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"
54 #include "llvm/Target/TargetOptions.h"
56 using namespace dwarf;
58 STATISTIC(NumTailCalls, "Number of tail calls");
60 // Forward declarations.
61 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
64 static SDValue Insert128BitVector(SDValue Result,
70 static SDValue Extract128BitVector(SDValue Vec,
75 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
76 /// sets things up to match to an AVX VEXTRACTF128 instruction or a
77 /// simple subregister reference. Idx is an index in the 128 bits we
78 /// want. It need not be aligned to a 128-bit bounday. That makes
79 /// lowering EXTRACT_VECTOR_ELT operations easier.
80 static SDValue Extract128BitVector(SDValue Vec,
84 EVT VT = Vec.getValueType();
85 assert(VT.getSizeInBits() == 256 && "Unexpected vector size!");
86 EVT ElVT = VT.getVectorElementType();
87 int Factor = VT.getSizeInBits()/128;
88 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
89 VT.getVectorNumElements()/Factor);
91 // Extract from UNDEF is UNDEF.
92 if (Vec.getOpcode() == ISD::UNDEF)
93 return DAG.getNode(ISD::UNDEF, dl, ResultVT);
95 if (isa<ConstantSDNode>(Idx)) {
96 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
98 // Extract the relevant 128 bits. Generate an EXTRACT_SUBVECTOR
99 // we can match to VEXTRACTF128.
100 unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
102 // This is the index of the first element of the 128-bit chunk
104 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
107 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
108 SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
117 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
118 /// sets things up to match to an AVX VINSERTF128 instruction or a
119 /// simple superregister reference. Idx is an index in the 128 bits
120 /// we want. It need not be aligned to a 128-bit bounday. That makes
121 /// lowering INSERT_VECTOR_ELT operations easier.
122 static SDValue Insert128BitVector(SDValue Result,
127 if (isa<ConstantSDNode>(Idx)) {
128 EVT VT = Vec.getValueType();
129 assert(VT.getSizeInBits() == 128 && "Unexpected vector size!");
131 EVT ElVT = VT.getVectorElementType();
132 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
133 EVT ResultVT = Result.getValueType();
135 // Insert the relevant 128 bits.
136 unsigned ElemsPerChunk = 128/ElVT.getSizeInBits();
138 // This is the index of the first element of the 128-bit chunk
140 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/128)
143 SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
144 Result = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
152 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
153 const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
154 bool is64Bit = Subtarget->is64Bit();
156 if (Subtarget->isTargetEnvMacho()) {
158 return new X8664_MachoTargetObjectFile();
159 return new TargetLoweringObjectFileMachO();
162 if (Subtarget->isTargetELF())
163 return new TargetLoweringObjectFileELF();
164 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
165 return new TargetLoweringObjectFileCOFF();
166 llvm_unreachable("unknown subtarget type");
169 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
170 : TargetLowering(TM, createTLOF(TM)) {
171 Subtarget = &TM.getSubtarget<X86Subtarget>();
172 X86ScalarSSEf64 = Subtarget->hasXMMInt() || Subtarget->hasAVX();
173 X86ScalarSSEf32 = Subtarget->hasXMM() || Subtarget->hasAVX();
174 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
176 RegInfo = TM.getRegisterInfo();
177 TD = getTargetData();
179 // Set up the TargetLowering object.
180 static MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
182 // X86 is weird, it always uses i8 for shift amounts and setcc results.
183 setBooleanContents(ZeroOrOneBooleanContent);
185 // For 64-bit since we have so many registers use the ILP scheduler, for
186 // 32-bit code use the register pressure specific scheduling.
187 if (Subtarget->is64Bit())
188 setSchedulingPreference(Sched::ILP);
190 setSchedulingPreference(Sched::RegPressure);
191 setStackPointerRegisterToSaveRestore(X86StackPtr);
193 if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
194 // Setup Windows compiler runtime calls.
195 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
196 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
197 setLibcallName(RTLIB::SREM_I64, "_allrem");
198 setLibcallName(RTLIB::UREM_I64, "_aullrem");
199 setLibcallName(RTLIB::MUL_I64, "_allmul");
200 setLibcallName(RTLIB::FPTOUINT_F64_I64, "_ftol2");
201 setLibcallName(RTLIB::FPTOUINT_F32_I64, "_ftol2");
202 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
203 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
204 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
205 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
206 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
207 setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::C);
208 setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::C);
211 if (Subtarget->isTargetDarwin()) {
212 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
213 setUseUnderscoreSetJmp(false);
214 setUseUnderscoreLongJmp(false);
215 } else if (Subtarget->isTargetMingw()) {
216 // MS runtime is weird: it exports _setjmp, but longjmp!
217 setUseUnderscoreSetJmp(true);
218 setUseUnderscoreLongJmp(false);
220 setUseUnderscoreSetJmp(true);
221 setUseUnderscoreLongJmp(true);
224 // Set up the register classes.
225 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
226 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
227 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
228 if (Subtarget->is64Bit())
229 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
231 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
233 // We don't accept any truncstore of integer registers.
234 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
235 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
236 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
237 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
238 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
239 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
241 // SETOEQ and SETUNE require checking two conditions.
242 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
243 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
244 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
245 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
246 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
247 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
249 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
251 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
252 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
253 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
255 if (Subtarget->is64Bit()) {
256 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
257 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
258 } else if (!UseSoftFloat) {
259 // We have an algorithm for SSE2->double, and we turn this into a
260 // 64-bit FILD followed by conditional FADD for other targets.
261 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
262 // We have an algorithm for SSE2, and we turn this into a 64-bit
263 // FILD for other targets.
264 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
267 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
269 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
270 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
273 // SSE has no i16 to fp conversion, only i32
274 if (X86ScalarSSEf32) {
275 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
276 // f32 and f64 cases are Legal, f80 case is not
277 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
279 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
280 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
283 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
284 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
287 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
288 // are Legal, f80 is custom lowered.
289 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
290 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
292 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
294 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
295 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
297 if (X86ScalarSSEf32) {
298 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
299 // f32 and f64 cases are Legal, f80 case is not
300 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
302 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
303 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
306 // Handle FP_TO_UINT by promoting the destination to a larger signed
308 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
309 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
310 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
312 if (Subtarget->is64Bit()) {
313 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
314 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
315 } else if (!UseSoftFloat) {
316 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
317 // Expand FP_TO_UINT into a select.
318 // FIXME: We would like to use a Custom expander here eventually to do
319 // the optimal thing for SSE vs. the default expansion in the legalizer.
320 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
322 // With SSE3 we can use fisttpll to convert to a signed i64; without
323 // SSE, we're stuck with a fistpll.
324 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
327 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
328 if (!X86ScalarSSEf64) {
329 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
330 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
331 if (Subtarget->is64Bit()) {
332 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
333 // Without SSE, i64->f64 goes through memory.
334 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
338 // Scalar integer divide and remainder are lowered to use operations that
339 // produce two results, to match the available instructions. This exposes
340 // the two-result form to trivial CSE, which is able to combine x/y and x%y
341 // into a single instruction.
343 // Scalar integer multiply-high is also lowered to use two-result
344 // operations, to match the available instructions. However, plain multiply
345 // (low) operations are left as Legal, as there are single-result
346 // instructions for this in x86. Using the two-result multiply instructions
347 // when both high and low results are needed must be arranged by dagcombine.
348 for (unsigned i = 0, e = 4; i != e; ++i) {
350 setOperationAction(ISD::MULHS, VT, Expand);
351 setOperationAction(ISD::MULHU, VT, Expand);
352 setOperationAction(ISD::SDIV, VT, Expand);
353 setOperationAction(ISD::UDIV, VT, Expand);
354 setOperationAction(ISD::SREM, VT, Expand);
355 setOperationAction(ISD::UREM, VT, Expand);
357 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
358 setOperationAction(ISD::ADDC, VT, Custom);
359 setOperationAction(ISD::ADDE, VT, Custom);
360 setOperationAction(ISD::SUBC, VT, Custom);
361 setOperationAction(ISD::SUBE, VT, Custom);
364 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
365 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
366 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
367 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
368 if (Subtarget->is64Bit())
369 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
370 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
371 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
372 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
373 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
374 setOperationAction(ISD::FREM , MVT::f32 , Expand);
375 setOperationAction(ISD::FREM , MVT::f64 , Expand);
376 setOperationAction(ISD::FREM , MVT::f80 , Expand);
377 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
379 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
380 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
381 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
382 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
383 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
384 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
385 if (Subtarget->is64Bit()) {
386 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
387 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
390 if (Subtarget->hasPOPCNT()) {
391 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
393 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
394 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
395 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
396 if (Subtarget->is64Bit())
397 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
400 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
401 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
403 // These should be promoted to a larger select which is supported.
404 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
405 // X86 wants to expand cmov itself.
406 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
407 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
408 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
409 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
410 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
411 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
412 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
413 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
414 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
415 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
416 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
417 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
418 if (Subtarget->is64Bit()) {
419 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
420 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
422 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
425 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
426 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
427 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
428 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
429 if (Subtarget->is64Bit())
430 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
431 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
432 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
433 if (Subtarget->is64Bit()) {
434 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
435 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
436 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
437 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
438 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
440 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
441 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
442 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
443 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
444 if (Subtarget->is64Bit()) {
445 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
446 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
447 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
450 if (Subtarget->hasXMM())
451 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
453 setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
454 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
456 // On X86 and X86-64, atomic operations are lowered to locked instructions.
457 // Locked instructions, in turn, have implicit fence semantics (all memory
458 // operations are flushed before issuing the locked instruction, and they
459 // are not buffered), so we can fold away the common pattern of
460 // fence-atomic-fence.
461 setShouldFoldAtomicFences(true);
463 // Expand certain atomics
464 for (unsigned i = 0, e = 4; i != e; ++i) {
466 setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
467 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
468 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
471 if (!Subtarget->is64Bit()) {
472 setOperationAction(ISD::ATOMIC_LOAD, MVT::i64, Custom);
473 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
474 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
475 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
476 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
477 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
478 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
479 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
482 if (Subtarget->hasCmpxchg16b()) {
483 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i128, Custom);
486 // FIXME - use subtarget debug flags
487 if (!Subtarget->isTargetDarwin() &&
488 !Subtarget->isTargetELF() &&
489 !Subtarget->isTargetCygMing()) {
490 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
493 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
494 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
495 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
496 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
497 if (Subtarget->is64Bit()) {
498 setExceptionPointerRegister(X86::RAX);
499 setExceptionSelectorRegister(X86::RDX);
501 setExceptionPointerRegister(X86::EAX);
502 setExceptionSelectorRegister(X86::EDX);
504 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
505 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
507 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
508 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
510 setOperationAction(ISD::TRAP, MVT::Other, Legal);
512 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
513 setOperationAction(ISD::VASTART , MVT::Other, Custom);
514 setOperationAction(ISD::VAEND , MVT::Other, Expand);
515 if (Subtarget->is64Bit()) {
516 setOperationAction(ISD::VAARG , MVT::Other, Custom);
517 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
519 setOperationAction(ISD::VAARG , MVT::Other, Expand);
520 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
523 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
524 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
526 if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
527 setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
528 MVT::i64 : MVT::i32, Custom);
529 else if (EnableSegmentedStacks)
530 setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
531 MVT::i64 : MVT::i32, Custom);
533 setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
534 MVT::i64 : MVT::i32, Expand);
536 if (!UseSoftFloat && X86ScalarSSEf64) {
537 // f32 and f64 use SSE.
538 // Set up the FP register classes.
539 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
540 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
542 // Use ANDPD to simulate FABS.
543 setOperationAction(ISD::FABS , MVT::f64, Custom);
544 setOperationAction(ISD::FABS , MVT::f32, Custom);
546 // Use XORP to simulate FNEG.
547 setOperationAction(ISD::FNEG , MVT::f64, Custom);
548 setOperationAction(ISD::FNEG , MVT::f32, Custom);
550 // Use ANDPD and ORPD to simulate FCOPYSIGN.
551 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
552 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
554 // Lower this to FGETSIGNx86 plus an AND.
555 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
556 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
558 // We don't support sin/cos/fmod
559 setOperationAction(ISD::FSIN , MVT::f64, Expand);
560 setOperationAction(ISD::FCOS , MVT::f64, Expand);
561 setOperationAction(ISD::FSIN , MVT::f32, Expand);
562 setOperationAction(ISD::FCOS , MVT::f32, Expand);
564 // Expand FP immediates into loads from the stack, except for the special
566 addLegalFPImmediate(APFloat(+0.0)); // xorpd
567 addLegalFPImmediate(APFloat(+0.0f)); // xorps
568 } else if (!UseSoftFloat && X86ScalarSSEf32) {
569 // Use SSE for f32, x87 for f64.
570 // Set up the FP register classes.
571 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
572 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
574 // Use ANDPS to simulate FABS.
575 setOperationAction(ISD::FABS , MVT::f32, Custom);
577 // Use XORP to simulate FNEG.
578 setOperationAction(ISD::FNEG , MVT::f32, Custom);
580 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
582 // Use ANDPS and ORPS to simulate FCOPYSIGN.
583 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
584 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
586 // We don't support sin/cos/fmod
587 setOperationAction(ISD::FSIN , MVT::f32, Expand);
588 setOperationAction(ISD::FCOS , MVT::f32, Expand);
590 // Special cases we handle for FP constants.
591 addLegalFPImmediate(APFloat(+0.0f)); // xorps
592 addLegalFPImmediate(APFloat(+0.0)); // FLD0
593 addLegalFPImmediate(APFloat(+1.0)); // FLD1
594 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
595 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
598 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
599 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
601 } else if (!UseSoftFloat) {
602 // f32 and f64 in x87.
603 // Set up the FP register classes.
604 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
605 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
607 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
608 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
609 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
610 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
613 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
614 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
616 addLegalFPImmediate(APFloat(+0.0)); // FLD0
617 addLegalFPImmediate(APFloat(+1.0)); // FLD1
618 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
619 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
620 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
621 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
622 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
623 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
626 // We don't support FMA.
627 setOperationAction(ISD::FMA, MVT::f64, Expand);
628 setOperationAction(ISD::FMA, MVT::f32, Expand);
630 // Long double always uses X87.
632 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
633 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
634 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
636 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
637 addLegalFPImmediate(TmpFlt); // FLD0
639 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
642 APFloat TmpFlt2(+1.0);
643 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
645 addLegalFPImmediate(TmpFlt2); // FLD1
646 TmpFlt2.changeSign();
647 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
651 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
652 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
655 setOperationAction(ISD::FMA, MVT::f80, Expand);
658 // Always use a library call for pow.
659 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
660 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
661 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
663 setOperationAction(ISD::FLOG, MVT::f80, Expand);
664 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
665 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
666 setOperationAction(ISD::FEXP, MVT::f80, Expand);
667 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
669 // First set operation action for all vector types to either promote
670 // (for widening) or expand (for scalarization). Then we will selectively
671 // turn on ones that can be effectively codegen'd.
672 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
673 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
674 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
675 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
676 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
677 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
678 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
679 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
680 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
681 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
682 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
683 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
684 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
685 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
686 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
687 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
688 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
689 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
690 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
691 setOperationAction(ISD::INSERT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
692 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
693 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
694 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
695 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
696 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
697 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
698 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
699 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
700 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
701 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
702 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
703 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
704 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
705 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
706 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
707 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
708 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
709 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
710 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
711 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
712 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
713 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
714 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
715 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
716 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
717 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
718 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
719 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
720 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
721 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
722 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
723 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
724 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
725 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
726 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
727 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
728 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
729 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
730 setTruncStoreAction((MVT::SimpleValueType)VT,
731 (MVT::SimpleValueType)InnerVT, Expand);
732 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
733 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
734 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
737 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
738 // with -msoft-float, disable use of MMX as well.
739 if (!UseSoftFloat && Subtarget->hasMMX()) {
740 addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass);
741 // No operations on x86mmx supported, everything uses intrinsics.
744 // MMX-sized vectors (other than x86mmx) are expected to be expanded
745 // into smaller operations.
746 setOperationAction(ISD::MULHS, MVT::v8i8, Expand);
747 setOperationAction(ISD::MULHS, MVT::v4i16, Expand);
748 setOperationAction(ISD::MULHS, MVT::v2i32, Expand);
749 setOperationAction(ISD::MULHS, MVT::v1i64, Expand);
750 setOperationAction(ISD::AND, MVT::v8i8, Expand);
751 setOperationAction(ISD::AND, MVT::v4i16, Expand);
752 setOperationAction(ISD::AND, MVT::v2i32, Expand);
753 setOperationAction(ISD::AND, MVT::v1i64, Expand);
754 setOperationAction(ISD::OR, MVT::v8i8, Expand);
755 setOperationAction(ISD::OR, MVT::v4i16, Expand);
756 setOperationAction(ISD::OR, MVT::v2i32, Expand);
757 setOperationAction(ISD::OR, MVT::v1i64, Expand);
758 setOperationAction(ISD::XOR, MVT::v8i8, Expand);
759 setOperationAction(ISD::XOR, MVT::v4i16, Expand);
760 setOperationAction(ISD::XOR, MVT::v2i32, Expand);
761 setOperationAction(ISD::XOR, MVT::v1i64, Expand);
762 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Expand);
763 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Expand);
764 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2i32, Expand);
765 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Expand);
766 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
767 setOperationAction(ISD::SELECT, MVT::v8i8, Expand);
768 setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
769 setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
770 setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
771 setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
772 setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
773 setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
774 setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
776 if (!UseSoftFloat && Subtarget->hasXMM()) {
777 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
779 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
780 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
781 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
782 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
783 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
784 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
785 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
786 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
787 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
788 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
789 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
790 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
793 if (!UseSoftFloat && Subtarget->hasXMMInt()) {
794 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
796 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
797 // registers cannot be used even for integer operations.
798 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
799 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
800 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
801 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
803 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
804 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
805 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
806 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
807 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
808 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
809 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
810 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
811 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
812 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
813 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
814 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
815 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
816 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
817 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
818 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
820 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
821 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
822 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
823 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
825 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
826 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
827 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
828 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
829 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
831 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
832 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
833 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
834 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
835 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
837 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
838 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
839 EVT VT = (MVT::SimpleValueType)i;
840 // Do not attempt to custom lower non-power-of-2 vectors
841 if (!isPowerOf2_32(VT.getVectorNumElements()))
843 // Do not attempt to custom lower non-128-bit vectors
844 if (!VT.is128BitVector())
846 setOperationAction(ISD::BUILD_VECTOR,
847 VT.getSimpleVT().SimpleTy, Custom);
848 setOperationAction(ISD::VECTOR_SHUFFLE,
849 VT.getSimpleVT().SimpleTy, Custom);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
851 VT.getSimpleVT().SimpleTy, Custom);
854 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
855 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
856 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
857 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
858 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
859 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
861 if (Subtarget->is64Bit()) {
862 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
863 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
866 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
867 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
868 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
871 // Do not attempt to promote non-128-bit vectors
872 if (!VT.is128BitVector())
875 setOperationAction(ISD::AND, SVT, Promote);
876 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
877 setOperationAction(ISD::OR, SVT, Promote);
878 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
879 setOperationAction(ISD::XOR, SVT, Promote);
880 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
881 setOperationAction(ISD::LOAD, SVT, Promote);
882 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
883 setOperationAction(ISD::SELECT, SVT, Promote);
884 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
887 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
889 // Custom lower v2i64 and v2f64 selects.
890 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
891 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
892 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
893 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
895 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
896 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
899 if (Subtarget->hasSSE41() || Subtarget->hasAVX()) {
900 setOperationAction(ISD::FFLOOR, MVT::f32, Legal);
901 setOperationAction(ISD::FCEIL, MVT::f32, Legal);
902 setOperationAction(ISD::FTRUNC, MVT::f32, Legal);
903 setOperationAction(ISD::FRINT, MVT::f32, Legal);
904 setOperationAction(ISD::FNEARBYINT, MVT::f32, Legal);
905 setOperationAction(ISD::FFLOOR, MVT::f64, Legal);
906 setOperationAction(ISD::FCEIL, MVT::f64, Legal);
907 setOperationAction(ISD::FTRUNC, MVT::f64, Legal);
908 setOperationAction(ISD::FRINT, MVT::f64, Legal);
909 setOperationAction(ISD::FNEARBYINT, MVT::f64, Legal);
911 // FIXME: Do we need to handle scalar-to-vector here?
912 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
914 // Can turn SHL into an integer multiply.
915 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
916 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
918 // i8 and i16 vectors are custom , because the source register and source
919 // source memory operand types are not the same width. f32 vectors are
920 // custom since the immediate controlling the insert encodes additional
922 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
923 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
924 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
925 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
927 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
928 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
929 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
930 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
932 if (Subtarget->is64Bit()) {
933 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
934 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
938 if (Subtarget->hasSSE2() || Subtarget->hasAVX()) {
939 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
940 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
941 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
942 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
944 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
945 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
946 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
948 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
949 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
952 if (Subtarget->hasSSE42() || Subtarget->hasAVX())
953 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
955 if (!UseSoftFloat && Subtarget->hasAVX()) {
956 addRegisterClass(MVT::v32i8, X86::VR256RegisterClass);
957 addRegisterClass(MVT::v16i16, X86::VR256RegisterClass);
958 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
959 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
960 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
961 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
963 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
964 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
965 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
967 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
968 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
969 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
970 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
971 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
972 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
974 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
975 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
976 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
977 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
978 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
979 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
981 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
982 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
983 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
985 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f64, Custom);
986 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i64, Custom);
987 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f32, Custom);
988 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i32, Custom);
989 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i8, Custom);
990 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i16, Custom);
992 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
993 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
994 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
995 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
997 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
998 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
999 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1000 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1002 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1003 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1005 setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
1006 setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
1007 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
1008 setOperationAction(ISD::VSETCC, MVT::v4i64, Custom);
1010 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1011 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1012 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1014 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1015 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1016 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1017 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1019 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1020 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1021 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1022 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1024 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1025 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1026 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1027 // Don't lower v32i8 because there is no 128-bit byte mul
1029 // Custom lower several nodes for 256-bit types.
1030 for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1031 i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++i) {
1032 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
1035 // Extract subvector is special because the value type
1036 // (result) is 128-bit but the source is 256-bit wide.
1037 if (VT.is128BitVector())
1038 setOperationAction(ISD::EXTRACT_SUBVECTOR, SVT, Custom);
1040 // Do not attempt to custom lower other non-256-bit vectors
1041 if (!VT.is256BitVector())
1044 setOperationAction(ISD::BUILD_VECTOR, SVT, Custom);
1045 setOperationAction(ISD::VECTOR_SHUFFLE, SVT, Custom);
1046 setOperationAction(ISD::INSERT_VECTOR_ELT, SVT, Custom);
1047 setOperationAction(ISD::EXTRACT_VECTOR_ELT, SVT, Custom);
1048 setOperationAction(ISD::SCALAR_TO_VECTOR, SVT, Custom);
1049 setOperationAction(ISD::INSERT_SUBVECTOR, SVT, Custom);
1052 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1053 for (unsigned i = (unsigned)MVT::v32i8; i != (unsigned)MVT::v4i64; ++i) {
1054 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
1057 // Do not attempt to promote non-256-bit vectors
1058 if (!VT.is256BitVector())
1061 setOperationAction(ISD::AND, SVT, Promote);
1062 AddPromotedToType (ISD::AND, SVT, MVT::v4i64);
1063 setOperationAction(ISD::OR, SVT, Promote);
1064 AddPromotedToType (ISD::OR, SVT, MVT::v4i64);
1065 setOperationAction(ISD::XOR, SVT, Promote);
1066 AddPromotedToType (ISD::XOR, SVT, MVT::v4i64);
1067 setOperationAction(ISD::LOAD, SVT, Promote);
1068 AddPromotedToType (ISD::LOAD, SVT, MVT::v4i64);
1069 setOperationAction(ISD::SELECT, SVT, Promote);
1070 AddPromotedToType (ISD::SELECT, SVT, MVT::v4i64);
1074 // SIGN_EXTEND_INREGs are evaluated by the extend type. Handle the expansion
1075 // of this type with custom code.
1076 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1077 VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE; VT++) {
1078 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT, Custom);
1081 // We want to custom lower some of our intrinsics.
1082 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1085 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1086 // handle type legalization for these operations here.
1088 // FIXME: We really should do custom legalization for addition and
1089 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1090 // than generic legalization for 64-bit multiplication-with-overflow, though.
1091 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1092 // Add/Sub/Mul with overflow operations are custom lowered.
1094 setOperationAction(ISD::SADDO, VT, Custom);
1095 setOperationAction(ISD::UADDO, VT, Custom);
1096 setOperationAction(ISD::SSUBO, VT, Custom);
1097 setOperationAction(ISD::USUBO, VT, Custom);
1098 setOperationAction(ISD::SMULO, VT, Custom);
1099 setOperationAction(ISD::UMULO, VT, Custom);
1102 // There are no 8-bit 3-address imul/mul instructions
1103 setOperationAction(ISD::SMULO, MVT::i8, Expand);
1104 setOperationAction(ISD::UMULO, MVT::i8, Expand);
1106 if (!Subtarget->is64Bit()) {
1107 // These libcalls are not available in 32-bit.
1108 setLibcallName(RTLIB::SHL_I128, 0);
1109 setLibcallName(RTLIB::SRL_I128, 0);
1110 setLibcallName(RTLIB::SRA_I128, 0);
1113 // We have target-specific dag combine patterns for the following nodes:
1114 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1115 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1116 setTargetDAGCombine(ISD::BUILD_VECTOR);
1117 setTargetDAGCombine(ISD::SELECT);
1118 setTargetDAGCombine(ISD::SHL);
1119 setTargetDAGCombine(ISD::SRA);
1120 setTargetDAGCombine(ISD::SRL);
1121 setTargetDAGCombine(ISD::OR);
1122 setTargetDAGCombine(ISD::AND);
1123 setTargetDAGCombine(ISD::ADD);
1124 setTargetDAGCombine(ISD::SUB);
1125 setTargetDAGCombine(ISD::STORE);
1126 setTargetDAGCombine(ISD::ZERO_EXTEND);
1127 setTargetDAGCombine(ISD::SINT_TO_FP);
1128 if (Subtarget->is64Bit())
1129 setTargetDAGCombine(ISD::MUL);
1131 computeRegisterProperties();
1133 // On Darwin, -Os means optimize for size without hurting performance,
1134 // do not reduce the limit.
1135 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1136 maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1137 maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1138 maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1139 maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1140 maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1141 setPrefLoopAlignment(16);
1142 benefitFromCodePlacementOpt = true;
1144 setPrefFunctionAlignment(4);
1148 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1153 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1154 /// the desired ByVal argument alignment.
1155 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1158 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1159 if (VTy->getBitWidth() == 128)
1161 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1162 unsigned EltAlign = 0;
1163 getMaxByValAlign(ATy->getElementType(), EltAlign);
1164 if (EltAlign > MaxAlign)
1165 MaxAlign = EltAlign;
1166 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1167 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1168 unsigned EltAlign = 0;
1169 getMaxByValAlign(STy->getElementType(i), EltAlign);
1170 if (EltAlign > MaxAlign)
1171 MaxAlign = EltAlign;
1179 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1180 /// function arguments in the caller parameter area. For X86, aggregates
1181 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1182 /// are at 4-byte boundaries.
1183 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty) const {
1184 if (Subtarget->is64Bit()) {
1185 // Max of 8 and alignment of type.
1186 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1193 if (Subtarget->hasXMM())
1194 getMaxByValAlign(Ty, Align);
1198 /// getOptimalMemOpType - Returns the target specific optimal type for load
1199 /// and store operations as a result of memset, memcpy, and memmove
1200 /// lowering. If DstAlign is zero that means it's safe to destination
1201 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1202 /// means there isn't a need to check it against alignment requirement,
1203 /// probably because the source does not need to be loaded. If
1204 /// 'NonScalarIntSafe' is true, that means it's safe to return a
1205 /// non-scalar-integer type, e.g. empty string source, constant, or loaded
1206 /// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
1207 /// constant so it does not need to be loaded.
1208 /// It returns EVT::Other if the type should be determined using generic
1209 /// target-independent logic.
1211 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1212 unsigned DstAlign, unsigned SrcAlign,
1213 bool NonScalarIntSafe,
1215 MachineFunction &MF) const {
1216 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1217 // linux. This is because the stack realignment code can't handle certain
1218 // cases like PR2962. This should be removed when PR2962 is fixed.
1219 const Function *F = MF.getFunction();
1220 if (NonScalarIntSafe &&
1221 !F->hasFnAttr(Attribute::NoImplicitFloat)) {
1223 (Subtarget->isUnalignedMemAccessFast() ||
1224 ((DstAlign == 0 || DstAlign >= 16) &&
1225 (SrcAlign == 0 || SrcAlign >= 16))) &&
1226 Subtarget->getStackAlignment() >= 16) {
1227 if (Subtarget->hasSSE2())
1229 if (Subtarget->hasSSE1())
1231 } else if (!MemcpyStrSrc && Size >= 8 &&
1232 !Subtarget->is64Bit() &&
1233 Subtarget->getStackAlignment() >= 8 &&
1234 Subtarget->hasXMMInt()) {
1235 // Do not use f64 to lower memcpy if source is string constant. It's
1236 // better to use i32 to avoid the loads.
1240 if (Subtarget->is64Bit() && Size >= 8)
1245 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1246 /// current function. The returned value is a member of the
1247 /// MachineJumpTableInfo::JTEntryKind enum.
1248 unsigned X86TargetLowering::getJumpTableEncoding() const {
1249 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1251 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1252 Subtarget->isPICStyleGOT())
1253 return MachineJumpTableInfo::EK_Custom32;
1255 // Otherwise, use the normal jump table encoding heuristics.
1256 return TargetLowering::getJumpTableEncoding();
1260 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1261 const MachineBasicBlock *MBB,
1262 unsigned uid,MCContext &Ctx) const{
1263 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1264 Subtarget->isPICStyleGOT());
1265 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1267 return MCSymbolRefExpr::Create(MBB->getSymbol(),
1268 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1271 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1273 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1274 SelectionDAG &DAG) const {
1275 if (!Subtarget->is64Bit())
1276 // This doesn't have DebugLoc associated with it, but is not really the
1277 // same as a Register.
1278 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc(), getPointerTy());
1282 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1283 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1285 const MCExpr *X86TargetLowering::
1286 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1287 MCContext &Ctx) const {
1288 // X86-64 uses RIP relative addressing based on the jump table label.
1289 if (Subtarget->isPICStyleRIPRel())
1290 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1292 // Otherwise, the reference is relative to the PIC base.
1293 return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
1296 // FIXME: Why this routine is here? Move to RegInfo!
