746 case X86ISD::VUNPCKHPSY:
2747 case X86ISD::VUNPCKHPDY:
2748 case X86ISD::PUNPCKHWD:
2749 case X86ISD::PUNPCKHBW:
2750 case X86ISD::PUNPCKHDQ:
2751 case X86ISD::PUNPCKHQDQ:
2752 case X86ISD::VPERMILPS:
2753 case X86ISD::VPERMILPSY:
2754 case X86ISD::VPERMILPD:
2755 case X86ISD::VPERMILPDY:
2756 case X86ISD::VPERM2F128:
2762 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2763 SDValue V1, SelectionDAG &DAG) {
2765 default: llvm_unreachable("Unknown x86 shuffle node");
2766 case X86ISD::MOVSHDUP:
2767 case X86ISD::MOVSLDUP:
2768 case X86ISD::MOVDDUP:
2769 return DAG.getNode(Opc, dl, VT, V1);
2775 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2776 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2778 default: llvm_unreachable("Unknown x86 shuffle node");
2779 case X86ISD::PSHUFD:
2780 case X86ISD::PSHUFHW:
2781 case X86ISD::PSHUFLW:
2782 case X86ISD::VPERMILPS:
2783 case X86ISD::VPERMILPSY:
2784 case X86ISD::VPERMILPD:
2785 case X86ISD::VPERMILPDY:
2786 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2792 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2793 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2795 default: llvm_unreachable("Unknown x86 shuffle node");
2796 case X86ISD::PALIGN:
2797 case X86ISD::SHUFPD:
2798 case X86ISD::SHUFPS:
2799 case X86ISD::VPERM2F128:
2800 return DAG.getNode(Opc, dl, VT, V1, V2,
2801 DAG.getConstant(TargetMask, MVT::i8));
2806 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2807 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2809 default: llvm_unreachable("Unknown x86 shuffle node");
2810 case X86ISD::MOVLHPS:
2811 case X86ISD::MOVLHPD:
2812 case X86ISD::MOVHLPS:
2813 case X86ISD::MOVLPS:
2814 case X86ISD::MOVLPD:
2817 case X86ISD::UNPCKLPS:
2818 case X86ISD::UNPCKLPD:
2819 case X86ISD::VUNPCKLPSY:
2820 case X86ISD::VUNPCKLPDY:
2821 case X86ISD::PUNPCKLWD:
2822 case X86ISD::PUNPCKLBW:
2823 case X86ISD::PUNPCKLDQ:
2824 case X86ISD::PUNPCKLQDQ:
2825 case X86ISD::UNPCKHPS:
2826 case X86ISD::UNPCKHPD:
2827 case X86ISD::VUNPCKHPSY:
2828 case X86ISD::VUNPCKHPDY:
2829 case X86ISD::PUNPCKHWD:
2830 case X86ISD::PUNPCKHBW:
2831 case X86ISD::PUNPCKHDQ:
2832 case X86ISD::PUNPCKHQDQ:
2833 return DAG.getNode(Opc, dl, VT, V1, V2);
2838 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2839 MachineFunction &MF = DAG.getMachineFunction();
2840 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2841 int ReturnAddrIndex = FuncInfo->getRAIndex();
2843 if (ReturnAddrIndex == 0) {
2844 // Set up a frame object for the return address.
2845 uint64_t SlotSize = TD->getPointerSize();
2846 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2848 FuncInfo->setRAIndex(ReturnAddrIndex);
2851 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2855 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2856 bool hasSymbolicDisplacement) {
2857 // Offset should fit into 32 bit immediate field.
2858 if (!isInt<32>(Offset))
2861 // If we don't have a symbolic displacement - we don't have any extra
2863 if (!hasSymbolicDisplacement)
2866 // FIXME: Some tweaks might be needed for medium code model.
2867 if (M != CodeModel::Small && M != CodeModel::Kernel)
2870 // For small code model we assume that latest object is 16MB before end of 31
2871 // bits boundary. We may also accept pretty large negative constants knowing
2872 // that all objects are in the positive half of address space.
2873 if (M == CodeModel::Small && Offset < 16*1024*1024)
2876 // For kernel code model we know that all object resist in the negative half
2877 // of 32bits address space. We may not accept negative offsets, since they may
2878 // be just off and we may accept pretty large positive ones.
2879 if (M == CodeModel::Kernel && Offset > 0)
2885 /// isCalleePop - Determines whether the callee is required to pop its
2886 /// own arguments. Callee pop is necessary to support tail calls.
2887 bool X86::isCalleePop(CallingConv::ID CallingConv,
2888 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
2892 switch (CallingConv) {
2895 case CallingConv::X86_StdCall:
2897 case CallingConv::X86_FastCall:
2899 case CallingConv::X86_ThisCall:
2901 case CallingConv::Fast:
2903 case CallingConv::GHC:
2908 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2909 /// specific condition code, returning the condition code and the LHS/RHS of the
2910 /// comparison to make.
2911 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2912 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2914 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2915 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2916 // X > -1 -> X == 0, jump !sign.
2917 RHS = DAG.getConstant(0, RHS.getValueType());
2918 return X86::COND_NS;
2919 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2920 // X < 0 -> X == 0, jump on sign.
2922 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2924 RHS = DAG.getConstant(0, RHS.getValueType());
2925 return X86::COND_LE;
2929 switch (SetCCOpcode) {
2930 default: llvm_unreachable("Invalid integer condition!");
2931 case ISD::SETEQ: return X86::COND_E;
2932 case ISD::SETGT: return X86::COND_G;
2933 case ISD::SETGE: return X86::COND_GE;
2934 case ISD::SETLT: return X86::COND_L;
2935 case ISD::SETLE: return X86::COND_LE;
2936 case ISD::SETNE: return X86::COND_NE;
2937 case ISD::SETULT: return X86::COND_B;
2938 case ISD::SETUGT: return X86::COND_A;
2939 case ISD::SETULE: return X86::COND_BE;
2940 case ISD::SETUGE: return X86::COND_AE;
2944 // First determine if it is required or is profitable to flip the operands.
2946 // If LHS is a foldable load, but RHS is not, flip the condition.
2947 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
2948 !ISD::isNON_EXTLoad(RHS.getNode())) {
2949 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2950 std::swap(LHS, RHS);
2953 switch (SetCCOpcode) {
2959 std::swap(LHS, RHS);
2963 // On a floating point condition, the flags are set as follows:
2965 // 0 | 0 | 0 | X > Y
2966 // 0 | 0 | 1 | X < Y
2967 // 1 | 0 | 0 | X == Y
2968 // 1 | 1 | 1 | unordered
2969 switch (SetCCOpcode) {
2970 default: llvm_unreachable("Condcode should be pre-legalized away");
2972 case ISD::SETEQ: return X86::COND_E;
2973 case ISD::SETOLT: // flipped
2975 case ISD::SETGT: return X86::COND_A;
2976 case ISD::SETOLE: // flipped
2978 case ISD::SETGE: return X86::COND_AE;
2979 case ISD::SETUGT: // flipped
2981 case ISD::SETLT: return X86::COND_B;
2982 case ISD::SETUGE: // flipped
2984 case ISD::SETLE: return X86::COND_BE;
2986 case ISD::SETNE: return X86::COND_NE;
2987 case ISD::SETUO: return X86::COND_P;
2988 case ISD::SETO: return X86::COND_NP;
2990 case ISD::SETUNE: return X86::COND_INVALID;
2994 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2995 /// code. Current x86 isa includes the following FP cmov instructions:
2996 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2997 static bool hasFPCMov(unsigned X86CC) {
3013 /// isFPImmLegal - Returns true if the target can instruction select the
3014 /// specified FP immediate natively. If false, the legalizer will
3015 /// materialize the FP immediate as a load from a constant pool.
3016 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3017 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3018 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3024 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3025 /// the specified range (L, H].
3026 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3027 return (Val < 0) || (Val >= Low && Val < Hi);
3030 /// isUndefOrInRange - Return true if every element in Mask, begining
3031 /// from position Pos and ending in Pos+Size, falls within the specified
3032 /// range (L, L+Pos]. or is undef.
3033 static bool isUndefOrInRange(const SmallVectorImpl<int> &Mask,
3034 int Pos, int Size, int Low, int Hi) {
3035 for (int i = Pos, e = Pos+Size; i != e; ++i)
3036 if (!isUndefOrInRange(Mask[i], Low, Hi))
3041 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3042 /// specified value.
3043 static bool isUndefOrEqual(int Val, int CmpVal) {
3044 if (Val < 0 || Val == CmpVal)
3049 /// isSequentialOrUndefInRange - Return true if every element in Mask, begining
3050 /// from position Pos and ending in Pos+Size, falls within the specified
3051 /// sequential range (L, L+Pos]. or is undef.
3052 static bool isSequentialOrUndefInRange(const SmallVectorImpl<int> &Mask,
3053 int Pos, int Size, int Low) {
3054 for (int i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3055 if (!isUndefOrEqual(Mask[i], Low))
3060 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3061 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
3062 /// the second operand.
3063 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3064 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
3065 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3066 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3067 return (Mask[0] < 2 && Mask[1] < 2);
3071 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
3072 SmallVector<int, 8> M;
3074 return ::isPSHUFDMask(M, N->getValueType(0));
3077 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3078 /// is suitable for input to PSHUFHW.
3079 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3080 if (VT != MVT::v8i16)
3083 // Lower quadword copied in order or undef.
3084 for (int i = 0; i != 4; ++i)
3085 if (Mask[i] >= 0 && Mask[i] != i)
3088 // Upper quadword shuffled.
3089 for (int i = 4; i != 8; ++i)
3090 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
3096 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
3097 SmallVector<int, 8> M;
3099 return ::isPSHUFHWMask(M, N->getValueType(0));
3102 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3103 /// is suitable for input to PSHUFLW.
3104 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3105 if (VT != MVT::v8i16)
3108 // Upper quadword copied in order.
3109 for (int i = 4; i != 8; ++i)
3110 if (Mask[i] >= 0 && Mask[i] != i)
3113 // Lower quadword shuffled.
3114 for (int i = 0; i != 4; ++i)
3121 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
3122 SmallVector<int, 8> M;
3124 return ::isPSHUFLWMask(M, N->getValueType(0));
3127 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3128 /// is suitable for input to PALIGNR.
3129 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
3131 int i, e = VT.getVectorNumElements();
3132 if (VT.getSizeInBits() != 128 && VT.getSizeInBits() != 64)
3135 // Do not handle v2i64 / v2f64 shuffles with palignr.
3136 if (e < 4 || !hasSSSE3)
3139 for (i = 0; i != e; ++i)
3143 // All undef, not a palignr.
3147 // Make sure we're shifting in the right direction.
3151 int s = Mask[i] - i;
3153 // Check the rest of the elements to see if they are consecutive.
3154 for (++i; i != e; ++i) {
3156 if (m >= 0 && m != s+i)
3162 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3163 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
3164 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3165 int NumElems = VT.getVectorNumElements();
3166 if (NumElems != 2 && NumElems != 4)
3169 int Half = NumElems / 2;
3170 for (int i = 0; i < Half; ++i)
3171 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3173 for (int i = Half; i < NumElems; ++i)
3174 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3180 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3181 SmallVector<int, 8> M;
3183 return ::isSHUFPMask(M, N->getValueType(0));
3186 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3187 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3188 /// half elements to come from vector 1 (which would equal the dest.) and
3189 /// the upper half to come from vector 2.
3190 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3191 int NumElems = VT.getVectorNumElements();
3193 if (NumElems != 2 && NumElems != 4)
3196 int Half = NumElems / 2;
3197 for (int i = 0; i < Half; ++i)
3198 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3200 for (int i = Half; i < NumElems; ++i)
3201 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3206 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3207 SmallVector<int, 8> M;
3209 return isCommutedSHUFPMask(M, N->getValueType(0));
3212 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3213 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3214 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3215 EVT VT = N->getValueType(0);
3216 unsigned NumElems = VT.getVectorNumElements();
3218 if (VT.getSizeInBits() != 128)
3224 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3225 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3226 isUndefOrEqual(N->getMaskElt(1), 7) &&
3227 isUndefOrEqual(N->getMaskElt(2), 2) &&
3228 isUndefOrEqual(N->getMaskElt(3), 3);
3231 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3232 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3234 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3235 EVT VT = N->getValueType(0);
3236 unsigned NumElems = VT.getVectorNumElements();
3238 if (VT.getSizeInBits() != 128)
3244 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3245 isUndefOrEqual(N->getMaskElt(1), 3) &&
3246 isUndefOrEqual(N->getMaskElt(2), 2) &&
3247 isUndefOrEqual(N->getMaskElt(3), 3);
3250 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3251 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3252 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3253 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3255 if (NumElems != 2 && NumElems != 4)
3258 for (unsigned i = 0; i < NumElems/2; ++i)
3259 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3262 for (unsigned i = NumElems/2; i < NumElems; ++i)
3263 if (!isUndefOrEqual(N->getMaskElt(i), i))
3269 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3270 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3271 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3272 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3274 if ((NumElems != 2 && NumElems != 4)
3275 || N->getValueType(0).getSizeInBits() > 128)
3278 for (unsigned i = 0; i < NumElems/2; ++i)
3279 if (!isUndefOrEqual(N->getMaskElt(i), i))
3282 for (unsigned i = 0; i < NumElems/2; ++i)
3283 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3289 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3290 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3291 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3292 bool V2IsSplat = false) {
3293 int NumElts = VT.getVectorNumElements();
3295 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3296 "Unsupported vector type for unpckh");
3298 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8)
3301 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3302 // independently on 128-bit lanes.
3303 unsigned NumLanes = VT.getSizeInBits()/128;
3304 unsigned NumLaneElts = NumElts/NumLanes;
3307 unsigned End = NumLaneElts;
3308 for (unsigned s = 0; s < NumLanes; ++s) {
3309 for (unsigned i = Start, j = s * NumLaneElts;
3313 int BitI1 = Mask[i+1];
3314 if (!isUndefOrEqual(BitI, j))
3317 if (!isUndefOrEqual(BitI1, NumElts))
3320 if (!isUndefOrEqual(BitI1, j + NumElts))
3324 // Process the next 128 bits.
3325 Start += NumLaneElts;
3332 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3333 SmallVector<int, 8> M;
3335 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3338 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3339 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3340 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3341 bool V2IsSplat = false) {
3342 int NumElts = VT.getVectorNumElements();
3344 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3345 "Unsupported vector type for unpckh");
3347 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8)
3350 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3351 // independently on 128-bit lanes.
3352 unsigned NumLanes = VT.getSizeInBits()/128;
3353 unsigned NumLaneElts = NumElts/NumLanes;
3356 unsigned End = NumLaneElts;
3357 for (unsigned l = 0; l != NumLanes; ++l) {
3358 for (unsigned i = Start, j = (l*NumLaneElts)+NumLaneElts/2;
3359 i != End; i += 2, ++j) {
3361 int BitI1 = Mask[i+1];
3362 if (!isUndefOrEqual(BitI, j))
3365 if (isUndefOrEqual(BitI1, NumElts))
3368 if (!isUndefOrEqual(BitI1, j+NumElts))
3372 // Process the next 128 bits.
3373 Start += NumLaneElts;
3379 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3380 SmallVector<int, 8> M;
3382 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3385 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3386 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3388 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3389 int NumElems = VT.getVectorNumElements();
3390 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3393 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3394 // independently on 128-bit lanes.
3395 unsigned NumLanes = VT.getSizeInBits() / 128;
3396 unsigned NumLaneElts = NumElems / NumLanes;
3398 for (unsigned s = 0; s < NumLanes; ++s) {
3399 for (unsigned i = s * NumLaneElts, j = s * NumLaneElts;
3400 i != NumLaneElts * (s + 1);
3403 int BitI1 = Mask[i+1];
3405 if (!isUndefOrEqual(BitI, j))
3407 if (!isUndefOrEqual(BitI1, j))
3415 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3416 SmallVector<int, 8> M;
3418 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3421 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3422 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3424 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3425 int NumElems = VT.getVectorNumElements();
3426 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3429 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3431 int BitI1 = Mask[i+1];
3432 if (!isUndefOrEqual(BitI, j))
3434 if (!isUndefOrEqual(BitI1, j))
3440 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3441 SmallVector<int, 8> M;
3443 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3446 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3447 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3448 /// MOVSD, and MOVD, i.e. setting the lowest element.
3449 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3450 if (VT.getVectorElementType().getSizeInBits() < 32)
3453 int NumElts = VT.getVectorNumElements();
3455 if (!isUndefOrEqual(Mask[0], NumElts))
3458 for (int i = 1; i < NumElts; ++i)
3459 if (!isUndefOrEqual(Mask[i], i))
3465 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3466 SmallVector<int, 8> M;
3468 return ::isMOVLMask(M, N->getValueType(0));
3471 /// isVPERM2F128Mask - Match 256-bit shuffles where the elements are considered
3472 /// as permutations between 128-bit chunks or halves. As an example: this
3474 /// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15>
3475 /// The first half comes from the second half of V1 and the second half from the
3476 /// the second half of V2.
3477 static bool isVPERM2F128Mask(const SmallVectorImpl<int> &Mask, EVT VT,
3478 const X86Subtarget *Subtarget) {
3479 if (!Subtarget->hasAVX() || VT.getSizeInBits() != 256)
3482 // The shuffle result is divided into half A and half B. In total the two
3483 // sources have 4 halves, namely: C, D, E, F. The final values of A and
3484 // B must come from C, D, E or F.
3485 int HalfSize = VT.getVectorNumElements()/2;
3486 bool MatchA = false, MatchB = false;
3488 // Check if A comes from one of C, D, E, F.
3489 for (int Half = 0; Half < 4; ++Half) {
3490 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) {
3496 // Check if B comes from one of C, D, E, F.
3497 for (int Half = 0; Half < 4; ++Half) {
3498 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) {
3504 return MatchA && MatchB;
3507 /// getShuffleVPERM2F128Immediate - Return the appropriate immediate to shuffle
3508 /// the specified VECTOR_MASK mask with VPERM2F128 instructions.
3509 static unsigned getShuffleVPERM2F128Immediate(SDNode *N) {
3510 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3511 EVT VT = SVOp->getValueType(0);
3513 int HalfSize = VT.getVectorNumElements()/2;
3515 int FstHalf = 0, SndHalf = 0;
3516 for (int i = 0; i < HalfSize; ++i) {
3517 if (SVOp->getMaskElt(i) > 0) {
3518 FstHalf = SVOp->getMaskElt(i)/HalfSize;
3522 for (int i = HalfSize; i < HalfSize*2; ++i) {
3523 if (SVOp->getMaskElt(i) > 0) {
3524 SndHalf = SVOp->getMaskElt(i)/HalfSize;
3529 return (FstHalf | (SndHalf << 4));
3532 /// isVPERMILPDMask - Return true if the specified VECTOR_SHUFFLE operand
3533 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
3534 /// Note that VPERMIL mask matching is different depending whether theunderlying
3535 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
3536 /// to the same elements of the low, but to the higher half of the source.
3537 /// In VPERMILPD the two lanes could be shuffled independently of each other
3538 /// with the same restriction that lanes can't be crossed.
3539 static bool isVPERMILPDMask(const SmallVectorImpl<int> &Mask, EVT VT,
3540 const X86Subtarget *Subtarget) {
3541 int NumElts = VT.getVectorNumElements();
3542 int NumLanes = VT.getSizeInBits()/128;
3544 if (!Subtarget->hasAVX())
3547 // Match any permutation of 128-bit vector with 64-bit types
3548 if (NumLanes == 1 && NumElts != 2)
3551 // Only match 256-bit with 32 types
3552 if (VT.getSizeInBits() == 256 && NumElts != 4)
3555 // The mask on the high lane is independent of the low. Both can match
3556 // any element in inside its own lane, but can't cross.
3557 int LaneSize = NumElts/NumLanes;
3558 for (int l = 0; l < NumLanes; ++l)
3559 for (int i = l*LaneSize; i < LaneSize*(l+1); ++i) {
3560 int LaneStart = l*LaneSize;
3561 if (!isUndefOrInRange(Mask[i], LaneStart, LaneStart+LaneSize))
3568 /// isVPERMILPSMask - Return true if the specified VECTOR_SHUFFLE operand
3569 /// specifies a shuffle of elements that is suitable for input to VPERMILPS*.
3570 /// Note that VPERMIL mask matching is different depending whether theunderlying
3571 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
3572 /// to the same elements of the low, but to the higher half of the source.
3573 /// In VPERMILPD the two lanes could be shuffled independently of each other
3574 /// with the same restriction that lanes can't be crossed.
3575 static bool isVPERMILPSMask(const SmallVectorImpl<int> &Mask, EVT VT,
3576 const X86Subtarget *Subtarget) {
3577 unsigned NumElts = VT.getVectorNumElements();
3578 unsigned NumLanes = VT.getSizeInBits()/128;
3580 if (!Subtarget->hasAVX())
3583 // Match any permutation of 128-bit vector with 32-bit types
3584 if (NumLanes == 1 && NumElts != 4)
3587 // Only match 256-bit with 32 types
3588 if (VT.getSizeInBits() == 256 && NumElts != 8)
3591 // The mask on the high lane should be the same as the low. Actually,
3592 // they can differ if any of the corresponding index in a lane is undef
3593 // and the other stays in range.
3594 int LaneSize = NumElts/NumLanes;
3595 for (int i = 0; i < LaneSize; ++i) {
3596 int HighElt = i+LaneSize;
3597 bool HighValid = isUndefOrInRange(Mask[HighElt], LaneSize, NumElts);
3598 bool LowValid = isUndefOrInRange(Mask[i], 0, LaneSize);
3600 if (!HighValid || !LowValid)
3602 if (Mask[i] < 0 || Mask[HighElt] < 0)
3604 if (Mask[HighElt]-Mask[i] != LaneSize)
3611 /// getShuffleVPERMILPSImmediate - Return the appropriate immediate to shuffle
3612 /// the specified VECTOR_MASK mask with VPERMILPS* instructions.
3613 static unsigned getShuffleVPERMILPSImmediate(SDNode *N) {
3614 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3615 EVT VT = SVOp->getValueType(0);
3617 int NumElts = VT.getVectorNumElements();
3618 int NumLanes = VT.getSizeInBits()/128;
3619 int LaneSize = NumElts/NumLanes;
3621 // Although the mask is equal for both lanes do it twice to get the cases
3622 // where a mask will match because the same mask element is undef on the
3623 // first half but valid on the second. This would get pathological cases
3624 // such as: shuffle <u, 0, 1, 2, 4, 4, 5, 6>, which is completely valid.
3626 for (int l = 0; l < NumLanes; ++l) {
3627 for (int i = 0; i < LaneSize; ++i) {
3628 int MaskElt = SVOp->getMaskElt(i+(l*LaneSize));
3631 if (MaskElt >= LaneSize)
3632 MaskElt -= LaneSize;
3633 Mask |= MaskElt << (i*2);
3640 /// getShuffleVPERMILPDImmediate - Return the appropriate immediate to shuffle
3641 /// the specified VECTOR_MASK mask with VPERMILPD* instructions.
3642 static unsigned getShuffleVPERMILPDImmediate(SDNode *N) {
3643 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3644 EVT VT = SVOp->getValueType(0);
3646 int NumElts = VT.getVectorNumElements();
3647 int NumLanes = VT.getSizeInBits()/128;
3650 int LaneSize = NumElts/NumLanes;
3651 for (int l = 0; l < NumLanes; ++l)
3652 for (int i = l*LaneSize; i < LaneSize*(l+1); ++i) {
3653 int MaskElt = SVOp->getMaskElt(i);
3656 Mask |= (MaskElt-l*LaneSize) << i;
3662 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3663 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3664 /// element of vector 2 and the other elements to come from vector 1 in order.
3665 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3666 bool V2IsSplat = false, bool V2IsUndef = false) {
3667 int NumOps = VT.getVectorNumElements();
3668 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3671 if (!isUndefOrEqual(Mask[0], 0))
3674 for (int i = 1; i < NumOps; ++i)
3675 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3676 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3677 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3683 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3684 bool V2IsUndef = false) {
3685 SmallVector<int, 8> M;
3687 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3690 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3691 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3692 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
3693 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N,
3694 const X86Subtarget *Subtarget) {
3695 if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
3698 // The second vector must be undef
3699 if (N->getOperand(1).getOpcode() != ISD::UNDEF)
3702 EVT VT = N->getValueType(0);
3703 unsigned NumElems = VT.getVectorNumElements();
3705 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3706 (VT.getSizeInBits() == 256 && NumElems != 8))
3709 // "i+1" is the value the indexed mask element must have
3710 for (unsigned i = 0; i < NumElems; i += 2)
3711 if (!isUndefOrEqual(N->getMaskElt(i), i+1) ||
3712 !isUndefOrEqual(N->getMaskElt(i+1), i+1))
3718 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3719 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3720 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
3721 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N,
3722 const X86Subtarget *Subtarget) {
3723 if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
3726 // The second vector must be undef
3727 if (N->getOperand(1).getOpcode() != ISD::UNDEF)
3730 EVT VT = N->getValueType(0);
3731 unsigned NumElems = VT.getVectorNumElements();
3733 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3734 (VT.getSizeInBits() == 256 && NumElems != 8))
3737 // "i" is the value the indexed mask element must have
3738 for (unsigned i = 0; i < NumElems; i += 2)
3739 if (!isUndefOrEqual(N->getMaskElt(i), i) ||
3740 !isUndefOrEqual(N->getMaskElt(i+1), i))
3746 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3747 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
3748 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3749 int e = N->getValueType(0).getVectorNumElements() / 2;
3751 for (int i = 0; i < e; ++i)
3752 if (!isUndefOrEqual(N->getMaskElt(i), i))
3754 for (int i = 0; i < e; ++i)
3755 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3760 /// isVEXTRACTF128Index - Return true if the specified
3761 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3762 /// suitable for input to VEXTRACTF128.
3763 bool X86::isVEXTRACTF128Index(SDNode *N) {
3764 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3767 // The index should be aligned on a 128-bit boundary.
3769 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3771 unsigned VL = N->getValueType(0).getVectorNumElements();
3772 unsigned VBits = N->getValueType(0).getSizeInBits();
3773 unsigned ElSize = VBits / VL;
3774 bool Result = (Index * ElSize) % 128 == 0;
3779 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
3780 /// operand specifies a subvector insert that is suitable for input to
3782 bool X86::isVINSERTF128Index(SDNode *N) {
3783 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3786 // The index should be aligned on a 128-bit boundary.
3788 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3790 unsigned VL = N->getValueType(0).getVectorNumElements();
3791 unsigned VBits = N->getValueType(0).getSizeInBits();
3792 unsigned ElSize = VBits / VL;
3793 bool Result = (Index * ElSize) % 128 == 0;
3798 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3799 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3800 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3801 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3802 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3804 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3806 for (int i = 0; i < NumOperands; ++i) {
3807 int Val = SVOp->getMaskElt(NumOperands-i-1);
3808 if (Val < 0) Val = 0;
3809 if (Val >= NumOperands) Val -= NumOperands;
3811 if (i != NumOperands - 1)
3817 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3818 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3819 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3820 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3822 // 8 nodes, but we only care about the last 4.
3823 for (unsigned i = 7; i >= 4; --i) {
3824 int Val = SVOp->getMaskElt(i);
3833 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3834 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3835 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3836 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3838 // 8 nodes, but we only care about the first 4.
3839 for (int i = 3; i >= 0; --i) {
3840 int Val = SVOp->getMaskElt(i);
3849 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3850 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3851 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3852 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3853 EVT VVT = N->getValueType(0);
3854 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3858 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3859 Val = SVOp->getMaskElt(i);
3863 assert(Val - i > 0 && "PALIGNR imm should be positive");
3864 return (Val - i) * EltSize;
3867 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
3868 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
3870 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
3871 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3872 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
3875 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3877 EVT VecVT = N->getOperand(0).getValueType();
3878 EVT ElVT = VecVT.getVectorElementType();
3880 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3881 return Index / NumElemsPerChunk;
3884 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
3885 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
3887 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
3888 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
3889 llvm_unreachable("Illegal insert subvector for VINSERTF128");
3892 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
3894 EVT VecVT = N->getValueType(0);
3895 EVT ElVT = VecVT.getVectorElementType();
3897 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
3898 return Index / NumElemsPerChunk;
3901 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3903 bool X86::isZeroNode(SDValue Elt) {
3904 return ((isa<ConstantSDNode>(Elt) &&
3905 cast<ConstantSDNode>(Elt)->isNullValue()) ||
3906 (isa<ConstantFPSDNode>(Elt) &&
3907 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3910 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3911 /// their permute mask.
3912 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3913 SelectionDAG &DAG) {
3914 EVT VT = SVOp->getValueType(0);
3915 unsigned NumElems = VT.getVectorNumElements();
3916 SmallVector<int, 8> MaskVec;
3918 for (unsigned i = 0; i != NumElems; ++i) {
3919 int idx = SVOp->getMaskElt(i);
3921 MaskVec.push_back(idx);
3922 else if (idx < (int)NumElems)
3923 MaskVec.push_back(idx + NumElems);
3925 MaskVec.push_back(idx - NumElems);
3927 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3928 SVOp->getOperand(0), &MaskVec[0]);
3931 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3932 /// the two vector operands have swapped position.
3933 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3934 unsigned NumElems = VT.getVectorNumElements();
3935 for (unsigned i = 0; i != NumElems; ++i) {
3939 else if (idx < (int)NumElems)
3940 Mask[i] = idx + NumElems;
3942 Mask[i] = idx - NumElems;
3946 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3947 /// match movhlps. The lower half elements should come from upper half of
3948 /// V1 (and in order), and the upper half elements should come from the upper
3949 /// half of V2 (and in order).
3950 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3951 EVT VT = Op->getValueType(0);
3952 if (VT.getSizeInBits() != 128)
3954 if (VT.getVectorNumElements() != 4)
3956 for (unsigned i = 0, e = 2; i != e; ++i)
3957 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3959 for (unsigned i = 2; i != 4; ++i)
3960 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3965 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3966 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3968 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3969 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3971 N = N->getOperand(0).getNode();
3972 if (!ISD::isNON_EXTLoad(N))
3975 *LD = cast<LoadSDNode>(N);
3979 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3980 /// match movlp{s|d}. The lower half elements should come from lower half of
3981 /// V1 (and in order), and the upper half elements should come from the upper
3982 /// half of V2 (and in order). And since V1 will become the source of the
3983 /// MOVLP, it must be either a vector load or a scalar load to vector.
3984 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3985 ShuffleVectorSDNode *Op) {
3986 EVT VT = Op->getValueType(0);
3987 if (VT.getSizeInBits() != 128)
3990 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3992 // Is V2 is a vector load, don't do this transformation. We will try to use
3993 // load folding shufps op.
3994 if (ISD::isNON_EXTLoad(V2))
3997 unsigned NumElems = VT.getVectorNumElements();
3999 if (NumElems != 2 && NumElems != 4)
4001 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4002 if (!isUndefOrEqual(Op->getMaskElt(i), i))
4004 for (unsigned i = NumElems/2; i != NumElems; ++i)
4005 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
4010 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
4012 static bool isSplatVector(SDNode *N) {
4013 if (N->getOpcode() != ISD::BUILD_VECTOR)
4016 SDValue SplatValue = N->getOperand(0);
4017 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
4018 if (N->getOperand(i) != SplatValue)
4023 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
4024 /// to an zero vector.
4025 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
4026 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
4027 SDValue V1 = N->getOperand(0);
4028 SDValue V2 = N->getOperand(1);
4029 unsigned NumElems = N->getValueType(0).getVectorNumElements();
4030 for (unsigned i = 0; i != NumElems; ++i) {
4031 int Idx = N->getMaskElt(i);
4032 if (Idx >= (int)NumElems) {
4033 unsigned Opc = V2.getOpcode();
4034 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
4036 if (Opc != ISD::BUILD_VECTOR ||
4037 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
4039 } else if (Idx >= 0) {
4040 unsigned Opc = V1.getOpcode();
4041 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
4043 if (Opc != ISD::BUILD_VECTOR ||
4044 !X86::isZeroNode(V1.getOperand(Idx)))
4051 /// getZeroVector - Returns a vector of specified type with all zero elements.
4053 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
4055 assert(VT.isVector() && "Expected a vector type");
4057 // Always build SSE zero vectors as <4 x i32> bitcasted
4058 // to their dest type. This ensures they get CSE'd.
4060 if (VT.getSizeInBits() == 128) { // SSE
4061 if (HasSSE2) { // SSE2
4062 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4063 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4065 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4066 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4068 } else if (VT.getSizeInBits() == 256) { // AVX
4069 // 256-bit logic and arithmetic instructions in AVX are
4070 // all floating-point, no support for integer ops. Default
4071 // to emitting fp zeroed vectors then.
4072 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4073 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4074 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
4076 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4079 /// getOnesVector - Returns a vector of specified type with all bits set.
4080 /// Always build ones vectors as <4 x i32>. For 256-bit types, use two
4081 /// <4 x i32> inserted in a <8 x i32> appropriately. Then bitcast to their
4082 /// original type, ensuring they get CSE'd.
4083 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4084 assert(VT.isVector() && "Expected a vector type");
4085 assert((VT.is128BitVector() || VT.is256BitVector())
4086 && "Expected a 128-bit or 256-bit vector type");
4088 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
4089 SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
4090 Cst, Cst, Cst, Cst);
4092 if (VT.is256BitVector()) {
4093 SDValue InsV = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, MVT::v8i32),
4094 Vec, DAG.getConstant(0, MVT::i32), DAG, dl);
4095 Vec = Insert128BitVector(InsV, Vec,
4096 DAG.getConstant(4 /* NumElems/2 */, MVT::i32), DAG, dl);
4099 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4102 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
4103 /// that point to V2 points to its first element.
4104 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4105 EVT VT = SVOp->getValueType(0);
4106 unsigned NumElems = VT.getVectorNumElements();
4108 bool Changed = false;
4109 SmallVector<int, 8> MaskVec;
4110 SVOp->getMask(MaskVec);
4112 for (unsigned i = 0; i != NumElems; ++i) {
4113 if (MaskVec[i] > (int)NumElems) {
4114 MaskVec[i] = NumElems;
4119 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
4120 SVOp->getOperand(1), &MaskVec[0]);
4121 return SDValue(SVOp, 0);
4124 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4125 /// operation of specified width.
4126 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4128 unsigned NumElems = VT.getVectorNumElements();
4129 SmallVector<int, 8> Mask;
4130 Mask.push_back(NumElems);
4131 for (unsigned i = 1; i != NumElems; ++i)
4133 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4136 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4137 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4139 unsigned NumElems = VT.getVectorNumElements();
4140 SmallVector<int, 8> Mask;
4141 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4143 Mask.push_back(i + NumElems);
4145 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4148 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4149 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4151 unsigned NumElems = VT.getVectorNumElements();
4152 unsigned Half = NumElems/2;
4153 SmallVector<int, 8> Mask;
4154 for (unsigned i = 0; i != Half; ++i) {
4155 Mask.push_back(i + Half);
4156 Mask.push_back(i + NumElems + Half);
4158 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4161 // PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by
4162 // a generic shuffle instruction because the target has no such instructions.
4163 // Generate shuffles which repeat i16 and i8 several times until they can be
4164 // represented by v4f32 and then be manipulated by target suported shuffles.
4165 static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) {
4166 EVT VT = V.getValueType();
4167 int NumElems = VT.getVectorNumElements();
4168 DebugLoc dl = V.getDebugLoc();
4170 while (NumElems > 4) {
4171 if (EltNo < NumElems/2) {
4172 V = getUnpackl(DAG, dl, VT, V, V);
4174 V = getUnpackh(DAG, dl, VT, V, V);
4175 EltNo -= NumElems/2;
4182 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
4183 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
4184 EVT VT = V.getValueType();
4185 DebugLoc dl = V.getDebugLoc();
4186 assert((VT.getSizeInBits() == 128 || VT.getSizeInBits() == 256)
4187 && "Vector size not supported");
4189 bool Is128 = VT.getSizeInBits() == 128;
4190 EVT NVT = Is128 ? MVT::v4f32 : MVT::v8f32;
4191 V = DAG.getNode(ISD::BITCAST, dl, NVT, V);
4194 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
4195 V = DAG.getVectorShuffle(NVT, dl, V, DAG.getUNDEF(NVT), &SplatMask[0]);
4197 // The second half of indicies refer to the higher part, which is a
4198 // duplication of the lower one. This makes this shuffle a perfect match
4199 // for the VPERM instruction.
4200 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
4201 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
4202 V = DAG.getVectorShuffle(NVT, dl, V, DAG.getUNDEF(NVT), &SplatMask[0]);
4205 return DAG.getNode(ISD::BITCAST, dl, VT, V);
4208 /// PromoteSplat - Promote a splat of v4i32, v8i16 or v16i8 to v4f32 and
4209 /// v8i32, v16i16 or v32i8 to v8f32.
4210 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
4211 EVT SrcVT = SV->getValueType(0);
4212 SDValue V1 = SV->getOperand(0);
4213 DebugLoc dl = SV->getDebugLoc();
4215 int EltNo = SV->getSplatIndex();
4216 int NumElems = SrcVT.getVectorNumElements();
4217 unsigned Size = SrcVT.getSizeInBits();
4219 // Extract the 128-bit part containing the splat element and update
4220 // the splat element index when it refers to the higher register.