1297 std::pair<const TargetRegisterClass*, uint8_t>
1298 X86TargetLowering::findRepresentativeClass(EVT VT) const{
1299 const TargetRegisterClass *RRC = 0;
1301 switch (VT.getSimpleVT().SimpleTy) {
1303 return TargetLowering::findRepresentativeClass(VT);
1304 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1305 RRC = (Subtarget->is64Bit()
1306 ? X86::GR64RegisterClass : X86::GR32RegisterClass);
1309 RRC = X86::VR64RegisterClass;
1311 case MVT::f32: case MVT::f64:
1312 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1313 case MVT::v4f32: case MVT::v2f64:
1314 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1316 RRC = X86::VR128RegisterClass;
1319 return std::make_pair(RRC, Cost);
1322 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1323 unsigned &Offset) const {
1324 if (!Subtarget->isTargetLinux())
1327 if (Subtarget->is64Bit()) {
1328 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
1330 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
1343 //===----------------------------------------------------------------------===//
1344 // Return Value Calling Convention Implementation
1345 //===----------------------------------------------------------------------===//
1347 #include "X86GenCallingConv.inc"
1350 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
1351 MachineFunction &MF, bool isVarArg,
1352 const SmallVectorImpl<ISD::OutputArg> &Outs,
1353 LLVMContext &Context) const {
1354 SmallVector<CCValAssign, 16> RVLocs;
1355 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1357 return CCInfo.CheckReturn(Outs, RetCC_X86);
1361 X86TargetLowering::LowerReturn(SDValue Chain,
1362 CallingConv::ID CallConv, bool isVarArg,
1363 const SmallVectorImpl<ISD::OutputArg> &Outs,
1364 const SmallVectorImpl<SDValue> &OutVals,
1365 DebugLoc dl, SelectionDAG &DAG) const {
1366 MachineFunction &MF = DAG.getMachineFunction();
1367 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1369 SmallVector<CCValAssign, 16> RVLocs;
1370 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1371 RVLocs, *DAG.getContext());
1372 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1374 // Add the regs to the liveout set for the function.
1375 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1376 for (unsigned i = 0; i != RVLocs.size(); ++i)
1377 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1378 MRI.addLiveOut(RVLocs[i].getLocReg());
1382 SmallVector<SDValue, 6> RetOps;
1383 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1384 // Operand #1 = Bytes To Pop
1385 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(),
1388 // Copy the result values into the output registers.
1389 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1390 CCValAssign &VA = RVLocs[i];
1391 assert(VA.isRegLoc() && "Can only return in registers!");
1392 SDValue ValToCopy = OutVals[i];
1393 EVT ValVT = ValToCopy.getValueType();
1395 // If this is x86-64, and we disabled SSE, we can't return FP values,
1396 // or SSE or MMX vectors.
1397 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
1398 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
1399 (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
1400 report_fatal_error("SSE register return with SSE disabled");
1402 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
1403 // llvm-gcc has never done it right and no one has noticed, so this
1404 // should be OK for now.
1405 if (ValVT == MVT::f64 &&
1406 (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
1407 report_fatal_error("SSE2 register return with SSE2 disabled");
1409 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1410 // the RET instruction and handled by the FP Stackifier.
1411 if (VA.getLocReg() == X86::ST0 ||
1412 VA.getLocReg() == X86::ST1) {
1413 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1414 // change the value to the FP stack register class.
1415 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1416 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1417 RetOps.push_back(ValToCopy);
1418 // Don't emit a copytoreg.
1422 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1423 // which is returned in RAX / RDX.
1424 if (Subtarget->is64Bit()) {
1425 if (ValVT == MVT::x86mmx) {
1426 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1427 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
1428 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
1430 // If we don't have SSE2 available, convert to v4f32 so the generated
1431 // register is legal.
1432 if (!Subtarget->hasSSE2())
1433 ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
1438 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1439 Flag = Chain.getValue(1);
1442 // The x86-64 ABI for returning structs by value requires that we copy
1443 // the sret argument into %rax for the return. We saved the argument into
1444 // a virtual register in the entry block, so now we copy the value out
1446 if (Subtarget->is64Bit() &&
1447 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1448 MachineFunction &MF = DAG.getMachineFunction();
1449 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1450 unsigned Reg = FuncInfo->getSRetReturnReg();
1452 "SRetReturnReg should have been set in LowerFormalArguments().");
1453 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1455 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1456 Flag = Chain.getValue(1);
1458 // RAX now acts like a return value.
1459 MRI.addLiveOut(X86::RAX);
1462 RetOps[0] = Chain; // Update chain.
1464 // Add the flag if we have it.
1466 RetOps.push_back(Flag);
1468 return DAG.getNode(X86ISD::RET_FLAG, dl,
1469 MVT::Other, &RetOps[0], RetOps.size());
1472 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
1473 if (N->getNumValues() != 1)
1475 if (!N->hasNUsesOfValue(1, 0))
1478 SDNode *Copy = *N->use_begin();
1479 if (Copy->getOpcode() != ISD::CopyToReg &&
1480 Copy->getOpcode() != ISD::FP_EXTEND)
1483 bool HasRet = false;
1484 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
1486 if (UI->getOpcode() != X86ISD::RET_FLAG)
1495 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
1496 ISD::NodeType ExtendKind) const {
1498 // TODO: Is this also valid on 32-bit?
1499 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
1500 ReturnMVT = MVT::i8;
1502 ReturnMVT = MVT::i32;
1504 EVT MinVT = getRegisterType(Context, ReturnMVT);
1505 return VT.bitsLT(MinVT) ? MinVT : VT;
1508 /// LowerCallResult - Lower the result values of a call into the
1509 /// appropriate copies out of appropriate physical registers.
1512 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1513 CallingConv::ID CallConv, bool isVarArg,
1514 const SmallVectorImpl<ISD::InputArg> &Ins,
1515 DebugLoc dl, SelectionDAG &DAG,
1516 SmallVectorImpl<SDValue> &InVals) const {
1518 // Assign locations to each value returned by this call.
1519 SmallVector<CCValAssign, 16> RVLocs;
1520 bool Is64Bit = Subtarget->is64Bit();
1521 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
1522 getTargetMachine(), RVLocs, *DAG.getContext());
1523 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1525 // Copy all of the result registers out of their specified physreg.
1526 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1527 CCValAssign &VA = RVLocs[i];
1528 EVT CopyVT = VA.getValVT();
1530 // If this is x86-64, and we disabled SSE, we can't return FP values
1531 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1532 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
1533 report_fatal_error("SSE register return with SSE disabled");
1538 // If this is a call to a function that returns an fp value on the floating
1539 // point stack, we must guarantee the the value is popped from the stack, so
1540 // a CopyFromReg is not good enough - the copy instruction may be eliminated
1541 // if the return value is not used. We use the FpPOP_RETVAL instruction
1543 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
1544 // If we prefer to use the value in xmm registers, copy it out as f80 and
1545 // use a truncate to move it from fp stack reg to xmm reg.
1546 if (isScalarFPTypeInSSEReg(VA.getValVT())) CopyVT = MVT::f80;
1547 SDValue Ops[] = { Chain, InFlag };
1548 Chain = SDValue(DAG.getMachineNode(X86::FpPOP_RETVAL, dl, CopyVT,
1549 MVT::Other, MVT::Glue, Ops, 2), 1);
1550 Val = Chain.getValue(0);
1552 // Round the f80 to the right size, which also moves it to the appropriate
1554 if (CopyVT != VA.getValVT())
1555 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1556 // This truncation won't change the value.
1557 DAG.getIntPtrConstant(1));
1559 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1560 CopyVT, InFlag).getValue(1);
1561 Val = Chain.getValue(0);
1563 InFlag = Chain.getValue(2);
1564 InVals.push_back(Val);
1571 //===----------------------------------------------------------------------===//
1572 // C & StdCall & Fast Calling Convention implementation
1573 //===----------------------------------------------------------------------===//
1574 // StdCall calling convention seems to be standard for many Windows' API
1575 // routines and around. It differs from C calling convention just a little:
1576 // callee should clean up the stack, not caller. Symbols should be also
1577 // decorated in some fancy way :) It doesn't support any vector arguments.
1578 // For info on fast calling convention see Fast Calling Convention (tail call)
1579 // implementation LowerX86_32FastCCCallTo.
1581 /// CallIsStructReturn - Determines whether a call uses struct return
1583 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1587 return Outs[0].Flags.isSRet();
1590 /// ArgsAreStructReturn - Determines whether a function uses struct
1591 /// return semantics.
1593 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1597 return Ins[0].Flags.isSRet();
1600 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1601 /// by "Src" to address "Dst" with size and alignment information specified by
1602 /// the specific parameter attribute. The copy will be passed as a byval
1603 /// function parameter.
1605 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1606 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1608 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1610 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1611 /*isVolatile*/false, /*AlwaysInline=*/true,
1612 MachinePointerInfo(), MachinePointerInfo());
1615 /// IsTailCallConvention - Return true if the calling convention is one that
1616 /// supports tail call optimization.
1617 static bool IsTailCallConvention(CallingConv::ID CC) {
1618 return (CC == CallingConv::Fast || CC == CallingConv::GHC);
1621 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
1622 if (!CI->isTailCall())
1626 CallingConv::ID CalleeCC = CS.getCallingConv();
1627 if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
1633 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1634 /// a tailcall target by changing its ABI.
1635 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1636 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
1640 X86TargetLowering::LowerMemArgument(SDValue Chain,
1641 CallingConv::ID CallConv,
1642 const SmallVectorImpl<ISD::InputArg> &Ins,
1643 DebugLoc dl, SelectionDAG &DAG,
1644 const CCValAssign &VA,
1645 MachineFrameInfo *MFI,
1647 // Create the nodes corresponding to a load from this parameter slot.
1648 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1649 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1650 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1653 // If value is passed by pointer we have address passed instead of the value
1655 if (VA.getLocInfo() == CCValAssign::Indirect)
1656 ValVT = VA.getLocVT();
1658 ValVT = VA.getValVT();
1660 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1661 // changed with more analysis.
1662 // In case of tail call optimization mark all arguments mutable. Since they
1663 // could be overwritten by lowering of arguments in case of a tail call.
1664 if (Flags.isByVal()) {
1665 unsigned Bytes = Flags.getByValSize();
1666 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
1667 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
1668 return DAG.getFrameIndex(FI, getPointerTy());
1670 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1671 VA.getLocMemOffset(), isImmutable);
1672 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1673 return DAG.getLoad(ValVT, dl, Chain, FIN,
1674 MachinePointerInfo::getFixedStack(FI),
1680 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1681 CallingConv::ID CallConv,
1683 const SmallVectorImpl<ISD::InputArg> &Ins,
1686 SmallVectorImpl<SDValue> &InVals)
1688 MachineFunction &MF = DAG.getMachineFunction();
1689 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1691 const Function* Fn = MF.getFunction();
1692 if (Fn->hasExternalLinkage() &&
1693 Subtarget->isTargetCygMing() &&
1694 Fn->getName() == "main")
1695 FuncInfo->setForceFramePointer(true);
1697 MachineFrameInfo *MFI = MF.getFrameInfo();
1698 bool Is64Bit = Subtarget->is64Bit();
1699 bool IsWin64 = Subtarget->isTargetWin64();
1701 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
1702 "Var args not supported with calling convention fastcc or ghc");
1704 // Assign locations to all of the incoming arguments.
1705 SmallVector<CCValAssign, 16> ArgLocs;
1706 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
1707 ArgLocs, *DAG.getContext());
1709 // Allocate shadow area for Win64
1711 CCInfo.AllocateStack(32, 8);
1714 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
1716 unsigned LastVal = ~0U;
1718 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1719 CCValAssign &VA = ArgLocs[i];
1720 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1722 assert(VA.getValNo() != LastVal &&
1723 "Don't support value assigned to multiple locs yet");
1724 LastVal = VA.getValNo();
1726 if (VA.isRegLoc()) {
1727 EVT RegVT = VA.getLocVT();
1728 TargetRegisterClass *RC = NULL;
1729 if (RegVT == MVT::i32)
1730 RC = X86::GR32RegisterClass;
1731 else if (Is64Bit && RegVT == MVT::i64)
1732 RC = X86::GR64RegisterClass;
1733 else if (RegVT == MVT::f32)
1734 RC = X86::FR32RegisterClass;
1735 else if (RegVT == MVT::f64)
1736 RC = X86::FR64RegisterClass;
1737 else if (RegVT.isVector() && RegVT.getSizeInBits() == 256)
1738 RC = X86::VR256RegisterClass;
1739 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1740 RC = X86::VR128RegisterClass;
1741 else if (RegVT == MVT::x86mmx)
1742 RC = X86::VR64RegisterClass;
1744 llvm_unreachable("Unknown argument type!");
1746 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1747 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1749 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1750 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1752 if (VA.getLocInfo() == CCValAssign::SExt)
1753 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1754 DAG.getValueType(VA.getValVT()));
1755 else if (VA.getLocInfo() == CCValAssign::ZExt)
1756 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1757 DAG.getValueType(VA.getValVT()));
1758 else if (VA.getLocInfo() == CCValAssign::BCvt)
1759 ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
1761 if (VA.isExtInLoc()) {
1762 // Handle MMX values passed in XMM regs.
1763 if (RegVT.isVector()) {
1764 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
1767 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1770 assert(VA.isMemLoc());
1771 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1774 // If value is passed via pointer - do a load.
1775 if (VA.getLocInfo() == CCValAssign::Indirect)
1776 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
1777 MachinePointerInfo(), false, false, 0);
1779 InVals.push_back(ArgValue);
1782 // The x86-64 ABI for returning structs by value requires that we copy
1783 // the sret argument into %rax for the return. Save the argument into
1784 // a virtual register so that we can access it from the return points.
1785 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1786 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1787 unsigned Reg = FuncInfo->getSRetReturnReg();
1789 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1790 FuncInfo->setSRetReturnReg(Reg);
1792 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1793 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1796 unsigned StackSize = CCInfo.getNextStackOffset();
1797 // Align stack specially for tail calls.
1798 if (FuncIsMadeTailCallSafe(CallConv))
1799 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1801 // If the function takes variable number of arguments, make a frame index for
1802 // the start of the first vararg value... for expansion of llvm.va_start.
1804 if (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
1805 CallConv != CallingConv::X86_ThisCall)) {
1806 FuncInfo->setVarArgsFrameIndex(MFI->CreateFixedObject(1, StackSize,true));
1809 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1811 // FIXME: We should really autogenerate these arrays
1812 static const unsigned GPR64ArgRegsWin64[] = {
1813 X86::RCX, X86::RDX, X86::R8, X86::R9
1815 static const unsigned GPR64ArgRegs64Bit[] = {
1816 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1818 static const unsigned XMMArgRegs64Bit[] = {
1819 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1820 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1822 const unsigned *GPR64ArgRegs;
1823 unsigned NumXMMRegs = 0;
1826 // The XMM registers which might contain var arg parameters are shadowed
1827 // in their paired GPR. So we only need to save the GPR to their home
1829 TotalNumIntRegs = 4;
1830 GPR64ArgRegs = GPR64ArgRegsWin64;
1832 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1833 GPR64ArgRegs = GPR64ArgRegs64Bit;
1835 NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs64Bit, TotalNumXMMRegs);
1837 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1840 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1841 assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
1842 "SSE register cannot be used when SSE is disabled!");
1843 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1844 "SSE register cannot be used when SSE is disabled!");
1845 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
1846 // Kernel mode asks for SSE to be disabled, so don't push them
1848 TotalNumXMMRegs = 0;
1851 const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
1852 // Get to the caller-allocated home save location. Add 8 to account
1853 // for the return address.
1854 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
1855 FuncInfo->setRegSaveFrameIndex(
1856 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
1857 // Fixup to set vararg frame on shadow area (4 x i64).
1859 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
1861 // For X86-64, if there are vararg parameters that are passed via
1862 // registers, then we must store them to their spots on the stack so they
1863 // may be loaded by deferencing the result of va_next.
1864 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
1865 FuncInfo->setVarArgsFPOffset(TotalNumIntRegs * 8 + NumXMMRegs * 16);
1866 FuncInfo->setRegSaveFrameIndex(
1867 MFI->CreateStackObject(TotalNumIntRegs * 8 + TotalNumXMMRegs * 16, 16,
1871 // Store the integer parameter registers.
1872 SmallVector<SDValue, 8> MemOps;
1873 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
1875 unsigned Offset = FuncInfo->getVarArgsGPOffset();
1876 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1877 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1878 DAG.getIntPtrConstant(Offset));
1879 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1880 X86::GR64RegisterClass);
1881 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1883 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1884 MachinePointerInfo::getFixedStack(
1885 FuncInfo->getRegSaveFrameIndex(), Offset),
1887 MemOps.push_back(Store);
1891 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1892 // Now store the XMM (fp + vector) parameter registers.
1893 SmallVector<SDValue, 11> SaveXMMOps;
1894 SaveXMMOps.push_back(Chain);
1896 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1897 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1898 SaveXMMOps.push_back(ALVal);
1900 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1901 FuncInfo->getRegSaveFrameIndex()));
1902 SaveXMMOps.push_back(DAG.getIntPtrConstant(
1903 FuncInfo->getVarArgsFPOffset()));
1905 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1906 unsigned VReg = MF.addLiveIn(XMMArgRegs64Bit[NumXMMRegs],
1907 X86::VR128RegisterClass);
1908 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1909 SaveXMMOps.push_back(Val);
1911 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1913 &SaveXMMOps[0], SaveXMMOps.size()));
1916 if (!MemOps.empty())
1917 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1918 &MemOps[0], MemOps.size());
1922 // Some CCs need callee pop.
1923 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt)) {
1924 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
1926 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
1927 // If this is an sret function, the return should pop the hidden pointer.
1928 if (!Is64Bit && !IsTailCallConvention(CallConv) && ArgsAreStructReturn(Ins))
1929 FuncInfo->setBytesToPopOnReturn(4);
1933 // RegSaveFrameIndex is X86-64 only.
1934 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
1935 if (CallConv == CallingConv::X86_FastCall ||
1936 CallConv == CallingConv::X86_ThisCall)
1937 // fastcc functions can't have varargs.
1938 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
1941 FuncInfo->setArgumentStackSize(StackSize);
1947 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1948 SDValue StackPtr, SDValue Arg,
1949 DebugLoc dl, SelectionDAG &DAG,
1950 const CCValAssign &VA,
1951 ISD::ArgFlagsTy Flags) const {
1952 unsigned LocMemOffset = VA.getLocMemOffset();
1953 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1954 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1955 if (Flags.isByVal())
1956 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1958 return DAG.getStore(Chain, dl, Arg, PtrOff,
1959 MachinePointerInfo::getStack(LocMemOffset),
1963 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1964 /// optimization is performed and it is required.
1966 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1967 SDValue &OutRetAddr, SDValue Chain,
1968 bool IsTailCall, bool Is64Bit,
1969 int FPDiff, DebugLoc dl) const {
1970 // Adjust the Return address stack slot.
1971 EVT VT = getPointerTy();
1972 OutRetAddr = getReturnAddressFrameIndex(DAG);
1974 // Load the "old" Return address.
1975 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
1977 return SDValue(OutRetAddr.getNode(), 1);
1980 /// EmitTailCallStoreRetAddr - Emit a store of the return address if tail call
1981 /// optimization is performed and it is required (FPDiff!=0).
1983 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1984 SDValue Chain, SDValue RetAddrFrIdx,
1985 bool Is64Bit, int FPDiff, DebugLoc dl) {
1986 // Store the return address to the appropriate stack slot.
1987 if (!FPDiff) return Chain;
1988 // Calculate the new stack slot for the return address.
1989 int SlotSize = Is64Bit ? 8 : 4;
1990 int NewReturnAddrFI =
1991 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false);
1992 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1993 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1994 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1995 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2001 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
2002 CallingConv::ID CallConv, bool isVarArg,
2004 const SmallVectorImpl<ISD::OutputArg> &Outs,
2005 const SmallVectorImpl<SDValue> &OutVals,
2006 const SmallVectorImpl<ISD::InputArg> &Ins,
2007 DebugLoc dl, SelectionDAG &DAG,
2008 SmallVectorImpl<SDValue> &InVals) const {
2009 MachineFunction &MF = DAG.getMachineFunction();
2010 bool Is64Bit = Subtarget->is64Bit();
2011 bool IsWin64 = Subtarget->isTargetWin64();
2012 bool IsStructRet = CallIsStructReturn(Outs);
2013 bool IsSibcall = false;
2016 // Check if it's really possible to do a tail call.
2017 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2018 isVarArg, IsStructRet, MF.getFunction()->hasStructRetAttr(),
2019 Outs, OutVals, Ins, DAG);
2021 // Sibcalls are automatically detected tailcalls which do not require
2023 if (!GuaranteedTailCallOpt && isTailCall)
2030 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2031 "Var args not supported with calling convention fastcc or ghc");
2033 // Analyze operands of the call, assigning locations to each operand.
2034 SmallVector<CCValAssign, 16> ArgLocs;
2035 CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(),
2036 ArgLocs, *DAG.getContext());
2038 // Allocate shadow area for Win64
2040 CCInfo.AllocateStack(32, 8);
2043 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2045 // Get a count of how many bytes are to be pushed on the stack.
2046 unsigned NumBytes = CCInfo.getNextStackOffset();
2048 // This is a sibcall. The memory operands are available in caller's
2049 // own caller's stack.
2051 else if (GuaranteedTailCallOpt && IsTailCallConvention(CallConv))
2052 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2055 if (isTailCall && !IsSibcall) {
2056 // Lower arguments at fp - stackoffset + fpdiff.
2057 unsigned NumBytesCallerPushed =
2058 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
2059 FPDiff = NumBytesCallerPushed - NumBytes;
2061 // Set the delta of movement of the returnaddr stackslot.
2062 // But only set if delta is greater than previous delta.
2063 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
2064 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
2068 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
2070 SDValue RetAddrFrIdx;
2071 // Load return address for tail calls.
2072 if (isTailCall && FPDiff)
2073 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2074 Is64Bit, FPDiff, dl);
2076 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
2077 SmallVector<SDValue, 8> MemOpChains;
2080 // Walk the register/memloc assignments, inserting copies/loads. In the case
2081 // of tail call optimization arguments are handle later.
2082 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2083 CCValAssign &VA = ArgLocs[i];
2084 EVT RegVT = VA.getLocVT();
2085 SDValue Arg = OutVals[i];
2086 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2087 bool isByVal = Flags.isByVal();
2089 // Promote the value if needed.
2090 switch (VA.getLocInfo()) {
2091 default: llvm_unreachable("Unknown loc info!");
2092 case CCValAssign::Full: break;
2093 case CCValAssign::SExt:
2094 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
2096 case CCValAssign::ZExt:
2097 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
2099 case CCValAssign::AExt:
2100 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
2101 // Special case: passing MMX values in XMM registers.
2102 Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
2103 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
2104 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
2106 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
2108 case CCValAssign::BCvt:
2109 Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
2111 case CCValAssign::Indirect: {
2112 // Store the argument.
2113 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
2114 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
2115 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
2116 MachinePointerInfo::getFixedStack(FI),
2123 if (VA.isRegLoc()) {
2124 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
2125 if (isVarArg && IsWin64) {
2126 // Win64 ABI requires argument XMM reg to be copied to the corresponding
2127 // shadow reg if callee is a varargs function.
2128 unsigned ShadowReg = 0;
2129 switch (VA.getLocReg()) {
2130 case X86::XMM0: ShadowReg = X86::RCX; break;
2131 case X86::XMM1: ShadowReg = X86::RDX; break;
2132 case X86::XMM2: ShadowReg = X86::R8; break;
2133 case X86::XMM3: ShadowReg = X86::R9; break;
2136 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
2138 } else if (!IsSibcall && (!isTailCall || isByVal)) {
2139 assert(VA.isMemLoc());
2140 if (StackPtr.getNode() == 0)
2141 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
2142 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
2143 dl, DAG, VA, Flags));
2147 if (!MemOpChains.empty())
2148 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2149 &MemOpChains[0], MemOpChains.size());
2151 // Build a sequence of copy-to-reg nodes chained together with token chain
2152 // and flag operands which copy the outgoing args into registers.
2154 // Tail call byval lowering might overwrite argument registers so in case of
2155 // tail call optimization the copies to registers are lowered later.
2157 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2158 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2159 RegsToPass[i].second, InFlag);
2160 InFlag = Chain.getValue(1);
2163 if (Subtarget->isPICStyleGOT()) {
2164 // ELF / PIC requires GOT in the EBX register before function calls via PLT
2167 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
2168 DAG.getNode(X86ISD::GlobalBaseReg,
2169 DebugLoc(), getPointerTy()),
2171 InFlag = Chain.getValue(1);
2173 // If we are tail calling and generating PIC/GOT style code load the
2174 // address of the callee into ECX. The value in ecx is used as target of
2175 // the tail jump. This is done to circumvent the ebx/callee-saved problem
2176 // for tail calls on PIC/GOT architectures. Normally we would just put the
2177 // address of GOT into ebx and then call target@PLT. But for tail calls
2178 // ebx would be restored (since ebx is callee saved) before jumping to the
2181 // Note: The actual moving to ECX is done further down.
2182 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2183 if (G && !G->getGlobal()->hasHiddenVisibility() &&
2184 !G->getGlobal()->hasProtectedVisibility())
2185 Callee = LowerGlobalAddress(Callee, DAG);
2186 else if (isa<ExternalSymbolSDNode>(Callee))
2187 Callee = LowerExternalSymbol(Callee, DAG);
2191 if (Is64Bit && isVarArg && !IsWin64) {
2192 // From AMD64 ABI document:
2193 // For calls that may call functions that use varargs or stdargs
2194 // (prototype-less calls or calls to functions containing ellipsis (...) in
2195 // the declaration) %al is used as hidden argument to specify the number
2196 // of SSE registers used. The contents of %al do not need to match exactly
2197 // the number of registers, but must be an ubound on the number of SSE
2198 // registers used and is in the range 0 - 8 inclusive.
2200 // Count the number of XMM registers allocated.
2201 static const unsigned XMMArgRegs[] = {
2202 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2203 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2205 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
2206 assert((Subtarget->hasXMM() || !NumXMMRegs)
2207 && "SSE registers cannot be used when SSE is disabled");
2209 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
2210 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
2211 InFlag = Chain.getValue(1);
2215 // For tail calls lower the arguments to the 'real' stack slot.
2217 // Force all the incoming stack arguments to be loaded from the stack
2218 // before any new outgoing arguments are stored to the stack, because the
2219 // outgoing stack slots may alias the incoming argument stack slots, and
2220 // the alias isn't otherwise explicit. This is slightly more conservative
2221 // than necessary, because it means that each store effectively depends
2222 // on every argument instead of just those arguments it would clobber.
2223 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
2225 SmallVector<SDValue, 8> MemOpChains2;
2228 // Do not flag preceding copytoreg stuff together with the following stuff.
2230 if (GuaranteedTailCallOpt) {
2231 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2232 CCValAssign &VA = ArgLocs[i];
2235 assert(VA.isMemLoc());
2236 SDValue Arg = OutVals[i];
2237 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2238 // Create frame index.
2239 int32_t Offset = VA.getLocMemOffset()+FPDiff;
2240 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
2241 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
2242 FIN = DAG.getFrameIndex(FI, getPointerTy());
2244 if (Flags.isByVal()) {
2245 // Copy relative to framepointer.
2246 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2247 if (StackPtr.getNode() == 0)
2248 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2250 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2252 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2256 // Store relative to framepointer.
2257 MemOpChains2.push_back(
2258 DAG.getStore(ArgChain, dl, Arg, FIN,
2259 MachinePointerInfo::getFixedStack(FI),
2265 if (!MemOpChains2.empty())
2266 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2267 &MemOpChains2[0], MemOpChains2.size());
2269 // Copy arguments to their registers.
2270 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2271 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2272 RegsToPass[i].second, InFlag);
2273 InFlag = Chain.getValue(1);
2277 // Store the return address to the appropriate stack slot.
2278 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2282 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2283 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2284 // In the 64-bit large code model, we have to make all calls
2285 // through a register, since the call instruction's 32-bit
2286 // pc-relative offset may not be large enough to hold the whole
2288 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2289 // If the callee is a GlobalAddress node (quite common, every direct call
2290 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2293 // We should use extra load for direct calls to dllimported functions in
2295 const GlobalValue *GV = G->getGlobal();
2296 if (!GV->hasDLLImportLinkage()) {
2297 unsigned char OpFlags = 0;
2298 bool ExtraLoad = false;
2299 unsigned WrapperKind = ISD::DELETED_NODE;
2301 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2302 // external symbols most go through the PLT in PIC mode. If the symbol
2303 // has hidden or protected visibility, or if it is static or local, then
2304 // we don't need to use the PLT - we can directly call it.
2305 if (Subtarget->isTargetELF() &&
2306 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2307 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2308 OpFlags = X86II::MO_PLT;
2309 } else if (Subtarget->isPICStyleStubAny() &&
2310 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2311 (!Subtarget->getTargetTriple().isMacOSX() ||
2312 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2313 // PC-relative references to external symbols should go through $stub,
2314 // unless we're building with the leopard linker or later, which
2315 // automatically synthesizes these stubs.
2316 OpFlags = X86II::MO_DARWIN_STUB;
2317 } else if (Subtarget->isPICStyleRIPRel() &&
2318 isa<Function>(GV) &&
2319 cast<Function>(GV)->hasFnAttr(Attribute::NonLazyBind)) {
2320 // If the function is marked as non-lazy, generate an indirect call
2321 // which loads from the GOT directly. This avoids runtime overhead
2322 // at the cost of eager binding (and one extra byte of encoding).
2323 OpFlags = X86II::MO_GOTPCREL;
2324 WrapperKind = X86ISD::WrapperRIP;
2328 Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(),
2329 G->getOffset(), OpFlags);
2331 // Add a wrapper if needed.
2332 if (WrapperKind != ISD::DELETED_NODE)
2333 Callee = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Callee);
2334 // Add extra indirection if needed.
2336 Callee = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Callee,
2337 MachinePointerInfo::getGOT(),
2340 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2341 unsigned char OpFlags = 0;
2343 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
2344 // external symbols should go through the PLT.
2345 if (Subtarget->isTargetELF() &&
2346 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2347 OpFlags = X86II::MO_PLT;
2348 } else if (Subtarget->isPICStyleStubAny() &&
2349 (!Subtarget->getTargetTriple().isMacOSX() ||
2350 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
2351 // PC-relative references to external symbols should go through $stub,
2352 // unless we're building with the leopard linker or later, which
2353 // automatically synthesizes these stubs.
2354 OpFlags = X86II::MO_DARWIN_STUB;
2357 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2361 // Returns a chain & a flag for retval copy to use.
2362 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
2363 SmallVector<SDValue, 8> Ops;
2365 if (!IsSibcall && isTailCall) {
2366 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2367 DAG.getIntPtrConstant(0, true), InFlag);
2368 InFlag = Chain.getValue(1);
2371 Ops.push_back(Chain);
2372 Ops.push_back(Callee);
2375 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2377 // Add argument registers to the end of the list so that they are known live
2379 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2380 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2381 RegsToPass[i].second.getValueType()));
2383 // Add an implicit use GOT pointer in EBX.
2384 if (!isTailCall && Subtarget->isPICStyleGOT())
2385 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2387 // Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
2388 if (Is64Bit && isVarArg && !IsWin64)
2389 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2391 if (InFlag.getNode())
2392 Ops.push_back(InFlag);
2396 //// If this is the first return lowered for this function, add the regs
2397 //// to the liveout set for the function.
2398 // This isn't right, although it's probably harmless on x86; liveouts
2399 // should be computed from returns not tail calls. Consider a void
2400 // function making a tail call to a function returning int.
2401 return DAG.getNode(X86ISD::TC_RETURN, dl,
2402 NodeTys, &Ops[0], Ops.size());
2405 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2406 InFlag = Chain.getValue(1);
2408 // Create the CALLSEQ_END node.
2409 unsigned NumBytesForCalleeToPush;
2410 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg, GuaranteedTailCallOpt))
2411 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2412 else if (!Is64Bit && !IsTailCallConvention(CallConv) && IsStructRet)
2413 // If this is a call to a struct-return function, the callee
2414 // pops the hidden struct pointer, so we have to push it back.
2415 // This is common for Darwin/X86, Linux & Mingw32 targets.
2416 NumBytesForCalleeToPush = 4;
2418 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2420 // Returns a flag for retval copy to use.
2422 Chain = DAG.getCALLSEQ_END(Chain,
2423 DAG.getIntPtrConstant(NumBytes, true),
2424 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2427 InFlag = Chain.getValue(1);
2430 // Handle result values, copying them out of physregs into vregs that we
2432 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2433 Ins, dl, DAG, InVals);
2437 //===----------------------------------------------------------------------===//
2438 // Fast Calling Convention (tail call) implementation
2439 //===----------------------------------------------------------------------===//
2441 // Like std call, callee cleans arguments, convention except that ECX is
2442 // reserved for storing the tail called function address. Only 2 registers are
2443 // free for argument passing (inreg). Tail call optimization is performed
2445 // * tailcallopt is enabled
2446 // * caller/callee are fastcc
2447 // On X86_64 architecture with GOT-style position independent code only local
2448 // (within module) calls are supported at the moment.
2449 // To keep the stack aligned according to platform abi the function
2450 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2451 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2452 // If a tail called function callee has more arguments than the caller the
2453 // caller needs to make sure that there is room to move the RETADDR to. This is
2454 // achieved by reserving an area the size of the argument delta right after the
2455 // original REtADDR, but before the saved framepointer or the spilled registers
2456 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2468 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2469 /// for a 16 byte align requirement.
2471 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2472 SelectionDAG& DAG) const {
2473 MachineFunction &MF = DAG.getMachineFunction();
2474 const TargetMachine &TM = MF.getTarget();
2475 const TargetFrameLowering &TFI = *TM.getFrameLowering();
2476 unsigned StackAlignment = TFI.getStackAlignment();
2477 uint64_t AlignMask = StackAlignment - 1;
2478 int64_t Offset = StackSize;
2479 uint64_t SlotSize = TD->getPointerSize();
2480 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2481 // Number smaller than 12 so just add the difference.
2482 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2484 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2485 Offset = ((~AlignMask) & Offset) + StackAlignment +
2486 (StackAlignment-SlotSize);
2491 /// MatchingStackOffset - Return true if the given stack call argument is
2492 /// already available in the same position (relatively) of the caller's
2493 /// incoming argument stack.
2495 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2496 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2497 const X86InstrInfo *TII) {
2498 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
2500 if (Arg.getOpcode() == ISD::CopyFromReg) {
2501 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2502 if (!TargetRegisterInfo::isVirtualRegister(VR))
2504 MachineInstr *Def = MRI->getVRegDef(VR);
2507 if (!Flags.isByVal()) {
2508 if (!TII->isLoadFromStackSlot(Def, FI))
2511 unsigned Opcode = Def->getOpcode();
2512 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2513 Def->getOperand(1).isFI()) {
2514 FI = Def->getOperand(1).getIndex();
2515 Bytes = Flags.getByValSize();
2519 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
2520 if (Flags.isByVal())
2521 // ByVal argument is passed in as a pointer but it's now being
2522 // dereferenced. e.g.
2523 // define @foo(%struct.X* %A) {
2524 // tail call @bar(%struct.X* byval %A)
2527 SDValue Ptr = Ld->getBasePtr();
2528 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2531 FI = FINode->getIndex();
2532 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
2533 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
2534 FI = FINode->getIndex();
2535 Bytes = Flags.getByValSize();
2539 assert(FI != INT_MAX);
2540 if (!MFI->isFixedObjectIndex(FI))
2542 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
2545 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2546 /// for tail call optimization. Targets which want to do tail call
2547 /// optimization should implement this function.
2549 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2550 CallingConv::ID CalleeCC,
2552 bool isCalleeStructRet,
2553 bool isCallerStructRet,
2554 const SmallVectorImpl<ISD::OutputArg> &Outs,
2555 const SmallVectorImpl<SDValue> &OutVals,
2556 const SmallVectorImpl<ISD::InputArg> &Ins,
2557 SelectionDAG& DAG) const {
2558 if (!IsTailCallConvention(CalleeCC) &&
2559 CalleeCC != CallingConv::C)
2562 // If -tailcallopt is specified, make fastcc functions tail-callable.
2563 const MachineFunction &MF = DAG.getMachineFunction();
2564 const Function *CallerF = DAG.getMachineFunction().getFunction();
2565 CallingConv::ID CallerCC = CallerF->getCallingConv();
2566 bool CCMatch = CallerCC == CalleeCC;
2568 if (GuaranteedTailCallOpt) {
2569 if (IsTailCallConvention(CalleeCC) && CCMatch)
2574 // Look for obvious safe cases to perform tail call optimization that do not
2575 // require ABI changes. This is what gcc calls sibcall.
2577 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
2578 // emit a special epilogue.
2579 if (RegInfo->needsStackRealignment(MF))
2582 // Also avoid sibcall optimization if either caller or callee uses struct
2583 // return semantics.
2584 if (isCalleeStructRet || isCallerStructRet)
2587 // An stdcall caller is expected to clean up its arguments; the callee
2588 // isn't going to do that.
2589 if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
2592 // Do not sibcall optimize vararg calls unless all arguments are passed via
2594 if (isVarArg && !Outs.empty()) {
2596 // Optimizing for varargs on Win64 is unlikely to be safe without
2597 // additional testing.
2598 if (Subtarget->isTargetWin64())
2601 SmallVector<CCValAssign, 16> ArgLocs;
2602 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2603 getTargetMachine(), ArgLocs, *DAG.getContext());
2605 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2606 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
2607 if (!ArgLocs[i].isRegLoc())
2611 // If the call result is in ST0 / ST1, it needs to be popped off the x87 stack.
2612 // Therefore if it's not used by the call it is not safe to optimize this into
2614 bool Unused = false;
2615 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
2622 SmallVector<CCValAssign, 16> RVLocs;
2623 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(),
2624 getTargetMachine(), RVLocs, *DAG.getContext());
2625 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2626 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2627 CCValAssign &VA = RVLocs[i];
2628 if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
2633 // If the calling conventions do not match, then we'd better make sure the
2634 // results are returned in the same way as what the caller expects.
2636 SmallVector<CCValAssign, 16> RVLocs1;
2637 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
2638 getTargetMachine(), RVLocs1, *DAG.getContext());
2639 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
2641 SmallVector<CCValAssign, 16> RVLocs2;
2642 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
2643 getTargetMachine(), RVLocs2, *DAG.getContext());
2644 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
2646 if (RVLocs1.size() != RVLocs2.size())
2648 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
2649 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
2651 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
2653 if (RVLocs1[i].isRegLoc()) {
2654 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
2657 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
2663 // If the callee takes no arguments then go on to check the results of the
2665 if (!Outs.empty()) {
2666 // Check if stack adjustment is needed. For now, do not do this if any
2667 // argument is passed on the stack.
2668 SmallVector<CCValAssign, 16> ArgLocs;
2669 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
2670 getTargetMachine(), ArgLocs, *DAG.getContext());
2672 // Allocate shadow area for Win64
2673 if (Subtarget->isTargetWin64()) {
2674 CCInfo.AllocateStack(32, 8);
2677 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2678 if (CCInfo.getNextStackOffset()) {
2679 MachineFunction &MF = DAG.getMachineFunction();
2680 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2683 // Check if the arguments are already laid out in the right way as
2684 // the caller's fixed stack objects.
2685 MachineFrameInfo *MFI = MF.getFrameInfo();
2686 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2687 const X86InstrInfo *TII =
2688 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2689 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2690 CCValAssign &VA = ArgLocs[i];
2691 SDValue Arg = OutVals[i];
2692 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2693 if (VA.getLocInfo() == CCValAssign::Indirect)
2695 if (!VA.isRegLoc()) {
2696 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2703 // If the tailcall address may be in a register, then make sure it's
2704 // possible to register allocate for it. In 32-bit, the call address can
2705 // only target EAX, EDX, or ECX since the tail call must be scheduled after
2706 // callee-saved registers are restored. These happen to be the same
2707 // registers used to pass 'inreg' arguments so watch out for those.
2708 if (!Subtarget->is64Bit() &&
2709 !isa<GlobalAddressSDNode>(Callee) &&
2710 !isa<ExternalSymbolSDNode>(Callee)) {
2711 unsigned NumInRegs = 0;
2712 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2713 CCValAssign &VA = ArgLocs[i];
2716 unsigned Reg = VA.getLocReg();
2719 case X86::EAX: case X86::EDX: case X86::ECX:
2720 if (++NumInRegs == 3)
2732 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo) const {
2733 return X86::createFastISel(funcInfo);
2737 //===----------------------------------------------------------------------===//
2738 // Other Lowering Hooks
2739 //===----------------------------------------------------------------------===//
2741 static bool MayFoldLoad(SDValue Op) {
2742 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
2745 static bool MayFoldIntoStore(SDValue Op) {
2746 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
2749 static bool isTargetShuffle(unsigned Opcode) {
2751 default: return false;
2752 case X86ISD::PSHUFD:
2753 case X86ISD::PSHUFHW:
2754 case X86ISD::PSHUFLW:
2755 case X86ISD::SHUFPD:
2756 case X86ISD::PALIGN:
2757 case X86ISD::SHUFPS:
2758 case X86ISD::MOVLHPS:
2759 case X86ISD::MOVLHPD:
2760 case X86ISD::MOVHLPS:
2761 case X86ISD::MOVLPS:
2762 case X86ISD::MOVLPD:
2763 case X86ISD::MOVSHDUP:
2764 case X86ISD::MOVSLDUP:
2765 case X86ISD::MOVDDUP:
2768 case X86ISD::UNPCKLPS:
2769 case X86ISD::UNPCKLPD:
2770 case X86ISD::VUNPCKLPSY:
2771 case X86ISD::VUNPCKLPDY:
2772 case X86ISD::PUNPCKLWD:
2773 case X86ISD::PUNPCKLBW:
2774 case X86ISD::PUNPCKLDQ:
2775 case X86ISD::PUNPCKLQDQ:
2776 case X86ISD::UNPCKHPS:
2777 case X86ISD::UNPCKHPD:
2778 case X86ISD::VUNPCKHPSY:
2779 case X86ISD::VUNPCKHPDY:
2780 case X86ISD::PUNPCKHWD:
2781 case X86ISD::PUNPCKHBW:
2782 case X86ISD::PUNPCKHDQ:
2783 case X86ISD::PUNPCKHQDQ:
2784 case X86ISD::VPERMILPS:
2785 case X86ISD::VPERMILPSY:
2786 case X86ISD::VPERMILPD:
2787 case X86ISD::VPERMILPDY:
2788 case X86ISD::VPERM2F128:
2794 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2795 SDValue V1, SelectionDAG &DAG) {
2797 default: llvm_unreachable("Unknown x86 shuffle node");
2798 case X86ISD::MOVSHDUP:
2799 case X86ISD::MOVSLDUP:
2800 case X86ISD::MOVDDUP:
2801 return DAG.getNode(Opc, dl, VT, V1);
2807 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2808 SDValue V1, unsigned TargetMask, SelectionDAG &DAG) {
2810 default: llvm_unreachable("Unknown x86 shuffle node");
2811 case X86ISD::PSHUFD:
2812 case X86ISD::PSHUFHW:
2813 case X86ISD::PSHUFLW:
2814 case X86ISD::VPERMILPS:
2815 case X86ISD::VPERMILPSY:
2816 case X86ISD::VPERMILPD:
2817 case X86ISD::VPERMILPDY:
2818 return DAG.getNode(Opc, dl, VT, V1, DAG.getConstant(TargetMask, MVT::i8));
2824 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2825 SDValue V1, SDValue V2, unsigned TargetMask, SelectionDAG &DAG) {
2827 default: llvm_unreachable("Unknown x86 shuffle node");
2828 case X86ISD::PALIGN:
2829 case X86ISD::SHUFPD:
2830 case X86ISD::SHUFPS:
2831 case X86ISD::VPERM2F128:
2832 return DAG.getNode(Opc, dl, VT, V1, V2,
2833 DAG.getConstant(TargetMask, MVT::i8));
2838 static SDValue getTargetShuffleNode(unsigned Opc, DebugLoc dl, EVT VT,
2839 SDValue V1, SDValue V2, SelectionDAG &DAG) {
2841 default: llvm_unreachable("Unknown x86 shuffle node");
2842 case X86ISD::MOVLHPS:
2843 case X86ISD::MOVLHPD:
2844 case X86ISD::MOVHLPS:
2845 case X86ISD::MOVLPS:
2846 case X86ISD::MOVLPD:
2849 case X86ISD::UNPCKLPS:
2850 case X86ISD::UNPCKLPD:
2851 case X86ISD::VUNPCKLPSY:
2852 case X86ISD::VUNPCKLPDY:
2853 case X86ISD::PUNPCKLWD:
2854 case X86ISD::PUNPCKLBW:
2855 case X86ISD::PUNPCKLDQ:
2856 case X86ISD::PUNPCKLQDQ:
2857 case X86ISD::UNPCKHPS:
2858 case X86ISD::UNPCKHPD:
2859 case X86ISD::VUNPCKHPSY:
2860 case X86ISD::VUNPCKHPDY:
2861 case X86ISD::PUNPCKHWD:
2862 case X86ISD::PUNPCKHBW:
2863 case X86ISD::PUNPCKHDQ:
2864 case X86ISD::PUNPCKHQDQ:
2865 return DAG.getNode(Opc, dl, VT, V1, V2);
2870 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
2871 MachineFunction &MF = DAG.getMachineFunction();
2872 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2873 int ReturnAddrIndex = FuncInfo->getRAIndex();
2875 if (ReturnAddrIndex == 0) {
2876 // Set up a frame object for the return address.
2877 uint64_t SlotSize = TD->getPointerSize();
2878 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2880 FuncInfo->setRAIndex(ReturnAddrIndex);
2883 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2887 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2888 bool hasSymbolicDisplacement) {
2889 // Offset should fit into 32 bit immediate field.
2890 if (!isInt<32>(Offset))
2893 // If we don't have a symbolic displacement - we don't have any extra
2895 if (!hasSymbolicDisplacement)
2898 // FIXME: Some tweaks might be needed for medium code model.
2899 if (M != CodeModel::Small && M != CodeModel::Kernel)
2902 // For small code model we assume that latest object is 16MB before end of 31
2903 // bits boundary. We may also accept pretty large negative constants knowing
2904 // that all objects are in the positive half of address space.
2905 if (M == CodeModel::Small && Offset < 16*1024*1024)
2908 // For kernel code model we know that all object resist in the negative half
2909 // of 32bits address space. We may not accept negative offsets, since they may
2910 // be just off and we may accept pretty large positive ones.
2911 if (M == CodeModel::Kernel && Offset > 0)
2917 /// isCalleePop - Determines whether the callee is required to pop its
2918 /// own arguments. Callee pop is necessary to support tail calls.
2919 bool X86::isCalleePop(CallingConv::ID CallingConv,
2920 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
2924 switch (CallingConv) {
2927 case CallingConv::X86_StdCall:
2929 case CallingConv::X86_FastCall:
2931 case CallingConv::X86_ThisCall:
2933 case CallingConv::Fast:
2935 case CallingConv::GHC:
2940 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2941 /// specific condition code, returning the condition code and the LHS/RHS of the
2942 /// comparison to make.
2943 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2944 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2946 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2947 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2948 // X > -1 -> X == 0, jump !sign.
2949 RHS = DAG.getConstant(0, RHS.getValueType());
2950 return X86::COND_NS;
2951 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2952 // X < 0 -> X == 0, jump on sign.
2954 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2956 RHS = DAG.getConstant(0, RHS.getValueType());
2957 return X86::COND_LE;
2961 switch (SetCCOpcode) {
2962 default: llvm_unreachable("Invalid integer condition!");
2963 case ISD::SETEQ: return X86::COND_E;
2964 case ISD::SETGT: return X86::COND_G;
2965 case ISD::SETGE: return X86::COND_GE;
2966 case ISD::SETLT: return X86::COND_L;
2967 case ISD::SETLE: return X86::COND_LE;
2968 case ISD::SETNE: return X86::COND_NE;
2969 case ISD::SETULT: return X86::COND_B;
2970 case ISD::SETUGT: return X86::COND_A;
2971 case ISD::SETULE: return X86::COND_BE;
2972 case ISD::SETUGE: return X86::COND_AE;
2976 // First determine if it is required or is profitable to flip the operands.
2978 // If LHS is a foldable load, but RHS is not, flip the condition.
2979 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
2980 !ISD::isNON_EXTLoad(RHS.getNode())) {
2981 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2982 std::swap(LHS, RHS);
2985 switch (SetCCOpcode) {
2991 std::swap(LHS, RHS);
2995 // On a floating point condition, the flags are set as follows:
2997 // 0 | 0 | 0 | X > Y
2998 // 0 | 0 | 1 | X < Y
2999 // 1 | 0 | 0 | X == Y
3000 // 1 | 1 | 1 | unordered
3001 switch (SetCCOpcode) {
3002 default: llvm_unreachable("Condcode should be pre-legalized away");
3004 case ISD::SETEQ: return X86::COND_E;
3005 case ISD::SETOLT: // flipped
3007 case ISD::SETGT: return X86::COND_A;
3008 case ISD::SETOLE: // flipped
3010 case ISD::SETGE: return X86::COND_AE;
3011 case ISD::SETUGT: // flipped
3013 case ISD::SETLT: return X86::COND_B;
3014 case ISD::SETUGE: // flipped
3016 case ISD::SETLE: return X86::COND_BE;
3018 case ISD::SETNE: return X86::COND_NE;
3019 case ISD::SETUO: return X86::COND_P;
3020 case ISD::SETO: return X86::COND_NP;
3022 case ISD::SETUNE: return X86::COND_INVALID;
3026 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
3027 /// code. Current x86 isa includes the following FP cmov instructions:
3028 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3029 static bool hasFPCMov(unsigned X86CC) {
3045 /// isFPImmLegal - Returns true if the target can instruction select the
3046 /// specified FP immediate natively. If false, the legalizer will
3047 /// materialize the FP immediate as a load from a constant pool.
3048 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3049 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3050 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3056 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
3057 /// the specified range (L, H].
3058 static bool isUndefOrInRange(int Val, int Low, int Hi) {
3059 return (Val < 0) || (Val >= Low && Val < Hi);
3062 /// isUndefOrInRange - Return true if every element in Mask, begining
3063 /// from position Pos and ending in Pos+Size, falls within the specified
3064 /// range (L, L+Pos]. or is undef.
3065 static bool isUndefOrInRange(const SmallVectorImpl<int> &Mask,
3066 int Pos, int Size, int Low, int Hi) {
3067 for (int i = Pos, e = Pos+Size; i != e; ++i)
3068 if (!isUndefOrInRange(Mask[i], Low, Hi))
3073 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
3074 /// specified value.
3075 static bool isUndefOrEqual(int Val, int CmpVal) {
3076 if (Val < 0 || Val == CmpVal)
3081 /// isSequentialOrUndefInRange - Return true if every element in Mask, begining
3082 /// from position Pos and ending in Pos+Size, falls within the specified
3083 /// sequential range (L, L+Pos]. or is undef.
3084 static bool isSequentialOrUndefInRange(const SmallVectorImpl<int> &Mask,
3085 int Pos, int Size, int Low) {
3086 for (int i = Pos, e = Pos+Size; i != e; ++i, ++Low)
3087 if (!isUndefOrEqual(Mask[i], Low))
3092 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
3093 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
3094 /// the second operand.
3095 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3096 if (VT == MVT::v4f32 || VT == MVT::v4i32 )
3097 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
3098 if (VT == MVT::v2f64 || VT == MVT::v2i64)
3099 return (Mask[0] < 2 && Mask[1] < 2);
3103 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
3104 SmallVector<int, 8> M;
3106 return ::isPSHUFDMask(M, N->getValueType(0));
3109 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
3110 /// is suitable for input to PSHUFHW.
3111 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3112 if (VT != MVT::v8i16)
3115 // Lower quadword copied in order or undef.
3116 for (int i = 0; i != 4; ++i)
3117 if (Mask[i] >= 0 && Mask[i] != i)
3120 // Upper quadword shuffled.
3121 for (int i = 4; i != 8; ++i)
3122 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
3128 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
3129 SmallVector<int, 8> M;
3131 return ::isPSHUFHWMask(M, N->getValueType(0));
3134 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
3135 /// is suitable for input to PSHUFLW.
3136 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3137 if (VT != MVT::v8i16)
3140 // Upper quadword copied in order.
3141 for (int i = 4; i != 8; ++i)
3142 if (Mask[i] >= 0 && Mask[i] != i)
3145 // Lower quadword shuffled.
3146 for (int i = 0; i != 4; ++i)
3153 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
3154 SmallVector<int, 8> M;
3156 return ::isPSHUFLWMask(M, N->getValueType(0));
3159 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
3160 /// is suitable for input to PALIGNR.
3161 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
3163 int i, e = VT.getVectorNumElements();
3164 if (VT.getSizeInBits() != 128 && VT.getSizeInBits() != 64)
3167 // Do not handle v2i64 / v2f64 shuffles with palignr.
3168 if (e < 4 || !hasSSSE3)
3171 for (i = 0; i != e; ++i)
3175 // All undef, not a palignr.
3179 // Make sure we're shifting in the right direction.
3183 int s = Mask[i] - i;
3185 // Check the rest of the elements to see if they are consecutive.
3186 for (++i; i != e; ++i) {
3188 if (m >= 0 && m != s+i)
3194 /// isVSHUFPSYMask - Return true if the specified VECTOR_SHUFFLE operand
3195 /// specifies a shuffle of elements that is suitable for input to 256-bit
3197 static bool isVSHUFPSYMask(const SmallVectorImpl<int> &Mask, EVT VT,
3198 const X86Subtarget *Subtarget) {
3199 int NumElems = VT.getVectorNumElements();
3201 if (!Subtarget->hasAVX() || VT.getSizeInBits() != 256)
3207 // VSHUFPSY divides the resulting vector into 4 chunks.
3208 // The sources are also splitted into 4 chunks, and each destination
3209 // chunk must come from a different source chunk.
3211 // SRC1 => X7 X6 X5 X4 X3 X2 X1 X0
3212 // SRC2 => Y7 Y6 Y5 Y4 Y3 Y2 Y1 Y9
3214 // DST => Y7..Y4, Y7..Y4, X7..X4, X7..X4,
3215 // Y3..Y0, Y3..Y0, X3..X0, X3..X0
3217 int QuarterSize = NumElems/4;
3218 int HalfSize = QuarterSize*2;
3219 for (int i = 0; i < QuarterSize; ++i)
3220 if (!isUndefOrInRange(Mask[i], 0, HalfSize))
3222 for (int i = QuarterSize; i < QuarterSize*2; ++i)
3223 if (!isUndefOrInRange(Mask[i], NumElems, NumElems+HalfSize))
3226 // The mask of the second half must be the same as the first but with
3227 // the appropriate offsets. This works in the same way as VPERMILPS
3228 // works with masks.
3229 for (int i = QuarterSize*2; i < QuarterSize*3; ++i) {
3230 if (!isUndefOrInRange(Mask[i], HalfSize, NumElems))
3232 int FstHalfIdx = i-HalfSize;
3233 if (Mask[FstHalfIdx] < 0)
3235 if (!isUndefOrEqual(Mask[i], Mask[FstHalfIdx]+HalfSize))
3238 for (int i = QuarterSize*3; i < NumElems; ++i) {
3239 if (!isUndefOrInRange(Mask[i], NumElems+HalfSize, NumElems*2))
3241 int FstHalfIdx = i-HalfSize;
3242 if (Mask[FstHalfIdx] < 0)
3244 if (!isUndefOrEqual(Mask[i], Mask[FstHalfIdx]+HalfSize))
3252 /// getShuffleVSHUFPSYImmediate - Return the appropriate immediate to shuffle
3253 /// the specified VECTOR_MASK mask with VSHUFPSY instruction.
3254 static unsigned getShuffleVSHUFPSYImmediate(SDNode *N) {
3255 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3256 EVT VT = SVOp->getValueType(0);
3257 int NumElems = VT.getVectorNumElements();
3259 assert(NumElems == 8 && VT.getSizeInBits() == 256 &&
3260 "Only supports v8i32 and v8f32 types");
3262 int HalfSize = NumElems/2;
3264 for (int i = 0; i != NumElems ; ++i) {
3265 if (SVOp->getMaskElt(i) < 0)
3267 // The mask of the first half must be equal to the second one.
3268 unsigned Shamt = (i%HalfSize)*2;
3269 unsigned Elt = SVOp->getMaskElt(i) % HalfSize;
3270 Mask |= Elt << Shamt;
3276 /// isVSHUFPDYMask - Return true if the specified VECTOR_SHUFFLE operand
3277 /// specifies a shuffle of elements that is suitable for input to 256-bit
3278 /// VSHUFPDY. This shuffle doesn't have the same restriction as the PS
3279 /// version and the mask of the second half isn't binded with the first
3281 static bool isVSHUFPDYMask(const SmallVectorImpl<int> &Mask, EVT VT,
3282 const X86Subtarget *Subtarget) {
3283 int NumElems = VT.getVectorNumElements();
3285 if (!Subtarget->hasAVX() || VT.getSizeInBits() != 256)
3291 // VSHUFPSY divides the resulting vector into 4 chunks.
3292 // The sources are also splitted into 4 chunks, and each destination
3293 // chunk must come from a different source chunk.
3295 // SRC1 => X3 X2 X1 X0
3296 // SRC2 => Y3 Y2 Y1 Y0
3298 // DST => Y2..Y3, X2..X3, Y1..Y0, X1..X0
3300 int QuarterSize = NumElems/4;
3301 int HalfSize = QuarterSize*2;
3302 for (int i = 0; i < QuarterSize; ++i)
3303 if (!isUndefOrInRange(Mask[i], 0, HalfSize))
3305 for (int i = QuarterSize; i < QuarterSize*2; ++i)
3306 if (!isUndefOrInRange(Mask[i], NumElems, NumElems+HalfSize))
3308 for (int i = QuarterSize*2; i < QuarterSize*3; ++i)
3309 if (!isUndefOrInRange(Mask[i], HalfSize, NumElems))
3311 for (int i = QuarterSize*3; i < NumElems; ++i)
3312 if (!isUndefOrInRange(Mask[i], NumElems+HalfSize, NumElems*2))
3318 /// getShuffleVSHUFPDYImmediate - Return the appropriate immediate to shuffle
3319 /// the specified VECTOR_MASK mask with VSHUFPDY instruction.
3320 static unsigned getShuffleVSHUFPDYImmediate(SDNode *N) {
3321 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3322 EVT VT = SVOp->getValueType(0);
3323 int NumElems = VT.getVectorNumElements();
3325 assert(NumElems == 4 && VT.getSizeInBits() == 256 &&
3326 "Only supports v4i64 and v4f64 types");
3328 int HalfSize = NumElems/2;
3330 for (int i = 0; i != NumElems ; ++i) {
3331 if (SVOp->getMaskElt(i) < 0)
3333 int Elt = SVOp->getMaskElt(i) % HalfSize;
3340 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
3341 /// specifies a shuffle of elements that is suitable for input to 128-bit
3342 /// SHUFPS and SHUFPD.
3343 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3344 int NumElems = VT.getVectorNumElements();
3346 if (VT.getSizeInBits() != 128)
3349 if (NumElems != 2 && NumElems != 4)
3352 int Half = NumElems / 2;
3353 for (int i = 0; i < Half; ++i)
3354 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3356 for (int i = Half; i < NumElems; ++i)
3357 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3363 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
3364 SmallVector<int, 8> M;
3366 return ::isSHUFPMask(M, N->getValueType(0));
3369 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
3370 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
3371 /// half elements to come from vector 1 (which would equal the dest.) and
3372 /// the upper half to come from vector 2.
3373 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3374 int NumElems = VT.getVectorNumElements();
3376 if (NumElems != 2 && NumElems != 4)
3379 int Half = NumElems / 2;
3380 for (int i = 0; i < Half; ++i)
3381 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
3383 for (int i = Half; i < NumElems; ++i)
3384 if (!isUndefOrInRange(Mask[i], 0, NumElems))
3389 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
3390 SmallVector<int, 8> M;
3392 return isCommutedSHUFPMask(M, N->getValueType(0));
3395 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
3396 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
3397 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
3398 EVT VT = N->getValueType(0);
3399 unsigned NumElems = VT.getVectorNumElements();
3401 if (VT.getSizeInBits() != 128)
3407 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
3408 return isUndefOrEqual(N->getMaskElt(0), 6) &&
3409 isUndefOrEqual(N->getMaskElt(1), 7) &&
3410 isUndefOrEqual(N->getMaskElt(2), 2) &&
3411 isUndefOrEqual(N->getMaskElt(3), 3);
3414 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
3415 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
3417 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
3418 EVT VT = N->getValueType(0);
3419 unsigned NumElems = VT.getVectorNumElements();
3421 if (VT.getSizeInBits() != 128)
3427 return isUndefOrEqual(N->getMaskElt(0), 2) &&
3428 isUndefOrEqual(N->getMaskElt(1), 3) &&
3429 isUndefOrEqual(N->getMaskElt(2), 2) &&
3430 isUndefOrEqual(N->getMaskElt(3), 3);
3433 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
3434 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
3435 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
3436 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3438 if (NumElems != 2 && NumElems != 4)
3441 for (unsigned i = 0; i < NumElems/2; ++i)
3442 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
3445 for (unsigned i = NumElems/2; i < NumElems; ++i)
3446 if (!isUndefOrEqual(N->getMaskElt(i), i))
3452 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
3453 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
3454 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
3455 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3457 if ((NumElems != 2 && NumElems != 4)
3458 || N->getValueType(0).getSizeInBits() > 128)
3461 for (unsigned i = 0; i < NumElems/2; ++i)
3462 if (!isUndefOrEqual(N->getMaskElt(i), i))
3465 for (unsigned i = 0; i < NumElems/2; ++i)
3466 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
3472 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
3473 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
3474 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3475 bool V2IsSplat = false) {
3476 int NumElts = VT.getVectorNumElements();
3478 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3479 "Unsupported vector type for unpckh");
3481 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8)
3484 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3485 // independently on 128-bit lanes.