4222 unsigned Idx = (EltNo > NumElems/2) ? NumElems/2 : 0;
4223 V1 = Extract128BitVector(V1, DAG.getConstant(Idx, MVT::i32), DAG, dl);
4225 EltNo -= NumElems/2;
4228 // Make this 128-bit vector duplicate i8 and i16 elements
4229 EVT EltVT = SrcVT.getVectorElementType();
4230 if (NumElems > 4 && (EltVT == MVT::i8 || EltVT == MVT::i16))
4231 V1 = PromoteSplati8i16(V1, DAG, EltNo);
4233 // Recreate the 256-bit vector and place the same 128-bit vector
4234 // into the low and high part. This is necessary because we want
4235 // to use VPERM to shuffle the v8f32 vector, and VPERM only shuffles
4236 // inside each separate v4f32 lane.
4238 SDValue InsV = Insert128BitVector(DAG.getUNDEF(SrcVT), V1,
4239 DAG.getConstant(0, MVT::i32), DAG, dl);
4240 V1 = Insert128BitVector(InsV, V1,
4241 DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
4244 return getLegalSplat(DAG, V1, EltNo);
4247 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4248 /// vector of zero or undef vector. This produces a shuffle where the low
4249 /// element of V2 is swizzled into the zero/undef vector, landing at element
4250 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4251 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4252 bool isZero, bool HasSSE2,
4253 SelectionDAG &DAG) {
4254 EVT VT = V2.getValueType();
4256 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
4257 unsigned NumElems = VT.getVectorNumElements();
4258 SmallVector<int, 16> MaskVec;
4259 for (unsigned i = 0; i != NumElems; ++i)
4260 // If this is the insertion idx, put the low elt of V2 here.
4261 MaskVec.push_back(i == Idx ? NumElems : i);
4262 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
4265 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4266 /// element of the result of the vector shuffle.
4267 static SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
4270 return SDValue(); // Limit search depth.
4272 SDValue V = SDValue(N, 0);
4273 EVT VT = V.getValueType();
4274 unsigned Opcode = V.getOpcode();
4276 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4277 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4278 Index = SV->getMaskElt(Index);
4281 return DAG.getUNDEF(VT.getVectorElementType());
4283 int NumElems = VT.getVectorNumElements();
4284 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
4285 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
4288 // Recurse into target specific vector shuffles to find scalars.
4289 if (isTargetShuffle(Opcode)) {
4290 int NumElems = VT.getVectorNumElements();
4291 SmallVector<unsigned, 16> ShuffleMask;
4295 case X86ISD::SHUFPS:
4296 case X86ISD::SHUFPD:
4297 ImmN = N->getOperand(N->getNumOperands()-1);
4298 DecodeSHUFPSMask(NumElems,
4299 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4302 case X86ISD::PUNPCKHBW:
4303 case X86ISD::PUNPCKHWD:
4304 case X86ISD::PUNPCKHDQ:
4305 case X86ISD::PUNPCKHQDQ:
4306 DecodePUNPCKHMask(NumElems, ShuffleMask);
4308 case X86ISD::UNPCKHPS:
4309 case X86ISD::UNPCKHPD:
4310 case X86ISD::VUNPCKHPSY:
4311 case X86ISD::VUNPCKHPDY:
4312 DecodeUNPCKHPMask(NumElems, ShuffleMask);
4314 case X86ISD::PUNPCKLBW:
4315 case X86ISD::PUNPCKLWD:
4316 case X86ISD::PUNPCKLDQ:
4317 case X86ISD::PUNPCKLQDQ:
4318 DecodePUNPCKLMask(VT, ShuffleMask);
4320 case X86ISD::UNPCKLPS:
4321 case X86ISD::UNPCKLPD:
4322 case X86ISD::VUNPCKLPSY:
4323 case X86ISD::VUNPCKLPDY:
4324 DecodeUNPCKLPMask(VT, ShuffleMask);
4326 case X86ISD::MOVHLPS:
4327 DecodeMOVHLPSMask(NumElems, ShuffleMask);
4329 case X86ISD::MOVLHPS:
4330 DecodeMOVLHPSMask(NumElems, ShuffleMask);
4332 case X86ISD::PSHUFD:
4333 ImmN = N->getOperand(N->getNumOperands()-1);
4334 DecodePSHUFMask(NumElems,
4335 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4338 case X86ISD::PSHUFHW:
4339 ImmN = N->getOperand(N->getNumOperands()-1);
4340 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4343 case X86ISD::PSHUFLW:
4344 ImmN = N->getOperand(N->getNumOperands()-1);
4345 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4349 case X86ISD::MOVSD: {
4350 // The index 0 always comes from the first element of the second source,
4351 // this is why MOVSS and MOVSD are used in the first place. The other
4352 // elements come from the other positions of the first source vector.
4353 unsigned OpNum = (Index == 0) ? 1 : 0;
4354 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
4357 case X86ISD::VPERMILPS:
4358 ImmN = N->getOperand(N->getNumOperands()-1);
4359 DecodeVPERMILPSMask(4, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4362 case X86ISD::VPERMILPSY:
4363 ImmN = N->getOperand(N->getNumOperands()-1);
4364 DecodeVPERMILPSMask(8, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4367 case X86ISD::VPERMILPD:
4368 ImmN = N->getOperand(N->getNumOperands()-1);
4369 DecodeVPERMILPDMask(2, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4372 case X86ISD::VPERMILPDY:
4373 ImmN = N->getOperand(N->getNumOperands()-1);
4374 DecodeVPERMILPDMask(4, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4377 case X86ISD::VPERM2F128:
4378 ImmN = N->getOperand(N->getNumOperands()-1);
4379 DecodeVPERM2F128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4383 assert("not implemented for target shuffle node");
4387 Index = ShuffleMask[Index];
4389 return DAG.getUNDEF(VT.getVectorElementType());
4391 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
4392 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
4396 // Actual nodes that may contain scalar elements
4397 if (Opcode == ISD::BITCAST) {
4398 V = V.getOperand(0);
4399 EVT SrcVT = V.getValueType();
4400 unsigned NumElems = VT.getVectorNumElements();
4402 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4406 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4407 return (Index == 0) ? V.getOperand(0)
4408 : DAG.getUNDEF(VT.getVectorElementType());
4410 if (V.getOpcode() == ISD::BUILD_VECTOR)
4411 return V.getOperand(Index);
4416 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
4417 /// shuffle operation which come from a consecutively from a zero. The
4418 /// search can start in two different directions, from left or right.
4420 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
4421 bool ZerosFromLeft, SelectionDAG &DAG) {
4424 while (i < NumElems) {
4425 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
4426 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
4427 if (!(Elt.getNode() &&
4428 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
4436 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
4437 /// MaskE correspond consecutively to elements from one of the vector operands,
4438 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
4440 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
4441 int OpIdx, int NumElems, unsigned &OpNum) {
4442 bool SeenV1 = false;
4443 bool SeenV2 = false;
4445 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
4446 int Idx = SVOp->getMaskElt(i);
4447 // Ignore undef indicies
4456 // Only accept consecutive elements from the same vector
4457 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
4461 OpNum = SeenV1 ? 0 : 1;
4465 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
4466 /// logical left shift of a vector.
4467 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4468 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4469 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4470 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4471 false /* check zeros from right */, DAG);
4477 // Considering the elements in the mask that are not consecutive zeros,
4478 // check if they consecutively come from only one of the source vectors.
4480 // V1 = {X, A, B, C} 0
4482 // vector_shuffle V1, V2 <1, 2, 3, X>
4484 if (!isShuffleMaskConsecutive(SVOp,
4485 0, // Mask Start Index
4486 NumElems-NumZeros-1, // Mask End Index
4487 NumZeros, // Where to start looking in the src vector
4488 NumElems, // Number of elements in vector
4489 OpSrc)) // Which source operand ?
4494 ShVal = SVOp->getOperand(OpSrc);
4498 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4499 /// logical left shift of a vector.
4500 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4501 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4502 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4503 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4504 true /* check zeros from left */, DAG);
4510 // Considering the elements in the mask that are not consecutive zeros,
4511 // check if they consecutively come from only one of the source vectors.
4513 // 0 { A, B, X, X } = V2
4515 // vector_shuffle V1, V2 <X, X, 4, 5>
4517 if (!isShuffleMaskConsecutive(SVOp,
4518 NumZeros, // Mask Start Index
4519 NumElems-1, // Mask End Index
4520 0, // Where to start looking in the src vector
4521 NumElems, // Number of elements in vector
4522 OpSrc)) // Which source operand ?
4527 ShVal = SVOp->getOperand(OpSrc);
4531 /// isVectorShift - Returns true if the shuffle can be implemented as a
4532 /// logical left or right shift of a vector.
4533 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4534 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4535 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4536 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4542 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4544 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4545 unsigned NumNonZero, unsigned NumZero,
4547 const TargetLowering &TLI) {
4551 DebugLoc dl = Op.getDebugLoc();
4554 for (unsigned i = 0; i < 16; ++i) {
4555 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4556 if (ThisIsNonZero && First) {
4558 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4560 V = DAG.getUNDEF(MVT::v8i16);
4565 SDValue ThisElt(0, 0), LastElt(0, 0);
4566 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4567 if (LastIsNonZero) {
4568 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4569 MVT::i16, Op.getOperand(i-1));
4571 if (ThisIsNonZero) {
4572 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4573 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4574 ThisElt, DAG.getConstant(8, MVT::i8));
4576 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4580 if (ThisElt.getNode())
4581 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4582 DAG.getIntPtrConstant(i/2));
4586 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4589 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4591 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4592 unsigned NumNonZero, unsigned NumZero,
4594 const TargetLowering &TLI) {
4598 DebugLoc dl = Op.getDebugLoc();
4601 for (unsigned i = 0; i < 8; ++i) {
4602 bool isNonZero = (NonZeros & (1 << i)) != 0;
4606 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4608 V = DAG.getUNDEF(MVT::v8i16);
4611 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4612 MVT::v8i16, V, Op.getOperand(i),
4613 DAG.getIntPtrConstant(i));
4620 /// getVShift - Return a vector logical shift node.
4622 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4623 unsigned NumBits, SelectionDAG &DAG,
4624 const TargetLowering &TLI, DebugLoc dl) {
4625 EVT ShVT = MVT::v2i64;
4626 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4627 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4628 return DAG.getNode(ISD::BITCAST, dl, VT,
4629 DAG.getNode(Opc, dl, ShVT, SrcOp,
4630 DAG.getConstant(NumBits,
4631 TLI.getShiftAmountTy(SrcOp.getValueType()))));
4635 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4636 SelectionDAG &DAG) const {
4638 // Check if the scalar load can be widened into a vector load. And if
4639 // the address is "base + cst" see if the cst can be "absorbed" into
4640 // the shuffle mask.
4641 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4642 SDValue Ptr = LD->getBasePtr();
4643 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4645 EVT PVT = LD->getValueType(0);
4646 if (PVT != MVT::i32 && PVT != MVT::f32)
4651 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4652 FI = FINode->getIndex();
4654 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4655 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4656 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4657 Offset = Ptr.getConstantOperandVal(1);
4658 Ptr = Ptr.getOperand(0);
4663 // FIXME: 256-bit vector instructions don't require a strict alignment,
4664 // improve this code to support it better.
4665 unsigned RequiredAlign = VT.getSizeInBits()/8;
4666 SDValue Chain = LD->getChain();
4667 // Make sure the stack object alignment is at least 16 or 32.
4668 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4669 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4670 if (MFI->isFixedObjectIndex(FI)) {
4671 // Can't change the alignment. FIXME: It's possible to compute
4672 // the exact stack offset and reference FI + adjust offset instead.
4673 // If someone *really* cares about this. That's the way to implement it.
4676 MFI->setObjectAlignment(FI, RequiredAlign);
4680 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4681 // Ptr + (Offset & ~15).
4684 if ((Offset % RequiredAlign) & 3)
4686 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4688 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4689 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4691 int EltNo = (Offset - StartOffset) >> 2;
4692 int NumElems = VT.getVectorNumElements();
4694 EVT CanonVT = VT.getSizeInBits() == 128 ? MVT::v4i32 : MVT::v8i32;
4695 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4696 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4697 LD->getPointerInfo().getWithOffset(StartOffset),
4700 // Canonicalize it to a v4i32 or v8i32 shuffle.
4701 SmallVector<int, 8> Mask;
4702 for (int i = 0; i < NumElems; ++i)
4703 Mask.push_back(EltNo);
4705 V1 = DAG.getNode(ISD::BITCAST, dl, CanonVT, V1);
4706 return DAG.getNode(ISD::BITCAST, dl, NVT,
4707 DAG.getVectorShuffle(CanonVT, dl, V1,
4708 DAG.getUNDEF(CanonVT),&Mask[0]));
4714 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4715 /// vector of type 'VT', see if the elements can be replaced by a single large
4716 /// load which has the same value as a build_vector whose operands are 'elts'.
4718 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4720 /// FIXME: we'd also like to handle the case where the last elements are zero
4721 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4722 /// There's even a handy isZeroNode for that purpose.
4723 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4724 DebugLoc &DL, SelectionDAG &DAG) {
4725 EVT EltVT = VT.getVectorElementType();
4726 unsigned NumElems = Elts.size();
4728 LoadSDNode *LDBase = NULL;
4729 unsigned LastLoadedElt = -1U;
4731 // For each element in the initializer, see if we've found a load or an undef.
4732 // If we don't find an initial load element, or later load elements are
4733 // non-consecutive, bail out.
4734 for (unsigned i = 0; i < NumElems; ++i) {
4735 SDValue Elt = Elts[i];
4737 if (!Elt.getNode() ||
4738 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4741 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4743 LDBase = cast<LoadSDNode>(Elt.getNode());
4747 if (Elt.getOpcode() == ISD::UNDEF)
4750 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4751 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4756 // If we have found an entire vector of loads and undefs, then return a large
4757 // load of the entire vector width starting at the base pointer. If we found
4758 // consecutive loads for the low half, generate a vzext_load node.
4759 if (LastLoadedElt == NumElems - 1) {
4760 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4761 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4762 LDBase->getPointerInfo(),
4763 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4764 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4765 LDBase->getPointerInfo(),
4766 LDBase->isVolatile(), LDBase->isNonTemporal(),
4767 LDBase->getAlignment());
4768 } else if (NumElems == 4 && LastLoadedElt == 1 &&
4769 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4770 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4771 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4772 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4774 LDBase->getMemOperand());
4775 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
4781 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
4782 DebugLoc dl = Op.getDebugLoc();
4784 EVT VT = Op.getValueType();
4785 EVT ExtVT = VT.getVectorElementType();
4786 unsigned NumElems = Op.getNumOperands();
4788 // Vectors containing all zeros can be matched by pxor and xorps later
4789 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
4790 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
4791 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
4792 if (Op.getValueType() == MVT::v4i32 ||
4793 Op.getValueType() == MVT::v8i32)
4796 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
4799 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
4800 // vectors or broken into v4i32 operations on 256-bit vectors.
4801 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
4802 if (Op.getValueType() == MVT::v4i32)
4805 return getOnesVector(Op.getValueType(), DAG, dl);
4808 unsigned EVTBits = ExtVT.getSizeInBits();
4810 unsigned NumZero = 0;
4811 unsigned NumNonZero = 0;
4812 unsigned NonZeros = 0;
4813 bool IsAllConstants = true;
4814 SmallSet<SDValue, 8> Values;
4815 for (unsigned i = 0; i < NumElems; ++i) {
4816 SDValue Elt = Op.getOperand(i);
4817 if (Elt.getOpcode() == ISD::UNDEF)
4820 if (Elt.getOpcode() != ISD::Constant &&
4821 Elt.getOpcode() != ISD::ConstantFP)
4822 IsAllConstants = false;
4823 if (X86::isZeroNode(Elt))
4826 NonZeros |= (1 << i);
4831 // All undef vector. Return an UNDEF. All zero vectors were handled above.
4832 if (NumNonZero == 0)
4833 return DAG.getUNDEF(VT);
4835 // Special case for single non-zero, non-undef, element.
4836 if (NumNonZero == 1) {
4837 unsigned Idx = CountTrailingZeros_32(NonZeros);
4838 SDValue Item = Op.getOperand(Idx);
4840 // If this is an insertion of an i64 value on x86-32, and if the top bits of
4841 // the value are obviously zero, truncate the value to i32 and do the
4842 // insertion that way. Only do this if the value is non-constant or if the
4843 // value is a constant being inserted into element 0. It is cheaper to do
4844 // a constant pool load than it is to do a movd + shuffle.
4845 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
4846 (!IsAllConstants || Idx == 0)) {
4847 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
4849 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
4850 EVT VecVT = MVT::v4i32;
4851 unsigned VecElts = 4;
4853 // Truncate the value (which may itself be a constant) to i32, and
4854 // convert it to a vector with movd (S2V+shuffle to zero extend).
4855 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
4856 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
4857 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4858 Subtarget->hasSSE2(), DAG);
4860 // Now we have our 32-bit value zero extended in the low element of
4861 // a vector. If Idx != 0, swizzle it into place.
4863 SmallVector<int, 4> Mask;
4864 Mask.push_back(Idx);
4865 for (unsigned i = 1; i != VecElts; ++i)
4867 Item = DAG.getVectorShuffle(VecVT, dl, Item,
4868 DAG.getUNDEF(Item.getValueType()),
4871 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
4875 // If we have a constant or non-constant insertion into the low element of
4876 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
4877 // the rest of the elements. This will be matched as movd/movq/movss/movsd
4878 // depending on what the source datatype is.
4881 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4882 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
4883 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
4884 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4885 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
4886 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
4888 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
4889 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
4890 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
4891 EVT MiddleVT = MVT::v4i32;
4892 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
4893 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
4894 Subtarget->hasSSE2(), DAG);
4895 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
4899 // Is it a vector logical left shift?
4900 if (NumElems == 2 && Idx == 1 &&
4901 X86::isZeroNode(Op.getOperand(0)) &&
4902 !X86::isZeroNode(Op.getOperand(1))) {
4903 unsigned NumBits = VT.getSizeInBits();
4904 return getVShift(true, VT,
4905 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4906 VT, Op.getOperand(1)),
4907 NumBits/2, DAG, *this, dl);
4910 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
4913 // Otherwise, if this is a vector with i32 or f32 elements, and the element
4914 // is a non-constant being inserted into an element other than the low one,
4915 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
4916 // movd/movss) to move this into the low element, then shuffle it into
4918 if (EVTBits == 32) {
4919 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
4921 // Turn it into a shuffle of zero and zero-extended scalar to vector.
4922 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
4923 Subtarget->hasSSE2(), DAG);
4924 SmallVector<int, 8> MaskVec;
4925 for (unsigned i = 0; i < NumElems; i++)
4926 MaskVec.push_back(i == Idx ? 0 : 1);
4927 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
4931 // Splat is obviously ok. Let legalizer expand it to a shuffle.
4932 if (Values.size() == 1) {
4933 if (EVTBits == 32) {
4934 // Instead of a shuffle like this:
4935 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
4936 // Check if it's possible to issue this instead.
4937 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
4938 unsigned Idx = CountTrailingZeros_32(NonZeros);
4939 SDValue Item = Op.getOperand(Idx);
4940 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
4941 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
4946 // A vector full of immediates; various special cases are already
4947 // handled, so this is best done with a single constant-pool load.
4951 // For AVX-length vectors, build the individual 128-bit pieces and use
4952 // shuffles to put them in place.
4953 if (VT.getSizeInBits() == 256 && !ISD::isBuildVectorAllZeros(Op.getNode())) {
4954 SmallVector<SDValue, 32> V;
4955 for (unsigned i = 0; i < NumElems; ++i)
4956 V.push_back(Op.getOperand(i));
4958 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
4960 // Build both the lower and upper subvector.
4961 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
4962 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
4965 // Recreate the wider vector with the lower and upper part.
4966 SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Lower,
4967 DAG.getConstant(0, MVT::i32), DAG, dl);
4968 return Insert128BitVector(Vec, Upper, DAG.getConstant(NumElems/2, MVT::i32),
4972 // Let legalizer expand 2-wide build_vectors.
4973 if (EVTBits == 64) {
4974 if (NumNonZero == 1) {
4975 // One half is zero or undef.
4976 unsigned Idx = CountTrailingZeros_32(NonZeros);
4977 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
4978 Op.getOperand(Idx));
4979 return getShuffleVectorZeroOrUndef(V2, Idx, true,
4980 Subtarget->hasSSE2(), DAG);
4985 // If element VT is < 32 bits, convert it to inserts into a zero vector.
4986 if (EVTBits == 8 && NumElems == 16) {
4987 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
4989 if (V.getNode()) return V;
4992 if (EVTBits == 16 && NumElems == 8) {
4993 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
4995 if (V.getNode()) return V;
4998 // If element VT is == 32 bits, turn it into a number of shuffles.
4999 SmallVector<SDValue, 8> V;
5001 if (NumElems == 4 && NumZero > 0) {
5002 for (unsigned i = 0; i < 4; ++i) {
5003 bool isZero = !(NonZeros & (1 << i));
5005 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5007 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5010 for (unsigned i = 0; i < 2; ++i) {
5011 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5014 V[i] = V[i*2]; // Must be a zero vector.
5017 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5020 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5023 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
5028 SmallVector<int, 8> MaskVec;
5029 bool Reverse = (NonZeros & 0x3) == 2;
5030 for (unsigned i = 0; i < 2; ++i)
5031 MaskVec.push_back(Reverse ? 1-i : i);
5032 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
5033 for (unsigned i = 0; i < 2; ++i)
5034 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
5035 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
5038 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
5039 // Check for a build vector of consecutive loads.
5040 for (unsigned i = 0; i < NumElems; ++i)
5041 V[i] = Op.getOperand(i);
5043 // Check for elements which are consecutive loads.
5044 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
5048 // For SSE 4.1, use insertps to put the high elements into the low element.
5049 if (getSubtarget()->hasSSE41()) {
5051 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
5052 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
5054 Result = DAG.getUNDEF(VT);
5056 for (unsigned i = 1; i < NumElems; ++i) {
5057 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
5058 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
5059 Op.getOperand(i), DAG.getIntPtrConstant(i));
5064 // Otherwise, expand into a number of unpckl*, start by extending each of
5065 // our (non-undef) elements to the full vector width with the element in the
5066 // bottom slot of the vector (which generates no code for SSE).
5067 for (unsigned i = 0; i < NumElems; ++i) {
5068 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
5069 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5071 V[i] = DAG.getUNDEF(VT);
5074 // Next, we iteratively mix elements, e.g. for v4f32:
5075 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
5076 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
5077 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
5078 unsigned EltStride = NumElems >> 1;
5079 while (EltStride != 0) {
5080 for (unsigned i = 0; i < EltStride; ++i) {
5081 // If V[i+EltStride] is undef and this is the first round of mixing,
5082 // then it is safe to just drop this shuffle: V[i] is already in the
5083 // right place, the one element (since it's the first round) being
5084 // inserted as undef can be dropped. This isn't safe for successive
5085 // rounds because they will permute elements within both vectors.
5086 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
5087 EltStride == NumElems/2)
5090 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
5099 // LowerMMXCONCAT_VECTORS - We support concatenate two MMX registers and place
5100 // them in a MMX register. This is better than doing a stack convert.
5101 static SDValue LowerMMXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5102 DebugLoc dl = Op.getDebugLoc();
5103 EVT ResVT = Op.getValueType();
5105 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
5106 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
5108 SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
5109 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
5110 InVec = Op.getOperand(1);
5111 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5112 unsigned NumElts = ResVT.getVectorNumElements();
5113 VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
5114 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
5115 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
5117 InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
5118 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
5119 Mask[0] = 0; Mask[1] = 2;
5120 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
5122 return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
5125 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
5126 // to create 256-bit vectors from two other 128-bit ones.
5127 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5128 DebugLoc dl = Op.getDebugLoc();
5129 EVT ResVT = Op.getValueType();
5131 assert(ResVT.getSizeInBits() == 256 && "Value type must be 256-bit wide");
5133 SDValue V1 = Op.getOperand(0);
5134 SDValue V2 = Op.getOperand(1);
5135 unsigned NumElems = ResVT.getVectorNumElements();
5137 SDValue V = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, ResVT), V1,
5138 DAG.getConstant(0, MVT::i32), DAG, dl);
5139 return Insert128BitVector(V, V2, DAG.getConstant(NumElems/2, MVT::i32),
5144 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
5145 EVT ResVT = Op.getValueType();
5147 assert(Op.getNumOperands() == 2);
5148 assert((ResVT.getSizeInBits() == 128 || ResVT.getSizeInBits() == 256) &&
5149 "Unsupported CONCAT_VECTORS for value type");
5151 // We support concatenate two MMX registers and place them in a MMX register.
5152 // This is better than doing a stack convert.
5153 if (ResVT.is128BitVector())
5154 return LowerMMXCONCAT_VECTORS(Op, DAG);
5156 // 256-bit AVX can use the vinsertf128 instruction to create 256-bit vectors
5157 // from two other 128-bit ones.
5158 return LowerAVXCONCAT_VECTORS(Op, DAG);
5161 // v8i16 shuffles - Prefer shuffles in the following order:
5162 // 1. [all] pshuflw, pshufhw, optional move
5163 // 2. [ssse3] 1 x pshufb
5164 // 3. [ssse3] 2 x pshufb + 1 x por
5165 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
5167 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
5168 SelectionDAG &DAG) const {
5169 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5170 SDValue V1 = SVOp->getOperand(0);
5171 SDValue V2 = SVOp->getOperand(1);
5172 DebugLoc dl = SVOp->getDebugLoc();
5173 SmallVector<int, 8> MaskVals;
5175 // Determine if more than 1 of the words in each of the low and high quadwords
5176 // of the result come from the same quadword of one of the two inputs. Undef
5177 // mask values count as coming from any quadword, for better codegen.
5178 SmallVector<unsigned, 4> LoQuad(4);
5179 SmallVector<unsigned, 4> HiQuad(4);
5180 BitVector InputQuads(4);
5181 for (unsigned i = 0; i < 8; ++i) {
5182 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
5183 int EltIdx = SVOp->getMaskElt(i);
5184 MaskVals.push_back(EltIdx);
5193 InputQuads.set(EltIdx / 4);
5196 int BestLoQuad = -1;
5197 unsigned MaxQuad = 1;
5198 for (unsigned i = 0; i < 4; ++i) {
5199 if (LoQuad[i] > MaxQuad) {
5201 MaxQuad = LoQuad[i];
5205 int BestHiQuad = -1;
5207 for (unsigned i = 0; i < 4; ++i) {
5208 if (HiQuad[i] > MaxQuad) {
5210 MaxQuad = HiQuad[i];
5214 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
5215 // of the two input vectors, shuffle them into one input vector so only a
5216 // single pshufb instruction is necessary. If There are more than 2 input
5217 // quads, disable the next transformation since it does not help SSSE3.
5218 bool V1Used = InputQuads[0] || InputQuads[1];
5219 bool V2Used = InputQuads[2] || InputQuads[3];
5220 if (Subtarget->hasSSSE3()) {
5221 if (InputQuads.count() == 2 && V1Used && V2Used) {
5222 BestLoQuad = InputQuads.find_first();
5223 BestHiQuad = InputQuads.find_next(BestLoQuad);
5225 if (InputQuads.count() > 2) {
5231 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
5232 // the shuffle mask. If a quad is scored as -1, that means that it contains
5233 // words from all 4 input quadwords.
5235 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
5236 SmallVector<int, 8> MaskV;
5237 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
5238 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
5239 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
5240 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
5241 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
5242 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
5244 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
5245 // source words for the shuffle, to aid later transformations.
5246 bool AllWordsInNewV = true;
5247 bool InOrder[2] = { true, true };
5248 for (unsigned i = 0; i != 8; ++i) {
5249 int idx = MaskVals[i];
5251 InOrder[i/4] = false;
5252 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
5254 AllWordsInNewV = false;
5258 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
5259 if (AllWordsInNewV) {
5260 for (int i = 0; i != 8; ++i) {
5261 int idx = MaskVals[i];
5264 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
5265 if ((idx != i) && idx < 4)
5267 if ((idx != i) && idx > 3)
5276 // If we've eliminated the use of V2, and the new mask is a pshuflw or
5277 // pshufhw, that's as cheap as it gets. Return the new shuffle.
5278 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
5279 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
5280 unsigned TargetMask = 0;
5281 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
5282 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
5283 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
5284 X86::getShufflePSHUFLWImmediate(NewV.getNode());
5285 V1 = NewV.getOperand(0);
5286 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
5290 // If we have SSSE3, and all words of the result are from 1 input vector,
5291 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
5292 // is present, fall back to case 4.
5293 if (Subtarget->hasSSSE3()) {
5294 SmallVector<SDValue,16> pshufbMask;
5296 // If we have elements from both input vectors, set the high bit of the
5297 // shuffle mask element to zero out elements that come from V2 in the V1
5298 // mask, and elements that come from V1 in the V2 mask, so that the two
5299 // results can be OR'd together.
5300 bool TwoInputs = V1Used && V2Used;
5301 for (unsigned i = 0; i != 8; ++i) {
5302 int EltIdx = MaskVals[i] * 2;
5303 if (TwoInputs && (EltIdx >= 16)) {
5304 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5305 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5308 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5309 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
5311 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
5312 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5313 DAG.getNode(ISD::BUILD_VECTOR, dl,
5314 MVT::v16i8, &pshufbMask[0], 16));
5316 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5318 // Calculate the shuffle mask for the second input, shuffle it, and
5319 // OR it with the first shuffled input.
5321 for (unsigned i = 0; i != 8; ++i) {
5322 int EltIdx = MaskVals[i] * 2;
5324 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5325 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5328 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5329 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
5331 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
5332 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5333 DAG.getNode(ISD::BUILD_VECTOR, dl,
5334 MVT::v16i8, &pshufbMask[0], 16));
5335 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5336 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5339 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
5340 // and update MaskVals with new element order.
5341 BitVector InOrder(8);
5342 if (BestLoQuad >= 0) {
5343 SmallVector<int, 8> MaskV;
5344 for (int i = 0; i != 4; ++i) {
5345 int idx = MaskVals[i];
5347 MaskV.push_back(-1);
5349 } else if ((idx / 4) == BestLoQuad) {
5350 MaskV.push_back(idx & 3);
5353 MaskV.push_back(-1);
5356 for (unsigned i = 4; i != 8; ++i)
5358 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5361 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5362 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
5364 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
5368 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
5369 // and update MaskVals with the new element order.
5370 if (BestHiQuad >= 0) {
5371 SmallVector<int, 8> MaskV;
5372 for (unsigned i = 0; i != 4; ++i)
5374 for (unsigned i = 4; i != 8; ++i) {
5375 int idx = MaskVals[i];
5377 MaskV.push_back(-1);
5379 } else if ((idx / 4) == BestHiQuad) {
5380 MaskV.push_back((idx & 3) + 4);
5383 MaskV.push_back(-1);
5386 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5389 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5390 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
5392 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
5396 // In case BestHi & BestLo were both -1, which means each quadword has a word
5397 // from each of the four input quadwords, calculate the InOrder bitvector now
5398 // before falling through to the insert/extract cleanup.
5399 if (BestLoQuad == -1 && BestHiQuad == -1) {
5401 for (int i = 0; i != 8; ++i)
5402 if (MaskVals[i] < 0 || MaskVals[i] == i)
5406 // The other elements are put in the right place using pextrw and pinsrw.
5407 for (unsigned i = 0; i != 8; ++i) {
5410 int EltIdx = MaskVals[i];
5413 SDValue ExtOp = (EltIdx < 8)
5414 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
5415 DAG.getIntPtrConstant(EltIdx))
5416 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
5417 DAG.getIntPtrConstant(EltIdx - 8));
5418 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
5419 DAG.getIntPtrConstant(i));
5424 // v16i8 shuffles - Prefer shuffles in the following order:
5425 // 1. [ssse3] 1 x pshufb
5426 // 2. [ssse3] 2 x pshufb + 1 x por
5427 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
5429 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
5431 const X86TargetLowering &TLI) {
5432 SDValue V1 = SVOp->getOperand(0);
5433 SDValue V2 = SVOp->getOperand(1);
5434 DebugLoc dl = SVOp->getDebugLoc();
5435 SmallVector<int, 16> MaskVals;
5436 SVOp->getMask(MaskVals);
5438 // If we have SSSE3, case 1 is generated when all result bytes come from
5439 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
5440 // present, fall back to case 3.
5441 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
5444 for (unsigned i = 0; i < 16; ++i) {
5445 int EltIdx = MaskVals[i];
5454 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
5455 if (TLI.getSubtarget()->hasSSSE3()) {
5456 SmallVector<SDValue,16> pshufbMask;
5458 // If all result elements are from one input vector, then only translate
5459 // undef mask values to 0x80 (zero out result) in the pshufb mask.
5461 // Otherwise, we have elements from both input vectors, and must zero out
5462 // elements that come from V2 in the first mask, and V1 in the second mask
5463 // so that we can OR them together.
5464 bool TwoInputs = !(V1Only || V2Only);
5465 for (unsigned i = 0; i != 16; ++i) {
5466 int EltIdx = MaskVals[i];
5467 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
5468 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5471 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5473 // If all the elements are from V2, assign it to V1 and return after
5474 // building the first pshufb.
5477 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5478 DAG.getNode(ISD::BUILD_VECTOR, dl,
5479 MVT::v16i8, &pshufbMask[0], 16));
5483 // Calculate the shuffle mask for the second input, shuffle it, and
5484 // OR it with the first shuffled input.
5486 for (unsigned i = 0; i != 16; ++i) {
5487 int EltIdx = MaskVals[i];
5489 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5492 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5494 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5495 DAG.getNode(ISD::BUILD_VECTOR, dl,
5496 MVT::v16i8, &pshufbMask[0], 16));
5497 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5500 // No SSSE3 - Calculate in place words and then fix all out of place words
5501 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
5502 // the 16 different words that comprise the two doublequadword input vectors.
5503 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5504 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
5505 SDValue NewV = V2Only ? V2 : V1;
5506 for (int i = 0; i != 8; ++i) {
5507 int Elt0 = MaskVals[i*2];
5508 int Elt1 = MaskVals[i*2+1];
5510 // This word of the result is all undef, skip it.
5511 if (Elt0 < 0 && Elt1 < 0)
5514 // This word of the result is already in the correct place, skip it.
5515 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
5517 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
5520 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
5521 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
5524 // If Elt0 and Elt1 are defined, are consecutive, and can be load
5525 // using a single extract together, load it and store it.
5526 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
5527 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5528 DAG.getIntPtrConstant(Elt1 / 2));
5529 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5530 DAG.getIntPtrConstant(i));
5534 // If Elt1 is defined, extract it from the appropriate source. If the
5535 // source byte is not also odd, shift the extracted word left 8 bits
5536 // otherwise clear the bottom 8 bits if we need to do an or.
5538 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5539 DAG.getIntPtrConstant(Elt1 / 2));
5540 if ((Elt1 & 1) == 0)
5541 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
5543 TLI.getShiftAmountTy(InsElt.getValueType())));
5545 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
5546 DAG.getConstant(0xFF00, MVT::i16));
5548 // If Elt0 is defined, extract it from the appropriate source. If the
5549 // source byte is not also even, shift the extracted word right 8 bits. If
5550 // Elt1 was also defined, OR the extracted values together before
5551 // inserting them in the result.
5553 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
5554 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
5555 if ((Elt0 & 1) != 0)
5556 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
5558 TLI.getShiftAmountTy(InsElt0.getValueType())));
5560 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
5561 DAG.getConstant(0x00FF, MVT::i16));
5562 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
5565 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5566 DAG.getIntPtrConstant(i));
5568 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
5571 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
5572 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
5573 /// done when every pair / quad of shuffle mask elements point to elements in
5574 /// the right sequence. e.g.
5575 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
5577 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
5578 SelectionDAG &DAG, DebugLoc dl) {
5579 EVT VT = SVOp->getValueType(0);
5580 SDValue V1 = SVOp->getOperand(0);
5581 SDValue V2 = SVOp->getOperand(1);
5582 unsigned NumElems = VT.getVectorNumElements();
5583 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
5585 switch (VT.getSimpleVT().SimpleTy) {
5586 default: assert(false && "Unexpected!");
5587 case MVT::v4f32: NewVT = MVT::v2f64; break;
5588 case MVT::v4i32: NewVT = MVT::v2i64; break;
5589 case MVT::v8i16: NewVT = MVT::v4i32; break;
5590 case MVT::v16i8: NewVT = MVT::v4i32; break;
5593 int Scale = NumElems / NewWidth;
5594 SmallVector<int, 8> MaskVec;
5595 for (unsigned i = 0; i < NumElems; i += Scale) {
5597 for (int j = 0; j < Scale; ++j) {
5598 int EltIdx = SVOp->getMaskElt(i+j);
5602 StartIdx = EltIdx - (EltIdx % Scale);
5603 if (EltIdx != StartIdx + j)
5607 MaskVec.push_back(-1);
5609 MaskVec.push_back(StartIdx / Scale);
5612 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
5613 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
5614 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
5617 /// getVZextMovL - Return a zero-extending vector move low node.
5619 static SDValue getVZextMovL(EVT VT, EVT OpVT,
5620 SDValue SrcOp, SelectionDAG &DAG,
5621 const X86Subtarget *Subtarget, DebugLoc dl) {
5622 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
5623 LoadSDNode *LD = NULL;
5624 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
5625 LD = dyn_cast<LoadSDNode>(SrcOp);
5627 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
5629 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
5630 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
5631 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
5632 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
5633 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
5635 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
5636 return DAG.getNode(ISD::BITCAST, dl, VT,
5637 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5638 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5646 return DAG.getNode(ISD::BITCAST, dl, VT,
5647 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5648 DAG.getNode(ISD::BITCAST, dl,
5652 /// areShuffleHalvesWithinDisjointLanes - Check whether each half of a vector
5653 /// shuffle node referes to only one lane in the sources.