3486 unsigned NumLanes = VT.getSizeInBits()/128;
3487 unsigned NumLaneElts = NumElts/NumLanes;
3490 unsigned End = NumLaneElts;
3491 for (unsigned s = 0; s < NumLanes; ++s) {
3492 for (unsigned i = Start, j = s * NumLaneElts;
3496 int BitI1 = Mask[i+1];
3497 if (!isUndefOrEqual(BitI, j))
3500 if (!isUndefOrEqual(BitI1, NumElts))
3503 if (!isUndefOrEqual(BitI1, j + NumElts))
3507 // Process the next 128 bits.
3508 Start += NumLaneElts;
3515 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3516 SmallVector<int, 8> M;
3518 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
3521 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
3522 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
3523 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
3524 bool V2IsSplat = false) {
3525 int NumElts = VT.getVectorNumElements();
3527 assert((VT.is128BitVector() || VT.is256BitVector()) &&
3528 "Unsupported vector type for unpckh");
3530 if (VT.getSizeInBits() == 256 && NumElts != 4 && NumElts != 8)
3533 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3534 // independently on 128-bit lanes.
3535 unsigned NumLanes = VT.getSizeInBits()/128;
3536 unsigned NumLaneElts = NumElts/NumLanes;
3539 unsigned End = NumLaneElts;
3540 for (unsigned l = 0; l != NumLanes; ++l) {
3541 for (unsigned i = Start, j = (l*NumLaneElts)+NumLaneElts/2;
3542 i != End; i += 2, ++j) {
3544 int BitI1 = Mask[i+1];
3545 if (!isUndefOrEqual(BitI, j))
3548 if (isUndefOrEqual(BitI1, NumElts))
3551 if (!isUndefOrEqual(BitI1, j+NumElts))
3555 // Process the next 128 bits.
3556 Start += NumLaneElts;
3562 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
3563 SmallVector<int, 8> M;
3565 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
3568 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
3569 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
3571 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3572 int NumElems = VT.getVectorNumElements();
3573 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3576 // For 256-bit i64/f64, use MOVDDUPY instead, so reject the matching pattern
3577 // FIXME: Need a better way to get rid of this, there's no latency difference
3578 // between UNPCKLPD and MOVDDUP, the later should always be checked first and
3579 // the former later. We should also remove the "_undef" special mask.
3580 if (NumElems == 4 && VT.getSizeInBits() == 256)
3583 // Handle 128 and 256-bit vector lengths. AVX defines UNPCK* to operate
3584 // independently on 128-bit lanes.
3585 unsigned NumLanes = VT.getSizeInBits() / 128;
3586 unsigned NumLaneElts = NumElems / NumLanes;
3588 for (unsigned s = 0; s < NumLanes; ++s) {
3589 for (unsigned i = s * NumLaneElts, j = s * NumLaneElts;
3590 i != NumLaneElts * (s + 1);
3593 int BitI1 = Mask[i+1];
3595 if (!isUndefOrEqual(BitI, j))
3597 if (!isUndefOrEqual(BitI1, j))
3605 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
3606 SmallVector<int, 8> M;
3608 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
3611 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
3612 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
3614 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
3615 int NumElems = VT.getVectorNumElements();
3616 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
3619 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
3621 int BitI1 = Mask[i+1];
3622 if (!isUndefOrEqual(BitI, j))
3624 if (!isUndefOrEqual(BitI1, j))
3630 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
3631 SmallVector<int, 8> M;
3633 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
3636 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
3637 /// specifies a shuffle of elements that is suitable for input to MOVSS,
3638 /// MOVSD, and MOVD, i.e. setting the lowest element.
3639 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
3640 if (VT.getVectorElementType().getSizeInBits() < 32)
3643 int NumElts = VT.getVectorNumElements();
3645 if (!isUndefOrEqual(Mask[0], NumElts))
3648 for (int i = 1; i < NumElts; ++i)
3649 if (!isUndefOrEqual(Mask[i], i))
3655 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
3656 SmallVector<int, 8> M;
3658 return ::isMOVLMask(M, N->getValueType(0));
3661 /// isVPERM2F128Mask - Match 256-bit shuffles where the elements are considered
3662 /// as permutations between 128-bit chunks or halves. As an example: this
3664 /// vector_shuffle <4, 5, 6, 7, 12, 13, 14, 15>
3665 /// The first half comes from the second half of V1 and the second half from the
3666 /// the second half of V2.
3667 static bool isVPERM2F128Mask(const SmallVectorImpl<int> &Mask, EVT VT,
3668 const X86Subtarget *Subtarget) {
3669 if (!Subtarget->hasAVX() || VT.getSizeInBits() != 256)
3672 // The shuffle result is divided into half A and half B. In total the two
3673 // sources have 4 halves, namely: C, D, E, F. The final values of A and
3674 // B must come from C, D, E or F.
3675 int HalfSize = VT.getVectorNumElements()/2;
3676 bool MatchA = false, MatchB = false;
3678 // Check if A comes from one of C, D, E, F.
3679 for (int Half = 0; Half < 4; ++Half) {
3680 if (isSequentialOrUndefInRange(Mask, 0, HalfSize, Half*HalfSize)) {
3686 // Check if B comes from one of C, D, E, F.
3687 for (int Half = 0; Half < 4; ++Half) {
3688 if (isSequentialOrUndefInRange(Mask, HalfSize, HalfSize, Half*HalfSize)) {
3694 return MatchA && MatchB;
3697 /// getShuffleVPERM2F128Immediate - Return the appropriate immediate to shuffle
3698 /// the specified VECTOR_MASK mask with VPERM2F128 instructions.
3699 static unsigned getShuffleVPERM2F128Immediate(SDNode *N) {
3700 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3701 EVT VT = SVOp->getValueType(0);
3703 int HalfSize = VT.getVectorNumElements()/2;
3705 int FstHalf = 0, SndHalf = 0;
3706 for (int i = 0; i < HalfSize; ++i) {
3707 if (SVOp->getMaskElt(i) > 0) {
3708 FstHalf = SVOp->getMaskElt(i)/HalfSize;
3712 for (int i = HalfSize; i < HalfSize*2; ++i) {
3713 if (SVOp->getMaskElt(i) > 0) {
3714 SndHalf = SVOp->getMaskElt(i)/HalfSize;
3719 return (FstHalf | (SndHalf << 4));
3722 /// isVPERMILPDMask - Return true if the specified VECTOR_SHUFFLE operand
3723 /// specifies a shuffle of elements that is suitable for input to VPERMILPD*.
3724 /// Note that VPERMIL mask matching is different depending whether theunderlying
3725 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
3726 /// to the same elements of the low, but to the higher half of the source.
3727 /// In VPERMILPD the two lanes could be shuffled independently of each other
3728 /// with the same restriction that lanes can't be crossed.
3729 static bool isVPERMILPDMask(const SmallVectorImpl<int> &Mask, EVT VT,
3730 const X86Subtarget *Subtarget) {
3731 int NumElts = VT.getVectorNumElements();
3732 int NumLanes = VT.getSizeInBits()/128;
3734 if (!Subtarget->hasAVX())
3737 // Match any permutation of 128-bit vector with 64-bit types
3738 if (NumLanes == 1 && NumElts != 2)
3741 // Only match 256-bit with 32 types
3742 if (VT.getSizeInBits() == 256 && NumElts != 4)
3745 // The mask on the high lane is independent of the low. Both can match
3746 // any element in inside its own lane, but can't cross.
3747 int LaneSize = NumElts/NumLanes;
3748 for (int l = 0; l < NumLanes; ++l)
3749 for (int i = l*LaneSize; i < LaneSize*(l+1); ++i) {
3750 int LaneStart = l*LaneSize;
3751 if (!isUndefOrInRange(Mask[i], LaneStart, LaneStart+LaneSize))
3758 /// isVPERMILPSMask - Return true if the specified VECTOR_SHUFFLE operand
3759 /// specifies a shuffle of elements that is suitable for input to VPERMILPS*.
3760 /// Note that VPERMIL mask matching is different depending whether theunderlying
3761 /// type is 32 or 64. In the VPERMILPS the high half of the mask should point
3762 /// to the same elements of the low, but to the higher half of the source.
3763 /// In VPERMILPD the two lanes could be shuffled independently of each other
3764 /// with the same restriction that lanes can't be crossed.
3765 static bool isVPERMILPSMask(const SmallVectorImpl<int> &Mask, EVT VT,
3766 const X86Subtarget *Subtarget) {
3767 unsigned NumElts = VT.getVectorNumElements();
3768 unsigned NumLanes = VT.getSizeInBits()/128;
3770 if (!Subtarget->hasAVX())
3773 // Match any permutation of 128-bit vector with 32-bit types
3774 if (NumLanes == 1 && NumElts != 4)
3777 // Only match 256-bit with 32 types
3778 if (VT.getSizeInBits() == 256 && NumElts != 8)
3781 // The mask on the high lane should be the same as the low. Actually,
3782 // they can differ if any of the corresponding index in a lane is undef
3783 // and the other stays in range.
3784 int LaneSize = NumElts/NumLanes;
3785 for (int i = 0; i < LaneSize; ++i) {
3786 int HighElt = i+LaneSize;
3787 bool HighValid = isUndefOrInRange(Mask[HighElt], LaneSize, NumElts);
3788 bool LowValid = isUndefOrInRange(Mask[i], 0, LaneSize);
3790 if (!HighValid || !LowValid)
3792 if (Mask[i] < 0 || Mask[HighElt] < 0)
3794 if (Mask[HighElt]-Mask[i] != LaneSize)
3801 /// getShuffleVPERMILPSImmediate - Return the appropriate immediate to shuffle
3802 /// the specified VECTOR_MASK mask with VPERMILPS* instructions.
3803 static unsigned getShuffleVPERMILPSImmediate(SDNode *N) {
3804 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3805 EVT VT = SVOp->getValueType(0);
3807 int NumElts = VT.getVectorNumElements();
3808 int NumLanes = VT.getSizeInBits()/128;
3809 int LaneSize = NumElts/NumLanes;
3811 // Although the mask is equal for both lanes do it twice to get the cases
3812 // where a mask will match because the same mask element is undef on the
3813 // first half but valid on the second. This would get pathological cases
3814 // such as: shuffle <u, 0, 1, 2, 4, 4, 5, 6>, which is completely valid.
3816 for (int l = 0; l < NumLanes; ++l) {
3817 for (int i = 0; i < LaneSize; ++i) {
3818 int MaskElt = SVOp->getMaskElt(i+(l*LaneSize));
3821 if (MaskElt >= LaneSize)
3822 MaskElt -= LaneSize;
3823 Mask |= MaskElt << (i*2);
3830 /// getShuffleVPERMILPDImmediate - Return the appropriate immediate to shuffle
3831 /// the specified VECTOR_MASK mask with VPERMILPD* instructions.
3832 static unsigned getShuffleVPERMILPDImmediate(SDNode *N) {
3833 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3834 EVT VT = SVOp->getValueType(0);
3836 int NumElts = VT.getVectorNumElements();
3837 int NumLanes = VT.getSizeInBits()/128;
3840 int LaneSize = NumElts/NumLanes;
3841 for (int l = 0; l < NumLanes; ++l)
3842 for (int i = l*LaneSize; i < LaneSize*(l+1); ++i) {
3843 int MaskElt = SVOp->getMaskElt(i);
3846 Mask |= (MaskElt-l*LaneSize) << i;
3852 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
3853 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
3854 /// element of vector 2 and the other elements to come from vector 1 in order.
3855 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
3856 bool V2IsSplat = false, bool V2IsUndef = false) {
3857 int NumOps = VT.getVectorNumElements();
3858 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
3861 if (!isUndefOrEqual(Mask[0], 0))
3864 for (int i = 1; i < NumOps; ++i)
3865 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
3866 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
3867 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
3873 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
3874 bool V2IsUndef = false) {
3875 SmallVector<int, 8> M;
3877 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
3880 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3881 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
3882 /// Masks to match: <1, 1, 3, 3> or <1, 1, 3, 3, 5, 5, 7, 7>
3883 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N,
3884 const X86Subtarget *Subtarget) {
3885 if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
3888 // The second vector must be undef
3889 if (N->getOperand(1).getOpcode() != ISD::UNDEF)
3892 EVT VT = N->getValueType(0);
3893 unsigned NumElems = VT.getVectorNumElements();
3895 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3896 (VT.getSizeInBits() == 256 && NumElems != 8))
3899 // "i+1" is the value the indexed mask element must have
3900 for (unsigned i = 0; i < NumElems; i += 2)
3901 if (!isUndefOrEqual(N->getMaskElt(i), i+1) ||
3902 !isUndefOrEqual(N->getMaskElt(i+1), i+1))
3908 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3909 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
3910 /// Masks to match: <0, 0, 2, 2> or <0, 0, 2, 2, 4, 4, 6, 6>
3911 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N,
3912 const X86Subtarget *Subtarget) {
3913 if (!Subtarget->hasSSE3() && !Subtarget->hasAVX())
3916 // The second vector must be undef
3917 if (N->getOperand(1).getOpcode() != ISD::UNDEF)
3920 EVT VT = N->getValueType(0);
3921 unsigned NumElems = VT.getVectorNumElements();
3923 if ((VT.getSizeInBits() == 128 && NumElems != 4) ||
3924 (VT.getSizeInBits() == 256 && NumElems != 8))
3927 // "i" is the value the indexed mask element must have
3928 for (unsigned i = 0; i < NumElems; i += 2)
3929 if (!isUndefOrEqual(N->getMaskElt(i), i) ||
3930 !isUndefOrEqual(N->getMaskElt(i+1), i))
3936 /// isMOVDDUPYMask - Return true if the specified VECTOR_SHUFFLE operand
3937 /// specifies a shuffle of elements that is suitable for input to 256-bit
3938 /// version of MOVDDUP.
3939 static bool isMOVDDUPYMask(ShuffleVectorSDNode *N,
3940 const X86Subtarget *Subtarget) {
3941 EVT VT = N->getValueType(0);
3942 int NumElts = VT.getVectorNumElements();
3943 bool V2IsUndef = N->getOperand(1).getOpcode() == ISD::UNDEF;
3945 if (!Subtarget->hasAVX() || VT.getSizeInBits() != 256 ||
3946 !V2IsUndef || NumElts != 4)
3949 for (int i = 0; i != NumElts/2; ++i)
3950 if (!isUndefOrEqual(N->getMaskElt(i), 0))
3952 for (int i = NumElts/2; i != NumElts; ++i)
3953 if (!isUndefOrEqual(N->getMaskElt(i), NumElts/2))
3958 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
3959 /// specifies a shuffle of elements that is suitable for input to 128-bit
3960 /// version of MOVDDUP.
3961 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
3962 EVT VT = N->getValueType(0);
3964 if (VT.getSizeInBits() != 128)
3967 int e = VT.getVectorNumElements() / 2;
3968 for (int i = 0; i < e; ++i)
3969 if (!isUndefOrEqual(N->getMaskElt(i), i))
3971 for (int i = 0; i < e; ++i)
3972 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3977 /// isVEXTRACTF128Index - Return true if the specified
3978 /// EXTRACT_SUBVECTOR operand specifies a vector extract that is
3979 /// suitable for input to VEXTRACTF128.
3980 bool X86::isVEXTRACTF128Index(SDNode *N) {
3981 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
3984 // The index should be aligned on a 128-bit boundary.
3986 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
3988 unsigned VL = N->getValueType(0).getVectorNumElements();
3989 unsigned VBits = N->getValueType(0).getSizeInBits();
3990 unsigned ElSize = VBits / VL;
3991 bool Result = (Index * ElSize) % 128 == 0;
3996 /// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
3997 /// operand specifies a subvector insert that is suitable for input to
3999 bool X86::isVINSERTF128Index(SDNode *N) {
4000 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4003 // The index should be aligned on a 128-bit boundary.
4005 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4007 unsigned VL = N->getValueType(0).getVectorNumElements();
4008 unsigned VBits = N->getValueType(0).getSizeInBits();
4009 unsigned ElSize = VBits / VL;
4010 bool Result = (Index * ElSize) % 128 == 0;
4015 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
4016 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
4017 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
4018 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
4019 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
4021 unsigned Shift = (NumOperands == 4) ? 2 : 1;
4023 for (int i = 0; i < NumOperands; ++i) {
4024 int Val = SVOp->getMaskElt(NumOperands-i-1);
4025 if (Val < 0) Val = 0;
4026 if (Val >= NumOperands) Val -= NumOperands;
4028 if (i != NumOperands - 1)
4034 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
4035 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
4036 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
4037 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
4039 // 8 nodes, but we only care about the last 4.
4040 for (unsigned i = 7; i >= 4; --i) {
4041 int Val = SVOp->getMaskElt(i);
4050 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
4051 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
4052 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
4053 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
4055 // 8 nodes, but we only care about the first 4.
4056 for (int i = 3; i >= 0; --i) {
4057 int Val = SVOp->getMaskElt(i);
4066 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
4067 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
4068 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
4069 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
4070 EVT VVT = N->getValueType(0);
4071 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
4075 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
4076 Val = SVOp->getMaskElt(i);
4080 assert(Val - i > 0 && "PALIGNR imm should be positive");
4081 return (Val - i) * EltSize;
4084 /// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
4085 /// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
4087 unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
4088 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4089 llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
4092 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4094 EVT VecVT = N->getOperand(0).getValueType();
4095 EVT ElVT = VecVT.getVectorElementType();
4097 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
4098 return Index / NumElemsPerChunk;
4101 /// getInsertVINSERTF128Immediate - Return the appropriate immediate
4102 /// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
4104 unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
4105 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4106 llvm_unreachable("Illegal insert subvector for VINSERTF128");
4109 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4111 EVT VecVT = N->getValueType(0);
4112 EVT ElVT = VecVT.getVectorElementType();
4114 unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
4115 return Index / NumElemsPerChunk;
4118 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
4120 bool X86::isZeroNode(SDValue Elt) {
4121 return ((isa<ConstantSDNode>(Elt) &&
4122 cast<ConstantSDNode>(Elt)->isNullValue()) ||
4123 (isa<ConstantFPSDNode>(Elt) &&
4124 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
4127 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
4128 /// their permute mask.
4129 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
4130 SelectionDAG &DAG) {
4131 EVT VT = SVOp->getValueType(0);
4132 unsigned NumElems = VT.getVectorNumElements();
4133 SmallVector<int, 8> MaskVec;
4135 for (unsigned i = 0; i != NumElems; ++i) {
4136 int idx = SVOp->getMaskElt(i);
4138 MaskVec.push_back(idx);
4139 else if (idx < (int)NumElems)
4140 MaskVec.push_back(idx + NumElems);
4142 MaskVec.push_back(idx - NumElems);
4144 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
4145 SVOp->getOperand(0), &MaskVec[0]);
4148 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
4149 /// the two vector operands have swapped position.
4150 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
4151 unsigned NumElems = VT.getVectorNumElements();
4152 for (unsigned i = 0; i != NumElems; ++i) {
4156 else if (idx < (int)NumElems)
4157 Mask[i] = idx + NumElems;
4159 Mask[i] = idx - NumElems;
4163 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
4164 /// match movhlps. The lower half elements should come from upper half of
4165 /// V1 (and in order), and the upper half elements should come from the upper
4166 /// half of V2 (and in order).
4167 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
4168 EVT VT = Op->getValueType(0);
4169 if (VT.getSizeInBits() != 128)
4171 if (VT.getVectorNumElements() != 4)
4173 for (unsigned i = 0, e = 2; i != e; ++i)
4174 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
4176 for (unsigned i = 2; i != 4; ++i)
4177 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
4182 /// isScalarLoadToVector - Returns true if the node is a scalar load that
4183 /// is promoted to a vector. It also returns the LoadSDNode by reference if
4185 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
4186 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
4188 N = N->getOperand(0).getNode();
4189 if (!ISD::isNON_EXTLoad(N))
4192 *LD = cast<LoadSDNode>(N);
4196 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
4197 /// match movlp{s|d}. The lower half elements should come from lower half of
4198 /// V1 (and in order), and the upper half elements should come from the upper
4199 /// half of V2 (and in order). And since V1 will become the source of the
4200 /// MOVLP, it must be either a vector load or a scalar load to vector.
4201 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
4202 ShuffleVectorSDNode *Op) {
4203 EVT VT = Op->getValueType(0);
4204 if (VT.getSizeInBits() != 128)
4207 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
4209 // Is V2 is a vector load, don't do this transformation. We will try to use
4210 // load folding shufps op.
4211 if (ISD::isNON_EXTLoad(V2))
4214 unsigned NumElems = VT.getVectorNumElements();
4216 if (NumElems != 2 && NumElems != 4)
4218 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
4219 if (!isUndefOrEqual(Op->getMaskElt(i), i))
4221 for (unsigned i = NumElems/2; i != NumElems; ++i)
4222 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
4227 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
4229 static bool isSplatVector(SDNode *N) {
4230 if (N->getOpcode() != ISD::BUILD_VECTOR)
4233 SDValue SplatValue = N->getOperand(0);
4234 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
4235 if (N->getOperand(i) != SplatValue)
4240 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
4241 /// to an zero vector.
4242 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
4243 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
4244 SDValue V1 = N->getOperand(0);
4245 SDValue V2 = N->getOperand(1);
4246 unsigned NumElems = N->getValueType(0).getVectorNumElements();
4247 for (unsigned i = 0; i != NumElems; ++i) {
4248 int Idx = N->getMaskElt(i);
4249 if (Idx >= (int)NumElems) {
4250 unsigned Opc = V2.getOpcode();
4251 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
4253 if (Opc != ISD::BUILD_VECTOR ||
4254 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
4256 } else if (Idx >= 0) {
4257 unsigned Opc = V1.getOpcode();
4258 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
4260 if (Opc != ISD::BUILD_VECTOR ||
4261 !X86::isZeroNode(V1.getOperand(Idx)))
4268 /// getZeroVector - Returns a vector of specified type with all zero elements.
4270 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
4272 assert(VT.isVector() && "Expected a vector type");
4274 // Always build SSE zero vectors as <4 x i32> bitcasted
4275 // to their dest type. This ensures they get CSE'd.
4277 if (VT.getSizeInBits() == 128) { // SSE
4278 if (HasSSE2) { // SSE2
4279 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
4280 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4282 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4283 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4285 } else if (VT.getSizeInBits() == 256) { // AVX
4286 // 256-bit logic and arithmetic instructions in AVX are
4287 // all floating-point, no support for integer ops. Default
4288 // to emitting fp zeroed vectors then.
4289 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
4290 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4291 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
4293 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4296 /// getOnesVector - Returns a vector of specified type with all bits set.
4297 /// Always build ones vectors as <4 x i32>. For 256-bit types, use two
4298 /// <4 x i32> inserted in a <8 x i32> appropriately. Then bitcast to their
4299 /// original type, ensuring they get CSE'd.
4300 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
4301 assert(VT.isVector() && "Expected a vector type");
4302 assert((VT.is128BitVector() || VT.is256BitVector())
4303 && "Expected a 128-bit or 256-bit vector type");
4305 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
4306 SDValue Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
4307 Cst, Cst, Cst, Cst);
4309 if (VT.is256BitVector()) {
4310 SDValue InsV = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, MVT::v8i32),
4311 Vec, DAG.getConstant(0, MVT::i32), DAG, dl);
4312 Vec = Insert128BitVector(InsV, Vec,
4313 DAG.getConstant(4 /* NumElems/2 */, MVT::i32), DAG, dl);
4316 return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
4319 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
4320 /// that point to V2 points to its first element.
4321 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4322 EVT VT = SVOp->getValueType(0);
4323 unsigned NumElems = VT.getVectorNumElements();
4325 bool Changed = false;
4326 SmallVector<int, 8> MaskVec;
4327 SVOp->getMask(MaskVec);
4329 for (unsigned i = 0; i != NumElems; ++i) {
4330 if (MaskVec[i] > (int)NumElems) {
4331 MaskVec[i] = NumElems;
4336 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
4337 SVOp->getOperand(1), &MaskVec[0]);
4338 return SDValue(SVOp, 0);
4341 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
4342 /// operation of specified width.
4343 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4345 unsigned NumElems = VT.getVectorNumElements();
4346 SmallVector<int, 8> Mask;
4347 Mask.push_back(NumElems);
4348 for (unsigned i = 1; i != NumElems; ++i)
4350 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4353 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
4354 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4356 unsigned NumElems = VT.getVectorNumElements();
4357 SmallVector<int, 8> Mask;
4358 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4360 Mask.push_back(i + NumElems);
4362 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4365 /// getUnpackh - Returns a vector_shuffle node for an unpackh operation.
4366 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
4368 unsigned NumElems = VT.getVectorNumElements();
4369 unsigned Half = NumElems/2;
4370 SmallVector<int, 8> Mask;
4371 for (unsigned i = 0; i != Half; ++i) {
4372 Mask.push_back(i + Half);
4373 Mask.push_back(i + NumElems + Half);
4375 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4378 // PromoteSplati8i16 - All i16 and i8 vector types can't be used directly by
4379 // a generic shuffle instruction because the target has no such instructions.
4380 // Generate shuffles which repeat i16 and i8 several times until they can be
4381 // represented by v4f32 and then be manipulated by target suported shuffles.
4382 static SDValue PromoteSplati8i16(SDValue V, SelectionDAG &DAG, int &EltNo) {
4383 EVT VT = V.getValueType();
4384 int NumElems = VT.getVectorNumElements();
4385 DebugLoc dl = V.getDebugLoc();
4387 while (NumElems > 4) {
4388 if (EltNo < NumElems/2) {
4389 V = getUnpackl(DAG, dl, VT, V, V);
4391 V = getUnpackh(DAG, dl, VT, V, V);
4392 EltNo -= NumElems/2;
4399 /// getLegalSplat - Generate a legal splat with supported x86 shuffles
4400 static SDValue getLegalSplat(SelectionDAG &DAG, SDValue V, int EltNo) {
4401 EVT VT = V.getValueType();
4402 DebugLoc dl = V.getDebugLoc();
4403 assert((VT.getSizeInBits() == 128 || VT.getSizeInBits() == 256)
4404 && "Vector size not supported");
4406 if (VT.getSizeInBits() == 128) {
4407 V = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V);
4408 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
4409 V = DAG.getVectorShuffle(MVT::v4f32, dl, V, DAG.getUNDEF(MVT::v4f32),
4412 // To use VPERMILPS to splat scalars, the second half of indicies must
4413 // refer to the higher part, which is a duplication of the lower one,
4414 // because VPERMILPS can only handle in-lane permutations.
4415 int SplatMask[8] = { EltNo, EltNo, EltNo, EltNo,
4416 EltNo+4, EltNo+4, EltNo+4, EltNo+4 };
4418 V = DAG.getNode(ISD::BITCAST, dl, MVT::v8f32, V);
4419 V = DAG.getVectorShuffle(MVT::v8f32, dl, V, DAG.getUNDEF(MVT::v8f32),
4423 return DAG.getNode(ISD::BITCAST, dl, VT, V);
4426 /// PromoteSplat - Splat is promoted to target supported vector shuffles.
4427 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG) {
4428 EVT SrcVT = SV->getValueType(0);
4429 SDValue V1 = SV->getOperand(0);
4430 DebugLoc dl = SV->getDebugLoc();
4432 int EltNo = SV->getSplatIndex();
4433 int NumElems = SrcVT.getVectorNumElements();
4434 unsigned Size = SrcVT.getSizeInBits();
4436 assert(((Size == 128 && NumElems > 4) || Size == 256) &&
4437 "Unknown how to promote splat for type");
4439 // Extract the 128-bit part containing the splat element and update
4440 // the splat element index when it refers to the higher register.
4442 unsigned Idx = (EltNo > NumElems/2) ? NumElems/2 : 0;
4443 V1 = Extract128BitVector(V1, DAG.getConstant(Idx, MVT::i32), DAG, dl);
4445 EltNo -= NumElems/2;
4448 // All i16 and i8 vector types can't be used directly by a generic shuffle
4449 // instruction because the target has no such instruction. Generate shuffles
4450 // which repeat i16 and i8 several times until they fit in i32, and then can
4451 // be manipulated by target suported shuffles.
4452 EVT EltVT = SrcVT.getVectorElementType();
4453 if (EltVT == MVT::i8 || EltVT == MVT::i16)
4454 V1 = PromoteSplati8i16(V1, DAG, EltNo);
4456 // Recreate the 256-bit vector and place the same 128-bit vector
4457 // into the low and high part. This is necessary because we want
4458 // to use VPERM* to shuffle the vectors
4460 SDValue InsV = Insert128BitVector(DAG.getUNDEF(SrcVT), V1,
4461 DAG.getConstant(0, MVT::i32), DAG, dl);
4462 V1 = Insert128BitVector(InsV, V1,
4463 DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
4466 return getLegalSplat(DAG, V1, EltNo);
4469 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
4470 /// vector of zero or undef vector. This produces a shuffle where the low
4471 /// element of V2 is swizzled into the zero/undef vector, landing at element
4472 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4473 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4474 bool isZero, bool HasSSE2,
4475 SelectionDAG &DAG) {
4476 EVT VT = V2.getValueType();
4478 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
4479 unsigned NumElems = VT.getVectorNumElements();
4480 SmallVector<int, 16> MaskVec;
4481 for (unsigned i = 0; i != NumElems; ++i)
4482 // If this is the insertion idx, put the low elt of V2 here.
4483 MaskVec.push_back(i == Idx ? NumElems : i);
4484 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
4487 /// getShuffleScalarElt - Returns the scalar element that will make up the ith
4488 /// element of the result of the vector shuffle.
4489 static SDValue getShuffleScalarElt(SDNode *N, int Index, SelectionDAG &DAG,
4492 return SDValue(); // Limit search depth.
4494 SDValue V = SDValue(N, 0);
4495 EVT VT = V.getValueType();
4496 unsigned Opcode = V.getOpcode();
4498 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4499 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4500 Index = SV->getMaskElt(Index);
4503 return DAG.getUNDEF(VT.getVectorElementType());
4505 int NumElems = VT.getVectorNumElements();
4506 SDValue NewV = (Index < NumElems) ? SV->getOperand(0) : SV->getOperand(1);
4507 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG, Depth+1);
4510 // Recurse into target specific vector shuffles to find scalars.
4511 if (isTargetShuffle(Opcode)) {
4512 int NumElems = VT.getVectorNumElements();
4513 SmallVector<unsigned, 16> ShuffleMask;
4517 case X86ISD::SHUFPS:
4518 case X86ISD::SHUFPD:
4519 ImmN = N->getOperand(N->getNumOperands()-1);
4520 DecodeSHUFPSMask(NumElems,
4521 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4524 case X86ISD::PUNPCKHBW:
4525 case X86ISD::PUNPCKHWD:
4526 case X86ISD::PUNPCKHDQ:
4527 case X86ISD::PUNPCKHQDQ:
4528 DecodePUNPCKHMask(NumElems, ShuffleMask);
4530 case X86ISD::UNPCKHPS:
4531 case X86ISD::UNPCKHPD:
4532 case X86ISD::VUNPCKHPSY:
4533 case X86ISD::VUNPCKHPDY:
4534 DecodeUNPCKHPMask(NumElems, ShuffleMask);
4536 case X86ISD::PUNPCKLBW:
4537 case X86ISD::PUNPCKLWD:
4538 case X86ISD::PUNPCKLDQ:
4539 case X86ISD::PUNPCKLQDQ:
4540 DecodePUNPCKLMask(VT, ShuffleMask);
4542 case X86ISD::UNPCKLPS:
4543 case X86ISD::UNPCKLPD:
4544 case X86ISD::VUNPCKLPSY:
4545 case X86ISD::VUNPCKLPDY:
4546 DecodeUNPCKLPMask(VT, ShuffleMask);
4548 case X86ISD::MOVHLPS:
4549 DecodeMOVHLPSMask(NumElems, ShuffleMask);
4551 case X86ISD::MOVLHPS:
4552 DecodeMOVLHPSMask(NumElems, ShuffleMask);
4554 case X86ISD::PSHUFD:
4555 ImmN = N->getOperand(N->getNumOperands()-1);
4556 DecodePSHUFMask(NumElems,
4557 cast<ConstantSDNode>(ImmN)->getZExtValue(),
4560 case X86ISD::PSHUFHW:
4561 ImmN = N->getOperand(N->getNumOperands()-1);
4562 DecodePSHUFHWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4565 case X86ISD::PSHUFLW:
4566 ImmN = N->getOperand(N->getNumOperands()-1);
4567 DecodePSHUFLWMask(cast<ConstantSDNode>(ImmN)->getZExtValue(),
4571 case X86ISD::MOVSD: {
4572 // The index 0 always comes from the first element of the second source,
4573 // this is why MOVSS and MOVSD are used in the first place. The other
4574 // elements come from the other positions of the first source vector.
4575 unsigned OpNum = (Index == 0) ? 1 : 0;
4576 return getShuffleScalarElt(V.getOperand(OpNum).getNode(), Index, DAG,
4579 case X86ISD::VPERMILPS:
4580 ImmN = N->getOperand(N->getNumOperands()-1);
4581 DecodeVPERMILPSMask(4, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4584 case X86ISD::VPERMILPSY:
4585 ImmN = N->getOperand(N->getNumOperands()-1);
4586 DecodeVPERMILPSMask(8, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4589 case X86ISD::VPERMILPD:
4590 ImmN = N->getOperand(N->getNumOperands()-1);
4591 DecodeVPERMILPDMask(2, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4594 case X86ISD::VPERMILPDY:
4595 ImmN = N->getOperand(N->getNumOperands()-1);
4596 DecodeVPERMILPDMask(4, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4599 case X86ISD::VPERM2F128:
4600 ImmN = N->getOperand(N->getNumOperands()-1);
4601 DecodeVPERM2F128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(),
4605 assert("not implemented for target shuffle node");
4609 Index = ShuffleMask[Index];
4611 return DAG.getUNDEF(VT.getVectorElementType());
4613 SDValue NewV = (Index < NumElems) ? N->getOperand(0) : N->getOperand(1);
4614 return getShuffleScalarElt(NewV.getNode(), Index % NumElems, DAG,
4618 // Actual nodes that may contain scalar elements
4619 if (Opcode == ISD::BITCAST) {
4620 V = V.getOperand(0);
4621 EVT SrcVT = V.getValueType();
4622 unsigned NumElems = VT.getVectorNumElements();
4624 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4628 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4629 return (Index == 0) ? V.getOperand(0)
4630 : DAG.getUNDEF(VT.getVectorElementType());
4632 if (V.getOpcode() == ISD::BUILD_VECTOR)
4633 return V.getOperand(Index);
4638 /// getNumOfConsecutiveZeros - Return the number of elements of a vector
4639 /// shuffle operation which come from a consecutively from a zero. The
4640 /// search can start in two different directions, from left or right.
4642 unsigned getNumOfConsecutiveZeros(SDNode *N, int NumElems,
4643 bool ZerosFromLeft, SelectionDAG &DAG) {
4646 while (i < NumElems) {
4647 unsigned Index = ZerosFromLeft ? i : NumElems-i-1;
4648 SDValue Elt = getShuffleScalarElt(N, Index, DAG, 0);
4649 if (!(Elt.getNode() &&
4650 (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt))))
4658 /// isShuffleMaskConsecutive - Check if the shuffle mask indicies from MaskI to
4659 /// MaskE correspond consecutively to elements from one of the vector operands,
4660 /// starting from its index OpIdx. Also tell OpNum which source vector operand.
4662 bool isShuffleMaskConsecutive(ShuffleVectorSDNode *SVOp, int MaskI, int MaskE,
4663 int OpIdx, int NumElems, unsigned &OpNum) {
4664 bool SeenV1 = false;
4665 bool SeenV2 = false;
4667 for (int i = MaskI; i <= MaskE; ++i, ++OpIdx) {
4668 int Idx = SVOp->getMaskElt(i);
4669 // Ignore undef indicies
4678 // Only accept consecutive elements from the same vector
4679 if ((Idx % NumElems != OpIdx) || (SeenV1 && SeenV2))
4683 OpNum = SeenV1 ? 0 : 1;
4687 /// isVectorShiftRight - Returns true if the shuffle can be implemented as a
4688 /// logical left shift of a vector.
4689 static bool isVectorShiftRight(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4690 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4691 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4692 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4693 false /* check zeros from right */, DAG);
4699 // Considering the elements in the mask that are not consecutive zeros,
4700 // check if they consecutively come from only one of the source vectors.
4702 // V1 = {X, A, B, C} 0
4704 // vector_shuffle V1, V2 <1, 2, 3, X>
4706 if (!isShuffleMaskConsecutive(SVOp,
4707 0, // Mask Start Index
4708 NumElems-NumZeros-1, // Mask End Index
4709 NumZeros, // Where to start looking in the src vector
4710 NumElems, // Number of elements in vector
4711 OpSrc)) // Which source operand ?
4716 ShVal = SVOp->getOperand(OpSrc);
4720 /// isVectorShiftLeft - Returns true if the shuffle can be implemented as a
4721 /// logical left shift of a vector.
4722 static bool isVectorShiftLeft(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4723 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4724 unsigned NumElems = SVOp->getValueType(0).getVectorNumElements();
4725 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems,
4726 true /* check zeros from left */, DAG);
4732 // Considering the elements in the mask that are not consecutive zeros,
4733 // check if they consecutively come from only one of the source vectors.
4735 // 0 { A, B, X, X } = V2
4737 // vector_shuffle V1, V2 <X, X, 4, 5>
4739 if (!isShuffleMaskConsecutive(SVOp,
4740 NumZeros, // Mask Start Index
4741 NumElems-1, // Mask End Index
4742 0, // Where to start looking in the src vector
4743 NumElems, // Number of elements in vector
4744 OpSrc)) // Which source operand ?
4749 ShVal = SVOp->getOperand(OpSrc);
4753 /// isVectorShift - Returns true if the shuffle can be implemented as a
4754 /// logical left or right shift of a vector.
4755 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
4756 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
4757 if (isVectorShiftLeft(SVOp, DAG, isLeft, ShVal, ShAmt) ||
4758 isVectorShiftRight(SVOp, DAG, isLeft, ShVal, ShAmt))
4764 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
4766 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4767 unsigned NumNonZero, unsigned NumZero,
4769 const TargetLowering &TLI) {
4773 DebugLoc dl = Op.getDebugLoc();
4776 for (unsigned i = 0; i < 16; ++i) {
4777 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4778 if (ThisIsNonZero && First) {
4780 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4782 V = DAG.getUNDEF(MVT::v8i16);
4787 SDValue ThisElt(0, 0), LastElt(0, 0);
4788 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4789 if (LastIsNonZero) {
4790 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4791 MVT::i16, Op.getOperand(i-1));
4793 if (ThisIsNonZero) {
4794 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4795 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4796 ThisElt, DAG.getConstant(8, MVT::i8));
4798 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4802 if (ThisElt.getNode())
4803 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4804 DAG.getIntPtrConstant(i/2));
4808 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
4811 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
4813 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4814 unsigned NumNonZero, unsigned NumZero,
4816 const TargetLowering &TLI) {
4820 DebugLoc dl = Op.getDebugLoc();
4823 for (unsigned i = 0; i < 8; ++i) {
4824 bool isNonZero = (NonZeros & (1 << i)) != 0;
4828 V = getZeroVector(MVT::v8i16, true, DAG, dl);
4830 V = DAG.getUNDEF(MVT::v8i16);
4833 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4834 MVT::v8i16, V, Op.getOperand(i),
4835 DAG.getIntPtrConstant(i));
4842 /// getVShift - Return a vector logical shift node.
4844 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4845 unsigned NumBits, SelectionDAG &DAG,
4846 const TargetLowering &TLI, DebugLoc dl) {
4847 EVT ShVT = MVT::v2i64;
4848 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
4849 SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
4850 return DAG.getNode(ISD::BITCAST, dl, VT,
4851 DAG.getNode(Opc, dl, ShVT, SrcOp,
4852 DAG.getConstant(NumBits,
4853 TLI.getShiftAmountTy(SrcOp.getValueType()))));
4857 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
4858 SelectionDAG &DAG) const {
4860 // Check if the scalar load can be widened into a vector load. And if
4861 // the address is "base + cst" see if the cst can be "absorbed" into
4862 // the shuffle mask.
4863 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4864 SDValue Ptr = LD->getBasePtr();
4865 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4867 EVT PVT = LD->getValueType(0);
4868 if (PVT != MVT::i32 && PVT != MVT::f32)
4873 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4874 FI = FINode->getIndex();
4876 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4877 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4878 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4879 Offset = Ptr.getConstantOperandVal(1);
4880 Ptr = Ptr.getOperand(0);
4885 // FIXME: 256-bit vector instructions don't require a strict alignment,
4886 // improve this code to support it better.
4887 unsigned RequiredAlign = VT.getSizeInBits()/8;
4888 SDValue Chain = LD->getChain();
4889 // Make sure the stack object alignment is at least 16 or 32.
4890 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4891 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4892 if (MFI->isFixedObjectIndex(FI)) {
4893 // Can't change the alignment. FIXME: It's possible to compute
4894 // the exact stack offset and reference FI + adjust offset instead.
4895 // If someone *really* cares about this. That's the way to implement it.
4898 MFI->setObjectAlignment(FI, RequiredAlign);
4902 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4903 // Ptr + (Offset & ~15).
4906 if ((Offset % RequiredAlign) & 3)
4908 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4910 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
4911 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
4913 int EltNo = (Offset - StartOffset) >> 2;
4914 int NumElems = VT.getVectorNumElements();
4916 EVT CanonVT = VT.getSizeInBits() == 128 ? MVT::v4i32 : MVT::v8i32;
4917 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4918 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4919 LD->getPointerInfo().getWithOffset(StartOffset),
4922 // Canonicalize it to a v4i32 or v8i32 shuffle.
4923 SmallVector<int, 8> Mask;
4924 for (int i = 0; i < NumElems; ++i)
4925 Mask.push_back(EltNo);
4927 V1 = DAG.getNode(ISD::BITCAST, dl, CanonVT, V1);
4928 return DAG.getNode(ISD::BITCAST, dl, NVT,
4929 DAG.getVectorShuffle(CanonVT, dl, V1,
4930 DAG.getUNDEF(CanonVT),&Mask[0]));
4936 /// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
4937 /// vector of type 'VT', see if the elements can be replaced by a single large
4938 /// load which has the same value as a build_vector whose operands are 'elts'.
4940 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4942 /// FIXME: we'd also like to handle the case where the last elements are zero
4943 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4944 /// There's even a handy isZeroNode for that purpose.
4945 static SDValue EltsFromConsecutiveLoads(EVT VT, SmallVectorImpl<SDValue> &Elts,
4946 DebugLoc &DL, SelectionDAG &DAG) {
4947 EVT EltVT = VT.getVectorElementType();
4948 unsigned NumElems = Elts.size();
4950 LoadSDNode *LDBase = NULL;
4951 unsigned LastLoadedElt = -1U;
4953 // For each element in the initializer, see if we've found a load or an undef.
4954 // If we don't find an initial load element, or later load elements are
4955 // non-consecutive, bail out.
4956 for (unsigned i = 0; i < NumElems; ++i) {
4957 SDValue Elt = Elts[i];
4959 if (!Elt.getNode() ||
4960 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
4963 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
4965 LDBase = cast<LoadSDNode>(Elt.getNode());
4969 if (Elt.getOpcode() == ISD::UNDEF)
4972 LoadSDNode *LD = cast<LoadSDNode>(Elt);
4973 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
4978 // If we have found an entire vector of loads and undefs, then return a large
4979 // load of the entire vector width starting at the base pointer. If we found
4980 // consecutive loads for the low half, generate a vzext_load node.
4981 if (LastLoadedElt == NumElems - 1) {
4982 if (DAG.InferPtrAlignment(LDBase->getBasePtr()) >= 16)
4983 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4984 LDBase->getPointerInfo(),
4985 LDBase->isVolatile(), LDBase->isNonTemporal(), 0);
4986 return DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
4987 LDBase->getPointerInfo(),
4988 LDBase->isVolatile(), LDBase->isNonTemporal(),
4989 LDBase->getAlignment());
4990 } else if (NumElems == 4 && LastLoadedElt == 1 &&
4991 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
4992 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
4993 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
4994 SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
4996 LDBase->getMemOperand());
4997 return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
5003 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5004 DebugLoc dl = Op.getDebugLoc();
5006 EVT VT = Op.getValueType();
5007 EVT ExtVT = VT.getVectorElementType();
5008 unsigned NumElems = Op.getNumOperands();
5010 // Vectors containing all zeros can be matched by pxor and xorps later
5011 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5012 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5013 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5014 if (Op.getValueType() == MVT::v4i32 ||
5015 Op.getValueType() == MVT::v8i32)
5018 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
5021 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5022 // vectors or broken into v4i32 operations on 256-bit vectors.
5023 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5024 if (Op.getValueType() == MVT::v4i32)
5027 return getOnesVector(Op.getValueType(), DAG, dl);
5030 unsigned EVTBits = ExtVT.getSizeInBits();
5032 unsigned NumZero = 0;
5033 unsigned NumNonZero = 0;
5034 unsigned NonZeros = 0;
5035 bool IsAllConstants = true;
5036 SmallSet<SDValue, 8> Values;
5037 for (unsigned i = 0; i < NumElems; ++i) {
5038 SDValue Elt = Op.getOperand(i);
5039 if (Elt.getOpcode() == ISD::UNDEF)
5042 if (Elt.getOpcode() != ISD::Constant &&
5043 Elt.getOpcode() != ISD::ConstantFP)
5044 IsAllConstants = false;
5045 if (X86::isZeroNode(Elt))
5048 NonZeros |= (1 << i);
5053 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5054 if (NumNonZero == 0)
5055 return DAG.getUNDEF(VT);
5057 // Special case for single non-zero, non-undef, element.
5058 if (NumNonZero == 1) {
5059 unsigned Idx = CountTrailingZeros_32(NonZeros);
5060 SDValue Item = Op.getOperand(Idx);
5062 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5063 // the value are obviously zero, truncate the value to i32 and do the
5064 // insertion that way. Only do this if the value is non-constant or if the
5065 // value is a constant being inserted into element 0. It is cheaper to do
5066 // a constant pool load than it is to do a movd + shuffle.
5067 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5068 (!IsAllConstants || Idx == 0)) {
5069 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5071 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5072 EVT VecVT = MVT::v4i32;
5073 unsigned VecElts = 4;
5075 // Truncate the value (which may itself be a constant) to i32, and
5076 // convert it to a vector with movd (S2V+shuffle to zero extend).
5077 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5078 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5079 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
5080 Subtarget->hasSSE2(), DAG);
5082 // Now we have our 32-bit value zero extended in the low element of
5083 // a vector. If Idx != 0, swizzle it into place.
5085 SmallVector<int, 4> Mask;
5086 Mask.push_back(Idx);
5087 for (unsigned i = 1; i != VecElts; ++i)
5089 Item = DAG.getVectorShuffle(VecVT, dl, Item,
5090 DAG.getUNDEF(Item.getValueType()),
5093 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
5097 // If we have a constant or non-constant insertion into the low element of
5098 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5099 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5100 // depending on what the source datatype is.
5103 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5104 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5105 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5106 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5107 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5108 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
5110 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5111 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5112 assert(VT.getSizeInBits() == 128 && "Expected an SSE value type!");
5113 EVT MiddleVT = MVT::v4i32;
5114 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
5115 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
5116 Subtarget->hasSSE2(), DAG);
5117 return DAG.getNode(ISD::BITCAST, dl, VT, Item);
5121 // Is it a vector logical left shift?
5122 if (NumElems == 2 && Idx == 1 &&
5123 X86::isZeroNode(Op.getOperand(0)) &&
5124 !X86::isZeroNode(Op.getOperand(1))) {
5125 unsigned NumBits = VT.getSizeInBits();
5126 return getVShift(true, VT,
5127 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5128 VT, Op.getOperand(1)),
5129 NumBits/2, DAG, *this, dl);
5132 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
5135 // Otherwise, if this is a vector with i32 or f32 elements, and the element
5136 // is a non-constant being inserted into an element other than the low one,
5137 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
5138 // movd/movss) to move this into the low element, then shuffle it into
5140 if (EVTBits == 32) {
5141 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5143 // Turn it into a shuffle of zero and zero-extended scalar to vector.
5144 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
5145 Subtarget->hasSSE2(), DAG);
5146 SmallVector<int, 8> MaskVec;
5147 for (unsigned i = 0; i < NumElems; i++)
5148 MaskVec.push_back(i == Idx ? 0 : 1);
5149 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
5153 // Splat is obviously ok. Let legalizer expand it to a shuffle.
5154 if (Values.size() == 1) {
5155 if (EVTBits == 32) {
5156 // Instead of a shuffle like this:
5157 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
5158 // Check if it's possible to issue this instead.
5159 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
5160 unsigned Idx = CountTrailingZeros_32(NonZeros);
5161 SDValue Item = Op.getOperand(Idx);
5162 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
5163 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
5168 // A vector full of immediates; various special cases are already
5169 // handled, so this is best done with a single constant-pool load.
5173 // For AVX-length vectors, build the individual 128-bit pieces and use
5174 // shuffles to put them in place.
5175 if (VT.getSizeInBits() == 256 && !ISD::isBuildVectorAllZeros(Op.getNode())) {
5176 SmallVector<SDValue, 32> V;
5177 for (unsigned i = 0; i < NumElems; ++i)
5178 V.push_back(Op.getOperand(i));
5180 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
5182 // Build both the lower and upper subvector.
5183 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
5184 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
5187 // Recreate the wider vector with the lower and upper part.
5188 SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Lower,
5189 DAG.getConstant(0, MVT::i32), DAG, dl);
5190 return Insert128BitVector(Vec, Upper, DAG.getConstant(NumElems/2, MVT::i32),
5194 // Let legalizer expand 2-wide build_vectors.
5195 if (EVTBits == 64) {
5196 if (NumNonZero == 1) {
5197 // One half is zero or undef.
5198 unsigned Idx = CountTrailingZeros_32(NonZeros);
5199 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
5200 Op.getOperand(Idx));
5201 return getShuffleVectorZeroOrUndef(V2, Idx, true,
5202 Subtarget->hasSSE2(), DAG);
5207 // If element VT is < 32 bits, convert it to inserts into a zero vector.
5208 if (EVTBits == 8 && NumElems == 16) {
5209 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
5211 if (V.getNode()) return V;
5214 if (EVTBits == 16 && NumElems == 8) {
5215 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
5217 if (V.getNode()) return V;
5220 // If element VT is == 32 bits, turn it into a number of shuffles.
5221 SmallVector<SDValue, 8> V;
5223 if (NumElems == 4 && NumZero > 0) {
5224 for (unsigned i = 0; i < 4; ++i) {
5225 bool isZero = !(NonZeros & (1 << i));
5227 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
5229 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5232 for (unsigned i = 0; i < 2; ++i) {
5233 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
5236 V[i] = V[i*2]; // Must be a zero vector.
5239 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
5242 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
5245 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
5250 SmallVector<int, 8> MaskVec;
5251 bool Reverse = (NonZeros & 0x3) == 2;
5252 for (unsigned i = 0; i < 2; ++i)
5253 MaskVec.push_back(Reverse ? 1-i : i);
5254 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
5255 for (unsigned i = 0; i < 2; ++i)
5256 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
5257 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
5260 if (Values.size() > 1 && VT.getSizeInBits() == 128) {
5261 // Check for a build vector of consecutive loads.
5262 for (unsigned i = 0; i < NumElems; ++i)
5263 V[i] = Op.getOperand(i);
5265 // Check for elements which are consecutive loads.
5266 SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
5270 // For SSE 4.1, use insertps to put the high elements into the low element.
5271 if (getSubtarget()->hasSSE41()) {
5273 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
5274 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
5276 Result = DAG.getUNDEF(VT);
5278 for (unsigned i = 1; i < NumElems; ++i) {
5279 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
5280 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
5281 Op.getOperand(i), DAG.getIntPtrConstant(i));
5286 // Otherwise, expand into a number of unpckl*, start by extending each of
5287 // our (non-undef) elements to the full vector width with the element in the
5288 // bottom slot of the vector (which generates no code for SSE).
5289 for (unsigned i = 0; i < NumElems; ++i) {
5290 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
5291 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
5293 V[i] = DAG.getUNDEF(VT);
5296 // Next, we iteratively mix elements, e.g. for v4f32:
5297 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
5298 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
5299 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
5300 unsigned EltStride = NumElems >> 1;
5301 while (EltStride != 0) {
5302 for (unsigned i = 0; i < EltStride; ++i) {
5303 // If V[i+EltStride] is undef and this is the first round of mixing,
5304 // then it is safe to just drop this shuffle: V[i] is already in the
5305 // right place, the one element (since it's the first round) being
5306 // inserted as undef can be dropped. This isn't safe for successive
5307 // rounds because they will permute elements within both vectors.
5308 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
5309 EltStride == NumElems/2)
5312 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
5321 // LowerMMXCONCAT_VECTORS - We support concatenate two MMX registers and place
5322 // them in a MMX register. This is better than doing a stack convert.
5323 static SDValue LowerMMXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5324 DebugLoc dl = Op.getDebugLoc();
5325 EVT ResVT = Op.getValueType();
5327 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
5328 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
5330 SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
5331 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
5332 InVec = Op.getOperand(1);
5333 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
5334 unsigned NumElts = ResVT.getVectorNumElements();
5335 VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
5336 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
5337 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
5339 InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
5340 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
5341 Mask[0] = 0; Mask[1] = 2;
5342 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
5344 return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
5347 // LowerAVXCONCAT_VECTORS - 256-bit AVX can use the vinsertf128 instruction
5348 // to create 256-bit vectors from two other 128-bit ones.
5349 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
5350 DebugLoc dl = Op.getDebugLoc();
5351 EVT ResVT = Op.getValueType();
5353 assert(ResVT.getSizeInBits() == 256 && "Value type must be 256-bit wide");
5355 SDValue V1 = Op.getOperand(0);
5356 SDValue V2 = Op.getOperand(1);
5357 unsigned NumElems = ResVT.getVectorNumElements();
5359 SDValue V = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, ResVT), V1,
5360 DAG.getConstant(0, MVT::i32), DAG, dl);
5361 return Insert128BitVector(V, V2, DAG.getConstant(NumElems/2, MVT::i32),
5366 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) const {
5367 EVT ResVT = Op.getValueType();
5369 assert(Op.getNumOperands() == 2);
5370 assert((ResVT.getSizeInBits() == 128 || ResVT.getSizeInBits() == 256) &&
5371 "Unsupported CONCAT_VECTORS for value type");
5373 // We support concatenate two MMX registers and place them in a MMX register.
5374 // This is better than doing a stack convert.
5375 if (ResVT.is128BitVector())
5376 return LowerMMXCONCAT_VECTORS(Op, DAG);
5378 // 256-bit AVX can use the vinsertf128 instruction to create 256-bit vectors
5379 // from two other 128-bit ones.
5380 return LowerAVXCONCAT_VECTORS(Op, DAG);
5383 // v8i16 shuffles - Prefer shuffles in the following order:
5384 // 1. [all] pshuflw, pshufhw, optional move
5385 // 2. [ssse3] 1 x pshufb
5386 // 3. [ssse3] 2 x pshufb + 1 x por
5387 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
5389 X86TargetLowering::LowerVECTOR_SHUFFLEv8i16(SDValue Op,
5390 SelectionDAG &DAG) const {
5391 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5392 SDValue V1 = SVOp->getOperand(0);
5393 SDValue V2 = SVOp->getOperand(1);
5394 DebugLoc dl = SVOp->getDebugLoc();
5395 SmallVector<int, 8> MaskVals;
5397 // Determine if more than 1 of the words in each of the low and high quadwords
5398 // of the result come from the same quadword of one of the two inputs. Undef
5399 // mask values count as coming from any quadword, for better codegen.
5400 SmallVector<unsigned, 4> LoQuad(4);
5401 SmallVector<unsigned, 4> HiQuad(4);
5402 BitVector InputQuads(4);
5403 for (unsigned i = 0; i < 8; ++i) {
5404 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
5405 int EltIdx = SVOp->getMaskElt(i);
5406 MaskVals.push_back(EltIdx);
5415 InputQuads.set(EltIdx / 4);
5418 int BestLoQuad = -1;
5419 unsigned MaxQuad = 1;
5420 for (unsigned i = 0; i < 4; ++i) {
5421 if (LoQuad[i] > MaxQuad) {
5423 MaxQuad = LoQuad[i];
5427 int BestHiQuad = -1;
5429 for (unsigned i = 0; i < 4; ++i) {
5430 if (HiQuad[i] > MaxQuad) {
5432 MaxQuad = HiQuad[i];
5436 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
5437 // of the two input vectors, shuffle them into one input vector so only a
5438 // single pshufb instruction is necessary. If There are more than 2 input
5439 // quads, disable the next transformation since it does not help SSSE3.
5440 bool V1Used = InputQuads[0] || InputQuads[1];
5441 bool V2Used = InputQuads[2] || InputQuads[3];
5442 if (Subtarget->hasSSSE3()) {
5443 if (InputQuads.count() == 2 && V1Used && V2Used) {
5444 BestLoQuad = InputQuads.find_first();
5445 BestHiQuad = InputQuads.find_next(BestLoQuad);
5447 if (InputQuads.count() > 2) {
5453 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
5454 // the shuffle mask. If a quad is scored as -1, that means that it contains
5455 // words from all 4 input quadwords.
5457 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
5458 SmallVector<int, 8> MaskV;
5459 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
5460 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
5461 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
5462 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
5463 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
5464 NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
5466 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
5467 // source words for the shuffle, to aid later transformations.
5468 bool AllWordsInNewV = true;
5469 bool InOrder[2] = { true, true };
5470 for (unsigned i = 0; i != 8; ++i) {
5471 int idx = MaskVals[i];
5473 InOrder[i/4] = false;
5474 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
5476 AllWordsInNewV = false;
5480 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
5481 if (AllWordsInNewV) {
5482 for (int i = 0; i != 8; ++i) {
5483 int idx = MaskVals[i];
5486 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
5487 if ((idx != i) && idx < 4)
5489 if ((idx != i) && idx > 3)
5498 // If we've eliminated the use of V2, and the new mask is a pshuflw or
5499 // pshufhw, that's as cheap as it gets. Return the new shuffle.
5500 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
5501 unsigned Opc = pshufhw ? X86ISD::PSHUFHW : X86ISD::PSHUFLW;
5502 unsigned TargetMask = 0;
5503 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
5504 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
5505 TargetMask = pshufhw ? X86::getShufflePSHUFHWImmediate(NewV.getNode()):
5506 X86::getShufflePSHUFLWImmediate(NewV.getNode());
5507 V1 = NewV.getOperand(0);
5508 return getTargetShuffleNode(Opc, dl, MVT::v8i16, V1, TargetMask, DAG);
5512 // If we have SSSE3, and all words of the result are from 1 input vector,
5513 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
5514 // is present, fall back to case 4.
5515 if (Subtarget->hasSSSE3()) {
5516 SmallVector<SDValue,16> pshufbMask;
5518 // If we have elements from both input vectors, set the high bit of the
5519 // shuffle mask element to zero out elements that come from V2 in the V1
5520 // mask, and elements that come from V1 in the V2 mask, so that the two
5521 // results can be OR'd together.
5522 bool TwoInputs = V1Used && V2Used;
5523 for (unsigned i = 0; i != 8; ++i) {
5524 int EltIdx = MaskVals[i] * 2;
5525 if (TwoInputs && (EltIdx >= 16)) {
5526 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5527 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5530 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5531 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
5533 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
5534 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5535 DAG.getNode(ISD::BUILD_VECTOR, dl,
5536 MVT::v16i8, &pshufbMask[0], 16));
5538 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5540 // Calculate the shuffle mask for the second input, shuffle it, and
5541 // OR it with the first shuffled input.
5543 for (unsigned i = 0; i != 8; ++i) {
5544 int EltIdx = MaskVals[i] * 2;
5546 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5547 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5550 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5551 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
5553 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
5554 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5555 DAG.getNode(ISD::BUILD_VECTOR, dl,
5556 MVT::v16i8, &pshufbMask[0], 16));
5557 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5558 return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5561 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
5562 // and update MaskVals with new element order.