5654 static bool areShuffleHalvesWithinDisjointLanes(ShuffleVectorSDNode *SVOp) {
5655 EVT VT = SVOp->getValueType(0);
5656 int NumElems = VT.getVectorNumElements();
5657 int HalfSize = NumElems/2;
5658 SmallVector<int, 16> M;
5660 bool MatchA = false, MatchB = false;
5662 for (int l = 0; l < NumElems*2; l += HalfSize) {
5663 if (isUndefOrInRange(M, 0, HalfSize, l, l+HalfSize)) {
5669 for (int l = 0; l < NumElems*2; l += HalfSize) {
5670 if (isUndefOrInRange(M, HalfSize, HalfSize, l, l+HalfSize)) {
5676 return MatchA && MatchB;
5679 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
5680 /// which could not be matched by any known target speficic shuffle
5682 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5683 if (areShuffleHalvesWithinDisjointLanes(SVOp)) {
5684 // If each half of a vector shuffle node referes to only one lane in the
5685 // source vectors, extract each used 128-bit lane and shuffle them using
5686 // 128-bit shuffles. Then, concatenate the results. Otherwise leave
5687 // the work to the legalizer.
5688 DebugLoc dl = SVOp->getDebugLoc();
5689 EVT VT = SVOp->getValueType(0);
5690 int NumElems = VT.getVectorNumElements();
5691 int HalfSize = NumElems/2;
5693 // Extract the reference for each half
5694 int FstVecExtractIdx = 0, SndVecExtractIdx = 0;
5695 int FstVecOpNum = 0, SndVecOpNum = 0;
5696 for (int i = 0; i < HalfSize; ++i) {
5697 int Elt = SVOp->getMaskElt(i);
5698 if (SVOp->getMaskElt(i) < 0)
5700 FstVecOpNum = Elt/NumElems;
5701 FstVecExtractIdx = Elt % NumElems < HalfSize ? 0 : HalfSize;
5704 for (int i = HalfSize; i < NumElems; ++i) {
5705 int Elt = SVOp->getMaskElt(i);
5706 if (SVOp->getMaskElt(i) < 0)
5708 SndVecOpNum = Elt/NumElems;
5709 SndVecExtractIdx = Elt % NumElems < HalfSize ? 0 : HalfSize;
5713 // Extract the subvectors
5714 SDValue V1 = Extract128BitVector(SVOp->getOperand(FstVecOpNum),
5715 DAG.getConstant(FstVecExtractIdx, MVT::i32), DAG, dl);
5716 SDValue V2 = Extract128BitVector(SVOp->getOperand(SndVecOpNum),
5717 DAG.getConstant(SndVecExtractIdx, MVT::i32), DAG, dl);
5719 // Generate 128-bit shuffles
5720 SmallVector<int, 16> MaskV1, MaskV2;
5721 for (int i = 0; i < HalfSize; ++i) {
5722 int Elt = SVOp->getMaskElt(i);
5723 MaskV1.push_back(Elt < 0 ? Elt : Elt % HalfSize);
5725 for (int i = HalfSize; i < NumElems; ++i) {
5726 int Elt = SVOp->getMaskElt(i);
5727 MaskV2.push_back(Elt < 0 ? Elt : Elt % HalfSize);
5730 EVT NVT = V1.getValueType();
5731 V1 = DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &MaskV1[0]);
5732 V2 = DAG.getVectorShuffle(NVT, dl, V2, DAG.getUNDEF(NVT), &MaskV2[0]);
5734 // Concatenate the result back
5735 SDValue V = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), V1,
5736 DAG.getConstant(0, MVT::i32), DAG, dl);
5737 return Insert128BitVector(V, V2, DAG.getConstant(NumElems/2, MVT::i32),
5744 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
5745 /// 4 elements, and match them with several different shuffle types.
5747 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5748 SDValue V1 = SVOp->getOperand(0);
5749 SDValue V2 = SVOp->getOperand(1);
5750 DebugLoc dl = SVOp->getDebugLoc();
5751 EVT VT = SVOp->getValueType(0);
5753 assert(VT.getSizeInBits() == 128 && "Unsupported vector size");
5755 SmallVector<std::pair<int, int>, 8> Locs;
5757 SmallVector<int, 8> Mask1(4U, -1);
5758 SmallVector<int, 8> PermMask;
5759 SVOp->getMask(PermMask);
5763 for (unsigned i = 0; i != 4; ++i) {
5764 int Idx = PermMask[i];
5766 Locs[i] = std::make_pair(-1, -1);
5768 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5770 Locs[i] = std::make_pair(0, NumLo);
5774 Locs[i] = std::make_pair(1, NumHi);
5776 Mask1[2+NumHi] = Idx;
5782 if (NumLo <= 2 && NumHi <= 2) {
5783 // If no more than two elements come from either vector. This can be
5784 // implemented with two shuffles. First shuffle gather the elements.
5785 // The second shuffle, which takes the first shuffle as both of its
5786 // vector operands, put the elements into the right order.
5787 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5789 SmallVector<int, 8> Mask2(4U, -1);
5791 for (unsigned i = 0; i != 4; ++i) {
5792 if (Locs[i].first == -1)
5795 unsigned Idx = (i < 2) ? 0 : 4;
5796 Idx += Locs[i].first * 2 + Locs[i].second;
5801 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
5802 } else if (NumLo == 3 || NumHi == 3) {
5803 // Otherwise, we must have three elements from one vector, call it X, and
5804 // one element from the other, call it Y. First, use a shufps to build an
5805 // intermediate vector with the one element from Y and the element from X
5806 // that will be in the same half in the final destination (the indexes don't
5807 // matter). Then, use a shufps to build the final vector, taking the half
5808 // containing the element from Y from the intermediate, and the other half
5811 // Normalize it so the 3 elements come from V1.
5812 CommuteVectorShuffleMask(PermMask, VT);
5816 // Find the element from V2.
5818 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
5819 int Val = PermMask[HiIndex];
5826 Mask1[0] = PermMask[HiIndex];
5828 Mask1[2] = PermMask[HiIndex^1];
5830 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5833 Mask1[0] = PermMask[0];
5834 Mask1[1] = PermMask[1];
5835 Mask1[2] = HiIndex & 1 ? 6 : 4;
5836 Mask1[3] = HiIndex & 1 ? 4 : 6;
5837 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
5839 Mask1[0] = HiIndex & 1 ? 2 : 0;
5840 Mask1[1] = HiIndex & 1 ? 0 : 2;
5841 Mask1[2] = PermMask[2];
5842 Mask1[3] = PermMask[3];
5847 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
5851 // Break it into (shuffle shuffle_hi, shuffle_lo).
5854 SmallVector<int,8> LoMask(4U, -1);
5855 SmallVector<int,8> HiMask(4U, -1);
5857 SmallVector<int,8> *MaskPtr = &LoMask;
5858 unsigned MaskIdx = 0;
5861 for (unsigned i = 0; i != 4; ++i) {
5868 int Idx = PermMask[i];
5870 Locs[i] = std::make_pair(-1, -1);
5871 } else if (Idx < 4) {
5872 Locs[i] = std::make_pair(MaskIdx, LoIdx);
5873 (*MaskPtr)[LoIdx] = Idx;
5876 Locs[i] = std::make_pair(MaskIdx, HiIdx);
5877 (*MaskPtr)[HiIdx] = Idx;
5882 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
5883 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
5884 SmallVector<int, 8> MaskOps;
5885 for (unsigned i = 0; i != 4; ++i) {
5886 if (Locs[i].first == -1) {
5887 MaskOps.push_back(-1);
5889 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
5890 MaskOps.push_back(Idx);
5893 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
5896 static bool MayFoldVectorLoad(SDValue V) {
5897 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5898 V = V.getOperand(0);
5899 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5900 V = V.getOperand(0);
5906 // FIXME: the version above should always be used. Since there's
5907 // a bug where several vector shuffles can't be folded because the
5908 // DAG is not updated during lowering and a node claims to have two
5909 // uses while it only has one, use this version, and let isel match
5910 // another instruction if the load really happens to have more than
5911 // one use. Remove this version after this bug get fixed.
5912 // rdar://8434668, PR8156
5913 static bool RelaxedMayFoldVectorLoad(SDValue V) {
5914 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
5915 V = V.getOperand(0);
5916 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
5917 V = V.getOperand(0);
5918 if (ISD::isNormalLoad(V.getNode()))
5923 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
5924 /// a vector extract, and if both can be later optimized into a single load.
5925 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
5926 /// here because otherwise a target specific shuffle node is going to be
5927 /// emitted for this shuffle, and the optimization not done.
5928 /// FIXME: This is probably not the best approach, but fix the problem
5929 /// until the right path is decided.
5931 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
5932 const TargetLowering &TLI) {
5933 EVT VT = V.getValueType();
5934 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
5936 // Be sure that the vector shuffle is present in a pattern like this:
5937 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
5941 SDNode *N = *V.getNode()->use_begin();
5942 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
5945 SDValue EltNo = N->getOperand(1);
5946 if (!isa<ConstantSDNode>(EltNo))
5949 // If the bit convert changed the number of elements, it is unsafe
5950 // to examine the mask.
5951 bool HasShuffleIntoBitcast = false;
5952 if (V.getOpcode() == ISD::BITCAST) {
5953 EVT SrcVT = V.getOperand(0).getValueType();
5954 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
5956 V = V.getOperand(0);
5957 HasShuffleIntoBitcast = true;
5960 // Select the input vector, guarding against out of range extract vector.
5961 unsigned NumElems = VT.getVectorNumElements();
5962 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
5963 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
5964 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
5966 // Skip one more bit_convert if necessary
5967 if (V.getOpcode() == ISD::BITCAST)
5968 V = V.getOperand(0);
5970 if (ISD::isNormalLoad(V.getNode())) {
5971 // Is the original load suitable?
5972 LoadSDNode *LN0 = cast<LoadSDNode>(V);
5974 // FIXME: avoid the multi-use bug that is preventing lots of
5975 // of foldings to be detected, this is still wrong of course, but
5976 // give the temporary desired behavior, and if it happens that
5977 // the load has real more uses, during isel it will not fold, and
5978 // will generate poor code.
5979 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
5982 if (!HasShuffleIntoBitcast)
5985 // If there's a bitcast before the shuffle, check if the load type and
5986 // alignment is valid.
5987 unsigned Align = LN0->getAlignment();
5989 TLI.getTargetData()->getABITypeAlignment(
5990 VT.getTypeForEVT(*DAG.getContext()));
5992 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
6000 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
6001 EVT VT = Op.getValueType();
6003 // Canonizalize to v2f64.
6004 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
6005 return DAG.getNode(ISD::BITCAST, dl, VT,
6006 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
6011 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
6013 SDValue V1 = Op.getOperand(0);
6014 SDValue V2 = Op.getOperand(1);
6015 EVT VT = Op.getValueType();
6017 assert(VT != MVT::v2i64 && "unsupported shuffle type");
6019 if (HasSSE2 && VT == MVT::v2f64)
6020 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
6023 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V2, DAG);
6027 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
6028 SDValue V1 = Op.getOperand(0);
6029 SDValue V2 = Op.getOperand(1);
6030 EVT VT = Op.getValueType();
6032 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
6033 "unsupported shuffle type");
6035 if (V2.getOpcode() == ISD::UNDEF)
6039 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
6043 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
6044 SDValue V1 = Op.getOperand(0);
6045 SDValue V2 = Op.getOperand(1);
6046 EVT VT = Op.getValueType();
6047 unsigned NumElems = VT.getVectorNumElements();
6049 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
6050 // operand of these instructions is only memory, so check if there's a
6051 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
6053 bool CanFoldLoad = false;
6055 // Trivial case, when V2 comes from a load.
6056 if (MayFoldVectorLoad(V2))
6059 // When V1 is a load, it can be folded later into a store in isel, example:
6060 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
6062 // (MOVLPSmr addr:$src1, VR128:$src2)
6063 // So, recognize this potential and also use MOVLPS or MOVLPD
6064 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
6067 // Both of them can't be memory operations though.
6068 if (MayFoldVectorLoad(V1) && MayFoldVectorLoad(V2))
6069 CanFoldLoad = false;
6072 if (HasSSE2 && NumElems == 2)
6073 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
6076 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
6079 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6080 // movl and movlp will both match v2i64, but v2i64 is never matched by
6081 // movl earlier because we make it strict to avoid messing with the movlp load
6082 // folding logic (see the code above getMOVLP call). Match it here then,
6083 // this is horrible, but will stay like this until we move all shuffle
6084 // matching to x86 specific nodes. Note that for the 1st condition all
6085 // types are matched with movsd.
6086 if ((HasSSE2 && NumElems == 2) || !X86::isMOVLMask(SVOp))
6087 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
6089 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
6092 assert(VT != MVT::v4i32 && "unsupported shuffle type");
6094 // Invert the operand order and use SHUFPS to match it.
6095 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V2, V1,
6096 X86::getShuffleSHUFImmediate(SVOp), DAG);
6099 static inline unsigned getUNPCKLOpcode(EVT VT) {
6100 switch(VT.getSimpleVT().SimpleTy) {
6101 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
6102 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
6103 case MVT::v4f32: return X86ISD::UNPCKLPS;
6104 case MVT::v2f64: return X86ISD::UNPCKLPD;
6105 case MVT::v8i32: // Use fp unit for int unpack.
6106 case MVT::v8f32: return X86ISD::VUNPCKLPSY;
6107 case MVT::v4i64: // Use fp unit for int unpack.
6108 case MVT::v4f64: return X86ISD::VUNPCKLPDY;
6109 case MVT::v16i8: return X86ISD::PUNPCKLBW;
6110 case MVT::v8i16: return X86ISD::PUNPCKLWD;
6112 llvm_unreachable("Unknown type for unpckl");
6117 static inline unsigned getUNPCKHOpcode(EVT VT) {
6118 switch(VT.getSimpleVT().SimpleTy) {
6119 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
6120 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
6121 case MVT::v4f32: return X86ISD::UNPCKHPS;
6122 case MVT::v2f64: return X86ISD::UNPCKHPD;
6123 case MVT::v8i32: // Use fp unit for int unpack.
6124 case MVT::v8f32: return X86ISD::VUNPCKHPSY;
6125 case MVT::v4i64: // Use fp unit for int unpack.
6126 case MVT::v4f64: return X86ISD::VUNPCKHPDY;
6127 case MVT::v16i8: return X86ISD::PUNPCKHBW;
6128 case MVT::v8i16: return X86ISD::PUNPCKHWD;
6130 llvm_unreachable("Unknown type for unpckh");
6135 static inline unsigned getVPERMILOpcode(EVT VT) {
6136 switch(VT.getSimpleVT().SimpleTy) {
6138 case MVT::v4f32: return X86ISD::VPERMILPS;
6140 case MVT::v2f64: return X86ISD::VPERMILPD;
6142 case MVT::v8f32: return X86ISD::VPERMILPSY;
6144 case MVT::v4f64: return X86ISD::VPERMILPDY;
6146 llvm_unreachable("Unknown type for vpermil");
6152 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
6153 const TargetLowering &TLI,
6154 const X86Subtarget *Subtarget) {
6155 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6156 EVT VT = Op.getValueType();
6157 DebugLoc dl = Op.getDebugLoc();
6158 SDValue V1 = Op.getOperand(0);
6159 SDValue V2 = Op.getOperand(1);
6161 if (isZeroShuffle(SVOp))
6162 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
6164 // Handle splat operations
6165 if (SVOp->isSplat()) {
6166 unsigned NumElem = VT.getVectorNumElements();
6167 // Special case, this is the only place now where it's allowed to return
6168 // a vector_shuffle operation without using a target specific node, because
6169 // *hopefully* it will be optimized away by the dag combiner. FIXME: should
6170 // this be moved to DAGCombine instead?
6171 if (NumElem <= 4 && CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
6174 // Handle splats by matching through known shuffle masks
6175 if (VT.is128BitVector() && NumElem <= 4)
6178 // All i16 and i8 vector types can't be used directly by a generic shuffle
6179 // instruction because the target has no such instruction. Generate shuffles
6180 // which repeat i16 and i8 several times until they fit in i32, and then can
6181 // be manipulated by target suported shuffles. After the insertion of the
6182 // necessary shuffles, the result is bitcasted back to v4f32 or v8f32.
6183 return PromoteSplat(SVOp, DAG);
6186 // If the shuffle can be profitably rewritten as a narrower shuffle, then
6188 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
6189 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6190 if (NewOp.getNode())
6191 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
6192 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
6193 // FIXME: Figure out a cleaner way to do this.
6194 // Try to make use of movq to zero out the top part.
6195 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
6196 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6197 if (NewOp.getNode()) {
6198 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
6199 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
6200 DAG, Subtarget, dl);
6202 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
6203 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6204 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
6205 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
6206 DAG, Subtarget, dl);
6213 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
6214 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6215 SDValue V1 = Op.getOperand(0);
6216 SDValue V2 = Op.getOperand(1);
6217 EVT VT = Op.getValueType();
6218 DebugLoc dl = Op.getDebugLoc();
6219 unsigned NumElems = VT.getVectorNumElements();
6220 bool isMMX = VT.getSizeInBits() == 64;
6221 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
6222 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
6223 bool V1IsSplat = false;
6224 bool V2IsSplat = false;
6225 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
6226 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
6227 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
6228 MachineFunction &MF = DAG.getMachineFunction();
6229 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
6231 // Shuffle operations on MMX not supported.
6235 // Vector shuffle lowering takes 3 steps:
6237 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
6238 // narrowing and commutation of operands should be handled.
6239 // 2) Matching of shuffles with known shuffle masks to x86 target specific
6241 // 3) Rewriting of unmatched masks into new generic shuffle operations,
6242 // so the shuffle can be broken into other shuffles and the legalizer can
6243 // try the lowering again.
6245 // The general ideia is that no vector_shuffle operation should be left to
6246 // be matched during isel, all of them must be converted to a target specific
6249 // Normalize the input vectors. Here splats, zeroed vectors, profitable
6250 // narrowing and commutation of operands should be handled. The actual code
6251 // doesn't include all of those, work in progress...
6252 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
6253 if (NewOp.getNode())
6256 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
6257 // unpckh_undef). Only use pshufd if speed is more important than size.
6258 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
6259 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
6260 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
6261 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6263 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
6264 RelaxedMayFoldVectorLoad(V1))
6265 return getMOVDDup(Op, dl, V1, DAG);
6267 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
6268 return getMOVHighToLow(Op, dl, DAG);
6270 // Use to match splats
6271 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
6272 (VT == MVT::v2f64 || VT == MVT::v2i64))
6273 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6275 if (X86::isPSHUFDMask(SVOp)) {
6276 // The actual implementation will match the mask in the if above and then
6277 // during isel it can match several different instructions, not only pshufd
6278 // as its name says, sad but true, emulate the behavior for now...
6279 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
6280 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
6282 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
6284 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
6285 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
6287 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
6288 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V1,
6291 if (VT == MVT::v4f32)
6292 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V1,
6296 // Check if this can be converted into a logical shift.
6297 bool isLeft = false;
6300 bool isShift = getSubtarget()->hasSSE2() &&
6301 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
6302 if (isShift && ShVal.hasOneUse()) {
6303 // If the shifted value has multiple uses, it may be cheaper to use
6304 // v_set0 + movlhps or movhlps, etc.
6305 EVT EltVT = VT.getVectorElementType();
6306 ShAmt *= EltVT.getSizeInBits();
6307 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6310 if (X86::isMOVLMask(SVOp)) {
6313 if (ISD::isBuildVectorAllZeros(V1.getNode()))
6314 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
6315 if (!X86::isMOVLPMask(SVOp)) {
6316 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
6317 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
6319 if (VT == MVT::v4i32 || VT == MVT::v4f32)
6320 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
6324 // FIXME: fold these into legal mask.
6325 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
6326 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
6328 if (X86::isMOVHLPSMask(SVOp))
6329 return getMOVHighToLow(Op, dl, DAG);
6331 if (X86::isMOVSHDUPMask(SVOp, Subtarget))
6332 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
6334 if (X86::isMOVSLDUPMask(SVOp, Subtarget))
6335 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
6337 if (X86::isMOVLPMask(SVOp))
6338 return getMOVLP(Op, dl, DAG, HasSSE2);
6340 if (ShouldXformToMOVHLPS(SVOp) ||
6341 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
6342 return CommuteVectorShuffle(SVOp, DAG);
6345 // No better options. Use a vshl / vsrl.
6346 EVT EltVT = VT.getVectorElementType();
6347 ShAmt *= EltVT.getSizeInBits();
6348 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6351 bool Commuted = false;
6352 // FIXME: This should also accept a bitcast of a splat? Be careful, not
6353 // 1,1,1,1 -> v8i16 though.
6354 V1IsSplat = isSplatVector(V1.getNode());
6355 V2IsSplat = isSplatVector(V2.getNode());
6357 // Canonicalize the splat or undef, if present, to be on the RHS.
6358 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
6359 Op = CommuteVectorShuffle(SVOp, DAG);
6360 SVOp = cast<ShuffleVectorSDNode>(Op);
6361 V1 = SVOp->getOperand(0);
6362 V2 = SVOp->getOperand(1);
6363 std::swap(V1IsSplat, V2IsSplat);
6364 std::swap(V1IsUndef, V2IsUndef);
6368 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
6369 // Shuffling low element of v1 into undef, just return v1.
6372 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
6373 // the instruction selector will not match, so get a canonical MOVL with
6374 // swapped operands to undo the commute.
6375 return getMOVL(DAG, dl, VT, V2, V1);
6378 if (X86::isUNPCKLMask(SVOp))
6379 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
6381 if (X86::isUNPCKHMask(SVOp))
6382 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
6385 // Normalize mask so all entries that point to V2 points to its first
6386 // element then try to match unpck{h|l} again. If match, return a
6387 // new vector_shuffle with the corrected mask.
6388 SDValue NewMask = NormalizeMask(SVOp, DAG);
6389 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
6390 if (NSVOp != SVOp) {
6391 if (X86::isUNPCKLMask(NSVOp, true)) {
6393 } else if (X86::isUNPCKHMask(NSVOp, true)) {
6400 // Commute is back and try unpck* again.
6401 // FIXME: this seems wrong.
6402 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
6403 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
6405 if (X86::isUNPCKLMask(NewSVOp))
6406 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
6408 if (X86::isUNPCKHMask(NewSVOp))
6409 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
6412 // Normalize the node to match x86 shuffle ops if needed
6413 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
6414 return CommuteVectorShuffle(SVOp, DAG);
6416 // The checks below are all present in isShuffleMaskLegal, but they are
6417 // inlined here right now to enable us to directly emit target specific
6418 // nodes, and remove one by one until they don't return Op anymore.
6419 SmallVector<int, 16> M;
6422 if (isPALIGNRMask(M, VT, HasSSSE3))
6423 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
6424 X86::getShufflePALIGNRImmediate(SVOp),
6427 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
6428 SVOp->getSplatIndex() == 0 && V2IsUndef) {
6429 if (VT == MVT::v2f64)
6430 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
6431 if (VT == MVT::v2i64)
6432 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
6435 if (isPSHUFHWMask(M, VT))
6436 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
6437 X86::getShufflePSHUFHWImmediate(SVOp),
6440 if (isPSHUFLWMask(M, VT))
6441 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
6442 X86::getShufflePSHUFLWImmediate(SVOp),
6445 if (isSHUFPMask(M, VT)) {
6446 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
6447 if (VT == MVT::v4f32 || VT == MVT::v4i32)
6448 return getTargetShuffleNode(X86ISD::SHUFPS, dl, VT, V1, V2,
6450 if (VT == MVT::v2f64 || VT == MVT::v2i64)
6451 return getTargetShuffleNode(X86ISD::SHUFPD, dl, VT, V1, V2,
6455 if (X86::isUNPCKL_v_undef_Mask(SVOp))
6456 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
6457 if (X86::isUNPCKH_v_undef_Mask(SVOp))
6458 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6460 //===--------------------------------------------------------------------===//
6461 // Generate target specific nodes for 128 or 256-bit shuffles only
6462 // supported in the AVX instruction set.
6465 // Handle VPERMILPS* permutations
6466 if (isVPERMILPSMask(M, VT, Subtarget))
6467 return getTargetShuffleNode(getVPERMILOpcode(VT), dl, VT, V1,
6468 getShuffleVPERMILPSImmediate(SVOp), DAG);
6470 // Handle VPERMILPD* permutations
6471 if (isVPERMILPDMask(M, VT, Subtarget))
6472 return getTargetShuffleNode(getVPERMILOpcode(VT), dl, VT, V1,
6473 getShuffleVPERMILPDImmediate(SVOp), DAG);
6475 // Handle VPERM2F128 permutations
6476 if (isVPERM2F128Mask(M, VT, Subtarget))
6477 return getTargetShuffleNode(X86ISD::VPERM2F128, dl, VT, V1, V2,
6478 getShuffleVPERM2F128Immediate(SVOp), DAG);
6480 //===--------------------------------------------------------------------===//
6481 // Since no target specific shuffle was selected for this generic one,
6482 // lower it into other known shuffles. FIXME: this isn't true yet, but
6483 // this is the plan.
6486 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
6487 if (VT == MVT::v8i16) {
6488 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
6489 if (NewOp.getNode())
6493 if (VT == MVT::v16i8) {
6494 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
6495 if (NewOp.getNode())
6499 // Handle all 128-bit wide vectors with 4 elements, and match them with
6500 // several different shuffle types.
6501 if (NumElems == 4 && VT.getSizeInBits() == 128)
6502 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
6504 // Handle general 256-bit shuffles
6505 if (VT.is256BitVector())
6506 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
6512 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
6513 SelectionDAG &DAG) const {
6514 EVT VT = Op.getValueType();
6515 DebugLoc dl = Op.getDebugLoc();
6517 if (Op.getOperand(0).getValueType().getSizeInBits() != 128)
6520 if (VT.getSizeInBits() == 8) {
6521 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
6522 Op.getOperand(0), Op.getOperand(1));
6523 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6524 DAG.getValueType(VT));
6525 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6526 } else if (VT.getSizeInBits() == 16) {
6527 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6528 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
6530 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6531 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6532 DAG.getNode(ISD::BITCAST, dl,
6536 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
6537 Op.getOperand(0), Op.getOperand(1));
6538 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6539 DAG.getValueType(VT));
6540 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6541 } else if (VT == MVT::f32) {
6542 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
6543 // the result back to FR32 register. It's only worth matching if the
6544 // result has a single use which is a store or a bitcast to i32. And in
6545 // the case of a store, it's not worth it if the index is a constant 0,
6546 // because a MOVSSmr can be used instead, which is smaller and faster.
6547 if (!Op.hasOneUse())
6549 SDNode *User = *Op.getNode()->use_begin();
6550 if ((User->getOpcode() != ISD::STORE ||
6551 (isa<ConstantSDNode>(Op.getOperand(1)) &&
6552 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
6553 (User->getOpcode() != ISD::BITCAST ||
6554 User->getValueType(0) != MVT::i32))
6556 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6557 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
6560 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
6561 } else if (VT == MVT::i32) {
6562 // ExtractPS works with constant index.
6563 if (isa<ConstantSDNode>(Op.getOperand(1)))
6571 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6572 SelectionDAG &DAG) const {
6573 if (!isa<ConstantSDNode>(Op.getOperand(1)))
6576 SDValue Vec = Op.getOperand(0);
6577 EVT VecVT = Vec.getValueType();
6579 // If this is a 256-bit vector result, first extract the 128-bit vector and
6580 // then extract the element from the 128-bit vector.
6581 if (VecVT.getSizeInBits() == 256) {
6582 DebugLoc dl = Op.getNode()->getDebugLoc();
6583 unsigned NumElems = VecVT.getVectorNumElements();
6584 SDValue Idx = Op.getOperand(1);
6585 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
6587 // Get the 128-bit vector.
6588 bool Upper = IdxVal >= NumElems/2;
6589 Vec = Extract128BitVector(Vec,
6590 DAG.getConstant(Upper ? NumElems/2 : 0, MVT::i32), DAG, dl);
6592 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
6593 Upper ? DAG.getConstant(IdxVal-NumElems/2, MVT::i32) : Idx);
6596 assert(Vec.getValueSizeInBits() <= 128 && "Unexpected vector length");
6598 if (Subtarget->hasSSE41() || Subtarget->hasAVX()) {
6599 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
6604 EVT VT = Op.getValueType();
6605 DebugLoc dl = Op.getDebugLoc();
6606 // TODO: handle v16i8.
6607 if (VT.getSizeInBits() == 16) {
6608 SDValue Vec = Op.getOperand(0);
6609 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6611 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6612 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6613 DAG.getNode(ISD::BITCAST, dl,
6616 // Transform it so it match pextrw which produces a 32-bit result.
6617 EVT EltVT = MVT::i32;
6618 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
6619 Op.getOperand(0), Op.getOperand(1));
6620 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
6621 DAG.getValueType(VT));
6622 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6623 } else if (VT.getSizeInBits() == 32) {
6624 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6628 // SHUFPS the element to the lowest double word, then movss.
6629 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
6630 EVT VVT = Op.getOperand(0).getValueType();
6631 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6632 DAG.getUNDEF(VVT), Mask);
6633 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6634 DAG.getIntPtrConstant(0));
6635 } else if (VT.getSizeInBits() == 64) {
6636 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
6637 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
6638 // to match extract_elt for f64.
6639 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6643 // UNPCKHPD the element to the lowest double word, then movsd.
6644 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
6645 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
6646 int Mask[2] = { 1, -1 };
6647 EVT VVT = Op.getOperand(0).getValueType();
6648 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6649 DAG.getUNDEF(VVT), Mask);
6650 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6651 DAG.getIntPtrConstant(0));
6658 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
6659 SelectionDAG &DAG) const {
6660 EVT VT = Op.getValueType();
6661 EVT EltVT = VT.getVectorElementType();
6662 DebugLoc dl = Op.getDebugLoc();
6664 SDValue N0 = Op.getOperand(0);
6665 SDValue N1 = Op.getOperand(1);
6666 SDValue N2 = Op.getOperand(2);
6668 if (VT.getSizeInBits() == 256)
6671 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
6672 isa<ConstantSDNode>(N2)) {
6674 if (VT == MVT::v8i16)
6675 Opc = X86ISD::PINSRW;
6676 else if (VT == MVT::v16i8)
6677 Opc = X86ISD::PINSRB;
6679 Opc = X86ISD::PINSRB;
6681 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
6683 if (N1.getValueType() != MVT::i32)
6684 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6685 if (N2.getValueType() != MVT::i32)
6686 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6687 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
6688 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
6689 // Bits [7:6] of the constant are the source select. This will always be
6690 // zero here. The DAG Combiner may combine an extract_elt index into these
6691 // bits. For example (insert (extract, 3), 2) could be matched by putting
6692 // the '3' into bits [7:6] of X86ISD::INSERTPS.
6693 // Bits [5:4] of the constant are the destination select. This is the
6694 // value of the incoming immediate.
6695 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
6696 // combine either bitwise AND or insert of float 0.0 to set these bits.
6697 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
6698 // Create this as a scalar to vector..
6699 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
6700 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
6701 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
6702 // PINSR* works with constant index.
6709 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
6710 EVT VT = Op.getValueType();
6711 EVT EltVT = VT.getVectorElementType();
6713 DebugLoc dl = Op.getDebugLoc();
6714 SDValue N0 = Op.getOperand(0);
6715 SDValue N1 = Op.getOperand(1);
6716 SDValue N2 = Op.getOperand(2);
6718 // If this is a 256-bit vector result, first extract the 128-bit vector,
6719 // insert the element into the extracted half and then place it back.
6720 if (VT.getSizeInBits() == 256) {
6721 if (!isa<ConstantSDNode>(N2))
6724 // Get the desired 128-bit vector half.
6725 unsigned NumElems = VT.getVectorNumElements();
6726 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
6727 bool Upper = IdxVal >= NumElems/2;
6728 SDValue Ins128Idx = DAG.getConstant(Upper ? NumElems/2 : 0, MVT::i32);
6729 SDValue V = Extract128BitVector(N0, Ins128Idx, DAG, dl);
6731 // Insert the element into the desired half.
6732 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V,
6733 N1, Upper ? DAG.getConstant(IdxVal-NumElems/2, MVT::i32) : N2);
6735 // Insert the changed part back to the 256-bit vector
6736 return Insert128BitVector(N0, V, Ins128Idx, DAG, dl);
6739 if (Subtarget->hasSSE41() || Subtarget->hasAVX())
6740 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
6742 if (EltVT == MVT::i8)
6745 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
6746 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
6747 // as its second argument.
6748 if (N1.getValueType() != MVT::i32)
6749 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6750 if (N2.getValueType() != MVT::i32)
6751 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6752 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
6758 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6759 LLVMContext *Context = DAG.getContext();
6760 DebugLoc dl = Op.getDebugLoc();
6761 EVT OpVT = Op.getValueType();
6763 // If this is a 256-bit vector result, first insert into a 128-bit
6764 // vector and then insert into the 256-bit vector.
6765 if (OpVT.getSizeInBits() > 128) {
6766 // Insert into a 128-bit vector.
6767 EVT VT128 = EVT::getVectorVT(*Context,
6768 OpVT.getVectorElementType(),
6769 OpVT.getVectorNumElements() / 2);
6771 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
6773 // Insert the 128-bit vector.
6774 return Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, OpVT), Op,
6775 DAG.getConstant(0, MVT::i32),
6779 if (Op.getValueType() == MVT::v1i64 &&
6780 Op.getOperand(0).getValueType() == MVT::i64)
6781 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
6783 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
6784 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
6785 "Expected an SSE type!");
6786 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
6787 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
6790 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
6791 // a simple subregister reference or explicit instructions to grab
6792 // upper bits of a vector.
6794 X86TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6795 if (Subtarget->hasAVX()) {
6796 DebugLoc dl = Op.getNode()->getDebugLoc();
6797 SDValue Vec = Op.getNode()->getOperand(0);
6798 SDValue Idx = Op.getNode()->getOperand(1);
6800 if (Op.getNode()->getValueType(0).getSizeInBits() == 128
6801 && Vec.getNode()->getValueType(0).getSizeInBits() == 256) {
6802 return Extract128BitVector(Vec, Idx, DAG, dl);
6808 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
6809 // simple superregister reference or explicit instructions to insert
6810 // the upper bits of a vector.
6812 X86TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
6813 if (Subtarget->hasAVX()) {
6814 DebugLoc dl = Op.getNode()->getDebugLoc();
6815 SDValue Vec = Op.getNode()->getOperand(0);
6816 SDValue SubVec = Op.getNode()->getOperand(1);
6817 SDValue Idx = Op.getNode()->getOperand(2);
6819 if (Op.getNode()->getValueType(0).getSizeInBits() == 256
6820 && SubVec.getNode()->getValueType(0).getSizeInBits() == 128) {
6821 return Insert128BitVector(Vec, SubVec, Idx, DAG, dl);
6827 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
6828 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
6829 // one of the above mentioned nodes. It has to be wrapped because otherwise
6830 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
6831 // be used to form addressing mode. These wrapped nodes will be selected
6834 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
6835 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
6837 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6839 unsigned char OpFlag = 0;
6840 unsigned WrapperKind = X86ISD::Wrapper;
6841 CodeModel::Model M = getTargetMachine().getCodeModel();
6843 if (Subtarget->isPICStyleRIPRel() &&
6844 (M == CodeModel::Small || M == CodeModel::Kernel))
6845 WrapperKind = X86ISD::WrapperRIP;
6846 else if (Subtarget->isPICStyleGOT())
6847 OpFlag = X86II::MO_GOTOFF;
6848 else if (Subtarget->isPICStyleStubPIC())
6849 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6851 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
6853 CP->getOffset(), OpFlag);
6854 DebugLoc DL = CP->getDebugLoc();
6855 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6856 // With PIC, the address is actually $g + Offset.
6858 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6859 DAG.getNode(X86ISD::GlobalBaseReg,
6860 DebugLoc(), getPointerTy()),
6867 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
6868 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
6870 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6872 unsigned char OpFlag = 0;
6873 unsigned WrapperKind = X86ISD::Wrapper;
6874 CodeModel::Model M = getTargetMachine().getCodeModel();
6876 if (Subtarget->isPICStyleRIPRel() &&
6877 (M == CodeModel::Small || M == CodeModel::Kernel))
6878 WrapperKind = X86ISD::WrapperRIP;
6879 else if (Subtarget->isPICStyleGOT())
6880 OpFlag = X86II::MO_GOTOFF;
6881 else if (Subtarget->isPICStyleStubPIC())
6882 OpFlag = X86II::MO_PIC_BASE_OFFSET;
6884 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
6886 DebugLoc DL = JT->getDebugLoc();
6887 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6889 // With PIC, the address is actually $g + Offset.
6891 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6892 DAG.getNode(X86ISD::GlobalBaseReg,
6893 DebugLoc(), getPointerTy()),
6900 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
6901 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
6903 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
6905 unsigned char OpFlag = 0;
6906 unsigned WrapperKind = X86ISD::Wrapper;
6907 CodeModel::Model M = getTargetMachine().getCodeModel();
6909 if (Subtarget->isPICStyleRIPRel() &&
6910 (M == CodeModel::Small || M == CodeModel::Kernel)) {
6911 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
6912 OpFlag = X86II::MO_GOTPCREL;
6913 WrapperKind = X86ISD::WrapperRIP;
6914 } else if (Subtarget->isPICStyleGOT()) {
6915 OpFlag = X86II::MO_GOT;
6916 } else if (Subtarget->isPICStyleStubPIC()) {
6917 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
6918 } else if (Subtarget->isPICStyleStubNoDynamic()) {
6919 OpFlag = X86II::MO_DARWIN_NONLAZY;
6922 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
6924 DebugLoc DL = Op.getDebugLoc();
6925 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
6928 // With PIC, the address is actually $g + Offset.
6929 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
6930 !Subtarget->is64Bit()) {
6931 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
6932 DAG.getNode(X86ISD::GlobalBaseReg,
6933 DebugLoc(), getPointerTy()),
6937 // For symbols that require a load from a stub to get the address, emit the
6939 if (isGlobalStubReference(OpFlag))
6940 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
6941 MachinePointerInfo::getGOT(), false, false, 0);
6947 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
6948 // Create the TargetBlockAddressAddress node.
6949 unsigned char OpFlags =
6950 Subtarget->ClassifyBlockAddressReference();
6951 CodeModel::Model M = getTargetMachine().getCodeModel();
6952 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
6953 DebugLoc dl = Op.getDebugLoc();
6954 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
6955 /*isTarget=*/true, OpFlags);
6957 if (Subtarget->isPICStyleRIPRel() &&
6958 (M == CodeModel::Small || M == CodeModel::Kernel))
6959 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6961 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6963 // With PIC, the address is actually $g + Offset.
6964 if (isGlobalRelativeToPICBase(OpFlags)) {
6965 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6966 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
6974 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
6976 SelectionDAG &DAG) const {
6977 // Create the TargetGlobalAddress node, folding in the constant
6978 // offset if it is legal.
6979 unsigned char OpFlags =
6980 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
6981 CodeModel::Model M = getTargetMachine().getCodeModel();
6983 if (OpFlags == X86II::MO_NO_FLAG &&
6984 X86::isOffsetSuitableForCodeModel(Offset, M)) {
6985 // A direct static reference to a global.
6986 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
6989 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
6992 if (Subtarget->isPICStyleRIPRel() &&
6993 (M == CodeModel::Small || M == CodeModel::Kernel))
6994 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
6996 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
6998 // With PIC, the address is actually $g + Offset.
6999 if (isGlobalRelativeToPICBase(OpFlags)) {
7000 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7001 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
7005 // For globals that require a load from a stub to get the address, emit the
7007 if (isGlobalStubReference(OpFlags))
7008 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
7009 MachinePointerInfo::getGOT(), false, false, 0);
7011 // If there was a non-zero offset that we didn't fold, create an explicit
7014 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
7015 DAG.getConstant(Offset, getPointerTy()));
7021 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
7022 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
7023 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
7024 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
7028 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
7029 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
7030 unsigned char OperandFlags) {
7031 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7032 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7033 DebugLoc dl = GA->getDebugLoc();
7034 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7035 GA->getValueType(0),
7039 SDValue Ops[] = { Chain, TGA, *InFlag };
7040 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
7042 SDValue Ops[] = { Chain, TGA };
7043 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
7046 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
7047 MFI->setAdjustsStack(true);
7049 SDValue Flag = Chain.getValue(1);
7050 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
7053 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
7055 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7058 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
7059 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
7060 DAG.getNode(X86ISD::GlobalBaseReg,
7061 DebugLoc(), PtrVT), InFlag);
7062 InFlag = Chain.getValue(1);
7064 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
7067 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
7069 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7071 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
7072 X86::RAX, X86II::MO_TLSGD);
7075 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
7076 // "local exec" model.
7077 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7078 const EVT PtrVT, TLSModel::Model model,
7080 DebugLoc dl = GA->getDebugLoc();
7082 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
7083 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
7084 is64Bit ? 257 : 256));
7086 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
7087 DAG.getIntPtrConstant(0),
7088 MachinePointerInfo(Ptr), false, false, 0);
7090 unsigned char OperandFlags = 0;
7091 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
7093 unsigned WrapperKind = X86ISD::Wrapper;
7094 if (model == TLSModel::LocalExec) {
7095 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
7096 } else if (is64Bit) {
7097 assert(model == TLSModel::InitialExec);
7098 OperandFlags = X86II::MO_GOTTPOFF;
7099 WrapperKind = X86ISD::WrapperRIP;
7101 assert(model == TLSModel::InitialExec);
7102 OperandFlags = X86II::MO_INDNTPOFF;
7105 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
7107 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7108 GA->getValueType(0),
7109 GA->getOffset(), OperandFlags);
7110 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
7112 if (model == TLSModel::InitialExec)
7113 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
7114 MachinePointerInfo::getGOT(), false, false, 0);
7116 // The address of the thread local variable is the add of the thread
7117 // pointer with the offset of the variable.
7118 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
7122 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
7124 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
7125 const GlobalValue *GV = GA->getGlobal();
7127 if (Subtarget->isTargetELF()) {
7128 // TODO: implement the "local dynamic" model
7129 // TODO: implement the "initial exec"model for pic executables
7131 // If GV is an alias then use the aliasee for determining
7132 // thread-localness.
7133 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
7134 GV = GA->resolveAliasedGlobal(false);
7136 TLSModel::Model model
7137 = getTLSModel(GV, getTargetMachine().getRelocationModel());
7140 case TLSModel::GeneralDynamic:
7141 case TLSModel::LocalDynamic: // not implemented
7142 if (Subtarget->is64Bit())
7143 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
7144 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
7146 case TLSModel::InitialExec:
7147 case TLSModel::LocalExec:
7148 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
7149 Subtarget->is64Bit());
7151 } else if (Subtarget->isTargetDarwin()) {
7152 // Darwin only has one model of TLS. Lower to that.
7153 unsigned char OpFlag = 0;
7154 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
7155 X86ISD::WrapperRIP : X86ISD::Wrapper;
7157 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7159 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
7160 !Subtarget->is64Bit();
7162 OpFlag = X86II::MO_TLVP_PIC_BASE;
7164 OpFlag = X86II::MO_TLVP;
7165 DebugLoc DL = Op.getDebugLoc();
7166 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
7167 GA->getValueType(0),
7168 GA->getOffset(), OpFlag);
7169 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7171 // With PIC32, the address is actually $g + Offset.
7173 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7174 DAG.getNode(X86ISD::GlobalBaseReg,
7175 DebugLoc(), getPointerTy()),
7178 // Lowering the machine isd will make sure everything is in the right
7180 SDValue Chain = DAG.getEntryNode();
7181 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7182 SDValue Args[] = { Chain, Offset };
7183 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
7185 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
7186 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7187 MFI->setAdjustsStack(true);
7189 // And our return value (tls address) is in the standard call return value
7191 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
7192 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
7196 "TLS not implemented for this target.");
7198 llvm_unreachable("Unreachable");
7203 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values and
7204 /// take a 2 x i32 value to shift plus a shift amount.
7205 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const {
7206 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
7207 EVT VT = Op.getValueType();
7208 unsigned VTBits = VT.getSizeInBits();
7209 DebugLoc dl = Op.getDebugLoc();
7210 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
7211 SDValue ShOpLo = Op.getOperand(0);
7212 SDValue ShOpHi = Op.getOperand(1);
7213 SDValue ShAmt = Op.getOperand(2);
7214 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
7215 DAG.getConstant(VTBits - 1, MVT::i8))
7216 : DAG.getConstant(0, VT);
7219 if (Op.getOpcode() == ISD::SHL_PARTS) {
7220 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
7221 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
7223 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
7224 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
7227 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
7228 DAG.getConstant(VTBits, MVT::i8));
7229 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
7230 AndNode, DAG.getConstant(0, MVT::i8));
7233 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7234 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
7235 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
7237 if (Op.getOpcode() == ISD::SHL_PARTS) {
7238 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7239 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7241 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7242 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7245 SDValue Ops[2] = { Lo, Hi };
7246 return DAG.getMergeValues(Ops, 2, dl);
7249 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
7250 SelectionDAG &DAG) const {
7251 EVT SrcVT = Op.getOperand(0).getValueType();
7253 if (SrcVT.isVector())
7256 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
7257 "Unknown SINT_TO_FP to lower!");
7259 // These are really Legal; return the operand so the caller accepts it as
7261 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
7263 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
7264 Subtarget->is64Bit()) {
7268 DebugLoc dl = Op.getDebugLoc();
7269 unsigned Size = SrcVT.getSizeInBits()/8;
7270 MachineFunction &MF = DAG.getMachineFunction();
7271 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
7272 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7273 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7275 MachinePointerInfo::getFixedStack(SSFI),
7277 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
7280 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
7282 SelectionDAG &DAG) const {
7284 DebugLoc DL = Op.getDebugLoc();
7286 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
7288 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
7290 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
7292 unsigned ByteSize = SrcVT.getSizeInBits()/8;
7294 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
7295 MachineMemOperand *MMO;
7297 int SSFI = FI->getIndex();
7299 DAG.getMachineFunction()
7300 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7301 MachineMemOperand::MOLoad, ByteSize, ByteSize);
7303 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
7304 StackSlot = StackSlot.getOperand(1);
7306 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
7307 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
7309 Tys, Ops, array_lengthof(Ops),
7313 Chain = Result.getValue(1);
7314 SDValue InFlag = Result.getValue(2);
7316 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
7317 // shouldn't be necessary except that RFP cannot be live across
7318 // multiple blocks. When stackifier is fixed, they can be uncoupled.
7319 MachineFunction &MF = DAG.getMachineFunction();
7320 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
7321 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
7322 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7323 Tys = DAG.getVTList(MVT::Other);
7325 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
7327 MachineMemOperand *MMO =
7328 DAG.getMachineFunction()
7329 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7330 MachineMemOperand::MOStore, SSFISize, SSFISize);
7332 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
7333 Ops, array_lengthof(Ops),
7334 Op.getValueType(), MMO);
7335 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
7336 MachinePointerInfo::getFixedStack(SSFI),
7343 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
7344 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
7345 SelectionDAG &DAG) const {
7346 // This algorithm is not obvious. Here it is in C code, more or less:
7348 double uint64_to_double( uint32_t hi, uint32_t lo ) {
7349 static const __m128i exp = { 0x4330000045300000ULL, 0 };
7350 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
7352 // Copy ints to xmm registers.
7353 __m128i xh = _mm_cvtsi32_si128( hi );
7354 __m128i xl = _mm_cvtsi32_si128( lo );
7356 // Combine into low half of a single xmm register.
7357 __m128i x = _mm_unpacklo_epi32( xh, xl );
7361 // Merge in appropriate exponents to give the integer bits the right
7363 x = _mm_unpacklo_epi32( x, exp );
7365 // Subtract away the biases to deal with the IEEE-754 double precision
7367 d = _mm_sub_pd( (__m128d) x, bias );
7369 // All conversions up to here are exact. The correctly rounded result is
7370 // calculated using the current rounding mode using the following
7372 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
7373 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
7374 // store doesn't really need to be here (except
7375 // maybe to zero the other double)
7380 DebugLoc dl = Op.getDebugLoc();
7381 LLVMContext *Context = DAG.getContext();
7383 // Build some magic constants.
7384 std::vector<Constant*> CV0;
7385 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
7386 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
7387 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
7388 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
7389 Constant *C0 = ConstantVector::get(CV0);
7390 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
7392 std::vector<Constant*> CV1;
7394 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
7396 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
7397 Constant *C1 = ConstantVector::get(CV1);
7398 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
7400 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7401 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7403 DAG.getIntPtrConstant(1)));
7404 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7405 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7407 DAG.getIntPtrConstant(0)));
7408 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
7409 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
7410 MachinePointerInfo::getConstantPool(),
7412 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
7413 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
7414 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
7415 MachinePointerInfo::getConstantPool(),
7417 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
7419 // Add the halves; easiest way is to swap them into another reg first.
7420 int ShufMask[2] = { 1, -1 };
7421 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
7422 DAG.getUNDEF(MVT::v2f64), ShufMask);
7423 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
7424 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
7425 DAG.getIntPtrConstant(0));
7428 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
7429 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
7430 SelectionDAG &DAG) const {
7431 DebugLoc dl = Op.getDebugLoc();
7432 // FP constant to bias correct the final result.
7433 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
7436 // Load the 32-bit value into an XMM register.
7437 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7440 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7441 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
7442 DAG.getIntPtrConstant(0));
7444 // Or the load with the bias.
7445 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
7446 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7447 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7449 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7450 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7451 MVT::v2f64, Bias)));
7452 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7453 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
7454 DAG.getIntPtrConstant(0));
7456 // Subtract the bias.
7457 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
7459 // Handle final rounding.
7460 EVT DestVT = Op.getValueType();
7462 if (DestVT.bitsLT(MVT::f64)) {
7463 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
7464 DAG.getIntPtrConstant(0));
7465 } else if (DestVT.bitsGT(MVT::f64)) {
7466 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
7469 // Handle final rounding.
7473 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
7474 SelectionDAG &DAG) const {
7475 SDValue N0 = Op.getOperand(0);
7476 DebugLoc dl = Op.getDebugLoc();
7478 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
7479 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
7480 // the optimization here.
7481 if (DAG.SignBitIsZero(N0))
7482 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
7484 EVT SrcVT = N0.getValueType();
7485 EVT DstVT = Op.getValueType();
7486 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
7487 return LowerUINT_TO_FP_i64(Op, DAG);
7488 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
7489 return LowerUINT_TO_FP_i32(Op, DAG);
7491 // Make a 64-bit buffer, and use it to build an FILD.
7492 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
7493 if (SrcVT == MVT::i32) {
7494 SDValue WordOff = DAG.getConstant(4, getPointerTy());
7495 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
7496 getPointerTy(), StackSlot, WordOff);
7497 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7498 StackSlot, MachinePointerInfo(),
7500 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
7501 OffsetSlot, MachinePointerInfo(),
7503 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
7507 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
7508 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7509 StackSlot, MachinePointerInfo(),
7511 // For i64 source, we need to add the appropriate power of 2 if the input
7512 // was negative. This is the same as the optimization in
7513 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
7514 // we must be careful to do the computation in x87 extended precision, not
7515 // in SSE. (The generic code can't know it's OK to do this, or how to.)
7516 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
7517 MachineMemOperand *MMO =
7518 DAG.getMachineFunction()
7519 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7520 MachineMemOperand::MOLoad, 8, 8);
7522 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
7523 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
7524 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
7527 APInt FF(32, 0x5F800000ULL);
7529 // Check whether the sign bit is set.
7530 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
7531 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
7534 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
7535 SDValue FudgePtr = DAG.getConstantPool(
7536 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
7539 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
7540 SDValue Zero = DAG.getIntPtrConstant(0);
7541 SDValue Four = DAG.getIntPtrConstant(4);
7542 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
7544 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
7546 // Load the value out, extending it from f32 to f80.
7547 // FIXME: Avoid the extend by constructing the right constant pool?
7548 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
7549 FudgePtr, MachinePointerInfo::getConstantPool(),
7550 MVT::f32, false, false, 4);
7551 // Extend everything to 80 bits to force it to be done on x87.
7552 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
7553 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
7556 std::pair<SDValue,SDValue> X86TargetLowering::
7557 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
7558 DebugLoc DL = Op.getDebugLoc();
7560 EVT DstTy = Op.getValueType();
7563 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
7567 assert(DstTy.getSimpleVT() <= MVT::i64 &&
7568 DstTy.getSimpleVT() >= MVT::i16 &&
7569 "Unknown FP_TO_SINT to lower!");
7571 // These are really Legal.
7572 if (DstTy == MVT::i32 &&
7573 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7574 return std::make_pair(SDValue(), SDValue());
7575 if (Subtarget->is64Bit() &&
7576 DstTy == MVT::i64 &&
7577 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7578 return std::make_pair(SDValue(), SDValue());
7580 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
7582 MachineFunction &MF = DAG.getMachineFunction();
7583 unsigned MemSize = DstTy.getSizeInBits()/8;
7584 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7585 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7590 switch (DstTy.getSimpleVT().SimpleTy) {
7591 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
7592 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
7593 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
7594 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
7597 SDValue Chain = DAG.getEntryNode();
7598 SDValue Value = Op.getOperand(0);
7599 EVT TheVT = Op.getOperand(0).getValueType();
7600 if (isScalarFPTypeInSSEReg(TheVT)) {
7601 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
7602 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
7603 MachinePointerInfo::getFixedStack(SSFI),
7605 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
7607 Chain, StackSlot, DAG.getValueType(TheVT)
7610 MachineMemOperand *MMO =
7611 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7612 MachineMemOperand::MOLoad, MemSize, MemSize);
7613 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
7615 Chain = Value.getValue(1);
7616 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7617 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7620 MachineMemOperand *MMO =
7621 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7622 MachineMemOperand::MOStore, MemSize, MemSize);
7624 // Build the FP_TO_INT*_IN_MEM
7625 SDValue Ops[] = { Chain, Value, StackSlot };
7626 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
7627 Ops, 3, DstTy, MMO);
7629 return std::make_pair(FIST, StackSlot);
7632 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
7633 SelectionDAG &DAG) const {
7634 if (Op.getValueType().isVector())
7637 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
7638 SDValue FIST = Vals.first, StackSlot = Vals.second;
7639 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
7640 if (FIST.getNode() == 0) return Op;
7643 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7644 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7647 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
7648 SelectionDAG &DAG) const {
7649 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
7650 SDValue FIST = Vals.first, StackSlot = Vals.second;
7651 assert(FIST.getNode() && "Unexpected failure");
7654 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7655 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7658 SDValue X86TargetLowering::LowerFABS(SDValue Op,
7659 SelectionDAG &DAG) const {
7660 LLVMContext *Context = DAG.getContext();
7661 DebugLoc dl = Op.getDebugLoc();
7662 EVT VT = Op.getValueType();
7665 EltVT = VT.getVectorElementType();
7666 std::vector<Constant*> CV;
7667 if (EltVT == MVT::f64) {
7668 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
7672 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
7678 Constant *C = ConstantVector::get(CV);
7679 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7680 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7681 MachinePointerInfo::getConstantPool(),
7683 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
7686 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
7687 LLVMContext *Context = DAG.getContext();
7688 DebugLoc dl = Op.getDebugLoc();
7689 EVT VT = Op.getValueType();
7692 EltVT = VT.getVectorElementType();
7693 std::vector<Constant*> CV;
7694 if (EltVT == MVT::f64) {
7695 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
7699 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
7705 Constant *C = ConstantVector::get(CV);
7706 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7707 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7708 MachinePointerInfo::getConstantPool(),
7710 if (VT.isVector()) {
7711 return DAG.getNode(ISD::BITCAST, dl, VT,
7712 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
7713 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7715 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
7717 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
7721 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
7722 LLVMContext *Context = DAG.getContext();
7723 SDValue Op0 = Op.getOperand(0);
7724 SDValue Op1 = Op.getOperand(1);
7725 DebugLoc dl = Op.getDebugLoc();
7726 EVT VT = Op.getValueType();
7727 EVT SrcVT = Op1.getValueType();
7729 // If second operand is smaller, extend it first.
7730 if (SrcVT.bitsLT(VT)) {
7731 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
7734 // And if it is bigger, shrink it first.
7735 if (SrcVT.bitsGT(VT)) {
7736 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
7740 // At this point the operands and the result should have the same
7741 // type, and that won't be f80 since that is not custom lowered.
7743 // First get the sign bit of second operand.
7744 std::vector<Constant*> CV;
7745 if (SrcVT == MVT::f64) {
7746 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
7747 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7749 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
7750 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7751 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7752 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7754 Constant *C = ConstantVector::get(CV);
7755 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7756 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
7757 MachinePointerInfo::getConstantPool(),
7759 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
7761 // Shift sign bit right or left if the two operands have different types.
7762 if (SrcVT.bitsGT(VT)) {
7763 // Op0 is MVT::f32, Op1 is MVT::f64.
7764 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
7765 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
7766 DAG.getConstant(32, MVT::i32));
7767 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
7768 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
7769 DAG.getIntPtrConstant(0));
7772 // Clear first operand sign bit.
7774 if (VT == MVT::f64) {
7775 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
7776 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
7778 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
7779 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7780 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7781 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
7783 C = ConstantVector::get(CV);
7784 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7785 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7786 MachinePointerInfo::getConstantPool(),
7788 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
7790 // Or the value with the sign bit.
7791 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
7794 SDValue X86TargetLowering::LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) const {
7795 SDValue N0 = Op.getOperand(0);
7796 DebugLoc dl = Op.getDebugLoc();
7797 EVT VT = Op.getValueType();
7799 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
7800 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
7801 DAG.getConstant(1, VT));
7802 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
7805 /// Emit nodes that will be selected as "test Op0,Op0", or something
7807 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
7808 SelectionDAG &DAG) const {
7809 DebugLoc dl = Op.getDebugLoc();
7811 // CF and OF aren't always set the way we want. Determine which
7812 // of these we need.
7813 bool NeedCF = false;
7814 bool NeedOF = false;
7817 case X86::COND_A: case X86::COND_AE:
7818 case X86::COND_B: case X86::COND_BE:
7821 case X86::COND_G: case X86::COND_GE:
7822 case X86::COND_L: case X86::COND_LE:
7823 case X86::COND_O: case X86::COND_NO:
7828 // See if we can use the EFLAGS value from the operand instead of
7829 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
7830 // we prove that the arithmetic won't overflow, we can't use OF or CF.
7831 if (Op.getResNo() != 0 || NeedOF || NeedCF)
7832 // Emit a CMP with 0, which is the TEST pattern.
7833 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7834 DAG.getConstant(0, Op.getValueType()));
7836 unsigned Opcode = 0;
7837 unsigned NumOperands = 0;
7838 switch (Op.getNode()->getOpcode()) {
7840 // Due to an isel shortcoming, be conservative if this add is likely to be
7841 // selected as part of a load-modify-store instruction. When the root node
7842 // in a match is a store, isel doesn't know how to remap non-chain non-flag
7843 // uses of other nodes in the match, such as the ADD in this case. This
7844 // leads to the ADD being left around and reselected, with the result being
7845 // two adds in the output. Alas, even if none our users are stores, that
7846 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
7847 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
7848 // climbing the DAG back to the root, and it doesn't seem to be worth the
7850 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7851 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7852 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
7855 if (ConstantSDNode *C =
7856 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
7857 // An add of one will be selected as an INC.
7858 if (C->getAPIntValue() == 1) {
7859 Opcode = X86ISD::INC;
7864 // An add of negative one (subtract of one) will be selected as a DEC.
7865 if (C->getAPIntValue().isAllOnesValue()) {
7866 Opcode = X86ISD::DEC;
7872 // Otherwise use a regular EFLAGS-setting add.
7873 Opcode = X86ISD::ADD;
7877 // If the primary and result isn't used, don't bother using X86ISD::AND,
7878 // because a TEST instruction will be better.
7879 bool NonFlagUse = false;
7880 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7881 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
7883 unsigned UOpNo = UI.getOperandNo();
7884 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
7885 // Look pass truncate.
7886 UOpNo = User->use_begin().getOperandNo();
7887 User = *User->use_begin();
7890 if (User->getOpcode() != ISD::BRCOND &&
7891 User->getOpcode() != ISD::SETCC &&
7892 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
7905 // Due to the ISEL shortcoming noted above, be conservative if this op is
7906 // likely to be selected as part of a load-modify-store instruction.
7907 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
7908 UE = Op.getNode()->use_end(); UI != UE; ++UI)
7909 if (UI->getOpcode() == ISD::STORE)
7912 // Otherwise use a regular EFLAGS-setting instruction.
7913 switch (Op.getNode()->getOpcode()) {
7914 default: llvm_unreachable("unexpected operator!");
7915 case ISD::SUB: Opcode = X86ISD::SUB; break;
7916 case ISD::OR: Opcode = X86ISD::OR; break;
7917 case ISD::XOR: Opcode = X86ISD::XOR; break;
7918 case ISD::AND: Opcode = X86ISD::AND; break;
7930 return SDValue(Op.getNode(), 1);
7937 // Emit a CMP with 0, which is the TEST pattern.
7938 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
7939 DAG.getConstant(0, Op.getValueType()));
7941 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
7942 SmallVector<SDValue, 4> Ops;
7943 for (unsigned i = 0; i != NumOperands; ++i)
7944 Ops.push_back(Op.getOperand(i));
7946 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
7947 DAG.ReplaceAllUsesWith(Op, New);
7948 return SDValue(New.getNode(), 1);
7951 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
7953 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
7954 SelectionDAG &DAG) const {
7955 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
7956 if (C->getAPIntValue() == 0)
7957 return EmitTest(Op0, X86CC, DAG);
7959 DebugLoc dl = Op0.getDebugLoc();
7960 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
7963 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
7964 /// if it's possible.
7965 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
7966 DebugLoc dl, SelectionDAG &DAG) const {
7967 SDValue Op0 = And.getOperand(0);
7968 SDValue Op1 = And.getOperand(1);
7969 if (Op0.getOpcode() == ISD::TRUNCATE)
7970 Op0 = Op0.getOperand(0);
7971 if (Op1.getOpcode() == ISD::TRUNCATE)
7972 Op1 = Op1.getOperand(0);
7975 if (Op1.getOpcode() == ISD::SHL)
7976 std::swap(Op0, Op1);
7977 if (Op0.getOpcode() == ISD::SHL) {
7978 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
7979 if (And00C->getZExtValue() == 1) {
7980 // If we looked past a truncate, check that it's only truncating away
7982 unsigned BitWidth = Op0.getValueSizeInBits();
7983 unsigned AndBitWidth = And.getValueSizeInBits();
7984 if (BitWidth > AndBitWidth) {
7985 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
7986 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
7987 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
7991 RHS = Op0.getOperand(1);
7993 } else if (Op1.getOpcode() == ISD::Constant) {
7994 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
7995 SDValue AndLHS = Op0;
7996 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
7997 LHS = AndLHS.getOperand(0);
7998 RHS = AndLHS.getOperand(1);
8002 if (LHS.getNode()) {
8003 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
8004 // instruction. Since the shift amount is in-range-or-undefined, we know
8005 // that doing a bittest on the i32 value is ok. We extend to i32 because
8006 // the encoding for the i16 version is larger than the i32 version.
8007 // Also promote i16 to i32 for performance / code size reason.
8008 if (LHS.getValueType() == MVT::i8 ||
8009 LHS.getValueType() == MVT::i16)
8010 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
8012 // If the operand types disagree, extend the shift amount to match. Since
8013 // BT ignores high bits (like shifts) we can use anyextend.
8014 if (LHS.getValueType() != RHS.getValueType())
8015 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
8017 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
8018 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
8019 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8020 DAG.getConstant(Cond, MVT::i8), BT);
8026 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
8027 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
8028 SDValue Op0 = Op.getOperand(0);
8029 SDValue Op1 = Op.getOperand(1);
8030 DebugLoc dl = Op.getDebugLoc();
8031 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
8033 // Optimize to BT if possible.
8034 // Lower (X & (1 << N)) == 0 to BT(X, N).
8035 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
8036 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
8037 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
8038 Op1.getOpcode() == ISD::Constant &&
8039 cast<ConstantSDNode>(Op1)->isNullValue() &&
8040 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
8041 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
8042 if (NewSetCC.getNode())
8046 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
8048 if (Op1.getOpcode() == ISD::Constant &&
8049 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
8050 cast<ConstantSDNode>(Op1)->isNullValue()) &&
8051 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
8053 // If the input is a setcc, then reuse the input setcc or use a new one with
8054 // the inverted condition.
8055 if (Op0.getOpcode() == X86ISD::SETCC) {
8056 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
8057 bool Invert = (CC == ISD::SETNE) ^
8058 cast<ConstantSDNode>(Op1)->isNullValue();
8059 if (!Invert) return Op0;
8061 CCode = X86::GetOppositeBranchCondition(CCode);
8062 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8063 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
8067 bool isFP = Op1.getValueType().isFloatingPoint();
8068 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
8069 if (X86CC == X86::COND_INVALID)
8072 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
8073 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8074 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
8077 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
8079 SDValue Op0 = Op.getOperand(0);
8080 SDValue Op1 = Op.getOperand(1);
8081 SDValue CC = Op.getOperand(2);
8082 EVT VT = Op.getValueType();
8083 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
8084 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
8085 DebugLoc dl = Op.getDebugLoc();
8089 EVT EltVT = Op0.getValueType().getVectorElementType();
8090 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
8092 unsigned Opc = EltVT == MVT::f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
8095 switch (SetCCOpcode) {
8098 case ISD::SETEQ: SSECC = 0; break;
8100 case ISD::SETGT: Swap = true; // Fallthrough
8102 case ISD::SETOLT: SSECC = 1; break;
8104 case ISD::SETGE: Swap = true; // Fallthrough
8106 case ISD::SETOLE: SSECC = 2; break;
8107 case ISD::SETUO: SSECC = 3; break;
8109 case ISD::SETNE: SSECC = 4; break;
8110 case ISD::SETULE: Swap = true;
8111 case ISD::SETUGE: SSECC = 5; break;
8112 case ISD::SETULT: Swap = true;
8113 case ISD::SETUGT: SSECC = 6; break;
8114 case ISD::SETO: SSECC = 7; break;
8117 std::swap(Op0, Op1);
8119 // In the two special cases we can't handle, emit two comparisons.
8121 if (SetCCOpcode == ISD::SETUEQ) {
8123 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
8124 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
8125 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
8127 else if (SetCCOpcode == ISD::SETONE) {
8129 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
8130 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
8131 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
8133 llvm_unreachable("Illegal FP comparison");
8135 // Handle all other FP comparisons here.
8136 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
8139 if (!isFP && VT.getSizeInBits() == 256)
8142 // We are handling one of the integer comparisons here. Since SSE only has
8143 // GT and EQ comparisons for integer, swapping operands and multiple
8144 // operations may be required for some comparisons.
8145 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
8146 bool Swap = false, Invert = false, FlipSigns = false;
8148 switch (VT.getSimpleVT().SimpleTy) {
8150 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
8151 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
8152 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
8153 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
8156 switch (SetCCOpcode) {
8158 case ISD::SETNE: Invert = true;
8159 case ISD::SETEQ: Opc = EQOpc; break;
8160 case ISD::SETLT: Swap = true;
8161 case ISD::SETGT: Opc = GTOpc; break;
8162 case ISD::SETGE: Swap = true;
8163 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
8164 case ISD::SETULT: Swap = true;
8165 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
8166 case ISD::SETUGE: Swap = true;
8167 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
8170 std::swap(Op0, Op1);
8172 // Since SSE has no unsigned integer comparisons, we need to flip the sign
8173 // bits of the inputs before performing those operations.
8175 EVT EltVT = VT.getVectorElementType();
8176 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
8178 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
8179 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
8181 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
8182 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
8185 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
8187 // If the logical-not of the result is required, perform that now.
8189 Result = DAG.getNOT(dl, Result, VT);
8194 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
8195 static bool isX86LogicalCmp(SDValue Op) {
8196 unsigned Opc = Op.getNode()->getOpcode();
8197 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
8199 if (Op.getResNo() == 1 &&
8200 (Opc == X86ISD::ADD ||
8201 Opc == X86ISD::SUB ||
8202 Opc == X86ISD::ADC ||
8203 Opc == X86ISD::SBB ||
8204 Opc == X86ISD::SMUL ||
8205 Opc == X86ISD::UMUL ||
8206 Opc == X86ISD::INC ||
8207 Opc == X86ISD::DEC ||
8208 Opc == X86ISD::OR ||
8209 Opc == X86ISD::XOR ||
8210 Opc == X86ISD::AND))
8213 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
8219 static bool isZero(SDValue V) {
8220 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
8221 return C && C->isNullValue();
8224 static bool isAllOnes(SDValue V) {
8225 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
8226 return C && C->isAllOnesValue();
8229 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
8230 bool addTest = true;
8231 SDValue Cond = Op.getOperand(0);
8232 SDValue Op1 = Op.getOperand(1);
8233 SDValue Op2 = Op.getOperand(2);
8234 DebugLoc DL = Op.getDebugLoc();
8237 if (Cond.getOpcode() == ISD::SETCC) {
8238 SDValue NewCond = LowerSETCC(Cond, DAG);
8239 if (NewCond.getNode())
8243 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
8244 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
8245 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
8246 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
8247 if (Cond.getOpcode() == X86ISD::SETCC &&
8248 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
8249 isZero(Cond.getOperand(1).getOperand(1))) {
8250 SDValue Cmp = Cond.getOperand(1);
8252 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
8254 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
8255 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
8256 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
8258 SDValue CmpOp0 = Cmp.getOperand(0);
8259 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
8260 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
8262 SDValue Res = // Res = 0 or -1.
8263 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
8264 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
8266 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
8267 Res = DAG.getNOT(DL, Res, Res.getValueType());
8269 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
8270 if (N2C == 0 || !N2C->isNullValue())
8271 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
8276 // Look past (and (setcc_carry (cmp ...)), 1).
8277 if (Cond.getOpcode() == ISD::AND &&
8278 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
8279 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
8280 if (C && C->getAPIntValue() == 1)
8281 Cond = Cond.getOperand(0);
8284 // If condition flag is set by a X86ISD::CMP, then use it as the condition
8285 // setting operand in place of the X86ISD::SETCC.
8286 if (Cond.getOpcode() == X86ISD::SETCC ||
8287 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
8288 CC = Cond.getOperand(0);
8290 SDValue Cmp = Cond.getOperand(1);
8291 unsigned Opc = Cmp.getOpcode();
8292 EVT VT = Op.getValueType();
8294 bool IllegalFPCMov = false;
8295 if (VT.isFloatingPoint() && !VT.isVector() &&
8296 !isScalarFPTypeInSSEReg(VT)) // FPStack?
8297 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
8299 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
8300 Opc == X86ISD::BT) { // FIXME
8307 // Look pass the truncate.
8308 if (Cond.getOpcode() == ISD::TRUNCATE)
8309 Cond = Cond.getOperand(0);
8311 // We know the result of AND is compared against zero. Try to match
8313 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
8314 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
8315 if (NewSetCC.getNode()) {
8316 CC = NewSetCC.getOperand(0);
8317 Cond = NewSetCC.getOperand(1);
8324 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8325 Cond = EmitTest(Cond, X86::COND_NE, DAG);
8328 // a < b ? -1 : 0 -> RES = ~setcc_carry
8329 // a < b ? 0 : -1 -> RES = setcc_carry
8330 // a >= b ? -1 : 0 -> RES = setcc_carry
8331 // a >= b ? 0 : -1 -> RES = ~setcc_carry
8332 if (Cond.getOpcode() == X86ISD::CMP) {
8333 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
8335 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
8336 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
8337 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
8338 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
8339 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
8340 return DAG.getNOT(DL, Res, Res.getValueType());
8345 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
8346 // condition is true.
8347 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
8348 SDValue Ops[] = { Op2, Op1, CC, Cond };
8349 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
8352 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
8353 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
8354 // from the AND / OR.
8355 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
8356 Opc = Op.getOpcode();
8357 if (Opc != ISD::OR && Opc != ISD::AND)
8359 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
8360 Op.getOperand(0).hasOneUse() &&
8361 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
8362 Op.getOperand(1).hasOneUse());
8365 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
8366 // 1 and that the SETCC node has a single use.
8367 static bool isXor1OfSetCC(SDValue Op) {
8368 if (Op.getOpcode() != ISD::XOR)
8370 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8371 if (N1C && N1C->getAPIntValue() == 1) {
8372 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
8373 Op.getOperand(0).hasOneUse();
8378 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
8379 bool addTest = true;
8380 SDValue Chain = Op.getOperand(0);
8381 SDValue Cond = Op.getOperand(1);
8382 SDValue Dest = Op.getOperand(2);
8383 DebugLoc dl = Op.getDebugLoc();
8386 if (Cond.getOpcode() == ISD::SETCC) {
8387 SDValue NewCond = LowerSETCC(Cond, DAG);
8388 if (NewCond.getNode())
8392 // FIXME: LowerXALUO doesn't handle these!!
8393 else if (Cond.getOpcode() == X86ISD::ADD ||
8394 Cond.getOpcode() == X86ISD::SUB ||
8395 Cond.getOpcode() == X86ISD::SMUL ||
8396 Cond.getOpcode() == X86ISD::UMUL)
8397 Cond = LowerXALUO(Cond, DAG);
8400 // Look pass (and (setcc_carry (cmp ...)), 1).
8401 if (Cond.getOpcode() == ISD::AND &&
8402 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
8403 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
8404 if (C && C->getAPIntValue() == 1)
8405 Cond = Cond.getOperand(0);
8408 // If condition flag is set by a X86ISD::CMP, then use it as the condition
8409 // setting operand in place of the X86ISD::SETCC.
8410 if (Cond.getOpcode() == X86ISD::SETCC ||
8411 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
8412 CC = Cond.getOperand(0);
8414 SDValue Cmp = Cond.getOperand(1);
8415 unsigned Opc = Cmp.getOpcode();
8416 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
8417 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
8421 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
8425 // These can only come from an arithmetic instruction with overflow,
8426 // e.g. SADDO, UADDO.
8427 Cond = Cond.getNode()->getOperand(1);
8434 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
8435 SDValue Cmp = Cond.getOperand(0).getOperand(1);
8436 if (CondOpc == ISD::OR) {
8437 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
8438 // two branches instead of an explicit OR instruction with a
8440 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8441 isX86LogicalCmp(Cmp)) {
8442 CC = Cond.getOperand(0).getOperand(0);
8443 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8444 Chain, Dest, CC, Cmp);
8445 CC = Cond.getOperand(1).getOperand(0);
8449 } else { // ISD::AND
8450 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
8451 // two branches instead of an explicit AND instruction with a
8452 // separate test. However, we only do this if this block doesn't
8453 // have a fall-through edge, because this requires an explicit
8454 // jmp when the condition is false.
8455 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8456 isX86LogicalCmp(Cmp) &&
8457 Op.getNode()->hasOneUse()) {
8458 X86::CondCode CCode =
8459 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8460 CCode = X86::GetOppositeBranchCondition(CCode);
8461 CC = DAG.getConstant(CCode, MVT::i8);
8462 SDNode *User = *Op.getNode()->use_begin();
8463 // Look for an unconditional branch following this conditional branch.
8464 // We need this because we need to reverse the successors in order
8465 // to implement FCMP_OEQ.