5563 BitVector InOrder(8);
5564 if (BestLoQuad >= 0) {
5565 SmallVector<int, 8> MaskV;
5566 for (int i = 0; i != 4; ++i) {
5567 int idx = MaskVals[i];
5569 MaskV.push_back(-1);
5571 } else if ((idx / 4) == BestLoQuad) {
5572 MaskV.push_back(idx & 3);
5575 MaskV.push_back(-1);
5578 for (unsigned i = 4; i != 8; ++i)
5580 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5583 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5584 NewV = getTargetShuffleNode(X86ISD::PSHUFLW, dl, MVT::v8i16,
5586 X86::getShufflePSHUFLWImmediate(NewV.getNode()),
5590 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
5591 // and update MaskVals with the new element order.
5592 if (BestHiQuad >= 0) {
5593 SmallVector<int, 8> MaskV;
5594 for (unsigned i = 0; i != 4; ++i)
5596 for (unsigned i = 4; i != 8; ++i) {
5597 int idx = MaskVals[i];
5599 MaskV.push_back(-1);
5601 } else if ((idx / 4) == BestHiQuad) {
5602 MaskV.push_back((idx & 3) + 4);
5605 MaskV.push_back(-1);
5608 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
5611 if (NewV.getOpcode() == ISD::VECTOR_SHUFFLE && Subtarget->hasSSSE3())
5612 NewV = getTargetShuffleNode(X86ISD::PSHUFHW, dl, MVT::v8i16,
5614 X86::getShufflePSHUFHWImmediate(NewV.getNode()),
5618 // In case BestHi & BestLo were both -1, which means each quadword has a word
5619 // from each of the four input quadwords, calculate the InOrder bitvector now
5620 // before falling through to the insert/extract cleanup.
5621 if (BestLoQuad == -1 && BestHiQuad == -1) {
5623 for (int i = 0; i != 8; ++i)
5624 if (MaskVals[i] < 0 || MaskVals[i] == i)
5628 // The other elements are put in the right place using pextrw and pinsrw.
5629 for (unsigned i = 0; i != 8; ++i) {
5632 int EltIdx = MaskVals[i];
5635 SDValue ExtOp = (EltIdx < 8)
5636 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
5637 DAG.getIntPtrConstant(EltIdx))
5638 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
5639 DAG.getIntPtrConstant(EltIdx - 8));
5640 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
5641 DAG.getIntPtrConstant(i));
5646 // v16i8 shuffles - Prefer shuffles in the following order:
5647 // 1. [ssse3] 1 x pshufb
5648 // 2. [ssse3] 2 x pshufb + 1 x por
5649 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
5651 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
5653 const X86TargetLowering &TLI) {
5654 SDValue V1 = SVOp->getOperand(0);
5655 SDValue V2 = SVOp->getOperand(1);
5656 DebugLoc dl = SVOp->getDebugLoc();
5657 SmallVector<int, 16> MaskVals;
5658 SVOp->getMask(MaskVals);
5660 // If we have SSSE3, case 1 is generated when all result bytes come from
5661 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
5662 // present, fall back to case 3.
5663 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
5666 for (unsigned i = 0; i < 16; ++i) {
5667 int EltIdx = MaskVals[i];
5676 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
5677 if (TLI.getSubtarget()->hasSSSE3()) {
5678 SmallVector<SDValue,16> pshufbMask;
5680 // If all result elements are from one input vector, then only translate
5681 // undef mask values to 0x80 (zero out result) in the pshufb mask.
5683 // Otherwise, we have elements from both input vectors, and must zero out
5684 // elements that come from V2 in the first mask, and V1 in the second mask
5685 // so that we can OR them together.
5686 bool TwoInputs = !(V1Only || V2Only);
5687 for (unsigned i = 0; i != 16; ++i) {
5688 int EltIdx = MaskVals[i];
5689 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
5690 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5693 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
5695 // If all the elements are from V2, assign it to V1 and return after
5696 // building the first pshufb.
5699 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
5700 DAG.getNode(ISD::BUILD_VECTOR, dl,
5701 MVT::v16i8, &pshufbMask[0], 16));
5705 // Calculate the shuffle mask for the second input, shuffle it, and
5706 // OR it with the first shuffled input.
5708 for (unsigned i = 0; i != 16; ++i) {
5709 int EltIdx = MaskVals[i];
5711 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
5714 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
5716 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
5717 DAG.getNode(ISD::BUILD_VECTOR, dl,
5718 MVT::v16i8, &pshufbMask[0], 16));
5719 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
5722 // No SSSE3 - Calculate in place words and then fix all out of place words
5723 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
5724 // the 16 different words that comprise the two doublequadword input vectors.
5725 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
5726 V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
5727 SDValue NewV = V2Only ? V2 : V1;
5728 for (int i = 0; i != 8; ++i) {
5729 int Elt0 = MaskVals[i*2];
5730 int Elt1 = MaskVals[i*2+1];
5732 // This word of the result is all undef, skip it.
5733 if (Elt0 < 0 && Elt1 < 0)
5736 // This word of the result is already in the correct place, skip it.
5737 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
5739 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
5742 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
5743 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
5746 // If Elt0 and Elt1 are defined, are consecutive, and can be load
5747 // using a single extract together, load it and store it.
5748 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
5749 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5750 DAG.getIntPtrConstant(Elt1 / 2));
5751 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5752 DAG.getIntPtrConstant(i));
5756 // If Elt1 is defined, extract it from the appropriate source. If the
5757 // source byte is not also odd, shift the extracted word left 8 bits
5758 // otherwise clear the bottom 8 bits if we need to do an or.
5760 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
5761 DAG.getIntPtrConstant(Elt1 / 2));
5762 if ((Elt1 & 1) == 0)
5763 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
5765 TLI.getShiftAmountTy(InsElt.getValueType())));
5767 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
5768 DAG.getConstant(0xFF00, MVT::i16));
5770 // If Elt0 is defined, extract it from the appropriate source. If the
5771 // source byte is not also even, shift the extracted word right 8 bits. If
5772 // Elt1 was also defined, OR the extracted values together before
5773 // inserting them in the result.
5775 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
5776 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
5777 if ((Elt0 & 1) != 0)
5778 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
5780 TLI.getShiftAmountTy(InsElt0.getValueType())));
5782 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
5783 DAG.getConstant(0x00FF, MVT::i16));
5784 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
5787 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
5788 DAG.getIntPtrConstant(i));
5790 return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
5793 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
5794 /// ones, or rewriting v4i32 / v4f32 as 2 wide ones if possible. This can be
5795 /// done when every pair / quad of shuffle mask elements point to elements in
5796 /// the right sequence. e.g.
5797 /// vector_shuffle X, Y, <2, 3, | 10, 11, | 0, 1, | 14, 15>
5799 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
5800 SelectionDAG &DAG, DebugLoc dl) {
5801 EVT VT = SVOp->getValueType(0);
5802 SDValue V1 = SVOp->getOperand(0);
5803 SDValue V2 = SVOp->getOperand(1);
5804 unsigned NumElems = VT.getVectorNumElements();
5805 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
5807 switch (VT.getSimpleVT().SimpleTy) {
5808 default: assert(false && "Unexpected!");
5809 case MVT::v4f32: NewVT = MVT::v2f64; break;
5810 case MVT::v4i32: NewVT = MVT::v2i64; break;
5811 case MVT::v8i16: NewVT = MVT::v4i32; break;
5812 case MVT::v16i8: NewVT = MVT::v4i32; break;
5815 int Scale = NumElems / NewWidth;
5816 SmallVector<int, 8> MaskVec;
5817 for (unsigned i = 0; i < NumElems; i += Scale) {
5819 for (int j = 0; j < Scale; ++j) {
5820 int EltIdx = SVOp->getMaskElt(i+j);
5824 StartIdx = EltIdx - (EltIdx % Scale);
5825 if (EltIdx != StartIdx + j)
5829 MaskVec.push_back(-1);
5831 MaskVec.push_back(StartIdx / Scale);
5834 V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
5835 V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
5836 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
5839 /// getVZextMovL - Return a zero-extending vector move low node.
5841 static SDValue getVZextMovL(EVT VT, EVT OpVT,
5842 SDValue SrcOp, SelectionDAG &DAG,
5843 const X86Subtarget *Subtarget, DebugLoc dl) {
5844 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
5845 LoadSDNode *LD = NULL;
5846 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
5847 LD = dyn_cast<LoadSDNode>(SrcOp);
5849 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
5851 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
5852 if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
5853 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
5854 SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
5855 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
5857 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
5858 return DAG.getNode(ISD::BITCAST, dl, VT,
5859 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5860 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5868 return DAG.getNode(ISD::BITCAST, dl, VT,
5869 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
5870 DAG.getNode(ISD::BITCAST, dl,
5874 /// areShuffleHalvesWithinDisjointLanes - Check whether each half of a vector
5875 /// shuffle node referes to only one lane in the sources.
5876 static bool areShuffleHalvesWithinDisjointLanes(ShuffleVectorSDNode *SVOp) {
5877 EVT VT = SVOp->getValueType(0);
5878 int NumElems = VT.getVectorNumElements();
5879 int HalfSize = NumElems/2;
5880 SmallVector<int, 16> M;
5882 bool MatchA = false, MatchB = false;
5884 for (int l = 0; l < NumElems*2; l += HalfSize) {
5885 if (isUndefOrInRange(M, 0, HalfSize, l, l+HalfSize)) {
5891 for (int l = 0; l < NumElems*2; l += HalfSize) {
5892 if (isUndefOrInRange(M, HalfSize, HalfSize, l, l+HalfSize)) {
5898 return MatchA && MatchB;
5901 /// LowerVECTOR_SHUFFLE_256 - Handle all 256-bit wide vectors shuffles
5902 /// which could not be matched by any known target speficic shuffle
5904 LowerVECTOR_SHUFFLE_256(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5905 if (areShuffleHalvesWithinDisjointLanes(SVOp)) {
5906 // If each half of a vector shuffle node referes to only one lane in the
5907 // source vectors, extract each used 128-bit lane and shuffle them using
5908 // 128-bit shuffles. Then, concatenate the results. Otherwise leave
5909 // the work to the legalizer.
5910 DebugLoc dl = SVOp->getDebugLoc();
5911 EVT VT = SVOp->getValueType(0);
5912 int NumElems = VT.getVectorNumElements();
5913 int HalfSize = NumElems/2;
5915 // Extract the reference for each half
5916 int FstVecExtractIdx = 0, SndVecExtractIdx = 0;
5917 int FstVecOpNum = 0, SndVecOpNum = 0;
5918 for (int i = 0; i < HalfSize; ++i) {
5919 int Elt = SVOp->getMaskElt(i);
5920 if (SVOp->getMaskElt(i) < 0)
5922 FstVecOpNum = Elt/NumElems;
5923 FstVecExtractIdx = Elt % NumElems < HalfSize ? 0 : HalfSize;
5926 for (int i = HalfSize; i < NumElems; ++i) {
5927 int Elt = SVOp->getMaskElt(i);
5928 if (SVOp->getMaskElt(i) < 0)
5930 SndVecOpNum = Elt/NumElems;
5931 SndVecExtractIdx = Elt % NumElems < HalfSize ? 0 : HalfSize;
5935 // Extract the subvectors
5936 SDValue V1 = Extract128BitVector(SVOp->getOperand(FstVecOpNum),
5937 DAG.getConstant(FstVecExtractIdx, MVT::i32), DAG, dl);
5938 SDValue V2 = Extract128BitVector(SVOp->getOperand(SndVecOpNum),
5939 DAG.getConstant(SndVecExtractIdx, MVT::i32), DAG, dl);
5941 // Generate 128-bit shuffles
5942 SmallVector<int, 16> MaskV1, MaskV2;
5943 for (int i = 0; i < HalfSize; ++i) {
5944 int Elt = SVOp->getMaskElt(i);
5945 MaskV1.push_back(Elt < 0 ? Elt : Elt % HalfSize);
5947 for (int i = HalfSize; i < NumElems; ++i) {
5948 int Elt = SVOp->getMaskElt(i);
5949 MaskV2.push_back(Elt < 0 ? Elt : Elt % HalfSize);
5952 EVT NVT = V1.getValueType();
5953 V1 = DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &MaskV1[0]);
5954 V2 = DAG.getVectorShuffle(NVT, dl, V2, DAG.getUNDEF(NVT), &MaskV2[0]);
5956 // Concatenate the result back
5957 SDValue V = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), V1,
5958 DAG.getConstant(0, MVT::i32), DAG, dl);
5959 return Insert128BitVector(V, V2, DAG.getConstant(NumElems/2, MVT::i32),
5966 /// LowerVECTOR_SHUFFLE_128v4 - Handle all 128-bit wide vectors with
5967 /// 4 elements, and match them with several different shuffle types.
5969 LowerVECTOR_SHUFFLE_128v4(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
5970 SDValue V1 = SVOp->getOperand(0);
5971 SDValue V2 = SVOp->getOperand(1);
5972 DebugLoc dl = SVOp->getDebugLoc();
5973 EVT VT = SVOp->getValueType(0);
5975 assert(VT.getSizeInBits() == 128 && "Unsupported vector size");
5977 SmallVector<std::pair<int, int>, 8> Locs;
5979 SmallVector<int, 8> Mask1(4U, -1);
5980 SmallVector<int, 8> PermMask;
5981 SVOp->getMask(PermMask);
5985 for (unsigned i = 0; i != 4; ++i) {
5986 int Idx = PermMask[i];
5988 Locs[i] = std::make_pair(-1, -1);
5990 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
5992 Locs[i] = std::make_pair(0, NumLo);
5996 Locs[i] = std::make_pair(1, NumHi);
5998 Mask1[2+NumHi] = Idx;
6004 if (NumLo <= 2 && NumHi <= 2) {
6005 // If no more than two elements come from either vector. This can be
6006 // implemented with two shuffles. First shuffle gather the elements.
6007 // The second shuffle, which takes the first shuffle as both of its
6008 // vector operands, put the elements into the right order.
6009 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6011 SmallVector<int, 8> Mask2(4U, -1);
6013 for (unsigned i = 0; i != 4; ++i) {
6014 if (Locs[i].first == -1)
6017 unsigned Idx = (i < 2) ? 0 : 4;
6018 Idx += Locs[i].first * 2 + Locs[i].second;
6023 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
6024 } else if (NumLo == 3 || NumHi == 3) {
6025 // Otherwise, we must have three elements from one vector, call it X, and
6026 // one element from the other, call it Y. First, use a shufps to build an
6027 // intermediate vector with the one element from Y and the element from X
6028 // that will be in the same half in the final destination (the indexes don't
6029 // matter). Then, use a shufps to build the final vector, taking the half
6030 // containing the element from Y from the intermediate, and the other half
6033 // Normalize it so the 3 elements come from V1.
6034 CommuteVectorShuffleMask(PermMask, VT);
6038 // Find the element from V2.
6040 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
6041 int Val = PermMask[HiIndex];
6048 Mask1[0] = PermMask[HiIndex];
6050 Mask1[2] = PermMask[HiIndex^1];
6052 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6055 Mask1[0] = PermMask[0];
6056 Mask1[1] = PermMask[1];
6057 Mask1[2] = HiIndex & 1 ? 6 : 4;
6058 Mask1[3] = HiIndex & 1 ? 4 : 6;
6059 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
6061 Mask1[0] = HiIndex & 1 ? 2 : 0;
6062 Mask1[1] = HiIndex & 1 ? 0 : 2;
6063 Mask1[2] = PermMask[2];
6064 Mask1[3] = PermMask[3];
6069 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
6073 // Break it into (shuffle shuffle_hi, shuffle_lo).
6076 SmallVector<int,8> LoMask(4U, -1);
6077 SmallVector<int,8> HiMask(4U, -1);
6079 SmallVector<int,8> *MaskPtr = &LoMask;
6080 unsigned MaskIdx = 0;
6083 for (unsigned i = 0; i != 4; ++i) {
6090 int Idx = PermMask[i];
6092 Locs[i] = std::make_pair(-1, -1);
6093 } else if (Idx < 4) {
6094 Locs[i] = std::make_pair(MaskIdx, LoIdx);
6095 (*MaskPtr)[LoIdx] = Idx;
6098 Locs[i] = std::make_pair(MaskIdx, HiIdx);
6099 (*MaskPtr)[HiIdx] = Idx;
6104 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
6105 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
6106 SmallVector<int, 8> MaskOps;
6107 for (unsigned i = 0; i != 4; ++i) {
6108 if (Locs[i].first == -1) {
6109 MaskOps.push_back(-1);
6111 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
6112 MaskOps.push_back(Idx);
6115 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
6118 static bool MayFoldVectorLoad(SDValue V) {
6119 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
6120 V = V.getOperand(0);
6121 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
6122 V = V.getOperand(0);
6128 // FIXME: the version above should always be used. Since there's
6129 // a bug where several vector shuffles can't be folded because the
6130 // DAG is not updated during lowering and a node claims to have two
6131 // uses while it only has one, use this version, and let isel match
6132 // another instruction if the load really happens to have more than
6133 // one use. Remove this version after this bug get fixed.
6134 // rdar://8434668, PR8156
6135 static bool RelaxedMayFoldVectorLoad(SDValue V) {
6136 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
6137 V = V.getOperand(0);
6138 if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
6139 V = V.getOperand(0);
6140 if (ISD::isNormalLoad(V.getNode()))
6145 /// CanFoldShuffleIntoVExtract - Check if the current shuffle is used by
6146 /// a vector extract, and if both can be later optimized into a single load.
6147 /// This is done in visitEXTRACT_VECTOR_ELT and the conditions are checked
6148 /// here because otherwise a target specific shuffle node is going to be
6149 /// emitted for this shuffle, and the optimization not done.
6150 /// FIXME: This is probably not the best approach, but fix the problem
6151 /// until the right path is decided.
6153 bool CanXFormVExtractWithShuffleIntoLoad(SDValue V, SelectionDAG &DAG,
6154 const TargetLowering &TLI) {
6155 EVT VT = V.getValueType();
6156 ShuffleVectorSDNode *SVOp = dyn_cast<ShuffleVectorSDNode>(V);
6158 // Be sure that the vector shuffle is present in a pattern like this:
6159 // (vextract (v4f32 shuffle (load $addr), <1,u,u,u>), c) -> (f32 load $addr)
6163 SDNode *N = *V.getNode()->use_begin();
6164 if (N->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
6167 SDValue EltNo = N->getOperand(1);
6168 if (!isa<ConstantSDNode>(EltNo))
6171 // If the bit convert changed the number of elements, it is unsafe
6172 // to examine the mask.
6173 bool HasShuffleIntoBitcast = false;
6174 if (V.getOpcode() == ISD::BITCAST) {
6175 EVT SrcVT = V.getOperand(0).getValueType();
6176 if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
6178 V = V.getOperand(0);
6179 HasShuffleIntoBitcast = true;
6182 // Select the input vector, guarding against out of range extract vector.
6183 unsigned NumElems = VT.getVectorNumElements();
6184 unsigned Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
6185 int Idx = (Elt > NumElems) ? -1 : SVOp->getMaskElt(Elt);
6186 V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
6188 // Skip one more bit_convert if necessary
6189 if (V.getOpcode() == ISD::BITCAST)
6190 V = V.getOperand(0);
6192 if (ISD::isNormalLoad(V.getNode())) {
6193 // Is the original load suitable?
6194 LoadSDNode *LN0 = cast<LoadSDNode>(V);
6196 // FIXME: avoid the multi-use bug that is preventing lots of
6197 // of foldings to be detected, this is still wrong of course, but
6198 // give the temporary desired behavior, and if it happens that
6199 // the load has real more uses, during isel it will not fold, and
6200 // will generate poor code.
6201 if (!LN0 || LN0->isVolatile()) // || !LN0->hasOneUse()
6204 if (!HasShuffleIntoBitcast)
6207 // If there's a bitcast before the shuffle, check if the load type and
6208 // alignment is valid.
6209 unsigned Align = LN0->getAlignment();
6211 TLI.getTargetData()->getABITypeAlignment(
6212 VT.getTypeForEVT(*DAG.getContext()));
6214 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, VT))
6222 SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
6223 EVT VT = Op.getValueType();
6225 // Canonizalize to v2f64.
6226 V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
6227 return DAG.getNode(ISD::BITCAST, dl, VT,
6228 getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
6233 SDValue getMOVLowToHigh(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG,
6235 SDValue V1 = Op.getOperand(0);
6236 SDValue V2 = Op.getOperand(1);
6237 EVT VT = Op.getValueType();
6239 assert(VT != MVT::v2i64 && "unsupported shuffle type");
6241 if (HasSSE2 && VT == MVT::v2f64)
6242 return getTargetShuffleNode(X86ISD::MOVLHPD, dl, VT, V1, V2, DAG);
6244 // v4f32 or v4i32: canonizalized to v4f32 (which is legal for SSE1)
6245 return DAG.getNode(ISD::BITCAST, dl, VT,
6246 getTargetShuffleNode(X86ISD::MOVLHPS, dl, MVT::v4f32,
6247 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V1),
6248 DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, V2), DAG));
6252 SDValue getMOVHighToLow(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG) {
6253 SDValue V1 = Op.getOperand(0);
6254 SDValue V2 = Op.getOperand(1);
6255 EVT VT = Op.getValueType();
6257 assert((VT == MVT::v4i32 || VT == MVT::v4f32) &&
6258 "unsupported shuffle type");
6260 if (V2.getOpcode() == ISD::UNDEF)
6264 return getTargetShuffleNode(X86ISD::MOVHLPS, dl, VT, V1, V2, DAG);
6267 static inline unsigned getSHUFPOpcode(EVT VT) {
6268 switch(VT.getSimpleVT().SimpleTy) {
6269 case MVT::v8i32: // Use fp unit for int unpack.
6271 case MVT::v4i32: // Use fp unit for int unpack.
6272 case MVT::v4f32: return X86ISD::SHUFPS;
6273 case MVT::v4i64: // Use fp unit for int unpack.
6275 case MVT::v2i64: // Use fp unit for int unpack.
6276 case MVT::v2f64: return X86ISD::SHUFPD;
6278 llvm_unreachable("Unknown type for shufp*");
6284 SDValue getMOVLP(SDValue &Op, DebugLoc &dl, SelectionDAG &DAG, bool HasSSE2) {
6285 SDValue V1 = Op.getOperand(0);
6286 SDValue V2 = Op.getOperand(1);
6287 EVT VT = Op.getValueType();
6288 unsigned NumElems = VT.getVectorNumElements();
6290 // Use MOVLPS and MOVLPD in case V1 or V2 are loads. During isel, the second
6291 // operand of these instructions is only memory, so check if there's a
6292 // potencial load folding here, otherwise use SHUFPS or MOVSD to match the
6294 bool CanFoldLoad = false;
6296 // Trivial case, when V2 comes from a load.
6297 if (MayFoldVectorLoad(V2))
6300 // When V1 is a load, it can be folded later into a store in isel, example:
6301 // (store (v4f32 (X86Movlps (load addr:$src1), VR128:$src2)), addr:$src1)
6303 // (MOVLPSmr addr:$src1, VR128:$src2)
6304 // So, recognize this potential and also use MOVLPS or MOVLPD
6305 if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
6308 // Both of them can't be memory operations though.
6309 if (MayFoldVectorLoad(V1) && MayFoldVectorLoad(V2))
6310 CanFoldLoad = false;
6313 if (HasSSE2 && NumElems == 2)
6314 return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
6317 return getTargetShuffleNode(X86ISD::MOVLPS, dl, VT, V1, V2, DAG);
6320 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6321 // movl and movlp will both match v2i64, but v2i64 is never matched by
6322 // movl earlier because we make it strict to avoid messing with the movlp load
6323 // folding logic (see the code above getMOVLP call). Match it here then,
6324 // this is horrible, but will stay like this until we move all shuffle
6325 // matching to x86 specific nodes. Note that for the 1st condition all
6326 // types are matched with movsd.
6329 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
6330 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
6333 assert(VT != MVT::v4i32 && "unsupported shuffle type");
6335 // Invert the operand order and use SHUFPS to match it.
6336 return getTargetShuffleNode(getSHUFPOpcode(VT), dl, VT, V2, V1,
6337 X86::getShuffleSHUFImmediate(SVOp), DAG);
6340 static inline unsigned getUNPCKLOpcode(EVT VT) {
6341 switch(VT.getSimpleVT().SimpleTy) {
6342 case MVT::v4i32: return X86ISD::PUNPCKLDQ;
6343 case MVT::v2i64: return X86ISD::PUNPCKLQDQ;
6344 case MVT::v4f32: return X86ISD::UNPCKLPS;
6345 case MVT::v2f64: return X86ISD::UNPCKLPD;
6346 case MVT::v8i32: // Use fp unit for int unpack.
6347 case MVT::v8f32: return X86ISD::VUNPCKLPSY;
6348 case MVT::v4i64: // Use fp unit for int unpack.
6349 case MVT::v4f64: return X86ISD::VUNPCKLPDY;
6350 case MVT::v16i8: return X86ISD::PUNPCKLBW;
6351 case MVT::v8i16: return X86ISD::PUNPCKLWD;
6353 llvm_unreachable("Unknown type for unpckl");
6358 static inline unsigned getUNPCKHOpcode(EVT VT) {
6359 switch(VT.getSimpleVT().SimpleTy) {
6360 case MVT::v4i32: return X86ISD::PUNPCKHDQ;
6361 case MVT::v2i64: return X86ISD::PUNPCKHQDQ;
6362 case MVT::v4f32: return X86ISD::UNPCKHPS;
6363 case MVT::v2f64: return X86ISD::UNPCKHPD;
6364 case MVT::v8i32: // Use fp unit for int unpack.
6365 case MVT::v8f32: return X86ISD::VUNPCKHPSY;
6366 case MVT::v4i64: // Use fp unit for int unpack.
6367 case MVT::v4f64: return X86ISD::VUNPCKHPDY;
6368 case MVT::v16i8: return X86ISD::PUNPCKHBW;
6369 case MVT::v8i16: return X86ISD::PUNPCKHWD;
6371 llvm_unreachable("Unknown type for unpckh");
6376 static inline unsigned getVPERMILOpcode(EVT VT) {
6377 switch(VT.getSimpleVT().SimpleTy) {
6379 case MVT::v4f32: return X86ISD::VPERMILPS;
6381 case MVT::v2f64: return X86ISD::VPERMILPD;
6383 case MVT::v8f32: return X86ISD::VPERMILPSY;
6385 case MVT::v4f64: return X86ISD::VPERMILPDY;
6387 llvm_unreachable("Unknown type for vpermil");
6392 /// isVectorBroadcast - Check if the node chain is suitable to be xformed to
6393 /// a vbroadcast node. The nodes are suitable whenever we can fold a load coming
6394 /// from a 32 or 64 bit scalar. Update Op to the desired load to be folded.
6395 static bool isVectorBroadcast(SDValue &Op) {
6396 EVT VT = Op.getValueType();
6397 bool Is256 = VT.getSizeInBits() == 256;
6399 assert((VT.getSizeInBits() == 128 || Is256) &&
6400 "Unsupported type for vbroadcast node");
6403 if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
6404 V = V.getOperand(0);
6406 if (Is256 && !(V.hasOneUse() &&
6407 V.getOpcode() == ISD::INSERT_SUBVECTOR &&
6408 V.getOperand(0).getOpcode() == ISD::UNDEF))
6412 V = V.getOperand(1);
6417 // Check the source scalar_to_vector type. 256-bit broadcasts are
6418 // supported for 32/64-bit sizes, while 128-bit ones are only supported
6419 // for 32-bit scalars.
6420 if (V.getOpcode() != ISD::SCALAR_TO_VECTOR)
6423 unsigned ScalarSize = V.getOperand(0).getValueType().getSizeInBits();
6424 if (ScalarSize != 32 && ScalarSize != 64)
6426 if (!Is256 && ScalarSize == 64)
6429 V = V.getOperand(0);
6430 if (!MayFoldLoad(V))
6433 // Return the load node
6439 SDValue NormalizeVectorShuffle(SDValue Op, SelectionDAG &DAG,
6440 const TargetLowering &TLI,
6441 const X86Subtarget *Subtarget) {
6442 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6443 EVT VT = Op.getValueType();
6444 DebugLoc dl = Op.getDebugLoc();
6445 SDValue V1 = Op.getOperand(0);
6446 SDValue V2 = Op.getOperand(1);
6448 if (isZeroShuffle(SVOp))
6449 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
6451 // Handle splat operations
6452 if (SVOp->isSplat()) {
6453 unsigned NumElem = VT.getVectorNumElements();
6454 int Size = VT.getSizeInBits();
6455 // Special case, this is the only place now where it's allowed to return
6456 // a vector_shuffle operation without using a target specific node, because
6457 // *hopefully* it will be optimized away by the dag combiner. FIXME: should
6458 // this be moved to DAGCombine instead?
6459 if (NumElem <= 4 && CanXFormVExtractWithShuffleIntoLoad(Op, DAG, TLI))
6462 // Use vbroadcast whenever the splat comes from a foldable load
6463 if (Subtarget->hasAVX() && isVectorBroadcast(V1))
6464 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, V1);
6466 // Handle splats by matching through known shuffle masks
6467 if ((Size == 128 && NumElem <= 4) ||
6468 (Size == 256 && NumElem < 8))
6471 // All remaning splats are promoted to target supported vector shuffles.
6472 return PromoteSplat(SVOp, DAG);
6475 // If the shuffle can be profitably rewritten as a narrower shuffle, then
6477 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
6478 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6479 if (NewOp.getNode())
6480 return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
6481 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
6482 // FIXME: Figure out a cleaner way to do this.
6483 // Try to make use of movq to zero out the top part.
6484 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
6485 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6486 if (NewOp.getNode()) {
6487 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
6488 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
6489 DAG, Subtarget, dl);
6491 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
6492 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
6493 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
6494 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
6495 DAG, Subtarget, dl);
6502 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const {
6503 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
6504 SDValue V1 = Op.getOperand(0);
6505 SDValue V2 = Op.getOperand(1);
6506 EVT VT = Op.getValueType();
6507 DebugLoc dl = Op.getDebugLoc();
6508 unsigned NumElems = VT.getVectorNumElements();
6509 bool isMMX = VT.getSizeInBits() == 64;
6510 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
6511 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
6512 bool V1IsSplat = false;
6513 bool V2IsSplat = false;
6514 bool HasSSE2 = Subtarget->hasSSE2() || Subtarget->hasAVX();
6515 bool HasSSE3 = Subtarget->hasSSE3() || Subtarget->hasAVX();
6516 bool HasSSSE3 = Subtarget->hasSSSE3() || Subtarget->hasAVX();
6517 MachineFunction &MF = DAG.getMachineFunction();
6518 bool OptForSize = MF.getFunction()->hasFnAttr(Attribute::OptimizeForSize);
6520 // Shuffle operations on MMX not supported.