8466 if (User->getOpcode() == ISD::BR) {
8467 SDValue FalseBB = User->getOperand(1);
8469 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
8470 assert(NewBR == User);
8474 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8475 Chain, Dest, CC, Cmp);
8476 X86::CondCode CCode =
8477 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
8478 CCode = X86::GetOppositeBranchCondition(CCode);
8479 CC = DAG.getConstant(CCode, MVT::i8);
8485 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
8486 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
8487 // It should be transformed during dag combiner except when the condition
8488 // is set by a arithmetics with overflow node.
8489 X86::CondCode CCode =
8490 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8491 CCode = X86::GetOppositeBranchCondition(CCode);
8492 CC = DAG.getConstant(CCode, MVT::i8);
8493 Cond = Cond.getOperand(0).getOperand(1);
8499 // Look pass the truncate.
8500 if (Cond.getOpcode() == ISD::TRUNCATE)
8501 Cond = Cond.getOperand(0);
8503 // We know the result of AND is compared against zero. Try to match
8505 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
8506 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
8507 if (NewSetCC.getNode()) {
8508 CC = NewSetCC.getOperand(0);
8509 Cond = NewSetCC.getOperand(1);
8516 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8517 Cond = EmitTest(Cond, X86::COND_NE, DAG);
8519 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8520 Chain, Dest, CC, Cond);
8524 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
8525 // Calls to _alloca is needed to probe the stack when allocating more than 4k
8526 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
8527 // that the guard pages used by the OS virtual memory manager are allocated in
8528 // correct sequence.
8530 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
8531 SelectionDAG &DAG) const {
8532 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows()) &&
8533 "This should be used only on Windows targets");
8534 assert(!Subtarget->isTargetEnvMacho());
8535 DebugLoc dl = Op.getDebugLoc();
8538 SDValue Chain = Op.getOperand(0);
8539 SDValue Size = Op.getOperand(1);
8540 // FIXME: Ensure alignment here
8544 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
8545 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
8547 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
8548 Flag = Chain.getValue(1);
8550 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
8552 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
8553 Flag = Chain.getValue(1);
8555 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
8557 SDValue Ops1[2] = { Chain.getValue(0), Chain };
8558 return DAG.getMergeValues(Ops1, 2, dl);
8561 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
8562 MachineFunction &MF = DAG.getMachineFunction();
8563 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
8565 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8566 DebugLoc DL = Op.getDebugLoc();
8568 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
8569 // vastart just stores the address of the VarArgsFrameIndex slot into the
8570 // memory location argument.
8571 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8573 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
8574 MachinePointerInfo(SV), false, false, 0);
8578 // gp_offset (0 - 6 * 8)
8579 // fp_offset (48 - 48 + 8 * 16)
8580 // overflow_arg_area (point to parameters coming in memory).
8582 SmallVector<SDValue, 8> MemOps;
8583 SDValue FIN = Op.getOperand(1);
8585 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
8586 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
8588 FIN, MachinePointerInfo(SV), false, false, 0);
8589 MemOps.push_back(Store);
8592 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8593 FIN, DAG.getIntPtrConstant(4));
8594 Store = DAG.getStore(Op.getOperand(0), DL,
8595 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
8597 FIN, MachinePointerInfo(SV, 4), false, false, 0);
8598 MemOps.push_back(Store);
8600 // Store ptr to overflow_arg_area
8601 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8602 FIN, DAG.getIntPtrConstant(4));
8603 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8605 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
8606 MachinePointerInfo(SV, 8),
8608 MemOps.push_back(Store);
8610 // Store ptr to reg_save_area.
8611 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8612 FIN, DAG.getIntPtrConstant(8));
8613 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
8615 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
8616 MachinePointerInfo(SV, 16), false, false, 0);
8617 MemOps.push_back(Store);
8618 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
8619 &MemOps[0], MemOps.size());
8622 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
8623 assert(Subtarget->is64Bit() &&
8624 "LowerVAARG only handles 64-bit va_arg!");
8625 assert((Subtarget->isTargetLinux() ||
8626 Subtarget->isTargetDarwin()) &&
8627 "Unhandled target in LowerVAARG");
8628 assert(Op.getNode()->getNumOperands() == 4);
8629 SDValue Chain = Op.getOperand(0);
8630 SDValue SrcPtr = Op.getOperand(1);
8631 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8632 unsigned Align = Op.getConstantOperandVal(3);
8633 DebugLoc dl = Op.getDebugLoc();
8635 EVT ArgVT = Op.getNode()->getValueType(0);
8636 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
8637 uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
8640 // Decide which area this value should be read from.
8641 // TODO: Implement the AMD64 ABI in its entirety. This simple
8642 // selection mechanism works only for the basic types.
8643 if (ArgVT == MVT::f80) {
8644 llvm_unreachable("va_arg for f80 not yet implemented");
8645 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
8646 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
8647 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
8648 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
8650 llvm_unreachable("Unhandled argument type in LowerVAARG");
8654 // Sanity Check: Make sure using fp_offset makes sense.
8655 assert(!UseSoftFloat &&
8656 !(DAG.getMachineFunction()
8657 .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
8658 Subtarget->hasXMM());
8661 // Insert VAARG_64 node into the DAG
8662 // VAARG_64 returns two values: Variable Argument Address, Chain
8663 SmallVector<SDValue, 11> InstOps;
8664 InstOps.push_back(Chain);
8665 InstOps.push_back(SrcPtr);
8666 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
8667 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
8668 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
8669 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
8670 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
8671 VTs, &InstOps[0], InstOps.size(),
8673 MachinePointerInfo(SV),
8678 Chain = VAARG.getValue(1);
8680 // Load the next argument and return it
8681 return DAG.getLoad(ArgVT, dl,
8684 MachinePointerInfo(),
8688 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
8689 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
8690 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
8691 SDValue Chain = Op.getOperand(0);
8692 SDValue DstPtr = Op.getOperand(1);
8693 SDValue SrcPtr = Op.getOperand(2);
8694 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
8695 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
8696 DebugLoc DL = Op.getDebugLoc();
8698 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
8699 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
8701 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
8705 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
8706 DebugLoc dl = Op.getDebugLoc();
8707 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8709 default: return SDValue(); // Don't custom lower most intrinsics.
8710 // Comparison intrinsics.
8711 case Intrinsic::x86_sse_comieq_ss:
8712 case Intrinsic::x86_sse_comilt_ss:
8713 case Intrinsic::x86_sse_comile_ss:
8714 case Intrinsic::x86_sse_comigt_ss:
8715 case Intrinsic::x86_sse_comige_ss:
8716 case Intrinsic::x86_sse_comineq_ss:
8717 case Intrinsic::x86_sse_ucomieq_ss:
8718 case Intrinsic::x86_sse_ucomilt_ss:
8719 case Intrinsic::x86_sse_ucomile_ss:
8720 case Intrinsic::x86_sse_ucomigt_ss:
8721 case Intrinsic::x86_sse_ucomige_ss:
8722 case Intrinsic::x86_sse_ucomineq_ss:
8723 case Intrinsic::x86_sse2_comieq_sd:
8724 case Intrinsic::x86_sse2_comilt_sd:
8725 case Intrinsic::x86_sse2_comile_sd:
8726 case Intrinsic::x86_sse2_comigt_sd:
8727 case Intrinsic::x86_sse2_comige_sd:
8728 case Intrinsic::x86_sse2_comineq_sd:
8729 case Intrinsic::x86_sse2_ucomieq_sd:
8730 case Intrinsic::x86_sse2_ucomilt_sd:
8731 case Intrinsic::x86_sse2_ucomile_sd:
8732 case Intrinsic::x86_sse2_ucomigt_sd:
8733 case Intrinsic::x86_sse2_ucomige_sd:
8734 case Intrinsic::x86_sse2_ucomineq_sd: {
8736 ISD::CondCode CC = ISD::SETCC_INVALID;
8739 case Intrinsic::x86_sse_comieq_ss:
8740 case Intrinsic::x86_sse2_comieq_sd:
8744 case Intrinsic::x86_sse_comilt_ss:
8745 case Intrinsic::x86_sse2_comilt_sd:
8749 case Intrinsic::x86_sse_comile_ss:
8750 case Intrinsic::x86_sse2_comile_sd:
8754 case Intrinsic::x86_sse_comigt_ss:
8755 case Intrinsic::x86_sse2_comigt_sd:
8759 case Intrinsic::x86_sse_comige_ss:
8760 case Intrinsic::x86_sse2_comige_sd:
8764 case Intrinsic::x86_sse_comineq_ss:
8765 case Intrinsic::x86_sse2_comineq_sd:
8769 case Intrinsic::x86_sse_ucomieq_ss:
8770 case Intrinsic::x86_sse2_ucomieq_sd:
8771 Opc = X86ISD::UCOMI;
8774 case Intrinsic::x86_sse_ucomilt_ss:
8775 case Intrinsic::x86_sse2_ucomilt_sd:
8776 Opc = X86ISD::UCOMI;
8779 case Intrinsic::x86_sse_ucomile_ss:
8780 case Intrinsic::x86_sse2_ucomile_sd:
8781 Opc = X86ISD::UCOMI;
8784 case Intrinsic::x86_sse_ucomigt_ss:
8785 case Intrinsic::x86_sse2_ucomigt_sd:
8786 Opc = X86ISD::UCOMI;
8789 case Intrinsic::x86_sse_ucomige_ss:
8790 case Intrinsic::x86_sse2_ucomige_sd:
8791 Opc = X86ISD::UCOMI;
8794 case Intrinsic::x86_sse_ucomineq_ss:
8795 case Intrinsic::x86_sse2_ucomineq_sd:
8796 Opc = X86ISD::UCOMI;
8801 SDValue LHS = Op.getOperand(1);
8802 SDValue RHS = Op.getOperand(2);
8803 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
8804 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
8805 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
8806 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8807 DAG.getConstant(X86CC, MVT::i8), Cond);
8808 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8810 // ptest and testp intrinsics. The intrinsic these come from are designed to
8811 // return an integer value, not just an instruction so lower it to the ptest
8812 // or testp pattern and a setcc for the result.
8813 case Intrinsic::x86_sse41_ptestz:
8814 case Intrinsic::x86_sse41_ptestc:
8815 case Intrinsic::x86_sse41_ptestnzc:
8816 case Intrinsic::x86_avx_ptestz_256:
8817 case Intrinsic::x86_avx_ptestc_256:
8818 case Intrinsic::x86_avx_ptestnzc_256:
8819 case Intrinsic::x86_avx_vtestz_ps:
8820 case Intrinsic::x86_avx_vtestc_ps:
8821 case Intrinsic::x86_avx_vtestnzc_ps:
8822 case Intrinsic::x86_avx_vtestz_pd:
8823 case Intrinsic::x86_avx_vtestc_pd:
8824 case Intrinsic::x86_avx_vtestnzc_pd:
8825 case Intrinsic::x86_avx_vtestz_ps_256:
8826 case Intrinsic::x86_avx_vtestc_ps_256:
8827 case Intrinsic::x86_avx_vtestnzc_ps_256:
8828 case Intrinsic::x86_avx_vtestz_pd_256:
8829 case Intrinsic::x86_avx_vtestc_pd_256:
8830 case Intrinsic::x86_avx_vtestnzc_pd_256: {
8831 bool IsTestPacked = false;
8834 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
8835 case Intrinsic::x86_avx_vtestz_ps:
8836 case Intrinsic::x86_avx_vtestz_pd:
8837 case Intrinsic::x86_avx_vtestz_ps_256:
8838 case Intrinsic::x86_avx_vtestz_pd_256:
8839 IsTestPacked = true; // Fallthrough
8840 case Intrinsic::x86_sse41_ptestz:
8841 case Intrinsic::x86_avx_ptestz_256:
8843 X86CC = X86::COND_E;
8845 case Intrinsic::x86_avx_vtestc_ps:
8846 case Intrinsic::x86_avx_vtestc_pd:
8847 case Intrinsic::x86_avx_vtestc_ps_256:
8848 case Intrinsic::x86_avx_vtestc_pd_256:
8849 IsTestPacked = true; // Fallthrough
8850 case Intrinsic::x86_sse41_ptestc:
8851 case Intrinsic::x86_avx_ptestc_256:
8853 X86CC = X86::COND_B;
8855 case Intrinsic::x86_avx_vtestnzc_ps:
8856 case Intrinsic::x86_avx_vtestnzc_pd:
8857 case Intrinsic::x86_avx_vtestnzc_ps_256:
8858 case Intrinsic::x86_avx_vtestnzc_pd_256:
8859 IsTestPacked = true; // Fallthrough
8860 case Intrinsic::x86_sse41_ptestnzc:
8861 case Intrinsic::x86_avx_ptestnzc_256:
8863 X86CC = X86::COND_A;
8867 SDValue LHS = Op.getOperand(1);
8868 SDValue RHS = Op.getOperand(2);
8869 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
8870 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
8871 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
8872 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
8873 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
8876 // Fix vector shift instructions where the last operand is a non-immediate
8878 case Intrinsic::x86_sse2_pslli_w:
8879 case Intrinsic::x86_sse2_pslli_d:
8880 case Intrinsic::x86_sse2_pslli_q:
8881 case Intrinsic::x86_sse2_psrli_w:
8882 case Intrinsic::x86_sse2_psrli_d:
8883 case Intrinsic::x86_sse2_psrli_q:
8884 case Intrinsic::x86_sse2_psrai_w:
8885 case Intrinsic::x86_sse2_psrai_d:
8886 case Intrinsic::x86_mmx_pslli_w:
8887 case Intrinsic::x86_mmx_pslli_d:
8888 case Intrinsic::x86_mmx_pslli_q:
8889 case Intrinsic::x86_mmx_psrli_w:
8890 case Intrinsic::x86_mmx_psrli_d:
8891 case Intrinsic::x86_mmx_psrli_q:
8892 case Intrinsic::x86_mmx_psrai_w:
8893 case Intrinsic::x86_mmx_psrai_d: {
8894 SDValue ShAmt = Op.getOperand(2);
8895 if (isa<ConstantSDNode>(ShAmt))
8898 unsigned NewIntNo = 0;
8899 EVT ShAmtVT = MVT::v4i32;
8901 case Intrinsic::x86_sse2_pslli_w:
8902 NewIntNo = Intrinsic::x86_sse2_psll_w;
8904 case Intrinsic::x86_sse2_pslli_d:
8905 NewIntNo = Intrinsic::x86_sse2_psll_d;
8907 case Intrinsic::x86_sse2_pslli_q:
8908 NewIntNo = Intrinsic::x86_sse2_psll_q;
8910 case Intrinsic::x86_sse2_psrli_w:
8911 NewIntNo = Intrinsic::x86_sse2_psrl_w;
8913 case Intrinsic::x86_sse2_psrli_d:
8914 NewIntNo = Intrinsic::x86_sse2_psrl_d;
8916 case Intrinsic::x86_sse2_psrli_q:
8917 NewIntNo = Intrinsic::x86_sse2_psrl_q;
8919 case Intrinsic::x86_sse2_psrai_w:
8920 NewIntNo = Intrinsic::x86_sse2_psra_w;
8922 case Intrinsic::x86_sse2_psrai_d:
8923 NewIntNo = Intrinsic::x86_sse2_psra_d;
8926 ShAmtVT = MVT::v2i32;
8928 case Intrinsic::x86_mmx_pslli_w:
8929 NewIntNo = Intrinsic::x86_mmx_psll_w;
8931 case Intrinsic::x86_mmx_pslli_d:
8932 NewIntNo = Intrinsic::x86_mmx_psll_d;
8934 case Intrinsic::x86_mmx_pslli_q:
8935 NewIntNo = Intrinsic::x86_mmx_psll_q;
8937 case Intrinsic::x86_mmx_psrli_w:
8938 NewIntNo = Intrinsic::x86_mmx_psrl_w;
8940 case Intrinsic::x86_mmx_psrli_d:
8941 NewIntNo = Intrinsic::x86_mmx_psrl_d;
8943 case Intrinsic::x86_mmx_psrli_q:
8944 NewIntNo = Intrinsic::x86_mmx_psrl_q;
8946 case Intrinsic::x86_mmx_psrai_w:
8947 NewIntNo = Intrinsic::x86_mmx_psra_w;
8949 case Intrinsic::x86_mmx_psrai_d:
8950 NewIntNo = Intrinsic::x86_mmx_psra_d;
8952 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
8958 // The vector shift intrinsics with scalars uses 32b shift amounts but
8959 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
8963 ShOps[1] = DAG.getConstant(0, MVT::i32);
8964 if (ShAmtVT == MVT::v4i32) {
8965 ShOps[2] = DAG.getUNDEF(MVT::i32);
8966 ShOps[3] = DAG.getUNDEF(MVT::i32);
8967 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
8969 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
8970 // FIXME this must be lowered to get rid of the invalid type.
8973 EVT VT = Op.getValueType();
8974 ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
8975 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
8976 DAG.getConstant(NewIntNo, MVT::i32),
8977 Op.getOperand(1), ShAmt);
8982 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
8983 SelectionDAG &DAG) const {
8984 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8985 MFI->setReturnAddressIsTaken(true);
8987 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
8988 DebugLoc dl = Op.getDebugLoc();
8991 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
8993 DAG.getConstant(TD->getPointerSize(),
8994 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
8995 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
8996 DAG.getNode(ISD::ADD, dl, getPointerTy(),
8998 MachinePointerInfo(), false, false, 0);
9001 // Just load the return address.
9002 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
9003 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
9004 RetAddrFI, MachinePointerInfo(), false, false, 0);
9007 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
9008 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
9009 MFI->setFrameAddressIsTaken(true);
9011 EVT VT = Op.getValueType();
9012 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
9013 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9014 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
9015 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
9017 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
9018 MachinePointerInfo(),
9023 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
9024 SelectionDAG &DAG) const {
9025 return DAG.getIntPtrConstant(2*TD->getPointerSize());
9028 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
9029 MachineFunction &MF = DAG.getMachineFunction();
9030 SDValue Chain = Op.getOperand(0);
9031 SDValue Offset = Op.getOperand(1);
9032 SDValue Handler = Op.getOperand(2);
9033 DebugLoc dl = Op.getDebugLoc();
9035 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
9036 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
9038 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
9040 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
9041 DAG.getIntPtrConstant(TD->getPointerSize()));
9042 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
9043 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
9045 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
9046 MF.getRegInfo().addLiveOut(StoreAddrReg);
9048 return DAG.getNode(X86ISD::EH_RETURN, dl,
9050 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
9053 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
9054 SelectionDAG &DAG) const {
9055 SDValue Root = Op.getOperand(0);
9056 SDValue Trmp = Op.getOperand(1); // trampoline
9057 SDValue FPtr = Op.getOperand(2); // nested function
9058 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
9059 DebugLoc dl = Op.getDebugLoc();
9061 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
9063 if (Subtarget->is64Bit()) {
9064 SDValue OutChains[6];
9066 // Large code-model.
9067 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
9068 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
9070 const unsigned char N86R10 = X86_MC::getX86RegNum(X86::R10);
9071 const unsigned char N86R11 = X86_MC::getX86RegNum(X86::R11);
9073 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
9075 // Load the pointer to the nested function into R11.
9076 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
9077 SDValue Addr = Trmp;
9078 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
9079 Addr, MachinePointerInfo(TrmpAddr),
9082 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9083 DAG.getConstant(2, MVT::i64));
9084 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
9085 MachinePointerInfo(TrmpAddr, 2),
9088 // Load the 'nest' parameter value into R10.
9089 // R10 is specified in X86CallingConv.td
9090 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
9091 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9092 DAG.getConstant(10, MVT::i64));
9093 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
9094 Addr, MachinePointerInfo(TrmpAddr, 10),
9097 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9098 DAG.getConstant(12, MVT::i64));
9099 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
9100 MachinePointerInfo(TrmpAddr, 12),
9103 // Jump to the nested function.
9104 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
9105 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9106 DAG.getConstant(20, MVT::i64));
9107 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
9108 Addr, MachinePointerInfo(TrmpAddr, 20),
9111 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
9112 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9113 DAG.getConstant(22, MVT::i64));
9114 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
9115 MachinePointerInfo(TrmpAddr, 22),
9119 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
9120 return DAG.getMergeValues(Ops, 2, dl);
9122 const Function *Func =
9123 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
9124 CallingConv::ID CC = Func->getCallingConv();
9129 llvm_unreachable("Unsupported calling convention");
9130 case CallingConv::C:
9131 case CallingConv::X86_StdCall: {
9132 // Pass 'nest' parameter in ECX.
9133 // Must be kept in sync with X86CallingConv.td
9136 // Check that ECX wasn't needed by an 'inreg' parameter.
9137 FunctionType *FTy = Func->getFunctionType();
9138 const AttrListPtr &Attrs = Func->getAttributes();
9140 if (!Attrs.isEmpty() && !Func->isVarArg()) {
9141 unsigned InRegCount = 0;
9144 for (FunctionType::param_iterator I = FTy->param_begin(),
9145 E = FTy->param_end(); I != E; ++I, ++Idx)
9146 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
9147 // FIXME: should only count parameters that are lowered to integers.
9148 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
9150 if (InRegCount > 2) {
9151 report_fatal_error("Nest register in use - reduce number of inreg"
9157 case CallingConv::X86_FastCall:
9158 case CallingConv::X86_ThisCall:
9159 case CallingConv::Fast:
9160 // Pass 'nest' parameter in EAX.
9161 // Must be kept in sync with X86CallingConv.td
9166 SDValue OutChains[4];
9169 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9170 DAG.getConstant(10, MVT::i32));
9171 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
9173 // This is storing the opcode for MOV32ri.
9174 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
9175 const unsigned char N86Reg = X86_MC::getX86RegNum(NestReg);
9176 OutChains[0] = DAG.getStore(Root, dl,
9177 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
9178 Trmp, MachinePointerInfo(TrmpAddr),
9181 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9182 DAG.getConstant(1, MVT::i32));
9183 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
9184 MachinePointerInfo(TrmpAddr, 1),
9187 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
9188 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9189 DAG.getConstant(5, MVT::i32));
9190 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
9191 MachinePointerInfo(TrmpAddr, 5),
9194 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9195 DAG.getConstant(6, MVT::i32));
9196 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
9197 MachinePointerInfo(TrmpAddr, 6),
9201 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
9202 return DAG.getMergeValues(Ops, 2, dl);
9206 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
9207 SelectionDAG &DAG) const {
9209 The rounding mode is in bits 11:10 of FPSR, and has the following
9216 FLT_ROUNDS, on the other hand, expects the following:
9223 To perform the conversion, we do:
9224 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
9227 MachineFunction &MF = DAG.getMachineFunction();
9228 const TargetMachine &TM = MF.getTarget();
9229 const TargetFrameLowering &TFI = *TM.getFrameLowering();
9230 unsigned StackAlignment = TFI.getStackAlignment();
9231 EVT VT = Op.getValueType();
9232 DebugLoc DL = Op.getDebugLoc();
9234 // Save FP Control Word to stack slot
9235 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
9236 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
9239 MachineMemOperand *MMO =
9240 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
9241 MachineMemOperand::MOStore, 2, 2);
9243 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
9244 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
9245 DAG.getVTList(MVT::Other),
9246 Ops, 2, MVT::i16, MMO);
9248 // Load FP Control Word from stack slot
9249 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
9250 MachinePointerInfo(), false, false, 0);
9252 // Transform as necessary
9254 DAG.getNode(ISD::SRL, DL, MVT::i16,
9255 DAG.getNode(ISD::AND, DL, MVT::i16,
9256 CWD, DAG.getConstant(0x800, MVT::i16)),
9257 DAG.getConstant(11, MVT::i8));
9259 DAG.getNode(ISD::SRL, DL, MVT::i16,
9260 DAG.getNode(ISD::AND, DL, MVT::i16,
9261 CWD, DAG.getConstant(0x400, MVT::i16)),
9262 DAG.getConstant(9, MVT::i8));
9265 DAG.getNode(ISD::AND, DL, MVT::i16,
9266 DAG.getNode(ISD::ADD, DL, MVT::i16,
9267 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
9268 DAG.getConstant(1, MVT::i16)),
9269 DAG.getConstant(3, MVT::i16));
9272 return DAG.getNode((VT.getSizeInBits() < 16 ?
9273 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
9276 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
9277 EVT VT = Op.getValueType();
9279 unsigned NumBits = VT.getSizeInBits();
9280 DebugLoc dl = Op.getDebugLoc();
9282 Op = Op.getOperand(0);
9283 if (VT == MVT::i8) {
9284 // Zero extend to i32 since there is not an i8 bsr.
9286 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
9289 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
9290 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
9291 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
9293 // If src is zero (i.e. bsr sets ZF), returns NumBits.
9296 DAG.getConstant(NumBits+NumBits-1, OpVT),
9297 DAG.getConstant(X86::COND_E, MVT::i8),
9300 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
9302 // Finally xor with NumBits-1.
9303 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
9306 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
9310 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
9311 EVT VT = Op.getValueType();
9313 unsigned NumBits = VT.getSizeInBits();
9314 DebugLoc dl = Op.getDebugLoc();
9316 Op = Op.getOperand(0);
9317 if (VT == MVT::i8) {
9319 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
9322 // Issue a bsf (scan bits forward) which also sets EFLAGS.
9323 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
9324 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
9326 // If src is zero (i.e. bsf sets ZF), returns NumBits.
9329 DAG.getConstant(NumBits, OpVT),
9330 DAG.getConstant(X86::COND_E, MVT::i8),
9333 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
9336 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
9340 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) const {
9341 EVT VT = Op.getValueType();
9342 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
9343 DebugLoc dl = Op.getDebugLoc();
9345 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
9346 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
9347 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
9348 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
9349 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
9351 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
9352 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
9353 // return AloBlo + AloBhi + AhiBlo;
9355 SDValue A = Op.getOperand(0);
9356 SDValue B = Op.getOperand(1);
9358 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9359 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9360 A, DAG.getConstant(32, MVT::i32));
9361 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9362 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9363 B, DAG.getConstant(32, MVT::i32));
9364 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9365 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9367 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9368 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9370 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9371 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9373 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9374 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9375 AloBhi, DAG.getConstant(32, MVT::i32));
9376 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9377 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9378 AhiBlo, DAG.getConstant(32, MVT::i32));
9379 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
9380 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
9384 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
9386 EVT VT = Op.getValueType();
9387 DebugLoc dl = Op.getDebugLoc();
9388 SDValue R = Op.getOperand(0);
9389 SDValue Amt = Op.getOperand(1);
9390 LLVMContext *Context = DAG.getContext();
9392 if (!(Subtarget->hasSSE2() || Subtarget->hasAVX()))
9395 // Decompose 256-bit shifts into smaller 128-bit shifts.
9396 if (VT.getSizeInBits() == 256) {
9397 int NumElems = VT.getVectorNumElements();
9398 MVT EltVT = VT.getVectorElementType().getSimpleVT();
9399 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
9401 // Extract the two vectors
9402 SDValue V1 = Extract128BitVector(R, DAG.getConstant(0, MVT::i32), DAG, dl);
9403 SDValue V2 = Extract128BitVector(R, DAG.getConstant(NumElems/2, MVT::i32),
9406 // Recreate the shift amount vectors
9407 SmallVector<SDValue, 4> Amt1Csts;
9408 SmallVector<SDValue, 4> Amt2Csts;
9409 for (int i = 0; i < NumElems/2; ++i)
9410 Amt1Csts.push_back(Amt->getOperand(i));
9411 for (int i = NumElems/2; i < NumElems; ++i)
9412 Amt2Csts.push_back(Amt->getOperand(i));
9414 SDValue Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
9415 &Amt1Csts[0], NumElems/2);
9416 SDValue Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
9417 &Amt2Csts[0], NumElems/2);
9419 // Issue new vector shifts for the smaller types
9420 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
9421 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
9423 // Concatenate the result back
9424 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
9427 // Optimize shl/srl/sra with constant shift amount.
9428 if (isSplatVector(Amt.getNode())) {
9429 SDValue SclrAmt = Amt->getOperand(0);
9430 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
9431 uint64_t ShiftAmt = C->getZExtValue();
9433 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SHL)
9434 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9435 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9436 R, DAG.getConstant(ShiftAmt, MVT::i32));
9438 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SHL)
9439 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9440 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9441 R, DAG.getConstant(ShiftAmt, MVT::i32));
9443 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SHL)
9444 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9445 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9446 R, DAG.getConstant(ShiftAmt, MVT::i32));
9448 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SRL)
9449 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9450 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9451 R, DAG.getConstant(ShiftAmt, MVT::i32));
9453 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRL)
9454 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9455 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9456 R, DAG.getConstant(ShiftAmt, MVT::i32));
9458 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRL)
9459 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9460 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9461 R, DAG.getConstant(ShiftAmt, MVT::i32));
9463 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRA)
9464 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9465 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9466 R, DAG.getConstant(ShiftAmt, MVT::i32));
9468 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRA)
9469 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9470 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9471 R, DAG.getConstant(ShiftAmt, MVT::i32));
9475 // Lower SHL with variable shift amount.
9476 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
9477 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9478 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9479 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
9481 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
9483 std::vector<Constant*> CV(4, CI);
9484 Constant *C = ConstantVector::get(CV);
9485 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9486 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9487 MachinePointerInfo::getConstantPool(),
9490 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
9491 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
9492 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
9493 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
9495 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
9497 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9498 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9499 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
9501 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
9502 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
9504 std::vector<Constant*> CVM1(16, CM1);
9505 std::vector<Constant*> CVM2(16, CM2);
9506 Constant *C = ConstantVector::get(CVM1);
9507 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9508 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9509 MachinePointerInfo::getConstantPool(),
9512 // r = pblendv(r, psllw(r & (char16)15, 4), a);
9513 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9514 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9515 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9516 DAG.getConstant(4, MVT::i32));
9517 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9519 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9521 C = ConstantVector::get(CVM2);
9522 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9523 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9524 MachinePointerInfo::getConstantPool(),
9527 // r = pblendv(r, psllw(r & (char16)63, 2), a);
9528 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9529 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9530 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9531 DAG.getConstant(2, MVT::i32));
9532 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9534 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9536 // return pblendv(r, r+r, a);
9537 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
9538 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
9544 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
9545 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
9546 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
9547 // looks for this combo and may remove the "setcc" instruction if the "setcc"
9548 // has only one use.
9549 SDNode *N = Op.getNode();
9550 SDValue LHS = N->getOperand(0);
9551 SDValue RHS = N->getOperand(1);
9552 unsigned BaseOp = 0;
9554 DebugLoc DL = Op.getDebugLoc();
9555 switch (Op.getOpcode()) {
9556 default: llvm_unreachable("Unknown ovf instruction!");
9558 // A subtract of one will be selected as a INC. Note that INC doesn't
9559 // set CF, so we can't do this for UADDO.
9560 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9562 BaseOp = X86ISD::INC;
9566 BaseOp = X86ISD::ADD;
9570 BaseOp = X86ISD::ADD;
9574 // A subtract of one will be selected as a DEC. Note that DEC doesn't
9575 // set CF, so we can't do this for USUBO.
9576 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9578 BaseOp = X86ISD::DEC;
9582 BaseOp = X86ISD::SUB;
9586 BaseOp = X86ISD::SUB;
9590 BaseOp = X86ISD::SMUL;
9593 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
9594 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
9596 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
9599 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9600 DAG.getConstant(X86::COND_O, MVT::i32),
9601 SDValue(Sum.getNode(), 2));
9603 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
9607 // Also sets EFLAGS.
9608 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
9609 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
9612 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
9613 DAG.getConstant(Cond, MVT::i32),
9614 SDValue(Sum.getNode(), 1));
9616 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
9619 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const{
9620 DebugLoc dl = Op.getDebugLoc();
9621 SDNode* Node = Op.getNode();
9622 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
9623 EVT VT = Node->getValueType(0);
9625 if (Subtarget->hasSSE2() && VT.isVector()) {
9626 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
9627 ExtraVT.getScalarType().getSizeInBits();
9628 SDValue ShAmt = DAG.getConstant(BitsDiff, MVT::i32);
9630 unsigned SHLIntrinsicsID = 0;
9631 unsigned SRAIntrinsicsID = 0;
9632 switch (VT.getSimpleVT().SimpleTy) {
9636 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_q;
9637 SRAIntrinsicsID = 0;
9641 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_d;
9642 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_d;
9646 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_w;
9647 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_w;
9652 SDValue Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9653 DAG.getConstant(SHLIntrinsicsID, MVT::i32),
9654 Node->getOperand(0), ShAmt);
9656 // In case of 1 bit sext, no need to shr
9657 if (ExtraVT.getScalarType().getSizeInBits() == 1) return Tmp1;
9659 if (SRAIntrinsicsID) {
9660 Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9661 DAG.getConstant(SRAIntrinsicsID, MVT::i32),
9671 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
9672 DebugLoc dl = Op.getDebugLoc();
9674 // Go ahead and emit the fence on x86-64 even if we asked for no-sse2.
9675 // There isn't any reason to disable it if the target processor supports it.
9676 if (!Subtarget->hasSSE2() && !Subtarget->is64Bit()) {
9677 SDValue Chain = Op.getOperand(0);
9678 SDValue Zero = DAG.getConstant(0, MVT::i32);
9680 DAG.getRegister(X86::ESP, MVT::i32), // Base
9681 DAG.getTargetConstant(1, MVT::i8), // Scale
9682 DAG.getRegister(0, MVT::i32), // Index
9683 DAG.getTargetConstant(0, MVT::i32), // Disp
9684 DAG.getRegister(0, MVT::i32), // Segment.
9689 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
9690 array_lengthof(Ops));
9691 return SDValue(Res, 0);
9694 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
9696 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
9698 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
9699 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
9700 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
9701 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
9703 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
9704 if (!Op1 && !Op2 && !Op3 && Op4)
9705 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
9707 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
9708 if (Op1 && !Op2 && !Op3 && !Op4)
9709 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
9711 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
9713 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
9716 SDValue X86TargetLowering::LowerATOMIC_FENCE(SDValue Op,
9717 SelectionDAG &DAG) const {
9718 DebugLoc dl = Op.getDebugLoc();
9719 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
9720 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
9721 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
9722 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
9724 // The only fence that needs an instruction is a sequentially-consistent
9725 // cross-thread fence.
9726 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
9727 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
9728 // no-sse2). There isn't any reason to disable it if the target processor
9730 if (Subtarget->hasSSE2() || Subtarget->is64Bit())
9731 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
9733 SDValue Chain = Op.getOperand(0);
9734 SDValue Zero = DAG.getConstant(0, MVT::i32);
9736 DAG.getRegister(X86::ESP, MVT::i32), // Base
9737 DAG.getTargetConstant(1, MVT::i8), // Scale
9738 DAG.getRegister(0, MVT::i32), // Index
9739 DAG.getTargetConstant(0, MVT::i32), // Disp
9740 DAG.getRegister(0, MVT::i32), // Segment.
9745 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
9746 array_lengthof(Ops));
9747 return SDValue(Res, 0);
9750 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
9751 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
9755 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
9756 EVT T = Op.getValueType();
9757 DebugLoc DL = Op.getDebugLoc();
9760 switch(T.getSimpleVT().SimpleTy) {
9762 assert(false && "Invalid value type!");
9763 case MVT::i8: Reg = X86::AL; size = 1; break;
9764 case MVT::i16: Reg = X86::AX; size = 2; break;
9765 case MVT::i32: Reg = X86::EAX; size = 4; break;
9767 assert(Subtarget->is64Bit() && "Node not type legal!");
9768 Reg = X86::RAX; size = 8;
9771 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
9772 Op.getOperand(2), SDValue());
9773 SDValue Ops[] = { cpIn.getValue(0),
9776 DAG.getTargetConstant(size, MVT::i8),
9778 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9779 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
9780 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
9783 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
9787 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
9788 SelectionDAG &DAG) const {
9789 assert(Subtarget->is64Bit() && "Result not type legalized?");
9790 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9791 SDValue TheChain = Op.getOperand(0);
9792 DebugLoc dl = Op.getDebugLoc();
9793 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9794 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
9795 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
9797 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
9798 DAG.getConstant(32, MVT::i8));
9800 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
9803 return DAG.getMergeValues(Ops, 2, dl);
9806 SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
9807 SelectionDAG &DAG) const {
9808 EVT SrcVT = Op.getOperand(0).getValueType();
9809 EVT DstVT = Op.getValueType();
9810 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
9811 Subtarget->hasMMX() && "Unexpected custom BITCAST");
9812 assert((DstVT == MVT::i64 ||
9813 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
9814 "Unexpected custom BITCAST");
9815 // i64 <=> MMX conversions are Legal.
9816 if (SrcVT==MVT::i64 && DstVT.isVector())
9818 if (DstVT==MVT::i64 && SrcVT.isVector())
9820 // MMX <=> MMX conversions are Legal.
9821 if (SrcVT.isVector() && DstVT.isVector())
9823 // All other conversions need to be expanded.
9827 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
9828 SDNode *Node = Op.getNode();
9829 DebugLoc dl = Node->getDebugLoc();
9830 EVT T = Node->getValueType(0);
9831 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
9832 DAG.getConstant(0, T), Node->getOperand(2));
9833 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
9834 cast<AtomicSDNode>(Node)->getMemoryVT(),
9835 Node->getOperand(0),
9836 Node->getOperand(1), negOp,
9837 cast<AtomicSDNode>(Node)->getSrcValue(),
9838 cast<AtomicSDNode>(Node)->getAlignment(),
9839 cast<AtomicSDNode>(Node)->getOrdering(),
9840 cast<AtomicSDNode>(Node)->getSynchScope());
9843 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
9844 EVT VT = Op.getNode()->getValueType(0);
9846 // Let legalize expand this if it isn't a legal type yet.
9847 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
9850 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
9853 bool ExtraOp = false;
9854 switch (Op.getOpcode()) {
9855 default: assert(0 && "Invalid code");
9856 case ISD::ADDC: Opc = X86ISD::ADD; break;
9857 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
9858 case ISD::SUBC: Opc = X86ISD::SUB; break;
9859 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
9863 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9865 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
9866 Op.getOperand(1), Op.getOperand(2));
9869 /// LowerOperation - Provide custom lowering hooks for some operations.
9871 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
9872 switch (Op.getOpcode()) {
9873 default: llvm_unreachable("Should not custom lower this!");
9874 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
9875 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
9876 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op,DAG);
9877 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
9878 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
9879 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
9880 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
9881 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
9882 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
9883 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
9884 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
9885 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, DAG);
9886 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
9887 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
9888 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
9889 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
9890 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
9891 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
9892 case ISD::SHL_PARTS:
9893 case ISD::SRA_PARTS:
9894 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
9895 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
9896 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
9897 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
9898 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
9899 case ISD::FABS: return LowerFABS(Op, DAG);
9900 case ISD::FNEG: return LowerFNEG(Op, DAG);
9901 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
9902 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
9903 case ISD::SETCC: return LowerSETCC(Op, DAG);
9904 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
9905 case ISD::SELECT: return LowerSELECT(Op, DAG);
9906 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
9907 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
9908 case ISD::VASTART: return LowerVASTART(Op, DAG);
9909 case ISD::VAARG: return LowerVAARG(Op, DAG);
9910 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
9911 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
9912 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
9913 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
9914 case ISD::FRAME_TO_ARGS_OFFSET:
9915 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
9916 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
9917 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
9918 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
9919 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
9920 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
9921 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
9922 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
9925 case ISD::SHL: return LowerShift(Op, DAG);
9931 case ISD::UMULO: return LowerXALUO(Op, DAG);
9932 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
9933 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
9937 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
9941 void X86TargetLowering::
9942 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
9943 SelectionDAG &DAG, unsigned NewOp) const {
9944 EVT T = Node->getValueType(0);
9945 DebugLoc dl = Node->getDebugLoc();
9946 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
9948 SDValue Chain = Node->getOperand(0);
9949 SDValue In1 = Node->getOperand(1);
9950 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9951 Node->getOperand(2), DAG.getIntPtrConstant(0));
9952 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
9953 Node->getOperand(2), DAG.getIntPtrConstant(1));
9954 SDValue Ops[] = { Chain, In1, In2L, In2H };
9955 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
9957 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
9958 cast<MemSDNode>(Node)->getMemOperand());
9959 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
9960 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
9961 Results.push_back(Result.getValue(2));
9964 /// ReplaceNodeResults - Replace a node with an illegal result type
9965 /// with a new node built out of custom code.
9966 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
9967 SmallVectorImpl<SDValue>&Results,
9968 SelectionDAG &DAG) const {
9969 DebugLoc dl = N->getDebugLoc();
9970 switch (N->getOpcode()) {
9972 assert(false && "Do not know how to custom type legalize this operation!");
9974 case ISD::SIGN_EXTEND_INREG:
9979 // We don't want to expand or promote these.
9981 case ISD::FP_TO_SINT: {
9982 std::pair<SDValue,SDValue> Vals =
9983 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
9984 SDValue FIST = Vals.first, StackSlot = Vals.second;
9985 if (FIST.getNode() != 0) {
9986 EVT VT = N->getValueType(0);
9987 // Return a load from the stack slot.
9988 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
9989 MachinePointerInfo(), false, false, 0));
9993 case ISD::READCYCLECOUNTER: {
9994 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
9995 SDValue TheChain = N->getOperand(0);
9996 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
9997 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
9999 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
10001 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
10002 SDValue Ops[] = { eax, edx };
10003 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
10004 Results.push_back(edx.getValue(1));
10007 case ISD::ATOMIC_CMP_SWAP: {
10008 EVT T = N->getValueType(0);
10009 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
10010 SDValue cpInL, cpInH;
10011 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
10012 DAG.getConstant(0, MVT::i32));
10013 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
10014 DAG.getConstant(1, MVT::i32));
10015 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
10016 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
10017 cpInL.getValue(1));
10018 SDValue swapInL, swapInH;
10019 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
10020 DAG.getConstant(0, MVT::i32));
10021 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
10022 DAG.getConstant(1, MVT::i32));
10023 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
10024 cpInH.getValue(1));
10025 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
10026 swapInL.getValue(1));
10027 SDValue Ops[] = { swapInH.getValue(0),
10029 swapInH.getValue(1) };
10030 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
10031 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
10032 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
10034 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
10035 MVT::i32, Result.getValue(1));
10036 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
10037 MVT::i32, cpOutL.getValue(2));
10038 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
10039 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
10040 Results.push_back(cpOutH.getValue(1));
10043 case ISD::ATOMIC_LOAD_ADD:
10044 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
10046 case ISD::ATOMIC_LOAD_AND:
10047 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
10049 case ISD::ATOMIC_LOAD_NAND:
10050 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
10052 case ISD::ATOMIC_LOAD_OR:
10053 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
10055 case ISD::ATOMIC_LOAD_SUB:
10056 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
10058 case ISD::ATOMIC_LOAD_XOR:
10059 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
10061 case ISD::ATOMIC_SWAP:
10062 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
10067 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
10069 default: return NULL;
10070 case X86ISD::BSF: return "X86ISD::BSF";
10071 case X86ISD::BSR: return "X86ISD::BSR";
10072 case X86ISD::SHLD: return "X86ISD::SHLD";
10073 case X86ISD::SHRD: return "X86ISD::SHRD";
10074 case X86ISD::FAND: return "X86ISD::FAND";
10075 case X86ISD::FOR: return "X86ISD::FOR";
10076 case X86ISD::FXOR: return "X86ISD::FXOR";
10077 case X86ISD::FSRL: return "X86ISD::FSRL";
10078 case X86ISD::FILD: return "X86ISD::FILD";
10079 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
10080 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
10081 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
10082 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
10083 case X86ISD::FLD: return "X86ISD::FLD";
10084 case X86ISD::FST: return "X86ISD::FST";
10085 case X86ISD::CALL: return "X86ISD::CALL";
10086 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
10087 case X86ISD::BT: return "X86ISD::BT";
10088 case X86ISD::CMP: return "X86ISD::CMP";
10089 case X86ISD::COMI: return "X86ISD::COMI";
10090 case X86ISD::UCOMI: return "X86ISD::UCOMI";
10091 case X86ISD::SETCC: return "X86ISD::SETCC";
10092 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
10093 case X86ISD::FSETCCsd: return "X86ISD::FSETCCsd";
10094 case X86ISD::FSETCCss: return "X86ISD::FSETCCss";
10095 case X86ISD::CMOV: return "X86ISD::CMOV";
10096 case X86ISD::BRCOND: return "X86ISD::BRCOND";
10097 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
10098 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
10099 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
10100 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
10101 case X86ISD::Wrapper: return "X86ISD::Wrapper";
10102 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
10103 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
10104 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
10105 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
10106 case X86ISD::PINSRB: return "X86ISD::PINSRB";
10107 case X86ISD::PINSRW: return "X86ISD::PINSRW";
10108 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
10109 case X86ISD::ANDNP: return "X86ISD::ANDNP";
10110 case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
10111 case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
10112 case X86ISD::PSIGND: return "X86ISD::PSIGND";
10113 case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
10114 case X86ISD::FMAX: return "X86ISD::FMAX";
10115 case X86ISD::FMIN: return "X86ISD::FMIN";
10116 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
10117 case X86ISD::FRCP: return "X86ISD::FRCP";
10118 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
10119 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
10120 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
10121 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
10122 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
10123 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
10124 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
10125 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
10126 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
10127 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
10128 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
10129 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
10130 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
10131 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
10132 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
10133 case X86ISD::VSHL: return "X86ISD::VSHL";
10134 case X86ISD::VSRL: return "X86ISD::VSRL";
10135 case X86ISD::CMPPD: return "X86ISD::CMPPD";
10136 case X86ISD::CMPPS: return "X86ISD::CMPPS";
10137 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
10138 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
10139 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
10140 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
10141 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
10142 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
10143 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
10144 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
10145 case X86ISD::ADD: return "X86ISD::ADD";
10146 case X86ISD::SUB: return "X86ISD::SUB";
10147 case X86ISD::ADC: return "X86ISD::ADC";
10148 case X86ISD::SBB: return "X86ISD::SBB";
10149 case X86ISD::SMUL: return "X86ISD::SMUL";
10150 case X86ISD::UMUL: return "X86ISD::UMUL";
10151 case X86ISD::INC: return "X86ISD::INC";
10152 case X86ISD::DEC: return "X86ISD::DEC";
10153 case X86ISD::OR: return "X86ISD::OR";
10154 case X86ISD::XOR: return "X86ISD::XOR";
10155 case X86ISD::AND: return "X86ISD::AND";
10156 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
10157 case X86ISD::PTEST: return "X86ISD::PTEST";
10158 case X86ISD::TESTP: return "X86ISD::TESTP";
10159 case X86ISD::PALIGN: return "X86ISD::PALIGN";
10160 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
10161 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
10162 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
10163 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
10164 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
10165 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
10166 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
10167 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
10168 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
10169 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
10170 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
10171 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
10172 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
10173 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
10174 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
10175 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
10176 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
10177 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
10178 case X86ISD::MOVSD: return "X86ISD::MOVSD";
10179 case X86ISD::MOVSS: return "X86ISD::MOVSS";
10180 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
10181 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
10182 case X86ISD::VUNPCKLPDY: return "X86ISD::VUNPCKLPDY";
10183 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
10184 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
10185 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
10186 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
10187 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
10188 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
10189 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
10190 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
10191 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
10192 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
10193 case X86ISD::VPERMILPS: return "X86ISD::VPERMILPS";
10194 case X86ISD::VPERMILPSY: return "X86ISD::VPERMILPSY";
10195 case X86ISD::VPERMILPD: return "X86ISD::VPERMILPD";
10196 case X86ISD::VPERMILPDY: return "X86ISD::VPERMILPDY";
10197 case X86ISD::VPERM2F128: return "X86ISD::VPERM2F128";
10198 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
10199 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
10200 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
10201 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
10205 // isLegalAddressingMode - Return true if the addressing mode represented
10206 // by AM is legal for this target, for a load/store of the specified type.
10207 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
10209 // X86 supports extremely general addressing modes.
10210 CodeModel::Model M = getTargetMachine().getCodeModel();
10211 Reloc::Model R = getTargetMachine().getRelocationModel();
10213 // X86 allows a sign-extended 32-bit immediate field as a displacement.
10214 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
10219 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
10221 // If a reference to this global requires an extra load, we can't fold it.
10222 if (isGlobalStubReference(GVFlags))
10225 // If BaseGV requires a register for the PIC base, we cannot also have a
10226 // BaseReg specified.
10227 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
10230 // If lower 4G is not available, then we must use rip-relative addressing.
10231 if ((M != CodeModel::Small || R != Reloc::Static) &&
10232 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
10236 switch (AM.Scale) {
10242 // These scales always work.
10247 // These scales are formed with basereg+scalereg. Only accept if there is
10252 default: // Other stuff never works.
10260 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
10261 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10263 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
10264 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
10265 if (NumBits1 <= NumBits2)
10270 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
10271 if (!VT1.isInteger() || !VT2.isInteger())
10273 unsigned NumBits1 = VT1.getSizeInBits();
10274 unsigned NumBits2 = VT2.getSizeInBits();
10275 if (NumBits1 <= NumBits2)
10280 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
10281 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
10282 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
10285 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
10286 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
10287 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
10290 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
10291 // i16 instructions are longer (0x66 prefix) and potentially slower.
10292 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
10295 /// isShuffleMaskLegal - Targets can use this to indicate that they only
10296 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
10297 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
10298 /// are assumed to be legal.
10300 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
10302 // Very little shuffling can be done for 64-bit vectors right now.
10303 if (VT.getSizeInBits() == 64)
10304 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
10306 // FIXME: pshufb, blends, shifts.
10307 return (VT.getVectorNumElements() == 2 ||
10308 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
10309 isMOVLMask(M, VT) ||
10310 isSHUFPMask(M, VT) ||
10311 isPSHUFDMask(M, VT) ||
10312 isPSHUFHWMask(M, VT) ||
10313 isPSHUFLWMask(M, VT) ||
10314 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
10315 isUNPCKLMask(M, VT) ||
10316 isUNPCKHMask(M, VT) ||
10317 isUNPCKL_v_undef_Mask(M, VT) ||
10318 isUNPCKH_v_undef_Mask(M, VT));
10322 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
10324 unsigned NumElts = VT.getVectorNumElements();
10325 // FIXME: This collection of masks seems suspect.
10328 if (NumElts == 4 && VT.getSizeInBits() == 128) {
10329 return (isMOVLMask(Mask, VT) ||
10330 isCommutedMOVLMask(Mask, VT, true) ||
10331 isSHUFPMask(Mask, VT) ||
10332 isCommutedSHUFPMask(Mask, VT));
10337 //===----------------------------------------------------------------------===//
10338 // X86 Scheduler Hooks
10339 //===----------------------------------------------------------------------===//
10341 // private utility function
10342 MachineBasicBlock *
10343 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
10344 MachineBasicBlock *MBB,
10351 TargetRegisterClass *RC,
10352 bool invSrc) const {
10353 // For the atomic bitwise operator, we generate
10356 // ld t1 = [bitinstr.addr]
10357 // op t2 = t1, [bitinstr.val]
10359 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
10361 // fallthrough -->nextMBB
10362 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10363 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10364 MachineFunction::iterator MBBIter = MBB;
10367 /// First build the CFG
10368 MachineFunction *F = MBB->getParent();
10369 MachineBasicBlock *thisMBB = MBB;
10370 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10371 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10372 F->insert(MBBIter, newMBB);
10373 F->insert(MBBIter, nextMBB);
10375 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10376 nextMBB->splice(nextMBB->begin(), thisMBB,
10377 llvm::next(MachineBasicBlock::iterator(bInstr)),
10379 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10381 // Update thisMBB to fall through to newMBB
10382 thisMBB->addSuccessor(newMBB);
10384 // newMBB jumps to itself and fall through to nextMBB
10385 newMBB->addSuccessor(nextMBB);
10386 newMBB->addSuccessor(newMBB);
10388 // Insert instructions into newMBB based on incoming instruction
10389 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
10390 "unexpected number of operands");
10391 DebugLoc dl = bInstr->getDebugLoc();
10392 MachineOperand& destOper = bInstr->getOperand(0);
10393 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10394 int numArgs = bInstr->getNumOperands() - 1;
10395 for (int i=0; i < numArgs; ++i)
10396 argOpers[i] = &bInstr->getOperand(i+1);
10398 // x86 address has 4 operands: base, index, scale, and displacement
10399 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10400 int valArgIndx = lastAddrIndx + 1;
10402 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
10403 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
10404 for (int i=0; i <= lastAddrIndx; ++i)
10405 (*MIB).addOperand(*argOpers[i]);
10407 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
10409 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
10414 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
10415 assert((argOpers[valArgIndx]->isReg() ||
10416 argOpers[valArgIndx]->isImm()) &&
10417 "invalid operand");
10418 if (argOpers[valArgIndx]->isReg())
10419 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
10421 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
10423 (*MIB).addOperand(*argOpers[valArgIndx]);
10425 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
10428 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
10429 for (int i=0; i <= lastAddrIndx; ++i)
10430 (*MIB).addOperand(*argOpers[i]);
10432 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10433 (*MIB).setMemRefs(bInstr->memoperands_begin(),
10434 bInstr->memoperands_end());
10436 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
10437 MIB.addReg(EAXreg);
10440 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10442 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
10446 // private utility function: 64 bit atomics on 32 bit host.
10447 MachineBasicBlock *
10448 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
10449 MachineBasicBlock *MBB,
10454 bool invSrc) const {
10455 // For the atomic bitwise operator, we generate
10456 // thisMBB (instructions are in pairs, except cmpxchg8b)
10457 // ld t1,t2 = [bitinstr.addr]
10459 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
10460 // op t5, t6 <- out1, out2, [bitinstr.val]
10461 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
10462 // mov ECX, EBX <- t5, t6
10463 // mov EAX, EDX <- t1, t2
10464 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
10465 // mov t3, t4 <- EAX, EDX
10467 // result in out1, out2
10468 // fallthrough -->nextMBB
10470 const TargetRegisterClass *RC = X86::GR32RegisterClass;
10471 const unsigned LoadOpc = X86::MOV32rm;
10472 const unsigned NotOpc = X86::NOT32r;
10473 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10474 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10475 MachineFunction::iterator MBBIter = MBB;
10478 /// First build the CFG
10479 MachineFunction *F = MBB->getParent();
10480 MachineBasicBlock *thisMBB = MBB;
10481 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10482 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10483 F->insert(MBBIter, newMBB);
10484 F->insert(MBBIter, nextMBB);
10486 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10487 nextMBB->splice(nextMBB->begin(), thisMBB,
10488 llvm::next(MachineBasicBlock::iterator(bInstr)),
10490 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10492 // Update thisMBB to fall through to newMBB
10493 thisMBB->addSuccessor(newMBB);
10495 // newMBB jumps to itself and fall through to nextMBB
10496 newMBB->addSuccessor(nextMBB);
10497 newMBB->addSuccessor(newMBB);
10499 DebugLoc dl = bInstr->getDebugLoc();
10500 // Insert instructions into newMBB based on incoming instruction
10501 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
10502 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
10503 "unexpected number of operands");
10504 MachineOperand& dest1Oper = bInstr->getOperand(0);
10505 MachineOperand& dest2Oper = bInstr->getOperand(1);
10506 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10507 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
10508 argOpers[i] = &bInstr->getOperand(i+2);
10510 // We use some of the operands multiple times, so conservatively just
10511 // clear any kill flags that might be present.
10512 if (argOpers[i]->isReg() && argOpers[i]->isUse())
10513 argOpers[i]->setIsKill(false);
10516 // x86 address has 5 operands: base, index, scale, displacement, and segment.
10517 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10519 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
10520 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
10521 for (int i=0; i <= lastAddrIndx; ++i)
10522 (*MIB).addOperand(*argOpers[i]);
10523 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
10524 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
10525 // add 4 to displacement.
10526 for (int i=0; i <= lastAddrIndx-2; ++i)
10527 (*MIB).addOperand(*argOpers[i]);
10528 MachineOperand newOp3 = *(argOpers[3]);
10529 if (newOp3.isImm())
10530 newOp3.setImm(newOp3.getImm()+4);
10532 newOp3.setOffset(newOp3.getOffset()+4);
10533 (*MIB).addOperand(newOp3);
10534 (*MIB).addOperand(*argOpers[lastAddrIndx]);
10536 // t3/4 are defined later, at the bottom of the loop
10537 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
10538 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
10539 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
10540 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
10541 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
10542 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
10544 // The subsequent operations should be using the destination registers of
10545 //the PHI instructions.
10547 t1 = F->getRegInfo().createVirtualRegister(RC);
10548 t2 = F->getRegInfo().createVirtualRegister(RC);
10549 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
10550 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
10552 t1 = dest1Oper.getReg();
10553 t2 = dest2Oper.getReg();
10556 int valArgIndx = lastAddrIndx + 1;
10557 assert((argOpers[valArgIndx]->isReg() ||
10558 argOpers[valArgIndx]->isImm()) &&
10559 "invalid operand");
10560 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
10561 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
10562 if (argOpers[valArgIndx]->isReg())
10563 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
10565 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
10566 if (regOpcL != X86::MOV32rr)
10568 (*MIB).addOperand(*argOpers[valArgIndx]);
10569 assert(argOpers[valArgIndx + 1]->isReg() ==
10570 argOpers[valArgIndx]->isReg());
10571 assert(argOpers[valArgIndx + 1]->isImm() ==
10572 argOpers[valArgIndx]->isImm());
10573 if (argOpers[valArgIndx + 1]->isReg())
10574 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
10576 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
10577 if (regOpcH != X86::MOV32rr)
10579 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
10581 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
10583 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
10586 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
10588 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
10591 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
10592 for (int i=0; i <= lastAddrIndx; ++i)
10593 (*MIB).addOperand(*argOpers[i]);
10595 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10596 (*MIB).setMemRefs(bInstr->memoperands_begin(),
10597 bInstr->memoperands_end());
10599 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
10600 MIB.addReg(X86::EAX);
10601 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
10602 MIB.addReg(X86::EDX);
10605 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10607 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
10611 // private utility function
10612 MachineBasicBlock *
10613 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
10614 MachineBasicBlock *MBB,
10615 unsigned cmovOpc) const {
10616 // For the atomic min/max operator, we generate
10619 // ld t1 = [min/max.addr]
10620 // mov t2 = [min/max.val]
10622 // cmov[cond] t2 = t1
10624 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
10626 // fallthrough -->nextMBB
10628 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10629 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10630 MachineFunction::iterator MBBIter = MBB;
10633 /// First build the CFG
10634 MachineFunction *F = MBB->getParent();
10635 MachineBasicBlock *thisMBB = MBB;
10636 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10637 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10638 F->insert(MBBIter, newMBB);
10639 F->insert(MBBIter, nextMBB);
10641 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10642 nextMBB->splice(nextMBB->begin(), thisMBB,
10643 llvm::next(MachineBasicBlock::iterator(mInstr)),
10645 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10647 // Update thisMBB to fall through to newMBB
10648 thisMBB->addSuccessor(newMBB);
10650 // newMBB jumps to newMBB and fall through to nextMBB
10651 newMBB->addSuccessor(nextMBB);
10652 newMBB->addSuccessor(newMBB);
10654 DebugLoc dl = mInstr->getDebugLoc();
10655 // Insert instructions into newMBB based on incoming instruction
10656 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
10657 "unexpected number of operands");
10658 MachineOperand& destOper = mInstr->getOperand(0);
10659 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10660 int numArgs = mInstr->getNumOperands() - 1;
10661 for (int i=0; i < numArgs; ++i)
10662 argOpers[i] = &mInstr->getOperand(i+1);
10664 // x86 address has 4 operands: base, index, scale, and displacement
10665 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10666 int valArgIndx = lastAddrIndx + 1;
10668 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10669 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
10670 for (int i=0; i <= lastAddrIndx; ++i)
10671 (*MIB).addOperand(*argOpers[i]);
10673 // We only support register and immediate values
10674 assert((argOpers[valArgIndx]->isReg() ||
10675 argOpers[valArgIndx]->isImm()) &&
10676 "invalid operand");
10678 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10679 if (argOpers[valArgIndx]->isReg())
10680 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
10682 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
10683 (*MIB).addOperand(*argOpers[valArgIndx]);
10685 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
10688 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
10693 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
10694 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
10698 // Cmp and exchange if none has modified the memory location
10699 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
10700 for (int i=0; i <= lastAddrIndx; ++i)
10701 (*MIB).addOperand(*argOpers[i]);
10703 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10704 (*MIB).setMemRefs(mInstr->memoperands_begin(),
10705 mInstr->memoperands_end());
10707 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
10708 MIB.addReg(X86::EAX);
10711 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10713 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
10717 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
10718 // or XMM0_V32I8 in AVX all of this code can be replaced with that
10719 // in the .td file.
10720 MachineBasicBlock *
10721 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
10722 unsigned numArgs, bool memArg) const {
10723 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
10724 "Target must have SSE4.2 or AVX features enabled");
10726 DebugLoc dl = MI->getDebugLoc();
10727 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10729 if (!Subtarget->hasAVX()) {
10731 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
10733 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
10736 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
10738 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
10741 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
10742 for (unsigned i = 0; i < numArgs; ++i) {
10743 MachineOperand &Op = MI->getOperand(i+1);
10744 if (!(Op.isReg() && Op.isImplicit()))
10745 MIB.addOperand(Op);
10747 BuildMI(*BB, MI, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
10748 .addReg(X86::XMM0);
10750 MI->eraseFromParent();
10754 MachineBasicBlock *
10755 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
10756 DebugLoc dl = MI->getDebugLoc();
10757 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10759 // Address into RAX/EAX, other two args into ECX, EDX.
10760 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
10761 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
10762 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
10763 for (int i = 0; i < X86::AddrNumOperands; ++i)
10764 MIB.addOperand(MI->getOperand(i));
10766 unsigned ValOps = X86::AddrNumOperands;
10767 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10768 .addReg(MI->getOperand(ValOps).getReg());
10769 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
10770 .addReg(MI->getOperand(ValOps+1).getReg());
10772 // The instruction doesn't actually take any operands though.
10773 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
10775 MI->eraseFromParent(); // The pseudo is gone now.
10779 MachineBasicBlock *
10780 X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
10781 DebugLoc dl = MI->getDebugLoc();
10782 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10784 // First arg in ECX, the second in EAX.
10785 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
10786 .addReg(MI->getOperand(0).getReg());
10787 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
10788 .addReg(MI->getOperand(1).getReg());
10790 // The instruction doesn't actually take any operands though.
10791 BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
10793 MI->eraseFromParent(); // The pseudo is gone now.
10797 MachineBasicBlock *
10798 X86TargetLowering::EmitVAARG64WithCustomInserter(
10800 MachineBasicBlock *MBB) const {
10801 // Emit va_arg instruction on X86-64.
10803 // Operands to this pseudo-instruction:
10804 // 0 ) Output : destination address (reg)
10805 // 1-5) Input : va_list address (addr, i64mem)
10806 // 6 ) ArgSize : Size (in bytes) of vararg type
10807 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
10808 // 8 ) Align : Alignment of type
10809 // 9 ) EFLAGS (implicit-def)
10811 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
10812 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
10814 unsigned DestReg = MI->getOperand(0).getReg();
10815 MachineOperand &Base = MI->getOperand(1);
10816 MachineOperand &Scale = MI->getOperand(2);
10817 MachineOperand &Index = MI->getOperand(3);
10818 MachineOperand &Disp = MI->getOperand(4);
10819 MachineOperand &Segment = MI->getOperand(5);
10820 unsigned ArgSize = MI->getOperand(6).getImm();
10821 unsigned ArgMode = MI->getOperand(7).getImm();
10822 unsigned Align = MI->getOperand(8).getImm();
10824 // Memory Reference
10825 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
10826 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
10827 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
10829 // Machine Information
10830 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10831 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
10832 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
10833 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
10834 DebugLoc DL = MI->getDebugLoc();
10836 // struct va_list {
10839 // i64 overflow_area (address)
10840 // i64 reg_save_area (address)
10842 // sizeof(va_list) = 24
10843 // alignment(va_list) = 8
10845 unsigned TotalNumIntRegs = 6;
10846 unsigned TotalNumXMMRegs = 8;
10847 bool UseGPOffset = (ArgMode == 1);
10848 bool UseFPOffset = (ArgMode == 2);
10849 unsigned MaxOffset = TotalNumIntRegs * 8 +
10850 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
10852 /* Align ArgSize to a multiple of 8 */
10853 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
10854 bool NeedsAlign = (Align > 8);
10856 MachineBasicBlock *thisMBB = MBB;
10857 MachineBasicBlock *overflowMBB;
10858 MachineBasicBlock *offsetMBB;
10859 MachineBasicBlock *endMBB;
10861 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
10862 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
10863 unsigned OffsetReg = 0;
10865 if (!UseGPOffset && !UseFPOffset) {
10866 // If we only pull from the overflow region, we don't create a branch.
10867 // We don't need to alter control flow.
10868 OffsetDestReg = 0; // unused
10869 OverflowDestReg = DestReg;
10872 overflowMBB = thisMBB;
10875 // First emit code to check if gp_offset (or fp_offset) is below the bound.
10876 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
10877 // If not, pull from overflow_area. (branch to overflowMBB)
10882 // offsetMBB overflowMBB
10887 // Registers for the PHI in endMBB
10888 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
10889 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
10891 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10892 MachineFunction *MF = MBB->getParent();
10893 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10894 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10895 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
10897 MachineFunction::iterator MBBIter = MBB;
10900 // Insert the new basic blocks
10901 MF->insert(MBBIter, offsetMBB);
10902 MF->insert(MBBIter, overflowMBB);
10903 MF->insert(MBBIter, endMBB);
10905 // Transfer the remainder of MBB and its successor edges to endMBB.
10906 endMBB->splice(endMBB->begin(), thisMBB,
10907 llvm::next(MachineBasicBlock::iterator(MI)),
10909 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10911 // Make offsetMBB and overflowMBB successors of thisMBB
10912 thisMBB->addSuccessor(offsetMBB);
10913 thisMBB->addSuccessor(overflowMBB);
10915 // endMBB is a successor of both offsetMBB and overflowMBB
10916 offsetMBB->addSuccessor(endMBB);
10917 overflowMBB->addSuccessor(endMBB);
10919 // Load the offset value into a register
10920 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10921 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
10925 .addDisp(Disp, UseFPOffset ? 4 : 0)
10926 .addOperand(Segment)
10927 .setMemRefs(MMOBegin, MMOEnd);
10929 // Check if there is enough room left to pull this argument.
10930 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
10932 .addImm(MaxOffset + 8 - ArgSizeA8);
10934 // Branch to "overflowMBB" if offset >= max
10935 // Fall through to "offsetMBB" otherwise
10936 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
10937 .addMBB(overflowMBB);
10940 // In offsetMBB, emit code to use the reg_save_area.
10942 assert(OffsetReg != 0);
10944 // Read the reg_save_area address.
10945 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
10946 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
10951 .addOperand(Segment)
10952 .setMemRefs(MMOBegin, MMOEnd);
10954 // Zero-extend the offset
10955 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
10956 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
10959 .addImm(X86::sub_32bit);
10961 // Add the offset to the reg_save_area to get the final address.
10962 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
10963 .addReg(OffsetReg64)
10964 .addReg(RegSaveReg);
10966 // Compute the offset for the next argument
10967 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
10968 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
10970 .addImm(UseFPOffset ? 16 : 8);
10972 // Store it back into the va_list.
10973 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
10977 .addDisp(Disp, UseFPOffset ? 4 : 0)
10978 .addOperand(Segment)
10979 .addReg(NextOffsetReg)
10980 .setMemRefs(MMOBegin, MMOEnd);
10983 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
10988 // Emit code to use overflow area
10991 // Load the overflow_area address into a register.
10992 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
10993 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
10998 .addOperand(Segment)
10999 .setMemRefs(MMOBegin, MMOEnd);
11001 // If we need to align it, do so. Otherwise, just copy the address
11002 // to OverflowDestReg.
11004 // Align the overflow address
11005 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
11006 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
11008 // aligned_addr = (addr + (align-1)) & ~(align-1)
11009 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
11010 .addReg(OverflowAddrReg)
11013 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
11015 .addImm(~(uint64_t)(Align-1));
11017 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
11018 .addReg(OverflowAddrReg);
11021 // Compute the next overflow address after this argument.
11022 // (the overflow address should be kept 8-byte aligned)
11023 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
11024 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
11025 .addReg(OverflowDestReg)
11026 .addImm(ArgSizeA8);
11028 // Store the new overflow address.
11029 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
11034 .addOperand(Segment)
11035 .addReg(NextAddrReg)
11036 .setMemRefs(MMOBegin, MMOEnd);
11038 // If we branched, emit the PHI to the front of endMBB.
11040 BuildMI(*endMBB, endMBB->begin(), DL,
11041 TII->get(X86::PHI), DestReg)
11042 .addReg(OffsetDestReg).addMBB(offsetMBB)
11043 .addReg(OverflowDestReg).addMBB(overflowMBB);
11046 // Erase the pseudo instruction
11047 MI->eraseFromParent();
11052 MachineBasicBlock *
11053 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
11055 MachineBasicBlock *MBB) const {
11056 // Emit code to save XMM registers to the stack. The ABI says that the
11057 // number of registers to save is given in %al, so it's theoretically
11058 // possible to do an indirect jump trick to avoid saving all of them,
11059 // however this code takes a simpler approach and just executes all
11060 // of the stores if %al is non-zero. It's less code, and it's probably
11061 // easier on the hardware branch predictor, and stores aren't all that
11062 // expensive anyway.
11064 // Create the new basic blocks. One block contains all the XMM stores,
11065 // and one block is the final destination regardless of whether any
11066 // stores were performed.
11067 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
11068 MachineFunction *F = MBB->getParent();
11069 MachineFunction::iterator MBBIter = MBB;
11071 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
11072 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
11073 F->insert(MBBIter, XMMSaveMBB);
11074 F->insert(MBBIter, EndMBB);
11076 // Transfer the remainder of MBB and its successor edges to EndMBB.
11077 EndMBB->splice(EndMBB->begin(), MBB,
11078 llvm::next(MachineBasicBlock::iterator(MI)),
11080 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
11082 // The original block will now fall through to the XMM save block.
11083 MBB->addSuccessor(XMMSaveMBB);
11084 // The XMMSaveMBB will fall through to the end block.
11085 XMMSaveMBB->addSuccessor(EndMBB);
11087 // Now add the instructions.
11088 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11089 DebugLoc DL = MI->getDebugLoc();
11091 unsigned CountReg = MI->getOperand(0).getReg();
11092 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
11093 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
11095 if (!Subtarget->isTargetWin64()) {
11096 // If %al is 0, branch around the XMM save block.
11097 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
11098 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
11099 MBB->addSuccessor(EndMBB);
11102 // In the XMM save block, save all the XMM argument registers.
11103 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
11104 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
11105 MachineMemOperand *MMO =
11106 F->getMachineMemOperand(
11107 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
11108 MachineMemOperand::MOStore,
11109 /*Size=*/16, /*Align=*/16);
11110 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
11111 .addFrameIndex(RegSaveFrameIndex)
11112 .addImm(/*Scale=*/1)
11113 .addReg(/*IndexReg=*/0)
11114 .addImm(/*Disp=*/Offset)
11115 .addReg(/*Segment=*/0)
11116 .addReg(MI->getOperand(i).getReg())
11117 .addMemOperand(MMO);
11120 MI->eraseFromParent(); // The pseudo instruction is gone now.
11125 MachineBasicBlock *
11126 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
11127 MachineBasicBlock *BB) const {
11128 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11129 DebugLoc DL = MI->getDebugLoc();
11131 // To "insert" a SELECT_CC instruction, we actually have to insert the
11132 // diamond control-flow pattern. The incoming instruction knows the
11133 // destination vreg to set, the condition code register to branch on, the
11134 // true/false values to select between, and a branch opcode to use.
11135 const BasicBlock *LLVM_BB = BB->getBasicBlock();
11136 MachineFunction::iterator It = BB;
11142 // cmpTY ccX, r1, r2
11144 // fallthrough --> copy0MBB
11145 MachineBasicBlock *thisMBB = BB;
11146 MachineFunction *F = BB->getParent();
11147 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
11148 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
11149 F->insert(It, copy0MBB);
11150 F->insert(It, sinkMBB);
11152 // If the EFLAGS register isn't dead in the terminator, then claim that it's
11153 // live into the sink and copy blocks.
11154 const MachineFunction *MF = BB->getParent();
11155 const TargetRegisterInfo *TRI = MF->getTarget().getRegisterInfo();
11156 BitVector ReservedRegs = TRI->getReservedRegs(*MF);
11158 for (unsigned I = 0, E = MI->getNumOperands(); I != E; ++I) {
11159 const MachineOperand &MO = MI->getOperand(I);
11160 if (!MO.isReg() || !MO.isUse() || MO.isKill()) continue;
11161 unsigned Reg = MO.getReg();
11162 if (Reg != X86::EFLAGS) continue;
11163 copy0MBB->addLiveIn(Reg);
11164 sinkMBB->addLiveIn(Reg);
11167 // Transfer the remainder of BB and its successor edges to sinkMBB.
11168 sinkMBB->splice(sinkMBB->begin(), BB,
11169 llvm::next(MachineBasicBlock::iterator(MI)),
11171 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
11173 // Add the true and fallthrough blocks as its successors.
11174 BB->addSuccessor(copy0MBB);
11175 BB->addSuccessor(sinkMBB);
11177 // Create the conditional branch instruction.
11179 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
11180 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
11183 // %FalseValue = ...
11184 // # fallthrough to sinkMBB
11185 copy0MBB->addSuccessor(sinkMBB);
11188 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
11190 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
11191 TII->get(X86::PHI), MI->getOperand(0).getReg())
11192 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
11193 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
11195 MI->eraseFromParent(); // The pseudo instruction is gone now.
11199 MachineBasicBlock *
11200 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
11201 MachineBasicBlock *BB) const {
11202 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11203 DebugLoc DL = MI->getDebugLoc();
11205 assert(!Subtarget->isTargetEnvMacho());
11207 // The lowering is pretty easy: we're just emitting the call to _alloca. The
11208 // non-trivial part is impdef of ESP.
11210 if (Subtarget->isTargetWin64()) {
11211 if (Subtarget->isTargetCygMing()) {
11212 // ___chkstk(Mingw64):
11213 // Clobbers R10, R11, RAX and EFLAGS.
11215 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
11216 .addExternalSymbol("___chkstk")
11217 .addReg(X86::RAX, RegState::Implicit)
11218 .addReg(X86::RSP, RegState::Implicit)
11219 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
11220 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
11221 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11223 // __chkstk(MSVCRT): does not update stack pointer.
11224 // Clobbers R10, R11 and EFLAGS.
11225 // FIXME: RAX(allocated size) might be reused and not killed.
11226 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
11227 .addExternalSymbol("__chkstk")
11228 .addReg(X86::RAX, RegState::Implicit)
11229 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11230 // RAX has the offset to subtracted from RSP.
11231 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
11236 const char *StackProbeSymbol =
11237 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
11239 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
11240 .addExternalSymbol(StackProbeSymbol)
11241 .addReg(X86::EAX, RegState::Implicit)
11242 .addReg(X86::ESP, RegState::Implicit)
11243 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
11244 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
11245 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11248 MI->eraseFromParent(); // The pseudo instruction is gone now.