6524 // Vector shuffle lowering takes 3 steps:
6526 // 1) Normalize the input vectors. Here splats, zeroed vectors, profitable
6527 // narrowing and commutation of operands should be handled.
6528 // 2) Matching of shuffles with known shuffle masks to x86 target specific
6530 // 3) Rewriting of unmatched masks into new generic shuffle operations,
6531 // so the shuffle can be broken into other shuffles and the legalizer can
6532 // try the lowering again.
6534 // The general ideia is that no vector_shuffle operation should be left to
6535 // be matched during isel, all of them must be converted to a target specific
6538 // Normalize the input vectors. Here splats, zeroed vectors, profitable
6539 // narrowing and commutation of operands should be handled. The actual code
6540 // doesn't include all of those, work in progress...
6541 SDValue NewOp = NormalizeVectorShuffle(Op, DAG, *this, Subtarget);
6542 if (NewOp.getNode())
6545 // NOTE: isPSHUFDMask can also match both masks below (unpckl_undef and
6546 // unpckh_undef). Only use pshufd if speed is more important than size.
6547 if (OptForSize && X86::isUNPCKL_v_undef_Mask(SVOp))
6548 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
6549 if (OptForSize && X86::isUNPCKH_v_undef_Mask(SVOp))
6550 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6552 if (X86::isMOVDDUPMask(SVOp) && HasSSE3 && V2IsUndef &&
6553 RelaxedMayFoldVectorLoad(V1))
6554 return getMOVDDup(Op, dl, V1, DAG);
6556 if (X86::isMOVHLPS_v_undef_Mask(SVOp))
6557 return getMOVHighToLow(Op, dl, DAG);
6559 // Use to match splats
6560 if (HasSSE2 && X86::isUNPCKHMask(SVOp) && V2IsUndef &&
6561 (VT == MVT::v2f64 || VT == MVT::v2i64))
6562 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6564 if (X86::isPSHUFDMask(SVOp)) {
6565 // The actual implementation will match the mask in the if above and then
6566 // during isel it can match several different instructions, not only pshufd
6567 // as its name says, sad but true, emulate the behavior for now...
6568 if (X86::isMOVDDUPMask(SVOp) && ((VT == MVT::v4f32 || VT == MVT::v2i64)))
6569 return getTargetShuffleNode(X86ISD::MOVLHPS, dl, VT, V1, V1, DAG);
6571 unsigned TargetMask = X86::getShuffleSHUFImmediate(SVOp);
6573 if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
6574 return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
6576 return getTargetShuffleNode(getSHUFPOpcode(VT), dl, VT, V1, V1,
6580 // Check if this can be converted into a logical shift.
6581 bool isLeft = false;
6584 bool isShift = getSubtarget()->hasSSE2() &&
6585 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
6586 if (isShift && ShVal.hasOneUse()) {
6587 // If the shifted value has multiple uses, it may be cheaper to use
6588 // v_set0 + movlhps or movhlps, etc.
6589 EVT EltVT = VT.getVectorElementType();
6590 ShAmt *= EltVT.getSizeInBits();
6591 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6594 if (X86::isMOVLMask(SVOp)) {
6597 if (ISD::isBuildVectorAllZeros(V1.getNode()))
6598 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
6599 if (!X86::isMOVLPMask(SVOp)) {
6600 if (HasSSE2 && (VT == MVT::v2i64 || VT == MVT::v2f64))
6601 return getTargetShuffleNode(X86ISD::MOVSD, dl, VT, V1, V2, DAG);
6603 if (VT == MVT::v4i32 || VT == MVT::v4f32)
6604 return getTargetShuffleNode(X86ISD::MOVSS, dl, VT, V1, V2, DAG);
6608 // FIXME: fold these into legal mask.
6609 if (X86::isMOVLHPSMask(SVOp) && !X86::isUNPCKLMask(SVOp))
6610 return getMOVLowToHigh(Op, dl, DAG, HasSSE2);
6612 if (X86::isMOVHLPSMask(SVOp))
6613 return getMOVHighToLow(Op, dl, DAG);
6615 if (X86::isMOVSHDUPMask(SVOp, Subtarget))
6616 return getTargetShuffleNode(X86ISD::MOVSHDUP, dl, VT, V1, DAG);
6618 if (X86::isMOVSLDUPMask(SVOp, Subtarget))
6619 return getTargetShuffleNode(X86ISD::MOVSLDUP, dl, VT, V1, DAG);
6621 if (X86::isMOVLPMask(SVOp))
6622 return getMOVLP(Op, dl, DAG, HasSSE2);
6624 if (ShouldXformToMOVHLPS(SVOp) ||
6625 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
6626 return CommuteVectorShuffle(SVOp, DAG);
6629 // No better options. Use a vshl / vsrl.
6630 EVT EltVT = VT.getVectorElementType();
6631 ShAmt *= EltVT.getSizeInBits();
6632 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
6635 bool Commuted = false;
6636 // FIXME: This should also accept a bitcast of a splat? Be careful, not
6637 // 1,1,1,1 -> v8i16 though.
6638 V1IsSplat = isSplatVector(V1.getNode());
6639 V2IsSplat = isSplatVector(V2.getNode());
6641 // Canonicalize the splat or undef, if present, to be on the RHS.
6642 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
6643 Op = CommuteVectorShuffle(SVOp, DAG);
6644 SVOp = cast<ShuffleVectorSDNode>(Op);
6645 V1 = SVOp->getOperand(0);
6646 V2 = SVOp->getOperand(1);
6647 std::swap(V1IsSplat, V2IsSplat);
6648 std::swap(V1IsUndef, V2IsUndef);
6652 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
6653 // Shuffling low element of v1 into undef, just return v1.
6656 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
6657 // the instruction selector will not match, so get a canonical MOVL with
6658 // swapped operands to undo the commute.
6659 return getMOVL(DAG, dl, VT, V2, V1);
6662 if (X86::isUNPCKLMask(SVOp))
6663 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V2, DAG);
6665 if (X86::isUNPCKHMask(SVOp))
6666 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V2, DAG);
6669 // Normalize mask so all entries that point to V2 points to its first
6670 // element then try to match unpck{h|l} again. If match, return a
6671 // new vector_shuffle with the corrected mask.
6672 SDValue NewMask = NormalizeMask(SVOp, DAG);
6673 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
6674 if (NSVOp != SVOp) {
6675 if (X86::isUNPCKLMask(NSVOp, true)) {
6677 } else if (X86::isUNPCKHMask(NSVOp, true)) {
6684 // Commute is back and try unpck* again.
6685 // FIXME: this seems wrong.
6686 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
6687 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
6689 if (X86::isUNPCKLMask(NewSVOp))
6690 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V2, V1, DAG);
6692 if (X86::isUNPCKHMask(NewSVOp))
6693 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V2, V1, DAG);
6696 // Normalize the node to match x86 shuffle ops if needed
6697 if (V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
6698 return CommuteVectorShuffle(SVOp, DAG);
6700 // The checks below are all present in isShuffleMaskLegal, but they are
6701 // inlined here right now to enable us to directly emit target specific
6702 // nodes, and remove one by one until they don't return Op anymore.
6703 SmallVector<int, 16> M;
6706 if (isPALIGNRMask(M, VT, HasSSSE3))
6707 return getTargetShuffleNode(X86ISD::PALIGN, dl, VT, V1, V2,
6708 X86::getShufflePALIGNRImmediate(SVOp),
6711 if (ShuffleVectorSDNode::isSplatMask(&M[0], VT) &&
6712 SVOp->getSplatIndex() == 0 && V2IsUndef) {
6713 if (VT == MVT::v2f64)
6714 return getTargetShuffleNode(X86ISD::UNPCKLPD, dl, VT, V1, V1, DAG);
6715 if (VT == MVT::v2i64)
6716 return getTargetShuffleNode(X86ISD::PUNPCKLQDQ, dl, VT, V1, V1, DAG);
6719 if (isPSHUFHWMask(M, VT))
6720 return getTargetShuffleNode(X86ISD::PSHUFHW, dl, VT, V1,
6721 X86::getShufflePSHUFHWImmediate(SVOp),
6724 if (isPSHUFLWMask(M, VT))
6725 return getTargetShuffleNode(X86ISD::PSHUFLW, dl, VT, V1,
6726 X86::getShufflePSHUFLWImmediate(SVOp),
6729 if (isSHUFPMask(M, VT))
6730 return getTargetShuffleNode(getSHUFPOpcode(VT), dl, VT, V1, V2,
6731 X86::getShuffleSHUFImmediate(SVOp), DAG);
6733 if (X86::isUNPCKL_v_undef_Mask(SVOp))
6734 return getTargetShuffleNode(getUNPCKLOpcode(VT), dl, VT, V1, V1, DAG);
6735 if (X86::isUNPCKH_v_undef_Mask(SVOp))
6736 return getTargetShuffleNode(getUNPCKHOpcode(VT), dl, VT, V1, V1, DAG);
6738 //===--------------------------------------------------------------------===//
6739 // Generate target specific nodes for 128 or 256-bit shuffles only
6740 // supported in the AVX instruction set.
6743 // Handle VMOVDDUPY permutations
6744 if (isMOVDDUPYMask(SVOp, Subtarget))
6745 return getTargetShuffleNode(X86ISD::MOVDDUP, dl, VT, V1, DAG);
6747 // Handle VPERMILPS* permutations
6748 if (isVPERMILPSMask(M, VT, Subtarget))
6749 return getTargetShuffleNode(getVPERMILOpcode(VT), dl, VT, V1,
6750 getShuffleVPERMILPSImmediate(SVOp), DAG);
6752 // Handle VPERMILPD* permutations
6753 if (isVPERMILPDMask(M, VT, Subtarget))
6754 return getTargetShuffleNode(getVPERMILOpcode(VT), dl, VT, V1,
6755 getShuffleVPERMILPDImmediate(SVOp), DAG);
6757 // Handle VPERM2F128 permutations
6758 if (isVPERM2F128Mask(M, VT, Subtarget))
6759 return getTargetShuffleNode(X86ISD::VPERM2F128, dl, VT, V1, V2,
6760 getShuffleVPERM2F128Immediate(SVOp), DAG);
6762 // Handle VSHUFPSY permutations
6763 if (isVSHUFPSYMask(M, VT, Subtarget))
6764 return getTargetShuffleNode(getSHUFPOpcode(VT), dl, VT, V1, V2,
6765 getShuffleVSHUFPSYImmediate(SVOp), DAG);
6767 // Handle VSHUFPDY permutations
6768 if (isVSHUFPDYMask(M, VT, Subtarget))
6769 return getTargetShuffleNode(getSHUFPOpcode(VT), dl, VT, V1, V2,
6770 getShuffleVSHUFPDYImmediate(SVOp), DAG);
6772 //===--------------------------------------------------------------------===//
6773 // Since no target specific shuffle was selected for this generic one,
6774 // lower it into other known shuffles. FIXME: this isn't true yet, but
6775 // this is the plan.
6778 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
6779 if (VT == MVT::v8i16) {
6780 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(Op, DAG);
6781 if (NewOp.getNode())
6785 if (VT == MVT::v16i8) {
6786 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
6787 if (NewOp.getNode())
6791 // Handle all 128-bit wide vectors with 4 elements, and match them with
6792 // several different shuffle types.
6793 if (NumElems == 4 && VT.getSizeInBits() == 128)
6794 return LowerVECTOR_SHUFFLE_128v4(SVOp, DAG);
6796 // Handle general 256-bit shuffles
6797 if (VT.is256BitVector())
6798 return LowerVECTOR_SHUFFLE_256(SVOp, DAG);
6804 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
6805 SelectionDAG &DAG) const {
6806 EVT VT = Op.getValueType();
6807 DebugLoc dl = Op.getDebugLoc();
6809 if (Op.getOperand(0).getValueType().getSizeInBits() != 128)
6812 if (VT.getSizeInBits() == 8) {
6813 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
6814 Op.getOperand(0), Op.getOperand(1));
6815 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6816 DAG.getValueType(VT));
6817 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6818 } else if (VT.getSizeInBits() == 16) {
6819 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6820 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
6822 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6823 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6824 DAG.getNode(ISD::BITCAST, dl,
6828 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
6829 Op.getOperand(0), Op.getOperand(1));
6830 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
6831 DAG.getValueType(VT));
6832 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6833 } else if (VT == MVT::f32) {
6834 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
6835 // the result back to FR32 register. It's only worth matching if the
6836 // result has a single use which is a store or a bitcast to i32. And in
6837 // the case of a store, it's not worth it if the index is a constant 0,
6838 // because a MOVSSmr can be used instead, which is smaller and faster.
6839 if (!Op.hasOneUse())
6841 SDNode *User = *Op.getNode()->use_begin();
6842 if ((User->getOpcode() != ISD::STORE ||
6843 (isa<ConstantSDNode>(Op.getOperand(1)) &&
6844 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
6845 (User->getOpcode() != ISD::BITCAST ||
6846 User->getValueType(0) != MVT::i32))
6848 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6849 DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
6852 return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
6853 } else if (VT == MVT::i32) {
6854 // ExtractPS works with constant index.
6855 if (isa<ConstantSDNode>(Op.getOperand(1)))
6863 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
6864 SelectionDAG &DAG) const {
6865 if (!isa<ConstantSDNode>(Op.getOperand(1)))
6868 SDValue Vec = Op.getOperand(0);
6869 EVT VecVT = Vec.getValueType();
6871 // If this is a 256-bit vector result, first extract the 128-bit vector and
6872 // then extract the element from the 128-bit vector.
6873 if (VecVT.getSizeInBits() == 256) {
6874 DebugLoc dl = Op.getNode()->getDebugLoc();
6875 unsigned NumElems = VecVT.getVectorNumElements();
6876 SDValue Idx = Op.getOperand(1);
6877 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
6879 // Get the 128-bit vector.
6880 bool Upper = IdxVal >= NumElems/2;
6881 Vec = Extract128BitVector(Vec,
6882 DAG.getConstant(Upper ? NumElems/2 : 0, MVT::i32), DAG, dl);
6884 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
6885 Upper ? DAG.getConstant(IdxVal-NumElems/2, MVT::i32) : Idx);
6888 assert(Vec.getValueSizeInBits() <= 128 && "Unexpected vector length");
6890 if (Subtarget->hasSSE41() || Subtarget->hasAVX()) {
6891 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
6896 EVT VT = Op.getValueType();
6897 DebugLoc dl = Op.getDebugLoc();
6898 // TODO: handle v16i8.
6899 if (VT.getSizeInBits() == 16) {
6900 SDValue Vec = Op.getOperand(0);
6901 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6903 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
6904 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
6905 DAG.getNode(ISD::BITCAST, dl,
6908 // Transform it so it match pextrw which produces a 32-bit result.
6909 EVT EltVT = MVT::i32;
6910 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
6911 Op.getOperand(0), Op.getOperand(1));
6912 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
6913 DAG.getValueType(VT));
6914 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
6915 } else if (VT.getSizeInBits() == 32) {
6916 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6920 // SHUFPS the element to the lowest double word, then movss.
6921 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
6922 EVT VVT = Op.getOperand(0).getValueType();
6923 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6924 DAG.getUNDEF(VVT), Mask);
6925 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6926 DAG.getIntPtrConstant(0));
6927 } else if (VT.getSizeInBits() == 64) {
6928 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
6929 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
6930 // to match extract_elt for f64.
6931 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
6935 // UNPCKHPD the element to the lowest double word, then movsd.
6936 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
6937 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
6938 int Mask[2] = { 1, -1 };
6939 EVT VVT = Op.getOperand(0).getValueType();
6940 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
6941 DAG.getUNDEF(VVT), Mask);
6942 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
6943 DAG.getIntPtrConstant(0));
6950 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op,
6951 SelectionDAG &DAG) const {
6952 EVT VT = Op.getValueType();
6953 EVT EltVT = VT.getVectorElementType();
6954 DebugLoc dl = Op.getDebugLoc();
6956 SDValue N0 = Op.getOperand(0);
6957 SDValue N1 = Op.getOperand(1);
6958 SDValue N2 = Op.getOperand(2);
6960 if (VT.getSizeInBits() == 256)
6963 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
6964 isa<ConstantSDNode>(N2)) {
6966 if (VT == MVT::v8i16)
6967 Opc = X86ISD::PINSRW;
6968 else if (VT == MVT::v16i8)
6969 Opc = X86ISD::PINSRB;
6971 Opc = X86ISD::PINSRB;
6973 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
6975 if (N1.getValueType() != MVT::i32)
6976 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
6977 if (N2.getValueType() != MVT::i32)
6978 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
6979 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
6980 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
6981 // Bits [7:6] of the constant are the source select. This will always be
6982 // zero here. The DAG Combiner may combine an extract_elt index into these
6983 // bits. For example (insert (extract, 3), 2) could be matched by putting
6984 // the '3' into bits [7:6] of X86ISD::INSERTPS.
6985 // Bits [5:4] of the constant are the destination select. This is the
6986 // value of the incoming immediate.
6987 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
6988 // combine either bitwise AND or insert of float 0.0 to set these bits.
6989 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
6990 // Create this as a scalar to vector..
6991 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
6992 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
6993 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
6994 // PINSR* works with constant index.
7001 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
7002 EVT VT = Op.getValueType();
7003 EVT EltVT = VT.getVectorElementType();
7005 DebugLoc dl = Op.getDebugLoc();
7006 SDValue N0 = Op.getOperand(0);
7007 SDValue N1 = Op.getOperand(1);
7008 SDValue N2 = Op.getOperand(2);
7010 // If this is a 256-bit vector result, first extract the 128-bit vector,
7011 // insert the element into the extracted half and then place it back.
7012 if (VT.getSizeInBits() == 256) {
7013 if (!isa<ConstantSDNode>(N2))
7016 // Get the desired 128-bit vector half.
7017 unsigned NumElems = VT.getVectorNumElements();
7018 unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
7019 bool Upper = IdxVal >= NumElems/2;
7020 SDValue Ins128Idx = DAG.getConstant(Upper ? NumElems/2 : 0, MVT::i32);
7021 SDValue V = Extract128BitVector(N0, Ins128Idx, DAG, dl);
7023 // Insert the element into the desired half.
7024 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V,
7025 N1, Upper ? DAG.getConstant(IdxVal-NumElems/2, MVT::i32) : N2);
7027 // Insert the changed part back to the 256-bit vector
7028 return Insert128BitVector(N0, V, Ins128Idx, DAG, dl);
7031 if (Subtarget->hasSSE41() || Subtarget->hasAVX())
7032 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
7034 if (EltVT == MVT::i8)
7037 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
7038 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
7039 // as its second argument.
7040 if (N1.getValueType() != MVT::i32)
7041 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
7042 if (N2.getValueType() != MVT::i32)
7043 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
7044 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
7050 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
7051 LLVMContext *Context = DAG.getContext();
7052 DebugLoc dl = Op.getDebugLoc();
7053 EVT OpVT = Op.getValueType();
7055 // If this is a 256-bit vector result, first insert into a 128-bit
7056 // vector and then insert into the 256-bit vector.
7057 if (OpVT.getSizeInBits() > 128) {
7058 // Insert into a 128-bit vector.
7059 EVT VT128 = EVT::getVectorVT(*Context,
7060 OpVT.getVectorElementType(),
7061 OpVT.getVectorNumElements() / 2);
7063 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
7065 // Insert the 128-bit vector.
7066 return Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, OpVT), Op,
7067 DAG.getConstant(0, MVT::i32),
7071 if (Op.getValueType() == MVT::v1i64 &&
7072 Op.getOperand(0).getValueType() == MVT::i64)
7073 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
7075 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
7076 assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
7077 "Expected an SSE type!");
7078 return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
7079 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
7082 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
7083 // a simple subregister reference or explicit instructions to grab
7084 // upper bits of a vector.
7086 X86TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
7087 if (Subtarget->hasAVX()) {
7088 DebugLoc dl = Op.getNode()->getDebugLoc();
7089 SDValue Vec = Op.getNode()->getOperand(0);
7090 SDValue Idx = Op.getNode()->getOperand(1);
7092 if (Op.getNode()->getValueType(0).getSizeInBits() == 128
7093 && Vec.getNode()->getValueType(0).getSizeInBits() == 256) {
7094 return Extract128BitVector(Vec, Idx, DAG, dl);
7100 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
7101 // simple superregister reference or explicit instructions to insert
7102 // the upper bits of a vector.
7104 X86TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
7105 if (Subtarget->hasAVX()) {
7106 DebugLoc dl = Op.getNode()->getDebugLoc();
7107 SDValue Vec = Op.getNode()->getOperand(0);
7108 SDValue SubVec = Op.getNode()->getOperand(1);
7109 SDValue Idx = Op.getNode()->getOperand(2);
7111 if (Op.getNode()->getValueType(0).getSizeInBits() == 256
7112 && SubVec.getNode()->getValueType(0).getSizeInBits() == 128) {
7113 return Insert128BitVector(Vec, SubVec, Idx, DAG, dl);
7119 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
7120 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
7121 // one of the above mentioned nodes. It has to be wrapped because otherwise
7122 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
7123 // be used to form addressing mode. These wrapped nodes will be selected
7126 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
7127 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
7129 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7131 unsigned char OpFlag = 0;
7132 unsigned WrapperKind = X86ISD::Wrapper;
7133 CodeModel::Model M = getTargetMachine().getCodeModel();
7135 if (Subtarget->isPICStyleRIPRel() &&
7136 (M == CodeModel::Small || M == CodeModel::Kernel))
7137 WrapperKind = X86ISD::WrapperRIP;
7138 else if (Subtarget->isPICStyleGOT())
7139 OpFlag = X86II::MO_GOTOFF;
7140 else if (Subtarget->isPICStyleStubPIC())
7141 OpFlag = X86II::MO_PIC_BASE_OFFSET;
7143 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
7145 CP->getOffset(), OpFlag);
7146 DebugLoc DL = CP->getDebugLoc();
7147 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7148 // With PIC, the address is actually $g + Offset.
7150 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7151 DAG.getNode(X86ISD::GlobalBaseReg,
7152 DebugLoc(), getPointerTy()),
7159 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
7160 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
7162 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7164 unsigned char OpFlag = 0;
7165 unsigned WrapperKind = X86ISD::Wrapper;
7166 CodeModel::Model M = getTargetMachine().getCodeModel();
7168 if (Subtarget->isPICStyleRIPRel() &&
7169 (M == CodeModel::Small || M == CodeModel::Kernel))
7170 WrapperKind = X86ISD::WrapperRIP;
7171 else if (Subtarget->isPICStyleGOT())
7172 OpFlag = X86II::MO_GOTOFF;
7173 else if (Subtarget->isPICStyleStubPIC())
7174 OpFlag = X86II::MO_PIC_BASE_OFFSET;
7176 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
7178 DebugLoc DL = JT->getDebugLoc();
7179 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7181 // With PIC, the address is actually $g + Offset.
7183 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7184 DAG.getNode(X86ISD::GlobalBaseReg,
7185 DebugLoc(), getPointerTy()),
7192 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
7193 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
7195 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7197 unsigned char OpFlag = 0;
7198 unsigned WrapperKind = X86ISD::Wrapper;
7199 CodeModel::Model M = getTargetMachine().getCodeModel();
7201 if (Subtarget->isPICStyleRIPRel() &&
7202 (M == CodeModel::Small || M == CodeModel::Kernel)) {
7203 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
7204 OpFlag = X86II::MO_GOTPCREL;
7205 WrapperKind = X86ISD::WrapperRIP;
7206 } else if (Subtarget->isPICStyleGOT()) {
7207 OpFlag = X86II::MO_GOT;
7208 } else if (Subtarget->isPICStyleStubPIC()) {
7209 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
7210 } else if (Subtarget->isPICStyleStubNoDynamic()) {
7211 OpFlag = X86II::MO_DARWIN_NONLAZY;
7214 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
7216 DebugLoc DL = Op.getDebugLoc();
7217 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7220 // With PIC, the address is actually $g + Offset.
7221 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
7222 !Subtarget->is64Bit()) {
7223 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7224 DAG.getNode(X86ISD::GlobalBaseReg,
7225 DebugLoc(), getPointerTy()),
7229 // For symbols that require a load from a stub to get the address, emit the
7231 if (isGlobalStubReference(OpFlag))
7232 Result = DAG.getLoad(getPointerTy(), DL, DAG.getEntryNode(), Result,
7233 MachinePointerInfo::getGOT(), false, false, 0);
7239 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
7240 // Create the TargetBlockAddressAddress node.
7241 unsigned char OpFlags =
7242 Subtarget->ClassifyBlockAddressReference();
7243 CodeModel::Model M = getTargetMachine().getCodeModel();
7244 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
7245 DebugLoc dl = Op.getDebugLoc();
7246 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
7247 /*isTarget=*/true, OpFlags);
7249 if (Subtarget->isPICStyleRIPRel() &&
7250 (M == CodeModel::Small || M == CodeModel::Kernel))
7251 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
7253 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
7255 // With PIC, the address is actually $g + Offset.
7256 if (isGlobalRelativeToPICBase(OpFlags)) {
7257 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7258 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
7266 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
7268 SelectionDAG &DAG) const {
7269 // Create the TargetGlobalAddress node, folding in the constant
7270 // offset if it is legal.
7271 unsigned char OpFlags =
7272 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
7273 CodeModel::Model M = getTargetMachine().getCodeModel();
7275 if (OpFlags == X86II::MO_NO_FLAG &&
7276 X86::isOffsetSuitableForCodeModel(Offset, M)) {
7277 // A direct static reference to a global.
7278 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), Offset);
7281 Result = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
7284 if (Subtarget->isPICStyleRIPRel() &&
7285 (M == CodeModel::Small || M == CodeModel::Kernel))
7286 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
7288 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
7290 // With PIC, the address is actually $g + Offset.
7291 if (isGlobalRelativeToPICBase(OpFlags)) {
7292 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
7293 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
7297 // For globals that require a load from a stub to get the address, emit the
7299 if (isGlobalStubReference(OpFlags))
7300 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
7301 MachinePointerInfo::getGOT(), false, false, 0);
7303 // If there was a non-zero offset that we didn't fold, create an explicit
7306 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
7307 DAG.getConstant(Offset, getPointerTy()));
7313 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
7314 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
7315 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
7316 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
7320 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
7321 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
7322 unsigned char OperandFlags) {
7323 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7324 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7325 DebugLoc dl = GA->getDebugLoc();
7326 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7327 GA->getValueType(0),
7331 SDValue Ops[] = { Chain, TGA, *InFlag };
7332 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
7334 SDValue Ops[] = { Chain, TGA };
7335 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
7338 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
7339 MFI->setAdjustsStack(true);
7341 SDValue Flag = Chain.getValue(1);
7342 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
7345 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
7347 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7350 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
7351 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
7352 DAG.getNode(X86ISD::GlobalBaseReg,
7353 DebugLoc(), PtrVT), InFlag);
7354 InFlag = Chain.getValue(1);
7356 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
7359 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
7361 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7363 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
7364 X86::RAX, X86II::MO_TLSGD);
7367 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
7368 // "local exec" model.
7369 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
7370 const EVT PtrVT, TLSModel::Model model,
7372 DebugLoc dl = GA->getDebugLoc();
7374 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
7375 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
7376 is64Bit ? 257 : 256));
7378 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
7379 DAG.getIntPtrConstant(0),
7380 MachinePointerInfo(Ptr), false, false, 0);
7382 unsigned char OperandFlags = 0;
7383 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
7385 unsigned WrapperKind = X86ISD::Wrapper;
7386 if (model == TLSModel::LocalExec) {
7387 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
7388 } else if (is64Bit) {
7389 assert(model == TLSModel::InitialExec);
7390 OperandFlags = X86II::MO_GOTTPOFF;
7391 WrapperKind = X86ISD::WrapperRIP;
7393 assert(model == TLSModel::InitialExec);
7394 OperandFlags = X86II::MO_INDNTPOFF;
7397 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
7399 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
7400 GA->getValueType(0),
7401 GA->getOffset(), OperandFlags);
7402 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
7404 if (model == TLSModel::InitialExec)
7405 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
7406 MachinePointerInfo::getGOT(), false, false, 0);
7408 // The address of the thread local variable is the add of the thread
7409 // pointer with the offset of the variable.
7410 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
7414 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
7416 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
7417 const GlobalValue *GV = GA->getGlobal();
7419 if (Subtarget->isTargetELF()) {
7420 // TODO: implement the "local dynamic" model
7421 // TODO: implement the "initial exec"model for pic executables
7423 // If GV is an alias then use the aliasee for determining
7424 // thread-localness.
7425 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
7426 GV = GA->resolveAliasedGlobal(false);
7428 TLSModel::Model model
7429 = getTLSModel(GV, getTargetMachine().getRelocationModel());
7432 case TLSModel::GeneralDynamic:
7433 case TLSModel::LocalDynamic: // not implemented
7434 if (Subtarget->is64Bit())
7435 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
7436 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
7438 case TLSModel::InitialExec:
7439 case TLSModel::LocalExec:
7440 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
7441 Subtarget->is64Bit());
7443 } else if (Subtarget->isTargetDarwin()) {
7444 // Darwin only has one model of TLS. Lower to that.
7445 unsigned char OpFlag = 0;
7446 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
7447 X86ISD::WrapperRIP : X86ISD::Wrapper;
7449 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
7451 bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
7452 !Subtarget->is64Bit();
7454 OpFlag = X86II::MO_TLVP_PIC_BASE;
7456 OpFlag = X86II::MO_TLVP;
7457 DebugLoc DL = Op.getDebugLoc();
7458 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
7459 GA->getValueType(0),
7460 GA->getOffset(), OpFlag);
7461 SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
7463 // With PIC32, the address is actually $g + Offset.
7465 Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
7466 DAG.getNode(X86ISD::GlobalBaseReg,
7467 DebugLoc(), getPointerTy()),
7470 // Lowering the machine isd will make sure everything is in the right
7472 SDValue Chain = DAG.getEntryNode();
7473 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
7474 SDValue Args[] = { Chain, Offset };
7475 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
7477 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
7478 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
7479 MFI->setAdjustsStack(true);
7481 // And our return value (tls address) is in the standard call return value
7483 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
7484 return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
7488 "TLS not implemented for this target.");
7490 llvm_unreachable("Unreachable");
7495 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values and
7496 /// take a 2 x i32 value to shift plus a shift amount.