11252 MachineBasicBlock *
11253 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
11254 MachineBasicBlock *BB) const {
11255 // This is pretty easy. We're taking the value that we received from
11256 // our load from the relocation, sticking it in either RDI (x86-64)
11257 // or EAX and doing an indirect call. The return value will then
11258 // be in the normal return register.
11259 const X86InstrInfo *TII
11260 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
11261 DebugLoc DL = MI->getDebugLoc();
11262 MachineFunction *F = BB->getParent();
11264 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
11265 assert(MI->getOperand(3).isGlobal() && "This should be a global");
11267 if (Subtarget->is64Bit()) {
11268 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11269 TII->get(X86::MOV64rm), X86::RDI)
11271 .addImm(0).addReg(0)
11272 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11273 MI->getOperand(3).getTargetFlags())
11275 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
11276 addDirectMem(MIB, X86::RDI);
11277 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
11278 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11279 TII->get(X86::MOV32rm), X86::EAX)
11281 .addImm(0).addReg(0)
11282 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11283 MI->getOperand(3).getTargetFlags())
11285 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
11286 addDirectMem(MIB, X86::EAX);
11288 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11289 TII->get(X86::MOV32rm), X86::EAX)
11290 .addReg(TII->getGlobalBaseReg(F))
11291 .addImm(0).addReg(0)
11292 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11293 MI->getOperand(3).getTargetFlags())
11295 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
11296 addDirectMem(MIB, X86::EAX);
11299 MI->eraseFromParent(); // The pseudo instruction is gone now.
11303 MachineBasicBlock *
11304 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
11305 MachineBasicBlock *BB) const {
11306 switch (MI->getOpcode()) {
11307 default: assert(false && "Unexpected instr type to insert");
11308 case X86::TAILJMPd64:
11309 case X86::TAILJMPr64:
11310 case X86::TAILJMPm64:
11311 assert(!"TAILJMP64 would not be touched here.");
11312 case X86::TCRETURNdi64:
11313 case X86::TCRETURNri64:
11314 case X86::TCRETURNmi64:
11315 // Defs of TCRETURNxx64 has Win64's callee-saved registers, as subset.
11316 // On AMD64, additional defs should be added before register allocation.
11317 if (!Subtarget->isTargetWin64()) {
11318 MI->addRegisterDefined(X86::RSI);
11319 MI->addRegisterDefined(X86::RDI);
11320 MI->addRegisterDefined(X86::XMM6);
11321 MI->addRegisterDefined(X86::XMM7);
11322 MI->addRegisterDefined(X86::XMM8);
11323 MI->addRegisterDefined(X86::XMM9);
11324 MI->addRegisterDefined(X86::XMM10);
11325 MI->addRegisterDefined(X86::XMM11);
11326 MI->addRegisterDefined(X86::XMM12);
11327 MI->addRegisterDefined(X86::XMM13);
11328 MI->addRegisterDefined(X86::XMM14);
11329 MI->addRegisterDefined(X86::XMM15);
11332 case X86::WIN_ALLOCA:
11333 return EmitLoweredWinAlloca(MI, BB);
11334 case X86::TLSCall_32:
11335 case X86::TLSCall_64:
11336 return EmitLoweredTLSCall(MI, BB);
11337 case X86::CMOV_GR8:
11338 case X86::CMOV_FR32:
11339 case X86::CMOV_FR64:
11340 case X86::CMOV_V4F32:
11341 case X86::CMOV_V2F64:
11342 case X86::CMOV_V2I64:
11343 case X86::CMOV_V8F32:
11344 case X86::CMOV_V4F64:
11345 case X86::CMOV_V4I64:
11346 case X86::CMOV_GR16:
11347 case X86::CMOV_GR32:
11348 case X86::CMOV_RFP32:
11349 case X86::CMOV_RFP64:
11350 case X86::CMOV_RFP80:
11351 return EmitLoweredSelect(MI, BB);
11353 case X86::FP32_TO_INT16_IN_MEM:
11354 case X86::FP32_TO_INT32_IN_MEM:
11355 case X86::FP32_TO_INT64_IN_MEM:
11356 case X86::FP64_TO_INT16_IN_MEM:
11357 case X86::FP64_TO_INT32_IN_MEM:
11358 case X86::FP64_TO_INT64_IN_MEM:
11359 case X86::FP80_TO_INT16_IN_MEM:
11360 case X86::FP80_TO_INT32_IN_MEM:
11361 case X86::FP80_TO_INT64_IN_MEM: {
11362 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11363 DebugLoc DL = MI->getDebugLoc();
11365 // Change the floating point control register to use "round towards zero"
11366 // mode when truncating to an integer value.
11367 MachineFunction *F = BB->getParent();
11368 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
11369 addFrameReference(BuildMI(*BB, MI, DL,
11370 TII->get(X86::FNSTCW16m)), CWFrameIdx);
11372 // Load the old value of the high byte of the control word...
11374 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
11375 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
11378 // Set the high part to be round to zero...
11379 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
11382 // Reload the modified control word now...
11383 addFrameReference(BuildMI(*BB, MI, DL,
11384 TII->get(X86::FLDCW16m)), CWFrameIdx);
11386 // Restore the memory image of control word to original value
11387 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
11390 // Get the X86 opcode to use.
11392 switch (MI->getOpcode()) {
11393 default: llvm_unreachable("illegal opcode!");
11394 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
11395 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
11396 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
11397 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
11398 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
11399 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
11400 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
11401 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
11402 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
11406 MachineOperand &Op = MI->getOperand(0);
11408 AM.BaseType = X86AddressMode::RegBase;
11409 AM.Base.Reg = Op.getReg();
11411 AM.BaseType = X86AddressMode::FrameIndexBase;
11412 AM.Base.FrameIndex = Op.getIndex();
11414 Op = MI->getOperand(1);
11416 AM.Scale = Op.getImm();
11417 Op = MI->getOperand(2);
11419 AM.IndexReg = Op.getImm();
11420 Op = MI->getOperand(3);
11421 if (Op.isGlobal()) {
11422 AM.GV = Op.getGlobal();
11424 AM.Disp = Op.getImm();
11426 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
11427 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
11429 // Reload the original control word now.
11430 addFrameReference(BuildMI(*BB, MI, DL,
11431 TII->get(X86::FLDCW16m)), CWFrameIdx);
11433 MI->eraseFromParent(); // The pseudo instruction is gone now.
11436 // String/text processing lowering.
11437 case X86::PCMPISTRM128REG:
11438 case X86::VPCMPISTRM128REG:
11439 return EmitPCMP(MI, BB, 3, false /* in-mem */);
11440 case X86::PCMPISTRM128MEM:
11441 case X86::VPCMPISTRM128MEM:
11442 return EmitPCMP(MI, BB, 3, true /* in-mem */);
11443 case X86::PCMPESTRM128REG:
11444 case X86::VPCMPESTRM128REG:
11445 return EmitPCMP(MI, BB, 5, false /* in mem */);
11446 case X86::PCMPESTRM128MEM:
11447 case X86::VPCMPESTRM128MEM:
11448 return EmitPCMP(MI, BB, 5, true /* in mem */);
11450 // Thread synchronization.
11452 return EmitMonitor(MI, BB);
11454 return EmitMwait(MI, BB);
11456 // Atomic Lowering.
11457 case X86::ATOMAND32:
11458 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
11459 X86::AND32ri, X86::MOV32rm,
11461 X86::NOT32r, X86::EAX,
11462 X86::GR32RegisterClass);
11463 case X86::ATOMOR32:
11464 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
11465 X86::OR32ri, X86::MOV32rm,
11467 X86::NOT32r, X86::EAX,
11468 X86::GR32RegisterClass);
11469 case X86::ATOMXOR32:
11470 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
11471 X86::XOR32ri, X86::MOV32rm,
11473 X86::NOT32r, X86::EAX,
11474 X86::GR32RegisterClass);
11475 case X86::ATOMNAND32:
11476 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
11477 X86::AND32ri, X86::MOV32rm,
11479 X86::NOT32r, X86::EAX,
11480 X86::GR32RegisterClass, true);
11481 case X86::ATOMMIN32:
11482 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
11483 case X86::ATOMMAX32:
11484 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
11485 case X86::ATOMUMIN32:
11486 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
11487 case X86::ATOMUMAX32:
11488 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
11490 case X86::ATOMAND16:
11491 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
11492 X86::AND16ri, X86::MOV16rm,
11494 X86::NOT16r, X86::AX,
11495 X86::GR16RegisterClass);
11496 case X86::ATOMOR16:
11497 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
11498 X86::OR16ri, X86::MOV16rm,
11500 X86::NOT16r, X86::AX,
11501 X86::GR16RegisterClass);
11502 case X86::ATOMXOR16:
11503 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
11504 X86::XOR16ri, X86::MOV16rm,
11506 X86::NOT16r, X86::AX,
11507 X86::GR16RegisterClass);
11508 case X86::ATOMNAND16:
11509 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
11510 X86::AND16ri, X86::MOV16rm,
11512 X86::NOT16r, X86::AX,
11513 X86::GR16RegisterClass, true);
11514 case X86::ATOMMIN16:
11515 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
11516 case X86::ATOMMAX16:
11517 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
11518 case X86::ATOMUMIN16:
11519 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
11520 case X86::ATOMUMAX16:
11521 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
11523 case X86::ATOMAND8:
11524 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
11525 X86::AND8ri, X86::MOV8rm,
11527 X86::NOT8r, X86::AL,
11528 X86::GR8RegisterClass);
11530 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
11531 X86::OR8ri, X86::MOV8rm,
11533 X86::NOT8r, X86::AL,
11534 X86::GR8RegisterClass);
11535 case X86::ATOMXOR8:
11536 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
11537 X86::XOR8ri, X86::MOV8rm,
11539 X86::NOT8r, X86::AL,
11540 X86::GR8RegisterClass);
11541 case X86::ATOMNAND8:
11542 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
11543 X86::AND8ri, X86::MOV8rm,
11545 X86::NOT8r, X86::AL,
11546 X86::GR8RegisterClass, true);
11547 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
11548 // This group is for 64-bit host.
11549 case X86::ATOMAND64:
11550 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
11551 X86::AND64ri32, X86::MOV64rm,
11553 X86::NOT64r, X86::RAX,
11554 X86::GR64RegisterClass);
11555 case X86::ATOMOR64:
11556 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
11557 X86::OR64ri32, X86::MOV64rm,
11559 X86::NOT64r, X86::RAX,
11560 X86::GR64RegisterClass);
11561 case X86::ATOMXOR64:
11562 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
11563 X86::XOR64ri32, X86::MOV64rm,
11565 X86::NOT64r, X86::RAX,
11566 X86::GR64RegisterClass);
11567 case X86::ATOMNAND64:
11568 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
11569 X86::AND64ri32, X86::MOV64rm,
11571 X86::NOT64r, X86::RAX,
11572 X86::GR64RegisterClass, true);
11573 case X86::ATOMMIN64:
11574 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
11575 case X86::ATOMMAX64:
11576 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
11577 case X86::ATOMUMIN64:
11578 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
11579 case X86::ATOMUMAX64:
11580 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
11582 // This group does 64-bit operations on a 32-bit host.
11583 case X86::ATOMAND6432:
11584 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11585 X86::AND32rr, X86::AND32rr,
11586 X86::AND32ri, X86::AND32ri,
11588 case X86::ATOMOR6432:
11589 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11590 X86::OR32rr, X86::OR32rr,
11591 X86::OR32ri, X86::OR32ri,
11593 case X86::ATOMXOR6432:
11594 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11595 X86::XOR32rr, X86::XOR32rr,
11596 X86::XOR32ri, X86::XOR32ri,
11598 case X86::ATOMNAND6432:
11599 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11600 X86::AND32rr, X86::AND32rr,
11601 X86::AND32ri, X86::AND32ri,
11603 case X86::ATOMADD6432:
11604 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11605 X86::ADD32rr, X86::ADC32rr,
11606 X86::ADD32ri, X86::ADC32ri,
11608 case X86::ATOMSUB6432:
11609 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11610 X86::SUB32rr, X86::SBB32rr,
11611 X86::SUB32ri, X86::SBB32ri,
11613 case X86::ATOMSWAP6432:
11614 return EmitAtomicBit6432WithCustomInserter(MI, BB,
11615 X86::MOV32rr, X86::MOV32rr,
11616 X86::MOV32ri, X86::MOV32ri,
11618 case X86::VASTART_SAVE_XMM_REGS:
11619 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
11621 case X86::VAARG_64:
11622 return EmitVAARG64WithCustomInserter(MI, BB);
11626 //===----------------------------------------------------------------------===//
11627 // X86 Optimization Hooks
11628 //===----------------------------------------------------------------------===//
11630 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
11634 const SelectionDAG &DAG,
11635 unsigned Depth) const {
11636 unsigned Opc = Op.getOpcode();
11637 assert((Opc >= ISD::BUILTIN_OP_END ||
11638 Opc == ISD::INTRINSIC_WO_CHAIN ||
11639 Opc == ISD::INTRINSIC_W_CHAIN ||
11640 Opc == ISD::INTRINSIC_VOID) &&
11641 "Should use MaskedValueIsZero if you don't know whether Op"
11642 " is a target node!");
11644 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
11658 // These nodes' second result is a boolean.
11659 if (Op.getResNo() == 0)
11662 case X86ISD::SETCC:
11663 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
11664 Mask.getBitWidth() - 1);
11669 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
11670 unsigned Depth) const {
11671 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
11672 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
11673 return Op.getValueType().getScalarType().getSizeInBits();
11679 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
11680 /// node is a GlobalAddress + offset.
11681 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
11682 const GlobalValue* &GA,
11683 int64_t &Offset) const {
11684 if (N->getOpcode() == X86ISD::Wrapper) {
11685 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
11686 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
11687 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
11691 return TargetLowering::isGAPlusOffset(N, GA, Offset);
11694 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
11695 /// same as extracting the high 128-bit part of 256-bit vector and then
11696 /// inserting the result into the low part of a new 256-bit vector
11697 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
11698 EVT VT = SVOp->getValueType(0);
11699 int NumElems = VT.getVectorNumElements();
11701 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
11702 for (int i = 0, j = NumElems/2; i < NumElems/2; ++i, ++j)
11703 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
11704 SVOp->getMaskElt(j) >= 0)
11710 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
11711 /// same as extracting the low 128-bit part of 256-bit vector and then
11712 /// inserting the result into the high part of a new 256-bit vector
11713 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
11714 EVT VT = SVOp->getValueType(0);
11715 int NumElems = VT.getVectorNumElements();
11717 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
11718 for (int i = NumElems/2, j = 0; i < NumElems; ++i, ++j)
11719 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
11720 SVOp->getMaskElt(j) >= 0)
11726 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
11727 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
11728 TargetLowering::DAGCombinerInfo &DCI) {
11729 DebugLoc dl = N->getDebugLoc();
11730 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
11731 SDValue V1 = SVOp->getOperand(0);
11732 SDValue V2 = SVOp->getOperand(1);
11733 EVT VT = SVOp->getValueType(0);
11734 int NumElems = VT.getVectorNumElements();
11736 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
11737 V2.getOpcode() == ISD::CONCAT_VECTORS) {
11741 // V UNDEF BUILD_VECTOR UNDEF
11743 // CONCAT_VECTOR CONCAT_VECTOR
11746 // RESULT: V + zero extended
11748 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
11749 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
11750 V1.getOperand(1).getOpcode() != ISD::UNDEF)
11753 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
11756 // To match the shuffle mask, the first half of the mask should
11757 // be exactly the first vector, and all the rest a splat with the
11758 // first element of the second one.
11759 for (int i = 0; i < NumElems/2; ++i)
11760 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
11761 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
11764 // Emit a zeroed vector and insert the desired subvector on its
11766 SDValue Zeros = getZeroVector(VT, true /* HasSSE2 */, DAG, dl);
11767 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0),
11768 DAG.getConstant(0, MVT::i32), DAG, dl);
11769 return DCI.CombineTo(N, InsV);
11772 //===--------------------------------------------------------------------===//
11773 // Combine some shuffles into subvector extracts and inserts:
11776 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
11777 if (isShuffleHigh128VectorInsertLow(SVOp)) {
11778 SDValue V = Extract128BitVector(V1, DAG.getConstant(NumElems/2, MVT::i32),
11780 SDValue InsV = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT),
11781 V, DAG.getConstant(0, MVT::i32), DAG, dl);
11782 return DCI.CombineTo(N, InsV);
11785 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
11786 if (isShuffleLow128VectorInsertHigh(SVOp)) {
11787 SDValue V = Extract128BitVector(V1, DAG.getConstant(0, MVT::i32), DAG, dl);
11788 SDValue InsV = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT),
11789 V, DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
11790 return DCI.CombineTo(N, InsV);
11796 /// PerformShuffleCombine - Performs several different shuffle combines.
11797 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
11798 TargetLowering::DAGCombinerInfo &DCI,
11799 const X86Subtarget *Subtarget) {
11800 DebugLoc dl = N->getDebugLoc();
11801 EVT VT = N->getValueType(0);
11803 // Don't create instructions with illegal types after legalize types has run.
11804 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
11805 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
11808 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
11809 if (Subtarget->hasAVX() && VT.getSizeInBits() == 256 &&
11810 N->getOpcode() == ISD::VECTOR_SHUFFLE)
11811 return PerformShuffleCombine256(N, DAG, DCI);
11813 // Only handle 128 wide vector from here on.
11814 if (VT.getSizeInBits() != 128)
11817 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
11818 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
11819 // consecutive, non-overlapping, and in the right order.
11820 SmallVector<SDValue, 16> Elts;
11821 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
11822 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
11824 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
11827 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
11828 /// generation and convert it from being a bunch of shuffles and extracts
11829 /// to a simple store and scalar loads to extract the elements.
11830 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
11831 const TargetLowering &TLI) {
11832 SDValue InputVector = N->getOperand(0);
11834 // Only operate on vectors of 4 elements, where the alternative shuffling
11835 // gets to be more expensive.
11836 if (InputVector.getValueType() != MVT::v4i32)
11839 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
11840 // single use which is a sign-extend or zero-extend, and all elements are
11842 SmallVector<SDNode *, 4> Uses;
11843 unsigned ExtractedElements = 0;
11844 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
11845 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
11846 if (UI.getUse().getResNo() != InputVector.getResNo())
11849 SDNode *Extract = *UI;
11850 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
11853 if (Extract->getValueType(0) != MVT::i32)
11855 if (!Extract->hasOneUse())
11857 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
11858 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
11860 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
11863 // Record which element was extracted.
11864 ExtractedElements |=
11865 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
11867 Uses.push_back(Extract);
11870 // If not all the elements were used, this may not be worthwhile.
11871 if (ExtractedElements != 15)
11874 // Ok, we've now decided to do the transformation.
11875 DebugLoc dl = InputVector.getDebugLoc();
11877 // Store the value to a temporary stack slot.
11878 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
11879 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
11880 MachinePointerInfo(), false, false, 0);
11882 // Replace each use (extract) with a load of the appropriate element.
11883 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
11884 UE = Uses.end(); UI != UE; ++UI) {
11885 SDNode *Extract = *UI;
11887 // cOMpute the element's address.
11888 SDValue Idx = Extract->getOperand(1);
11890 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
11891 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
11892 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
11894 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
11895 StackPtr, OffsetVal);
11897 // Load the scalar.
11898 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
11899 ScalarAddr, MachinePointerInfo(),
11902 // Replace the exact with the load.
11903 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
11906 // The replacement was made in place; don't return anything.
11910 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
11911 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
11912 const X86Subtarget *Subtarget) {
11913 DebugLoc DL = N->getDebugLoc();
11914 SDValue Cond = N->getOperand(0);
11915 // Get the LHS/RHS of the select.
11916 SDValue LHS = N->getOperand(1);
11917 SDValue RHS = N->getOperand(2);
11919 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
11920 // instructions match the semantics of the common C idiom x<y?x:y but not
11921 // x<=y?x:y, because of how they handle negative zero (which can be
11922 // ignored in unsafe-math mode).
11923 if (Subtarget->hasSSE2() &&
11924 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
11925 Cond.getOpcode() == ISD::SETCC) {
11926 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
11928 unsigned Opcode = 0;
11929 // Check for x CC y ? x : y.
11930 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
11931 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
11935 // Converting this to a min would handle NaNs incorrectly, and swapping
11936 // the operands would cause it to handle comparisons between positive
11937 // and negative zero incorrectly.
11938 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11939 if (!UnsafeFPMath &&
11940 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11942 std::swap(LHS, RHS);
11944 Opcode = X86ISD::FMIN;
11947 // Converting this to a min would handle comparisons between positive
11948 // and negative zero incorrectly.
11949 if (!UnsafeFPMath &&
11950 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
11952 Opcode = X86ISD::FMIN;
11955 // Converting this to a min would handle both negative zeros and NaNs
11956 // incorrectly, but we can swap the operands to fix both.
11957 std::swap(LHS, RHS);
11961 Opcode = X86ISD::FMIN;
11965 // Converting this to a max would handle comparisons between positive
11966 // and negative zero incorrectly.
11967 if (!UnsafeFPMath &&
11968 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
11970 Opcode = X86ISD::FMAX;
11973 // Converting this to a max would handle NaNs incorrectly, and swapping
11974 // the operands would cause it to handle comparisons between positive
11975 // and negative zero incorrectly.
11976 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
11977 if (!UnsafeFPMath &&
11978 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
11980 std::swap(LHS, RHS);
11982 Opcode = X86ISD::FMAX;
11985 // Converting this to a max would handle both negative zeros and NaNs
11986 // incorrectly, but we can swap the operands to fix both.
11987 std::swap(LHS, RHS);
11991 Opcode = X86ISD::FMAX;
11994 // Check for x CC y ? y : x -- a min/max with reversed arms.
11995 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
11996 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
12000 // Converting this to a min would handle comparisons between positive
12001 // and negative zero incorrectly, and swapping the operands would
12002 // cause it to handle NaNs incorrectly.
12003 if (!UnsafeFPMath &&
12004 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
12005 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
12007 std::swap(LHS, RHS);
12009 Opcode = X86ISD::FMIN;
12012 // Converting this to a min would handle NaNs incorrectly.
12013 if (!UnsafeFPMath &&
12014 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
12016 Opcode = X86ISD::FMIN;
12019 // Converting this to a min would handle both negative zeros and NaNs
12020 // incorrectly, but we can swap the operands to fix both.
12021 std::swap(LHS, RHS);
12025 Opcode = X86ISD::FMIN;
12029 // Converting this to a max would handle NaNs incorrectly.
12030 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
12032 Opcode = X86ISD::FMAX;
12035 // Converting this to a max would handle comparisons between positive
12036 // and negative zero incorrectly, and swapping the operands would
12037 // cause it to handle NaNs incorrectly.
12038 if (!UnsafeFPMath &&
12039 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
12040 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
12042 std::swap(LHS, RHS);
12044 Opcode = X86ISD::FMAX;
12047 // Converting this to a max would handle both negative zeros and NaNs
12048 // incorrectly, but we can swap the operands to fix both.
12049 std::swap(LHS, RHS);
12053 Opcode = X86ISD::FMAX;
12059 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
12062 // If this is a select between two integer constants, try to do some
12064 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
12065 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
12066 // Don't do this for crazy integer types.
12067 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
12068 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
12069 // so that TrueC (the true value) is larger than FalseC.
12070 bool NeedsCondInvert = false;
12072 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
12073 // Efficiently invertible.
12074 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
12075 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
12076 isa<ConstantSDNode>(Cond.getOperand(1))))) {
12077 NeedsCondInvert = true;
12078 std::swap(TrueC, FalseC);
12081 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
12082 if (FalseC->getAPIntValue() == 0 &&
12083 TrueC->getAPIntValue().isPowerOf2()) {
12084 if (NeedsCondInvert) // Invert the condition if needed.
12085 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
12086 DAG.getConstant(1, Cond.getValueType()));
12088 // Zero extend the condition if needed.
12089 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
12091 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
12092 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
12093 DAG.getConstant(ShAmt, MVT::i8));
12096 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
12097 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
12098 if (NeedsCondInvert) // Invert the condition if needed.
12099 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
12100 DAG.getConstant(1, Cond.getValueType()));
12102 // Zero extend the condition if needed.
12103 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
12104 FalseC->getValueType(0), Cond);
12105 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12106 SDValue(FalseC, 0));
12109 // Optimize cases that will turn into an LEA instruction. This requires
12110 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
12111 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
12112 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
12113 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
12115 bool isFastMultiplier = false;
12117 switch ((unsigned char)Diff) {
12119 case 1: // result = add base, cond
12120 case 2: // result = lea base( , cond*2)
12121 case 3: // result = lea base(cond, cond*2)
12122 case 4: // result = lea base( , cond*4)
12123 case 5: // result = lea base(cond, cond*4)
12124 case 8: // result = lea base( , cond*8)
12125 case 9: // result = lea base(cond, cond*8)
12126 isFastMultiplier = true;
12131 if (isFastMultiplier) {
12132 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
12133 if (NeedsCondInvert) // Invert the condition if needed.
12134 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
12135 DAG.getConstant(1, Cond.getValueType()));
12137 // Zero extend the condition if needed.
12138 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
12140 // Scale the condition by the difference.
12142 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
12143 DAG.getConstant(Diff, Cond.getValueType()));
12145 // Add the base if non-zero.
12146 if (FalseC->getAPIntValue() != 0)
12147 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12148 SDValue(FalseC, 0));
12158 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
12159 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
12160 TargetLowering::DAGCombinerInfo &DCI) {
12161 DebugLoc DL = N->getDebugLoc();
12163 // If the flag operand isn't dead, don't touch this CMOV.
12164 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
12167 SDValue FalseOp = N->getOperand(0);
12168 SDValue TrueOp = N->getOperand(1);
12169 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
12170 SDValue Cond = N->getOperand(3);
12171 if (CC == X86::COND_E || CC == X86::COND_NE) {
12172 switch (Cond.getOpcode()) {
12176 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
12177 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
12178 return (CC == X86::COND_E) ? FalseOp : TrueOp;
12182 // If this is a select between two integer constants, try to do some
12183 // optimizations. Note that the operands are ordered the opposite of SELECT
12185 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
12186 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
12187 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
12188 // larger than FalseC (the false value).
12189 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
12190 CC = X86::GetOppositeBranchCondition(CC);
12191 std::swap(TrueC, FalseC);
12194 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
12195 // This is efficient for any integer data type (including i8/i16) and
12197 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
12198 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
12199 DAG.getConstant(CC, MVT::i8), Cond);
12201 // Zero extend the condition if needed.
12202 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
12204 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
12205 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
12206 DAG.getConstant(ShAmt, MVT::i8));
12207 if (N->getNumValues() == 2) // Dead flag value?
12208 return DCI.CombineTo(N, Cond, SDValue());
12212 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
12213 // for any integer data type, including i8/i16.
12214 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
12215 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
12216 DAG.getConstant(CC, MVT::i8), Cond);
12218 // Zero extend the condition if needed.
12219 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
12220 FalseC->getValueType(0), Cond);
12221 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12222 SDValue(FalseC, 0));
12224 if (N->getNumValues() == 2) // Dead flag value?
12225 return DCI.CombineTo(N, Cond, SDValue());
12229 // Optimize cases that will turn into an LEA instruction. This requires
12230 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
12231 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
12232 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
12233 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
12235 bool isFastMultiplier = false;
12237 switch ((unsigned char)Diff) {
12239 case 1: // result = add base, cond
12240 case 2: // result = lea base( , cond*2)
12241 case 3: // result = lea base(cond, cond*2)
12242 case 4: // result = lea base( , cond*4)
12243 case 5: // result = lea base(cond, cond*4)
12244 case 8: // result = lea base( , cond*8)
12245 case 9: // result = lea base(cond, cond*8)
12246 isFastMultiplier = true;
12251 if (isFastMultiplier) {
12252 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
12253 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
12254 DAG.getConstant(CC, MVT::i8), Cond);
12255 // Zero extend the condition if needed.
12256 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
12258 // Scale the condition by the difference.
12260 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
12261 DAG.getConstant(Diff, Cond.getValueType()));
12263 // Add the base if non-zero.
12264 if (FalseC->getAPIntValue() != 0)
12265 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12266 SDValue(FalseC, 0));
12267 if (N->getNumValues() == 2) // Dead flag value?
12268 return DCI.CombineTo(N, Cond, SDValue());
12278 /// PerformMulCombine - Optimize a single multiply with constant into two
12279 /// in order to implement it with two cheaper instructions, e.g.
12280 /// LEA + SHL, LEA + LEA.
12281 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
12282 TargetLowering::DAGCombinerInfo &DCI) {
12283 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
12286 EVT VT = N->getValueType(0);
12287 if (VT != MVT::i64)
12290 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
12293 uint64_t MulAmt = C->getZExtValue();
12294 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
12297 uint64_t MulAmt1 = 0;
12298 uint64_t MulAmt2 = 0;
12299 if ((MulAmt % 9) == 0) {
12301 MulAmt2 = MulAmt / 9;
12302 } else if ((MulAmt % 5) == 0) {
12304 MulAmt2 = MulAmt / 5;
12305 } else if ((MulAmt % 3) == 0) {
12307 MulAmt2 = MulAmt / 3;
12310 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
12311 DebugLoc DL = N->getDebugLoc();
12313 if (isPowerOf2_64(MulAmt2) &&
12314 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
12315 // If second multiplifer is pow2, issue it first. We want the multiply by
12316 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
12318 std::swap(MulAmt1, MulAmt2);
12321 if (isPowerOf2_64(MulAmt1))
12322 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
12323 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
12325 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
12326 DAG.getConstant(MulAmt1, VT));
12328 if (isPowerOf2_64(MulAmt2))
12329 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
12330 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
12332 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
12333 DAG.getConstant(MulAmt2, VT));
12335 // Do not add new nodes to DAG combiner worklist.
12336 DCI.CombineTo(N, NewMul, false);
12341 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
12342 SDValue N0 = N->getOperand(0);
12343 SDValue N1 = N->getOperand(1);
12344 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
12345 EVT VT = N0.getValueType();
12347 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
12348 // since the result of setcc_c is all zero's or all ones.
12349 if (N1C && N0.getOpcode() == ISD::AND &&
12350 N0.getOperand(1).getOpcode() == ISD::Constant) {
12351 SDValue N00 = N0.getOperand(0);
12352 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
12353 ((N00.getOpcode() == ISD::ANY_EXTEND ||
12354 N00.getOpcode() == ISD::ZERO_EXTEND) &&
12355 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
12356 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
12357 APInt ShAmt = N1C->getAPIntValue();
12358 Mask = Mask.shl(ShAmt);
12360 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
12361 N00, DAG.getConstant(Mask, VT));
12368 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
12370 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
12371 const X86Subtarget *Subtarget) {
12372 EVT VT = N->getValueType(0);
12373 if (!VT.isVector() && VT.isInteger() &&
12374 N->getOpcode() == ISD::SHL)
12375 return PerformSHLCombine(N, DAG);
12377 // On X86 with SSE2 support, we can transform this to a vector shift if
12378 // all elements are shifted by the same amount. We can't do this in legalize
12379 // because the a constant vector is typically transformed to a constant pool
12380 // so we have no knowledge of the shift amount.
12381 if (!(Subtarget->hasSSE2() || Subtarget->hasAVX()))
12384 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
12387 SDValue ShAmtOp = N->getOperand(1);
12388 EVT EltVT = VT.getVectorElementType();
12389 DebugLoc DL = N->getDebugLoc();
12390 SDValue BaseShAmt = SDValue();
12391 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
12392 unsigned NumElts = VT.getVectorNumElements();
12394 for (; i != NumElts; ++i) {
12395 SDValue Arg = ShAmtOp.getOperand(i);
12396 if (Arg.getOpcode() == ISD::UNDEF) continue;
12400 for (; i != NumElts; ++i) {
12401 SDValue Arg = ShAmtOp.getOperand(i);
12402 if (Arg.getOpcode() == ISD::UNDEF) continue;
12403 if (Arg != BaseShAmt) {
12407 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
12408 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
12409 SDValue InVec = ShAmtOp.getOperand(0);
12410 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
12411 unsigned NumElts = InVec.getValueType().getVectorNumElements();
12413 for (; i != NumElts; ++i) {
12414 SDValue Arg = InVec.getOperand(i);
12415 if (Arg.getOpcode() == ISD::UNDEF) continue;
12419 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
12420 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
12421 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
12422 if (C->getZExtValue() == SplatIdx)
12423 BaseShAmt = InVec.getOperand(1);
12426 if (BaseShAmt.getNode() == 0)
12427 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
12428 DAG.getIntPtrConstant(0));
12432 // The shift amount is an i32.
12433 if (EltVT.bitsGT(MVT::i32))
12434 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
12435 else if (EltVT.bitsLT(MVT::i32))
12436 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
12438 // The shift amount is identical so we can do a vector shift.
12439 SDValue ValOp = N->getOperand(0);
12440 switch (N->getOpcode()) {
12442 llvm_unreachable("Unknown shift opcode!");
12445 if (VT == MVT::v2i64)
12446 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12447 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
12449 if (VT == MVT::v4i32)
12450 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12451 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
12453 if (VT == MVT::v8i16)
12454 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12455 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
12459 if (VT == MVT::v4i32)
12460 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12461 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
12463 if (VT == MVT::v8i16)
12464 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12465 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
12469 if (VT == MVT::v2i64)
12470 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12471 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
12473 if (VT == MVT::v4i32)
12474 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12475 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
12477 if (VT == MVT::v8i16)
12478 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
12479 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
12487 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
12488 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
12489 // and friends. Likewise for OR -> CMPNEQSS.
12490 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
12491 TargetLowering::DAGCombinerInfo &DCI,
12492 const X86Subtarget *Subtarget) {
12495 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
12496 // we're requiring SSE2 for both.
12497 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
12498 SDValue N0 = N->getOperand(0);
12499 SDValue N1 = N->getOperand(1);
12500 SDValue CMP0 = N0->getOperand(1);
12501 SDValue CMP1 = N1->getOperand(1);
12502 DebugLoc DL = N->getDebugLoc();
12504 // The SETCCs should both refer to the same CMP.
12505 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
12508 SDValue CMP00 = CMP0->getOperand(0);
12509 SDValue CMP01 = CMP0->getOperand(1);
12510 EVT VT = CMP00.getValueType();
12512 if (VT == MVT::f32 || VT == MVT::f64) {
12513 bool ExpectingFlags = false;
12514 // Check for any users that want flags:
12515 for (SDNode::use_iterator UI = N->use_begin(),
12517 !ExpectingFlags && UI != UE; ++UI)
12518 switch (UI->getOpcode()) {
12523 ExpectingFlags = true;
12525 case ISD::CopyToReg:
12526 case ISD::SIGN_EXTEND:
12527 case ISD::ZERO_EXTEND:
12528 case ISD::ANY_EXTEND:
12532 if (!ExpectingFlags) {
12533 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
12534 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
12536 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
12537 X86::CondCode tmp = cc0;
12542 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
12543 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
12544 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
12545 X86ISD::NodeType NTOperator = is64BitFP ?
12546 X86ISD::FSETCCsd : X86ISD::FSETCCss;
12547 // FIXME: need symbolic constants for these magic numbers.
12548 // See X86ATTInstPrinter.cpp:printSSECC().
12549 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
12550 SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
12551 DAG.getConstant(x86cc, MVT::i8));
12552 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
12554 SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
12555 DAG.getConstant(1, MVT::i32));
12556 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
12557 return OneBitOfTruth;
12565 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
12566 /// so it can be folded inside ANDNP.
12567 static bool CanFoldXORWithAllOnes(const SDNode *N) {
12568 EVT VT = N->getValueType(0);
12570 // Match direct AllOnes for 128 and 256-bit vectors
12571 if (ISD::isBuildVectorAllOnes(N))
12574 // Look through a bit convert.
12575 if (N->getOpcode() == ISD::BITCAST)
12576 N = N->getOperand(0).getNode();
12578 // Sometimes the operand may come from a insert_subvector building a 256-bit
12580 if (VT.getSizeInBits() == 256 &&
12581 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
12582 SDValue V1 = N->getOperand(0);
12583 SDValue V2 = N->getOperand(1);
12585 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
12586 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
12587 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
12588 ISD::isBuildVectorAllOnes(V2.getNode()))
12595 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
12596 TargetLowering::DAGCombinerInfo &DCI,
12597 const X86Subtarget *Subtarget) {
12598 if (DCI.isBeforeLegalizeOps())
12601 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
12605 // Want to form ANDNP nodes:
12606 // 1) In the hopes of then easily combining them with OR and AND nodes
12607 // to form PBLEND/PSIGN.
12608 // 2) To match ANDN packed intrinsics
12609 EVT VT = N->getValueType(0);
12610 if (VT != MVT::v2i64 && VT != MVT::v4i64)
12613 SDValue N0 = N->getOperand(0);
12614 SDValue N1 = N->getOperand(1);
12615 DebugLoc DL = N->getDebugLoc();
12617 // Check LHS for vnot
12618 if (N0.getOpcode() == ISD::XOR &&
12619 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
12620 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
12621 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
12623 // Check RHS for vnot
12624 if (N1.getOpcode() == ISD::XOR &&
12625 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
12626 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
12627 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
12632 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
12633 TargetLowering::DAGCombinerInfo &DCI,
12634 const X86Subtarget *Subtarget) {
12635 if (DCI.isBeforeLegalizeOps())
12638 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
12642 EVT VT = N->getValueType(0);
12643 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
12646 SDValue N0 = N->getOperand(0);
12647 SDValue N1 = N->getOperand(1);
12649 // look for psign/blend
12650 if (Subtarget->hasSSSE3()) {
12651 if (VT == MVT::v2i64) {
12652 // Canonicalize pandn to RHS
12653 if (N0.getOpcode() == X86ISD::ANDNP)
12655 // or (and (m, x), (pandn m, y))
12656 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
12657 SDValue Mask = N1.getOperand(0);
12658 SDValue X = N1.getOperand(1);
12660 if (N0.getOperand(0) == Mask)
12661 Y = N0.getOperand(1);
12662 if (N0.getOperand(1) == Mask)
12663 Y = N0.getOperand(0);
12665 // Check to see if the mask appeared in both the AND and ANDNP and
12669 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
12670 if (Mask.getOpcode() != ISD::BITCAST ||
12671 X.getOpcode() != ISD::BITCAST ||
12672 Y.getOpcode() != ISD::BITCAST)
12675 // Look through mask bitcast.
12676 Mask = Mask.getOperand(0);
12677 EVT MaskVT = Mask.getValueType();
12679 // Validate that the Mask operand is a vector sra node. The sra node
12680 // will be an intrinsic.
12681 if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
12684 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
12685 // there is no psrai.b
12686 switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
12687 case Intrinsic::x86_sse2_psrai_w:
12688 case Intrinsic::x86_sse2_psrai_d:
12690 default: return SDValue();
12693 // Check that the SRA is all signbits.
12694 SDValue SraC = Mask.getOperand(2);
12695 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
12696 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
12697 if ((SraAmt + 1) != EltBits)
12700 DebugLoc DL = N->getDebugLoc();
12702 // Now we know we at least have a plendvb with the mask val. See if
12703 // we can form a psignb/w/d.
12704 // psign = x.type == y.type == mask.type && y = sub(0, x);
12705 X = X.getOperand(0);
12706 Y = Y.getOperand(0);
12707 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
12708 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
12709 X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
12712 case 8: Opc = X86ISD::PSIGNB; break;
12713 case 16: Opc = X86ISD::PSIGNW; break;
12714 case 32: Opc = X86ISD::PSIGND; break;
12718 SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
12719 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
12722 // PBLENDVB only available on SSE 4.1
12723 if (!Subtarget->hasSSE41())
12726 X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
12727 Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
12728 Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
12729 Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
12730 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
12735 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
12736 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
12738 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
12740 if (!N0.hasOneUse() || !N1.hasOneUse())
12743 SDValue ShAmt0 = N0.getOperand(1);
12744 if (ShAmt0.getValueType() != MVT::i8)
12746 SDValue ShAmt1 = N1.getOperand(1);
12747 if (ShAmt1.getValueType() != MVT::i8)
12749 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
12750 ShAmt0 = ShAmt0.getOperand(0);
12751 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
12752 ShAmt1 = ShAmt1.getOperand(0);
12754 DebugLoc DL = N->getDebugLoc();
12755 unsigned Opc = X86ISD::SHLD;
12756 SDValue Op0 = N0.getOperand(0);
12757 SDValue Op1 = N1.getOperand(0);
12758 if (ShAmt0.getOpcode() == ISD::SUB) {
12759 Opc = X86ISD::SHRD;
12760 std::swap(Op0, Op1);
12761 std::swap(ShAmt0, ShAmt1);
12764 unsigned Bits = VT.getSizeInBits();
12765 if (ShAmt1.getOpcode() == ISD::SUB) {
12766 SDValue Sum = ShAmt1.getOperand(0);
12767 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
12768 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
12769 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
12770 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
12771 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
12772 return DAG.getNode(Opc, DL, VT,
12774 DAG.getNode(ISD::TRUNCATE, DL,
12777 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
12778 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
12780 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
12781 return DAG.getNode(Opc, DL, VT,
12782 N0.getOperand(0), N1.getOperand(0),
12783 DAG.getNode(ISD::TRUNCATE, DL,
12790 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
12791 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
12792 const X86Subtarget *Subtarget) {
12793 StoreSDNode *St = cast<StoreSDNode>(N);
12794 EVT VT = St->getValue().getValueType();
12795 EVT StVT = St->getMemoryVT();
12796 DebugLoc dl = St->getDebugLoc();
12797 SDValue StoredVal = St->getOperand(1);
12798 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12800 // If we are saving a concatination of two XMM registers, perform two stores.
12801 // This is better in Sandy Bridge cause one 256-bit mem op is done via two
12802 // 128-bit ones. If in the future the cost becomes only one memory access the
12803 // first version would be better.
12804 if (VT.getSizeInBits() == 256 &&
12805 StoredVal.getNode()->getOpcode() == ISD::CONCAT_VECTORS &&
12806 StoredVal.getNumOperands() == 2) {
12808 SDValue Value0 = StoredVal.getOperand(0);
12809 SDValue Value1 = StoredVal.getOperand(1);
12811 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
12812 SDValue Ptr0 = St->getBasePtr();
12813 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
12815 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
12816 St->getPointerInfo(), St->isVolatile(),
12817 St->isNonTemporal(), St->getAlignment());
12818 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
12819 St->getPointerInfo(), St->isVolatile(),
12820 St->isNonTemporal(), St->getAlignment());
12821 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
12824 // Optimize trunc store (of multiple scalars) to shuffle and store.
12825 // First, pack all of the elements in one place. Next, store to memory
12826 // in fewer chunks.
12827 if (St->isTruncatingStore() && VT.isVector()) {
12828 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12829 unsigned NumElems = VT.getVectorNumElements();
12830 assert(StVT != VT && "Cannot truncate to the same type");
12831 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
12832 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
12834 // From, To sizes and ElemCount must be pow of two
12835 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
12836 // We are going to use the original vector elt for storing.
12837 // accumulated smaller vector elements must be a multiple of bigger size.
12838 if (0 != (NumElems * ToSz) % FromSz) return SDValue();
12839 unsigned SizeRatio = FromSz / ToSz;
12841 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
12843 // Create a type on which we perform the shuffle
12844 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
12845 StVT.getScalarType(), NumElems*SizeRatio);
12847 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
12849 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
12850 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
12851 for (unsigned i = 0; i < NumElems; i++ ) ShuffleVec[i] = i * SizeRatio;
12853 // Can't shuffle using an illegal type
12854 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
12856 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
12857 DAG.getUNDEF(WideVec.getValueType()),
12858 ShuffleVec.data());
12859 // At this point all of the data is stored at the bottom of the
12860 // register. We now need to save it to mem.
12862 // Find the largest store unit
12863 MVT StoreType = MVT::i8;
12864 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
12865 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
12866 MVT Tp = (MVT::SimpleValueType)tp;
12867 if (TLI.isTypeLegal(Tp) && StoreType.getSizeInBits() < NumElems * ToSz)
12871 // Bitcast the original vector into a vector of store-size units
12872 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
12873 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
12874 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
12875 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
12876 SmallVector<SDValue, 8> Chains;
12877 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
12878 TLI.getPointerTy());
12879 SDValue Ptr = St->getBasePtr();
12881 // Perform one or more big stores into memory.
12882 for (unsigned i = 0; i < (ToSz*NumElems)/StoreType.getSizeInBits() ; i++) {
12883 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
12884 StoreType, ShuffWide,
12885 DAG.getIntPtrConstant(i));
12886 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
12887 St->getPointerInfo(), St->isVolatile(),
12888 St->isNonTemporal(), St->getAlignment());
12889 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
12890 Chains.push_back(Ch);
12893 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0],
12898 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
12899 // the FP state in cases where an emms may be missing.
12900 // A preferable solution to the general problem is to figure out the right
12901 // places to insert EMMS. This qualifies as a quick hack.
12903 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
12904 if (VT.getSizeInBits() != 64)
12907 const Function *F = DAG.getMachineFunction().getFunction();
12908 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
12909 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
12910 && Subtarget->hasSSE2();
12911 if ((VT.isVector() ||
12912 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
12913 isa<LoadSDNode>(St->getValue()) &&
12914 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
12915 St->getChain().hasOneUse() && !St->isVolatile()) {
12916 SDNode* LdVal = St->getValue().getNode();
12917 LoadSDNode *Ld = 0;
12918 int TokenFactorIndex = -1;
12919 SmallVector<SDValue, 8> Ops;
12920 SDNode* ChainVal = St->getChain().getNode();
12921 // Must be a store of a load. We currently handle two cases: the load
12922 // is a direct child, and it's under an intervening TokenFactor. It is
12923 // possible to dig deeper under nested TokenFactors.
12924 if (ChainVal == LdVal)
12925 Ld = cast<LoadSDNode>(St->getChain());
12926 else if (St->getValue().hasOneUse() &&
12927 ChainVal->getOpcode() == ISD::TokenFactor) {
12928 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
12929 if (ChainVal->getOperand(i).getNode() == LdVal) {
12930 TokenFactorIndex = i;
12931 Ld = cast<LoadSDNode>(St->getValue());
12933 Ops.push_back(ChainVal->getOperand(i));
12937 if (!Ld || !ISD::isNormalLoad(Ld))
12940 // If this is not the MMX case, i.e. we are just turning i64 load/store
12941 // into f64 load/store, avoid the transformation if there are multiple
12942 // uses of the loaded value.
12943 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
12946 DebugLoc LdDL = Ld->getDebugLoc();
12947 DebugLoc StDL = N->getDebugLoc();
12948 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
12949 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
12951 if (Subtarget->is64Bit() || F64IsLegal) {
12952 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
12953 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
12954 Ld->getPointerInfo(), Ld->isVolatile(),
12955 Ld->isNonTemporal(), Ld->getAlignment());
12956 SDValue NewChain = NewLd.getValue(1);
12957 if (TokenFactorIndex != -1) {
12958 Ops.push_back(NewChain);
12959 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12962 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
12963 St->getPointerInfo(),
12964 St->isVolatile(), St->isNonTemporal(),
12965 St->getAlignment());
12968 // Otherwise, lower to two pairs of 32-bit loads / stores.
12969 SDValue LoAddr = Ld->getBasePtr();
12970 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
12971 DAG.getConstant(4, MVT::i32));
12973 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
12974 Ld->getPointerInfo(),
12975 Ld->isVolatile(), Ld->isNonTemporal(),
12976 Ld->getAlignment());
12977 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
12978 Ld->getPointerInfo().getWithOffset(4),
12979 Ld->isVolatile(), Ld->isNonTemporal(),
12980 MinAlign(Ld->getAlignment(), 4));
12982 SDValue NewChain = LoLd.getValue(1);
12983 if (TokenFactorIndex != -1) {
12984 Ops.push_back(LoLd);
12985 Ops.push_back(HiLd);
12986 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
12990 LoAddr = St->getBasePtr();
12991 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
12992 DAG.getConstant(4, MVT::i32));
12994 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
12995 St->getPointerInfo(),
12996 St->isVolatile(), St->isNonTemporal(),
12997 St->getAlignment());
12998 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
12999 St->getPointerInfo().getWithOffset(4),
13001 St->isNonTemporal(),
13002 MinAlign(St->getAlignment(), 4));
13003 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
13008 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
13009 /// X86ISD::FXOR nodes.
13010 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
13011 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
13012 // F[X]OR(0.0, x) -> x
13013 // F[X]OR(x, 0.0) -> x
13014 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
13015 if (C->getValueAPF().isPosZero())
13016 return N->getOperand(1);
13017 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
13018 if (C->getValueAPF().isPosZero())
13019 return N->getOperand(0);
13023 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
13024 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
13025 // FAND(0.0, x) -> 0.0
13026 // FAND(x, 0.0) -> 0.0
13027 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
13028 if (C->getValueAPF().isPosZero())
13029 return N->getOperand(0);
13030 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
13031 if (C->getValueAPF().isPosZero())
13032 return N->getOperand(1);
13036 static SDValue PerformBTCombine(SDNode *N,
13038 TargetLowering::DAGCombinerInfo &DCI) {
13039 // BT ignores high bits in the bit index operand.
13040 SDValue Op1 = N->getOperand(1);
13041 if (Op1.hasOneUse()) {
13042 unsigned BitWidth = Op1.getValueSizeInBits();
13043 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
13044 APInt KnownZero, KnownOne;
13045 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
13046 !DCI.isBeforeLegalizeOps());
13047 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13048 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
13049 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
13050 DCI.CommitTargetLoweringOpt(TLO);
13055 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
13056 SDValue Op = N->getOperand(0);
13057 if (Op.getOpcode() == ISD::BITCAST)
13058 Op = Op.getOperand(0);
13059 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
13060 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
13061 VT.getVectorElementType().getSizeInBits() ==
13062 OpVT.getVectorElementType().getSizeInBits()) {
13063 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
13068 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
13069 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
13070 // (and (i32 x86isd::setcc_carry), 1)
13071 // This eliminates the zext. This transformation is necessary because
13072 // ISD::SETCC is always legalized to i8.
13073 DebugLoc dl = N->getDebugLoc();
13074 SDValue N0 = N->getOperand(0);
13075 EVT VT = N->getValueType(0);
13076 if (N0.getOpcode() == ISD::AND &&
13078 N0.getOperand(0).hasOneUse()) {
13079 SDValue N00 = N0.getOperand(0);
13080 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
13082 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
13083 if (!C || C->getZExtValue() != 1)
13085 return DAG.getNode(ISD::AND, dl, VT,
13086 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
13087 N00.getOperand(0), N00.getOperand(1)),
13088 DAG.getConstant(1, VT));
13094 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
13095 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
13096 unsigned X86CC = N->getConstantOperandVal(0);
13097 SDValue EFLAG = N->getOperand(1);
13098 DebugLoc DL = N->getDebugLoc();
13100 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
13101 // a zext and produces an all-ones bit which is more useful than 0/1 in some
13103 if (X86CC == X86::COND_B)
13104 return DAG.getNode(ISD::AND, DL, MVT::i8,
13105 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
13106 DAG.getConstant(X86CC, MVT::i8), EFLAG),
13107 DAG.getConstant(1, MVT::i8));
13112 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
13113 const X86TargetLowering *XTLI) {
13114 SDValue Op0 = N->getOperand(0);
13115 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
13116 // a 32-bit target where SSE doesn't support i64->FP operations.
13117 if (Op0.getOpcode() == ISD::LOAD) {
13118 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
13119 EVT VT = Ld->getValueType(0);
13120 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
13121 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
13122 !XTLI->getSubtarget()->is64Bit() &&
13123 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
13124 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
13125 Ld->getChain(), Op0, DAG);
13126 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
13133 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
13134 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
13135 X86TargetLowering::DAGCombinerInfo &DCI) {
13136 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
13137 // the result is either zero or one (depending on the input carry bit).
13138 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
13139 if (X86::isZeroNode(N->getOperand(0)) &&
13140 X86::isZeroNode(N->getOperand(1)) &&
13141 // We don't have a good way to replace an EFLAGS use, so only do this when
13143 SDValue(N, 1).use_empty()) {
13144 DebugLoc DL = N->getDebugLoc();
13145 EVT VT = N->getValueType(0);
13146 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
13147 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
13148 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
13149 DAG.getConstant(X86::COND_B,MVT::i8),
13151 DAG.getConstant(1, VT));
13152 return DCI.CombineTo(N, Res1, CarryOut);
13158 // fold (add Y, (sete X, 0)) -> adc 0, Y
13159 // (add Y, (setne X, 0)) -> sbb -1, Y
13160 // (sub (sete X, 0), Y) -> sbb 0, Y
13161 // (sub (setne X, 0), Y) -> adc -1, Y
13162 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
13163 DebugLoc DL = N->getDebugLoc();
13165 // Look through ZExts.
13166 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
13167 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
13170 SDValue SetCC = Ext.getOperand(0);
13171 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
13174 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
13175 if (CC != X86::COND_E && CC != X86::COND_NE)
13178 SDValue Cmp = SetCC.getOperand(1);
13179 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
13180 !X86::isZeroNode(Cmp.getOperand(1)) ||
13181 !Cmp.getOperand(0).getValueType().isInteger())
13184 SDValue CmpOp0 = Cmp.getOperand(0);
13185 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
13186 DAG.getConstant(1, CmpOp0.getValueType()));
13188 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
13189 if (CC == X86::COND_NE)
13190 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
13191 DL, OtherVal.getValueType(), OtherVal,
13192 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
13193 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
13194 DL, OtherVal.getValueType(), OtherVal,
13195 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
13198 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG) {
13199 SDValue Op0 = N->getOperand(0);
13200 SDValue Op1 = N->getOperand(1);
13202 // X86 can't encode an immediate LHS of a sub. See if we can push the
13203 // negation into a preceding instruction.
13204 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
13205 uint64_t Op0C = C->getSExtValue();
13207 // If the RHS of the sub is a XOR with one use and a constant, invert the
13208 // immediate. Then add one to the LHS of the sub so we can turn
13209 // X-Y -> X+~Y+1, saving one register.
13210 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
13211 isa<ConstantSDNode>(Op1.getOperand(1))) {
13212 uint64_t XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getSExtValue();
13213 EVT VT = Op0.getValueType();
13214 SDValue NewXor = DAG.getNode(ISD::XOR, Op1.getDebugLoc(), VT,
13216 DAG.getConstant(~XorC, VT));
13217 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, NewXor,
13218 DAG.getConstant(Op0C+1, VT));
13222 return OptimizeConditionalInDecrement(N, DAG);
13225 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
13226 DAGCombinerInfo &DCI) const {
13227 SelectionDAG &DAG = DCI.DAG;
13228 switch (N->getOpcode()) {
13230 case ISD::EXTRACT_VECTOR_ELT:
13231 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
13232 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
13233 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
13234 case ISD::ADD: return OptimizeConditionalInDecrement(N, DAG);
13235 case ISD::SUB: return PerformSubCombine(N, DAG);
13236 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
13237 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
13240 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
13241 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
13242 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
13243 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
13244 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
13246 case X86ISD::FOR: return PerformFORCombine(N, DAG);
13247 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
13248 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
13249 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
13250 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
13251 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
13252 case X86ISD::SHUFPS: // Handle all target specific shuffles
13253 case X86ISD::SHUFPD:
13254 case X86ISD::PALIGN:
13255 case X86ISD::PUNPCKHBW:
13256 case X86ISD::PUNPCKHWD:
13257 case X86ISD::PUNPCKHDQ:
13258 case X86ISD::PUNPCKHQDQ:
13259 case X86ISD::UNPCKHPS:
13260 case X86ISD::UNPCKHPD:
13261 case X86ISD::VUNPCKHPSY:
13262 case X86ISD::VUNPCKHPDY:
13263 case X86ISD::PUNPCKLBW:
13264 case X86ISD::PUNPCKLWD:
13265 case X86ISD::PUNPCKLDQ:
13266 case X86ISD::PUNPCKLQDQ:
13267 case X86ISD::UNPCKLPS:
13268 case X86ISD::UNPCKLPD:
13269 case X86ISD::VUNPCKLPSY:
13270 case X86ISD::VUNPCKLPDY:
13271 case X86ISD::MOVHLPS:
13272 case X86ISD::MOVLHPS:
13273 case X86ISD::PSHUFD:
13274 case X86ISD::PSHUFHW:
13275 case X86ISD::PSHUFLW:
13276 case X86ISD::MOVSS:
13277 case X86ISD::MOVSD:
13278 case X86ISD::VPERMILPS:
13279 case X86ISD::VPERMILPSY:
13280 case X86ISD::VPERMILPD:
13281 case X86ISD::VPERMILPDY:
13282 case X86ISD::VPERM2F128:
13283 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
13289 /// isTypeDesirableForOp - Return true if the target has native support for
13290 /// the specified value type and it is 'desirable' to use the type for the
13291 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
13292 /// instruction encodings are longer and some i16 instructions are slow.
13293 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
13294 if (!isTypeLegal(VT))
13296 if (VT != MVT::i16)
13303 case ISD::SIGN_EXTEND:
13304 case ISD::ZERO_EXTEND:
13305 case ISD::ANY_EXTEND:
13318 /// IsDesirableToPromoteOp - This method query the target whether it is
13319 /// beneficial for dag combiner to promote the specified node. If true, it
13320 /// should return the desired promotion type by reference.
13321 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
13322 EVT VT = Op.getValueType();
13323 if (VT != MVT::i16)
13326 bool Promote = false;
13327 bool Commute = false;
13328 switch (Op.getOpcode()) {
13331 LoadSDNode *LD = cast<LoadSDNode>(Op);
13332 // If the non-extending load has a single use and it's not live out, then it
13333 // might be folded.
13334 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
13335 Op.hasOneUse()*/) {
13336 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13337 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
13338 // The only case where we'd want to promote LOAD (rather then it being
13339 // promoted as an operand is when it's only use is liveout.
13340 if (UI->getOpcode() != ISD::CopyToReg)
13347 case ISD::SIGN_EXTEND:
13348 case ISD::ZERO_EXTEND:
13349 case ISD::ANY_EXTEND:
13354 SDValue N0 = Op.getOperand(0);
13355 // Look out for (store (shl (load), x)).
13356 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
13369 SDValue N0 = Op.getOperand(0);
13370 SDValue N1 = Op.getOperand(1);
13371 if (!Commute && MayFoldLoad(N1))
13373 // Avoid disabling potential load folding opportunities.
13374 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
13376 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
13386 //===----------------------------------------------------------------------===//
13387 // X86 Inline Assembly Support
13388 //===----------------------------------------------------------------------===//
13390 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
13391 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
13393 std::string AsmStr = IA->getAsmString();
13395 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
13396 SmallVector<StringRef, 4> AsmPieces;
13397 SplitString(AsmStr, AsmPieces, ";\n");
13399 switch (AsmPieces.size()) {
13400 default: return false;
13402 AsmStr = AsmPieces[0];
13404 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
13406 // FIXME: this should verify that we are targeting a 486 or better. If not,
13407 // we will turn this bswap into something that will be lowered to logical ops
13408 // instead of emitting the bswap asm. For now, we don't support 486 or lower
13409 // so don't worry about this.
13411 if (AsmPieces.size() == 2 &&
13412 (AsmPieces[0] == "bswap" ||
13413 AsmPieces[0] == "bswapq" ||
13414 AsmPieces[0] == "bswapl") &&
13415 (AsmPieces[1] == "$0" ||
13416 AsmPieces[1] == "${0:q}")) {
13417 // No need to check constraints, nothing other than the equivalent of
13418 // "=r,0" would be valid here.
13419 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
13420 if (!Ty || Ty->getBitWidth() % 16 != 0)
13422 return IntrinsicLowering::LowerToByteSwap(CI);
13424 // rorw $$8, ${0:w} --> llvm.bswap.i16
13425 if (CI->getType()->isIntegerTy(16) &&
13426 AsmPieces.size() == 3 &&
13427 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
13428 AsmPieces[1] == "$$8," &&
13429 AsmPieces[2] == "${0:w}" &&
13430 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
13432 const std::string &ConstraintsStr = IA->getConstraintString();
13433 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
13434 std::sort(AsmPieces.begin(), AsmPieces.end());
13435 if (AsmPieces.size() == 4 &&
13436 AsmPieces[0] == "~{cc}" &&
13437 AsmPieces[1] == "~{dirflag}" &&
13438 AsmPieces[2] == "~{flags}" &&
13439 AsmPieces[3] == "~{fpsr}") {
13440 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
13441 if (!Ty || Ty->getBitWidth() % 16 != 0)
13443 return IntrinsicLowering::LowerToByteSwap(CI);
13448 if (CI->getType()->isIntegerTy(32) &&
13449 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
13450 SmallVector<StringRef, 4> Words;
13451 SplitString(AsmPieces[0], Words, " \t,");
13452 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
13453 Words[2] == "${0:w}") {
13455 SplitString(AsmPieces[1], Words, " \t,");
13456 if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
13457 Words[2] == "$0") {
13459 SplitString(AsmPieces[2], Words, " \t,");
13460 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
13461 Words[2] == "${0:w}") {
13463 const std::string &ConstraintsStr = IA->getConstraintString();
13464 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
13465 std::sort(AsmPieces.begin(), AsmPieces.end());
13466 if (AsmPieces.size() == 4 &&
13467 AsmPieces[0] == "~{cc}" &&
13468 AsmPieces[1] == "~{dirflag}" &&
13469 AsmPieces[2] == "~{flags}" &&
13470 AsmPieces[3] == "~{fpsr}") {
13471 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
13472 if (!Ty || Ty->getBitWidth() % 16 != 0)
13474 return IntrinsicLowering::LowerToByteSwap(CI);
13481 if (CI->getType()->isIntegerTy(64)) {
13482 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
13483 if (Constraints.size() >= 2 &&
13484 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
13485 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
13486 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
13487 SmallVector<StringRef, 4> Words;
13488 SplitString(AsmPieces[0], Words, " \t");
13489 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
13491 SplitString(AsmPieces[1], Words, " \t");
13492 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
13494 SplitString(AsmPieces[2], Words, " \t,");
13495 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
13496 Words[2] == "%edx") {
13497 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
13498 if (!Ty || Ty->getBitWidth() % 16 != 0)
13500 return IntrinsicLowering::LowerToByteSwap(CI);
13513 /// getConstraintType - Given a constraint letter, return the type of
13514 /// constraint it is for this target.
13515 X86TargetLowering::ConstraintType
13516 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
13517 if (Constraint.size() == 1) {
13518 switch (Constraint[0]) {
13529 return C_RegisterClass;
13553 return TargetLowering::getConstraintType(Constraint);
13556 /// Examine constraint type and operand type and determine a weight value.
13557 /// This object must already have been set up with the operand type
13558 /// and the current alternative constraint selected.
13559 TargetLowering::ConstraintWeight
13560 X86TargetLowering::getSingleConstraintMatchWeight(
13561 AsmOperandInfo &info, const char *constraint) const {
13562 ConstraintWeight weight = CW_Invalid;
13563 Value *CallOperandVal = info.CallOperandVal;
13564 // If we don't have a value, we can't do a match,
13565 // but allow it at the lowest weight.
13566 if (CallOperandVal == NULL)
13568 Type *type = CallOperandVal->getType();
13569 // Look at the constraint type.
13570 switch (*constraint) {
13572 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
13583 if (CallOperandVal->getType()->isIntegerTy())
13584 weight = CW_SpecificReg;
13589 if (type->isFloatingPointTy())
13590 weight = CW_SpecificReg;
13593 if (type->isX86_MMXTy() && Subtarget->hasMMX())
13594 weight = CW_SpecificReg;
13598 if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
13599 weight = CW_Register;
13602 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
13603 if (C->getZExtValue() <= 31)
13604 weight = CW_Constant;
13608 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13609 if (C->getZExtValue() <= 63)
13610 weight = CW_Constant;
13614 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13615 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
13616 weight = CW_Constant;
13620 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13621 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
13622 weight = CW_Constant;
13626 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13627 if (C->getZExtValue() <= 3)
13628 weight = CW_Constant;
13632 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13633 if (C->getZExtValue() <= 0xff)
13634 weight = CW_Constant;
13639 if (dyn_cast<ConstantFP>(CallOperandVal)) {
13640 weight = CW_Constant;
13644 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13645 if ((C->getSExtValue() >= -0x80000000LL) &&
13646 (C->getSExtValue() <= 0x7fffffffLL))
13647 weight = CW_Constant;
13651 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
13652 if (C->getZExtValue() <= 0xffffffff)
13653 weight = CW_Constant;
13660 /// LowerXConstraint - try to replace an X constraint, which matches anything,
13661 /// with another that has more specific requirements based on the type of the
13662 /// corresponding operand.
13663 const char *X86TargetLowering::
13664 LowerXConstraint(EVT ConstraintVT) const {
13665 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
13666 // 'f' like normal targets.
13667 if (ConstraintVT.isFloatingPoint()) {
13668 if (Subtarget->hasXMMInt())
13670 if (Subtarget->hasXMM())
13674 return TargetLowering::LowerXConstraint(ConstraintVT);
13677 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
13678 /// vector. If it is invalid, don't add anything to Ops.
13679 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
13680 std::string &Constraint,
13681 std::vector<SDValue>&Ops,
13682 SelectionDAG &DAG) const {
13683 SDValue Result(0, 0);
13685 // Only support length 1 constraints for now.
13686 if (Constraint.length() > 1) return;
13688 char ConstraintLetter = Constraint[0];
13689 switch (ConstraintLetter) {
13692 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13693 if (C->getZExtValue() <= 31) {
13694 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13700 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13701 if (C->getZExtValue() <= 63) {
13702 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13708 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13709 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
13710 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13716 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13717 if (C->getZExtValue() <= 255) {
13718 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13724 // 32-bit signed value
13725 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13726 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
13727 C->getSExtValue())) {
13728 // Widen to 64 bits here to get it sign extended.
13729 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
13732 // FIXME gcc accepts some relocatable values here too, but only in certain
13733 // memory models; it's complicated.
13738 // 32-bit unsigned value
13739 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
13740 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
13741 C->getZExtValue())) {
13742 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
13746 // FIXME gcc accepts some relocatable values here too, but only in certain
13747 // memory models; it's complicated.
13751 // Literal immediates are always ok.
13752 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
13753 // Widen to 64 bits here to get it sign extended.
13754 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
13758 // In any sort of PIC mode addresses need to be computed at runtime by
13759 // adding in a register or some sort of table lookup. These can't
13760 // be used as immediates.
13761 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
13764 // If we are in non-pic codegen mode, we allow the address of a global (with
13765 // an optional displacement) to be used with 'i'.
13766 GlobalAddressSDNode *GA = 0;
13767 int64_t Offset = 0;
13769 // Match either (GA), (GA+C), (GA+C1+C2), etc.
13771 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
13772 Offset += GA->getOffset();
13774 } else if (Op.getOpcode() == ISD::ADD) {
13775 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
13776 Offset += C->getZExtValue();
13777 Op = Op.getOperand(0);
13780 } else if (Op.getOpcode() == ISD::SUB) {
13781 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
13782 Offset += -C->getZExtValue();
13783 Op = Op.getOperand(0);
13788 // Otherwise, this isn't something we can handle, reject it.
13792 const GlobalValue *GV = GA->getGlobal();
13793 // If we require an extra load to get this address, as in PIC mode, we
13794 // can't accept it.
13795 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
13796 getTargetMachine())))
13799 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
13800 GA->getValueType(0), Offset);
13805 if (Result.getNode()) {
13806 Ops.push_back(Result);
13809 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
13812 std::pair<unsigned, const TargetRegisterClass*>
13813 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
13815 // First, see if this is a constraint that directly corresponds to an LLVM
13817 if (Constraint.size() == 1) {
13818 // GCC Constraint Letters
13819 switch (Constraint[0]) {
13821 // TODO: Slight differences here in allocation order and leaving
13822 // RIP in the class. Do they matter any more here than they do
13823 // in the normal allocation?
13824 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
13825 if (Subtarget->is64Bit()) {
13826 if (VT == MVT::i32 || VT == MVT::f32)
13827 return std::make_pair(0U, X86::GR32RegisterClass);
13828 else if (VT == MVT::i16)
13829 return std::make_pair(0U, X86::GR16RegisterClass);
13830 else if (VT == MVT::i8 || VT == MVT::i1)
13831 return std::make_pair(0U, X86::GR8RegisterClass);
13832 else if (VT == MVT::i64 || VT == MVT::f64)
13833 return std::make_pair(0U, X86::GR64RegisterClass);
13836 // 32-bit fallthrough
13837 case 'Q': // Q_REGS
13838 if (VT == MVT::i32 || VT == MVT::f32)
13839 return std::make_pair(0U, X86::GR32_ABCDRegisterClass);
13840 else if (VT == MVT::i16)
13841 return std::make_pair(0U, X86::GR16_ABCDRegisterClass);
13842 else if (VT == MVT::i8 || VT == MVT::i1)
13843 return std::make_pair(0U, X86::GR8_ABCD_LRegisterClass);
13844 else if (VT == MVT::i64)
13845 return std::make_pair(0U, X86::GR64_ABCDRegisterClass);
13847 case 'r': // GENERAL_REGS
13848 case 'l': // INDEX_REGS
13849 if (VT == MVT::i8 || VT == MVT::i1)
13850 return std::make_pair(0U, X86::GR8RegisterClass);
13851 if (VT == MVT::i16)
13852 return std::make_pair(0U, X86::GR16RegisterClass);
13853 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
13854 return std::make_pair(0U, X86::GR32RegisterClass);
13855 return std::make_pair(0U, X86::GR64RegisterClass);
13856 case 'R': // LEGACY_REGS
13857 if (VT == MVT::i8 || VT == MVT::i1)
13858 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
13859 if (VT == MVT::i16)
13860 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
13861 if (VT == MVT::i32 || !Subtarget->is64Bit())
13862 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
13863 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
13864 case 'f': // FP Stack registers.
13865 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
13866 // value to the correct fpstack register class.
13867 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
13868 return std::make_pair(0U, X86::RFP32RegisterClass);
13869 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
13870 return std::make_pair(0U, X86::RFP64RegisterClass);
13871 return std::make_pair(0U, X86::RFP80RegisterClass);
13872 case 'y': // MMX_REGS if MMX allowed.
13873 if (!Subtarget->hasMMX()) break;
13874 return std::make_pair(0U, X86::VR64RegisterClass);
13875 case 'Y': // SSE_REGS if SSE2 allowed
13876 if (!Subtarget->hasXMMInt()) break;
13878 case 'x': // SSE_REGS if SSE1 allowed
13879 if (!Subtarget->hasXMM()) break;
13881 switch (VT.getSimpleVT().SimpleTy) {
13883 // Scalar SSE types.
13886 return std::make_pair(0U, X86::FR32RegisterClass);
13889 return std::make_pair(0U, X86::FR64RegisterClass);
13897 return std::make_pair(0U, X86::VR128RegisterClass);
13903 // Use the default implementation in TargetLowering to convert the register
13904 // constraint into a member of a register class.
13905 std::pair<unsigned, const TargetRegisterClass*> Res;
13906 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
13908 // Not found as a standard register?
13909 if (Res.second == 0) {
13910 // Map st(0) -> st(7) -> ST0
13911 if (Constraint.size() == 7 && Constraint[0] == '{' &&
13912 tolower(Constraint[1]) == 's' &&
13913 tolower(Constraint[2]) == 't' &&
13914 Constraint[3] == '(' &&
13915 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
13916 Constraint[5] == ')' &&
13917 Constraint[6] == '}') {
13919 Res.first = X86::ST0+Constraint[4]-'0';
13920 Res.second = X86::RFP80RegisterClass;
13924 // GCC allows "st(0)" to be called just plain "st".
13925 if (StringRef("{st}").equals_lower(Constraint)) {
13926 Res.first = X86::ST0;
13927 Res.second = X86::RFP80RegisterClass;
13932 if (StringRef("{flags}").equals_lower(Constraint)) {
13933 Res.first = X86::EFLAGS;
13934 Res.second = X86::CCRRegisterClass;
13938 // 'A' means EAX + EDX.
13939 if (Constraint == "A") {
13940 Res.first = X86::EAX;
13941 Res.second = X86::GR32_ADRegisterClass;
13947 // Otherwise, check to see if this is a register class of the wrong value
13948 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
13949 // turn into {ax},{dx}.
13950 if (Res.second->hasType(VT))
13951 return Res; // Correct type already, nothing to do.
13953 // All of the single-register GCC register classes map their values onto
13954 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
13955 // really want an 8-bit or 32-bit register, map to the appropriate register
13956 // class and return the appropriate register.
13957 if (Res.second == X86::GR16RegisterClass) {
13958 if (VT == MVT::i8) {
13959 unsigned DestReg = 0;
13960 switch (Res.first) {
13962 case X86::AX: DestReg = X86::AL; break;
13963 case X86::DX: DestReg = X86::DL; break;
13964 case X86::CX: DestReg = X86::CL; break;
13965 case X86::BX: DestReg = X86::BL; break;
13968 Res.first = DestReg;
13969 Res.second = X86::GR8RegisterClass;
13971 } else if (VT == MVT::i32) {
13972 unsigned DestReg = 0;
13973 switch (Res.first) {
13975 case X86::AX: DestReg = X86::EAX; break;
13976 case X86::DX: DestReg = X86::EDX; break;
13977 case X86::CX: DestReg = X86::ECX; break;
13978 case X86::BX: DestReg = X86::EBX; break;
13979 case X86::SI: DestReg = X86::ESI; break;
13980 case X86::DI: DestReg = X86::EDI; break;
13981 case X86::BP: DestReg = X86::EBP; break;
13982 case X86::SP: DestReg = X86::ESP; break;
13985 Res.first = DestReg;
13986 Res.second = X86::GR32RegisterClass;
13988 } else if (VT == MVT::i64) {
13989 unsigned DestReg = 0;
13990 switch (Res.first) {
13992 case X86::AX: DestReg = X86::RAX; break;
13993 case X86::DX: DestReg = X86::RDX; break;
13994 case X86::CX: DestReg = X86::RCX; break;
13995 case X86::BX: DestReg = X86::RBX; break;
13996 case X86::SI: DestReg = X86::RSI; break;
13997 case X86::DI: DestReg = X86::RDI; break;
13998 case X86::BP: DestReg = X86::RBP; break;
13999 case X86::SP: DestReg = X86::RSP; break;
14002 Res.first = DestReg;
14003 Res.second = X86::GR64RegisterClass;
14006 } else if (Res.second == X86::FR32RegisterClass ||
14007 Res.second == X86::FR64RegisterClass ||
14008 Res.second == X86::VR128RegisterClass) {
14009 // Handle references to XMM physical registers that got mapped into the
14010 // wrong class. This can happen with constraints like {xmm0} where the
14011 // target independent register mapper will just pick the first match it can
14012 // find, ignoring the required type.
14013 if (VT == MVT::f32)
14014 Res.second = X86::FR32RegisterClass;
14015 else if (VT == MVT::f64)
14016 Res.second = X86::FR64RegisterClass;
14017 else if (X86::VR128RegisterClass->hasType(VT))
14018 Res.second = X86::VR128RegisterClass;