7497 SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const {
7498 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
7499 EVT VT = Op.getValueType();
7500 unsigned VTBits = VT.getSizeInBits();
7501 DebugLoc dl = Op.getDebugLoc();
7502 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
7503 SDValue ShOpLo = Op.getOperand(0);
7504 SDValue ShOpHi = Op.getOperand(1);
7505 SDValue ShAmt = Op.getOperand(2);
7506 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
7507 DAG.getConstant(VTBits - 1, MVT::i8))
7508 : DAG.getConstant(0, VT);
7511 if (Op.getOpcode() == ISD::SHL_PARTS) {
7512 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
7513 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
7515 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
7516 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
7519 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
7520 DAG.getConstant(VTBits, MVT::i8));
7521 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
7522 AndNode, DAG.getConstant(0, MVT::i8));
7525 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
7526 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
7527 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
7529 if (Op.getOpcode() == ISD::SHL_PARTS) {
7530 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7531 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7533 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
7534 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
7537 SDValue Ops[2] = { Lo, Hi };
7538 return DAG.getMergeValues(Ops, 2, dl);
7541 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
7542 SelectionDAG &DAG) const {
7543 EVT SrcVT = Op.getOperand(0).getValueType();
7545 if (SrcVT.isVector())
7548 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
7549 "Unknown SINT_TO_FP to lower!");
7551 // These are really Legal; return the operand so the caller accepts it as
7553 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
7555 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
7556 Subtarget->is64Bit()) {
7560 DebugLoc dl = Op.getDebugLoc();
7561 unsigned Size = SrcVT.getSizeInBits()/8;
7562 MachineFunction &MF = DAG.getMachineFunction();
7563 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
7564 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7565 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7567 MachinePointerInfo::getFixedStack(SSFI),
7569 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
7572 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
7574 SelectionDAG &DAG) const {
7576 DebugLoc DL = Op.getDebugLoc();
7578 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
7580 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
7582 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
7584 unsigned ByteSize = SrcVT.getSizeInBits()/8;
7586 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
7587 MachineMemOperand *MMO;
7589 int SSFI = FI->getIndex();
7591 DAG.getMachineFunction()
7592 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7593 MachineMemOperand::MOLoad, ByteSize, ByteSize);
7595 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
7596 StackSlot = StackSlot.getOperand(1);
7598 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
7599 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
7601 Tys, Ops, array_lengthof(Ops),
7605 Chain = Result.getValue(1);
7606 SDValue InFlag = Result.getValue(2);
7608 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
7609 // shouldn't be necessary except that RFP cannot be live across
7610 // multiple blocks. When stackifier is fixed, they can be uncoupled.
7611 MachineFunction &MF = DAG.getMachineFunction();
7612 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
7613 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
7614 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7615 Tys = DAG.getVTList(MVT::Other);
7617 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
7619 MachineMemOperand *MMO =
7620 DAG.getMachineFunction()
7621 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7622 MachineMemOperand::MOStore, SSFISize, SSFISize);
7624 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
7625 Ops, array_lengthof(Ops),
7626 Op.getValueType(), MMO);
7627 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
7628 MachinePointerInfo::getFixedStack(SSFI),
7635 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
7636 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
7637 SelectionDAG &DAG) const {
7638 // This algorithm is not obvious. Here it is in C code, more or less:
7640 double uint64_to_double( uint32_t hi, uint32_t lo ) {
7641 static const __m128i exp = { 0x4330000045300000ULL, 0 };
7642 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
7644 // Copy ints to xmm registers.
7645 __m128i xh = _mm_cvtsi32_si128( hi );
7646 __m128i xl = _mm_cvtsi32_si128( lo );
7648 // Combine into low half of a single xmm register.
7649 __m128i x = _mm_unpacklo_epi32( xh, xl );
7653 // Merge in appropriate exponents to give the integer bits the right
7655 x = _mm_unpacklo_epi32( x, exp );
7657 // Subtract away the biases to deal with the IEEE-754 double precision
7659 d = _mm_sub_pd( (__m128d) x, bias );
7661 // All conversions up to here are exact. The correctly rounded result is
7662 // calculated using the current rounding mode using the following
7664 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
7665 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
7666 // store doesn't really need to be here (except
7667 // maybe to zero the other double)
7672 DebugLoc dl = Op.getDebugLoc();
7673 LLVMContext *Context = DAG.getContext();
7675 // Build some magic constants.
7676 std::vector<Constant*> CV0;
7677 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
7678 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
7679 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
7680 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
7681 Constant *C0 = ConstantVector::get(CV0);
7682 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
7684 std::vector<Constant*> CV1;
7686 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
7688 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
7689 Constant *C1 = ConstantVector::get(CV1);
7690 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
7692 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7693 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7695 DAG.getIntPtrConstant(1)));
7696 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7697 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7699 DAG.getIntPtrConstant(0)));
7700 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
7701 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
7702 MachinePointerInfo::getConstantPool(),
7704 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
7705 SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
7706 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
7707 MachinePointerInfo::getConstantPool(),
7709 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
7711 // Add the halves; easiest way is to swap them into another reg first.
7712 int ShufMask[2] = { 1, -1 };
7713 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
7714 DAG.getUNDEF(MVT::v2f64), ShufMask);
7715 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
7716 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
7717 DAG.getIntPtrConstant(0));
7720 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
7721 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
7722 SelectionDAG &DAG) const {
7723 DebugLoc dl = Op.getDebugLoc();
7724 // FP constant to bias correct the final result.
7725 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
7728 // Load the 32-bit value into an XMM register.
7729 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
7732 // Zero out the upper parts of the register.
7733 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget->hasSSE2(), DAG);
7735 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7736 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
7737 DAG.getIntPtrConstant(0));
7739 // Or the load with the bias.
7740 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
7741 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7742 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7744 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
7745 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
7746 MVT::v2f64, Bias)));
7747 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
7748 DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
7749 DAG.getIntPtrConstant(0));
7751 // Subtract the bias.
7752 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
7754 // Handle final rounding.
7755 EVT DestVT = Op.getValueType();
7757 if (DestVT.bitsLT(MVT::f64)) {
7758 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
7759 DAG.getIntPtrConstant(0));
7760 } else if (DestVT.bitsGT(MVT::f64)) {
7761 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
7764 // Handle final rounding.
7768 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
7769 SelectionDAG &DAG) const {
7770 SDValue N0 = Op.getOperand(0);
7771 DebugLoc dl = Op.getDebugLoc();
7773 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
7774 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
7775 // the optimization here.
7776 if (DAG.SignBitIsZero(N0))
7777 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
7779 EVT SrcVT = N0.getValueType();
7780 EVT DstVT = Op.getValueType();
7781 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
7782 return LowerUINT_TO_FP_i64(Op, DAG);
7783 else if (SrcVT == MVT::i32 && X86ScalarSSEf64)
7784 return LowerUINT_TO_FP_i32(Op, DAG);
7786 // Make a 64-bit buffer, and use it to build an FILD.
7787 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
7788 if (SrcVT == MVT::i32) {
7789 SDValue WordOff = DAG.getConstant(4, getPointerTy());
7790 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
7791 getPointerTy(), StackSlot, WordOff);
7792 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7793 StackSlot, MachinePointerInfo(),
7795 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
7796 OffsetSlot, MachinePointerInfo(),
7798 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
7802 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
7803 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
7804 StackSlot, MachinePointerInfo(),
7806 // For i64 source, we need to add the appropriate power of 2 if the input
7807 // was negative. This is the same as the optimization in
7808 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
7809 // we must be careful to do the computation in x87 extended precision, not
7810 // in SSE. (The generic code can't know it's OK to do this, or how to.)
7811 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
7812 MachineMemOperand *MMO =
7813 DAG.getMachineFunction()
7814 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7815 MachineMemOperand::MOLoad, 8, 8);
7817 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
7818 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
7819 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
7822 APInt FF(32, 0x5F800000ULL);
7824 // Check whether the sign bit is set.
7825 SDValue SignSet = DAG.getSetCC(dl, getSetCCResultType(MVT::i64),
7826 Op.getOperand(0), DAG.getConstant(0, MVT::i64),
7829 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
7830 SDValue FudgePtr = DAG.getConstantPool(
7831 ConstantInt::get(*DAG.getContext(), FF.zext(64)),
7834 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
7835 SDValue Zero = DAG.getIntPtrConstant(0);
7836 SDValue Four = DAG.getIntPtrConstant(4);
7837 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
7839 FudgePtr = DAG.getNode(ISD::ADD, dl, getPointerTy(), FudgePtr, Offset);
7841 // Load the value out, extending it from f32 to f80.
7842 // FIXME: Avoid the extend by constructing the right constant pool?
7843 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
7844 FudgePtr, MachinePointerInfo::getConstantPool(),
7845 MVT::f32, false, false, 4);
7846 // Extend everything to 80 bits to force it to be done on x87.
7847 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
7848 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add, DAG.getIntPtrConstant(0));
7851 std::pair<SDValue,SDValue> X86TargetLowering::
7852 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) const {
7853 DebugLoc DL = Op.getDebugLoc();
7855 EVT DstTy = Op.getValueType();
7858 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
7862 assert(DstTy.getSimpleVT() <= MVT::i64 &&
7863 DstTy.getSimpleVT() >= MVT::i16 &&
7864 "Unknown FP_TO_SINT to lower!");
7866 // These are really Legal.
7867 if (DstTy == MVT::i32 &&
7868 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7869 return std::make_pair(SDValue(), SDValue());
7870 if (Subtarget->is64Bit() &&
7871 DstTy == MVT::i64 &&
7872 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
7873 return std::make_pair(SDValue(), SDValue());
7875 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
7877 MachineFunction &MF = DAG.getMachineFunction();
7878 unsigned MemSize = DstTy.getSizeInBits()/8;
7879 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7880 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7885 switch (DstTy.getSimpleVT().SimpleTy) {
7886 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
7887 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
7888 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
7889 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
7892 SDValue Chain = DAG.getEntryNode();
7893 SDValue Value = Op.getOperand(0);
7894 EVT TheVT = Op.getOperand(0).getValueType();
7895 if (isScalarFPTypeInSSEReg(TheVT)) {
7896 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
7897 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
7898 MachinePointerInfo::getFixedStack(SSFI),
7900 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
7902 Chain, StackSlot, DAG.getValueType(TheVT)
7905 MachineMemOperand *MMO =
7906 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7907 MachineMemOperand::MOLoad, MemSize, MemSize);
7908 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, 3,
7910 Chain = Value.getValue(1);
7911 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
7912 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7915 MachineMemOperand *MMO =
7916 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
7917 MachineMemOperand::MOStore, MemSize, MemSize);
7919 // Build the FP_TO_INT*_IN_MEM
7920 SDValue Ops[] = { Chain, Value, StackSlot };
7921 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
7922 Ops, 3, DstTy, MMO);
7924 return std::make_pair(FIST, StackSlot);
7927 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
7928 SelectionDAG &DAG) const {
7929 if (Op.getValueType().isVector())
7932 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
7933 SDValue FIST = Vals.first, StackSlot = Vals.second;
7934 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
7935 if (FIST.getNode() == 0) return Op;
7938 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7939 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7942 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
7943 SelectionDAG &DAG) const {
7944 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
7945 SDValue FIST = Vals.first, StackSlot = Vals.second;
7946 assert(FIST.getNode() && "Unexpected failure");
7949 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
7950 FIST, StackSlot, MachinePointerInfo(), false, false, 0);
7953 SDValue X86TargetLowering::LowerFABS(SDValue Op,
7954 SelectionDAG &DAG) const {
7955 LLVMContext *Context = DAG.getContext();
7956 DebugLoc dl = Op.getDebugLoc();
7957 EVT VT = Op.getValueType();
7960 EltVT = VT.getVectorElementType();
7961 std::vector<Constant*> CV;
7962 if (EltVT == MVT::f64) {
7963 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
7967 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
7973 Constant *C = ConstantVector::get(CV);
7974 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
7975 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
7976 MachinePointerInfo::getConstantPool(),
7978 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
7981 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) const {
7982 LLVMContext *Context = DAG.getContext();
7983 DebugLoc dl = Op.getDebugLoc();
7984 EVT VT = Op.getValueType();
7987 EltVT = VT.getVectorElementType();
7988 std::vector<Constant*> CV;
7989 if (EltVT == MVT::f64) {
7990 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
7994 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
8000 Constant *C = ConstantVector::get(CV);
8001 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8002 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8003 MachinePointerInfo::getConstantPool(),
8005 if (VT.isVector()) {
8006 return DAG.getNode(ISD::BITCAST, dl, VT,
8007 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
8008 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
8010 DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
8012 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
8016 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
8017 LLVMContext *Context = DAG.getContext();
8018 SDValue Op0 = Op.getOperand(0);
8019 SDValue Op1 = Op.getOperand(1);
8020 DebugLoc dl = Op.getDebugLoc();
8021 EVT VT = Op.getValueType();
8022 EVT SrcVT = Op1.getValueType();
8024 // If second operand is smaller, extend it first.
8025 if (SrcVT.bitsLT(VT)) {
8026 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
8029 // And if it is bigger, shrink it first.
8030 if (SrcVT.bitsGT(VT)) {
8031 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
8035 // At this point the operands and the result should have the same
8036 // type, and that won't be f80 since that is not custom lowered.
8038 // First get the sign bit of second operand.
8039 std::vector<Constant*> CV;
8040 if (SrcVT == MVT::f64) {
8041 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
8042 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
8044 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
8045 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8046 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8047 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8049 Constant *C = ConstantVector::get(CV);
8050 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8051 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
8052 MachinePointerInfo::getConstantPool(),
8054 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
8056 // Shift sign bit right or left if the two operands have different types.
8057 if (SrcVT.bitsGT(VT)) {
8058 // Op0 is MVT::f32, Op1 is MVT::f64.
8059 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
8060 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
8061 DAG.getConstant(32, MVT::i32));
8062 SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
8063 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
8064 DAG.getIntPtrConstant(0));
8067 // Clear first operand sign bit.
8069 if (VT == MVT::f64) {
8070 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
8071 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
8073 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
8074 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8075 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8076 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
8078 C = ConstantVector::get(CV);
8079 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
8080 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
8081 MachinePointerInfo::getConstantPool(),
8083 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
8085 // Or the value with the sign bit.
8086 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
8089 SDValue X86TargetLowering::LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) const {
8090 SDValue N0 = Op.getOperand(0);
8091 DebugLoc dl = Op.getDebugLoc();
8092 EVT VT = Op.getValueType();
8094 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
8095 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
8096 DAG.getConstant(1, VT));
8097 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, VT));
8100 /// Emit nodes that will be selected as "test Op0,Op0", or something
8102 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
8103 SelectionDAG &DAG) const {
8104 DebugLoc dl = Op.getDebugLoc();
8106 // CF and OF aren't always set the way we want. Determine which
8107 // of these we need.
8108 bool NeedCF = false;
8109 bool NeedOF = false;
8112 case X86::COND_A: case X86::COND_AE:
8113 case X86::COND_B: case X86::COND_BE:
8116 case X86::COND_G: case X86::COND_GE:
8117 case X86::COND_L: case X86::COND_LE:
8118 case X86::COND_O: case X86::COND_NO:
8123 // See if we can use the EFLAGS value from the operand instead of
8124 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
8125 // we prove that the arithmetic won't overflow, we can't use OF or CF.
8126 if (Op.getResNo() != 0 || NeedOF || NeedCF)
8127 // Emit a CMP with 0, which is the TEST pattern.
8128 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
8129 DAG.getConstant(0, Op.getValueType()));
8131 unsigned Opcode = 0;
8132 unsigned NumOperands = 0;
8133 switch (Op.getNode()->getOpcode()) {
8135 // Due to an isel shortcoming, be conservative if this add is likely to be
8136 // selected as part of a load-modify-store instruction. When the root node
8137 // in a match is a store, isel doesn't know how to remap non-chain non-flag
8138 // uses of other nodes in the match, such as the ADD in this case. This
8139 // leads to the ADD being left around and reselected, with the result being
8140 // two adds in the output. Alas, even if none our users are stores, that
8141 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
8142 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
8143 // climbing the DAG back to the root, and it doesn't seem to be worth the
8145 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
8146 UE = Op.getNode()->use_end(); UI != UE; ++UI)
8147 if (UI->getOpcode() != ISD::CopyToReg && UI->getOpcode() != ISD::SETCC)
8150 if (ConstantSDNode *C =
8151 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
8152 // An add of one will be selected as an INC.
8153 if (C->getAPIntValue() == 1) {
8154 Opcode = X86ISD::INC;
8159 // An add of negative one (subtract of one) will be selected as a DEC.
8160 if (C->getAPIntValue().isAllOnesValue()) {
8161 Opcode = X86ISD::DEC;
8167 // Otherwise use a regular EFLAGS-setting add.
8168 Opcode = X86ISD::ADD;
8172 // If the primary and result isn't used, don't bother using X86ISD::AND,
8173 // because a TEST instruction will be better.
8174 bool NonFlagUse = false;
8175 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
8176 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
8178 unsigned UOpNo = UI.getOperandNo();
8179 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
8180 // Look pass truncate.
8181 UOpNo = User->use_begin().getOperandNo();
8182 User = *User->use_begin();
8185 if (User->getOpcode() != ISD::BRCOND &&
8186 User->getOpcode() != ISD::SETCC &&
8187 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
8200 // Due to the ISEL shortcoming noted above, be conservative if this op is
8201 // likely to be selected as part of a load-modify-store instruction.
8202 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
8203 UE = Op.getNode()->use_end(); UI != UE; ++UI)
8204 if (UI->getOpcode() == ISD::STORE)
8207 // Otherwise use a regular EFLAGS-setting instruction.
8208 switch (Op.getNode()->getOpcode()) {
8209 default: llvm_unreachable("unexpected operator!");
8210 case ISD::SUB: Opcode = X86ISD::SUB; break;
8211 case ISD::OR: Opcode = X86ISD::OR; break;
8212 case ISD::XOR: Opcode = X86ISD::XOR; break;
8213 case ISD::AND: Opcode = X86ISD::AND; break;
8225 return SDValue(Op.getNode(), 1);
8232 // Emit a CMP with 0, which is the TEST pattern.
8233 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
8234 DAG.getConstant(0, Op.getValueType()));
8236 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
8237 SmallVector<SDValue, 4> Ops;
8238 for (unsigned i = 0; i != NumOperands; ++i)
8239 Ops.push_back(Op.getOperand(i));
8241 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
8242 DAG.ReplaceAllUsesWith(Op, New);
8243 return SDValue(New.getNode(), 1);
8246 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
8248 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
8249 SelectionDAG &DAG) const {
8250 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
8251 if (C->getAPIntValue() == 0)
8252 return EmitTest(Op0, X86CC, DAG);
8254 DebugLoc dl = Op0.getDebugLoc();
8255 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
8258 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
8259 /// if it's possible.
8260 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
8261 DebugLoc dl, SelectionDAG &DAG) const {
8262 SDValue Op0 = And.getOperand(0);
8263 SDValue Op1 = And.getOperand(1);
8264 if (Op0.getOpcode() == ISD::TRUNCATE)
8265 Op0 = Op0.getOperand(0);
8266 if (Op1.getOpcode() == ISD::TRUNCATE)
8267 Op1 = Op1.getOperand(0);
8270 if (Op1.getOpcode() == ISD::SHL)
8271 std::swap(Op0, Op1);
8272 if (Op0.getOpcode() == ISD::SHL) {
8273 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
8274 if (And00C->getZExtValue() == 1) {
8275 // If we looked past a truncate, check that it's only truncating away
8277 unsigned BitWidth = Op0.getValueSizeInBits();
8278 unsigned AndBitWidth = And.getValueSizeInBits();
8279 if (BitWidth > AndBitWidth) {
8280 APInt Mask = APInt::getAllOnesValue(BitWidth), Zeros, Ones;
8281 DAG.ComputeMaskedBits(Op0, Mask, Zeros, Ones);
8282 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
8286 RHS = Op0.getOperand(1);
8288 } else if (Op1.getOpcode() == ISD::Constant) {
8289 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
8290 SDValue AndLHS = Op0;
8291 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
8292 LHS = AndLHS.getOperand(0);
8293 RHS = AndLHS.getOperand(1);
8297 if (LHS.getNode()) {
8298 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
8299 // instruction. Since the shift amount is in-range-or-undefined, we know
8300 // that doing a bittest on the i32 value is ok. We extend to i32 because
8301 // the encoding for the i16 version is larger than the i32 version.
8302 // Also promote i16 to i32 for performance / code size reason.
8303 if (LHS.getValueType() == MVT::i8 ||
8304 LHS.getValueType() == MVT::i16)
8305 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
8307 // If the operand types disagree, extend the shift amount to match. Since
8308 // BT ignores high bits (like shifts) we can use anyextend.
8309 if (LHS.getValueType() != RHS.getValueType())
8310 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
8312 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
8313 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
8314 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8315 DAG.getConstant(Cond, MVT::i8), BT);
8321 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
8322 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
8323 SDValue Op0 = Op.getOperand(0);
8324 SDValue Op1 = Op.getOperand(1);
8325 DebugLoc dl = Op.getDebugLoc();
8326 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
8328 // Optimize to BT if possible.
8329 // Lower (X & (1 << N)) == 0 to BT(X, N).
8330 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
8331 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
8332 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
8333 Op1.getOpcode() == ISD::Constant &&
8334 cast<ConstantSDNode>(Op1)->isNullValue() &&
8335 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
8336 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
8337 if (NewSetCC.getNode())
8341 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
8343 if (Op1.getOpcode() == ISD::Constant &&
8344 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
8345 cast<ConstantSDNode>(Op1)->isNullValue()) &&
8346 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
8348 // If the input is a setcc, then reuse the input setcc or use a new one with
8349 // the inverted condition.
8350 if (Op0.getOpcode() == X86ISD::SETCC) {
8351 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
8352 bool Invert = (CC == ISD::SETNE) ^
8353 cast<ConstantSDNode>(Op1)->isNullValue();
8354 if (!Invert) return Op0;
8356 CCode = X86::GetOppositeBranchCondition(CCode);
8357 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8358 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
8362 bool isFP = Op1.getValueType().isFloatingPoint();
8363 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
8364 if (X86CC == X86::COND_INVALID)
8367 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
8368 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
8369 DAG.getConstant(X86CC, MVT::i8), EFLAGS);
8372 // Lower256IntVETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
8373 // ones, and then concatenate the result back.
8374 static SDValue Lower256IntVETCC(SDValue Op, SelectionDAG &DAG) {
8375 EVT VT = Op.getValueType();
8377 assert(VT.getSizeInBits() == 256 && Op.getOpcode() == ISD::VSETCC &&
8378 "Unsupported value type for operation");
8380 int NumElems = VT.getVectorNumElements();
8381 DebugLoc dl = Op.getDebugLoc();
8382 SDValue CC = Op.getOperand(2);
8383 SDValue Idx0 = DAG.getConstant(0, MVT::i32);
8384 SDValue Idx1 = DAG.getConstant(NumElems/2, MVT::i32);
8386 // Extract the LHS vectors
8387 SDValue LHS = Op.getOperand(0);
8388 SDValue LHS1 = Extract128BitVector(LHS, Idx0, DAG, dl);
8389 SDValue LHS2 = Extract128BitVector(LHS, Idx1, DAG, dl);
8391 // Extract the RHS vectors
8392 SDValue RHS = Op.getOperand(1);
8393 SDValue RHS1 = Extract128BitVector(RHS, Idx0, DAG, dl);
8394 SDValue RHS2 = Extract128BitVector(RHS, Idx1, DAG, dl);
8396 // Issue the operation on the smaller types and concatenate the result back
8397 MVT EltVT = VT.getVectorElementType().getSimpleVT();
8398 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
8399 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
8400 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
8401 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
8405 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
8407 SDValue Op0 = Op.getOperand(0);
8408 SDValue Op1 = Op.getOperand(1);
8409 SDValue CC = Op.getOperand(2);
8410 EVT VT = Op.getValueType();
8411 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
8412 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
8413 DebugLoc dl = Op.getDebugLoc();
8417 EVT EltVT = Op0.getValueType().getVectorElementType();
8418 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
8420 unsigned Opc = EltVT == MVT::f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
8423 switch (SetCCOpcode) {
8426 case ISD::SETEQ: SSECC = 0; break;
8428 case ISD::SETGT: Swap = true; // Fallthrough
8430 case ISD::SETOLT: SSECC = 1; break;
8432 case ISD::SETGE: Swap = true; // Fallthrough
8434 case ISD::SETOLE: SSECC = 2; break;
8435 case ISD::SETUO: SSECC = 3; break;
8437 case ISD::SETNE: SSECC = 4; break;
8438 case ISD::SETULE: Swap = true;
8439 case ISD::SETUGE: SSECC = 5; break;
8440 case ISD::SETULT: Swap = true;
8441 case ISD::SETUGT: SSECC = 6; break;
8442 case ISD::SETO: SSECC = 7; break;
8445 std::swap(Op0, Op1);
8447 // In the two special cases we can't handle, emit two comparisons.
8449 if (SetCCOpcode == ISD::SETUEQ) {
8451 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
8452 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
8453 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
8455 else if (SetCCOpcode == ISD::SETONE) {
8457 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
8458 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
8459 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
8461 llvm_unreachable("Illegal FP comparison");
8463 // Handle all other FP comparisons here.
8464 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
8467 // Break 256-bit integer vector compare into smaller ones.
8468 if (!isFP && VT.getSizeInBits() == 256)
8469 return Lower256IntVETCC(Op, DAG);
8471 // We are handling one of the integer comparisons here. Since SSE only has
8472 // GT and EQ comparisons for integer, swapping operands and multiple
8473 // operations may be required for some comparisons.
8474 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
8475 bool Swap = false, Invert = false, FlipSigns = false;
8477 switch (VT.getSimpleVT().SimpleTy) {
8479 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
8480 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
8481 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
8482 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
8485 switch (SetCCOpcode) {
8487 case ISD::SETNE: Invert = true;
8488 case ISD::SETEQ: Opc = EQOpc; break;
8489 case ISD::SETLT: Swap = true;
8490 case ISD::SETGT: Opc = GTOpc; break;
8491 case ISD::SETGE: Swap = true;
8492 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
8493 case ISD::SETULT: Swap = true;
8494 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
8495 case ISD::SETUGE: Swap = true;
8496 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
8499 std::swap(Op0, Op1);
8501 // Since SSE has no unsigned integer comparisons, we need to flip the sign
8502 // bits of the inputs before performing those operations.
8504 EVT EltVT = VT.getVectorElementType();
8505 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
8507 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
8508 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
8510 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
8511 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
8514 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
8516 // If the logical-not of the result is required, perform that now.
8518 Result = DAG.getNOT(dl, Result, VT);
8523 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
8524 static bool isX86LogicalCmp(SDValue Op) {
8525 unsigned Opc = Op.getNode()->getOpcode();
8526 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
8528 if (Op.getResNo() == 1 &&
8529 (Opc == X86ISD::ADD ||
8530 Opc == X86ISD::SUB ||
8531 Opc == X86ISD::ADC ||
8532 Opc == X86ISD::SBB ||
8533 Opc == X86ISD::SMUL ||
8534 Opc == X86ISD::UMUL ||
8535 Opc == X86ISD::INC ||
8536 Opc == X86ISD::DEC ||
8537 Opc == X86ISD::OR ||
8538 Opc == X86ISD::XOR ||
8539 Opc == X86ISD::AND))
8542 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
8548 static bool isZero(SDValue V) {
8549 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
8550 return C && C->isNullValue();
8553 static bool isAllOnes(SDValue V) {
8554 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
8555 return C && C->isAllOnesValue();
8558 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
8559 bool addTest = true;
8560 SDValue Cond = Op.getOperand(0);
8561 SDValue Op1 = Op.getOperand(1);
8562 SDValue Op2 = Op.getOperand(2);
8563 DebugLoc DL = Op.getDebugLoc();
8566 if (Cond.getOpcode() == ISD::SETCC) {
8567 SDValue NewCond = LowerSETCC(Cond, DAG);
8568 if (NewCond.getNode())
8572 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
8573 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
8574 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
8575 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
8576 if (Cond.getOpcode() == X86ISD::SETCC &&
8577 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
8578 isZero(Cond.getOperand(1).getOperand(1))) {
8579 SDValue Cmp = Cond.getOperand(1);
8581 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
8583 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
8584 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
8585 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
8587 SDValue CmpOp0 = Cmp.getOperand(0);
8588 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
8589 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
8591 SDValue Res = // Res = 0 or -1.
8592 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
8593 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
8595 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
8596 Res = DAG.getNOT(DL, Res, Res.getValueType());
8598 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
8599 if (N2C == 0 || !N2C->isNullValue())
8600 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
8605 // Look past (and (setcc_carry (cmp ...)), 1).
8606 if (Cond.getOpcode() == ISD::AND &&
8607 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
8608 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
8609 if (C && C->getAPIntValue() == 1)
8610 Cond = Cond.getOperand(0);
8613 // If condition flag is set by a X86ISD::CMP, then use it as the condition
8614 // setting operand in place of the X86ISD::SETCC.
8615 if (Cond.getOpcode() == X86ISD::SETCC ||
8616 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
8617 CC = Cond.getOperand(0);
8619 SDValue Cmp = Cond.getOperand(1);
8620 unsigned Opc = Cmp.getOpcode();
8621 EVT VT = Op.getValueType();
8623 bool IllegalFPCMov = false;
8624 if (VT.isFloatingPoint() && !VT.isVector() &&
8625 !isScalarFPTypeInSSEReg(VT)) // FPStack?
8626 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
8628 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
8629 Opc == X86ISD::BT) { // FIXME
8636 // Look pass the truncate.
8637 if (Cond.getOpcode() == ISD::TRUNCATE)
8638 Cond = Cond.getOperand(0);
8640 // We know the result of AND is compared against zero. Try to match
8642 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
8643 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
8644 if (NewSetCC.getNode()) {
8645 CC = NewSetCC.getOperand(0);
8646 Cond = NewSetCC.getOperand(1);
8653 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8654 Cond = EmitTest(Cond, X86::COND_NE, DAG);
8657 // a < b ? -1 : 0 -> RES = ~setcc_carry
8658 // a < b ? 0 : -1 -> RES = setcc_carry
8659 // a >= b ? -1 : 0 -> RES = setcc_carry
8660 // a >= b ? 0 : -1 -> RES = ~setcc_carry
8661 if (Cond.getOpcode() == X86ISD::CMP) {
8662 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
8664 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
8665 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
8666 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
8667 DAG.getConstant(X86::COND_B, MVT::i8), Cond);
8668 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
8669 return DAG.getNOT(DL, Res, Res.getValueType());
8674 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
8675 // condition is true.
8676 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
8677 SDValue Ops[] = { Op2, Op1, CC, Cond };
8678 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
8681 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
8682 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
8683 // from the AND / OR.
8684 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
8685 Opc = Op.getOpcode();
8686 if (Opc != ISD::OR && Opc != ISD::AND)
8688 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
8689 Op.getOperand(0).hasOneUse() &&
8690 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
8691 Op.getOperand(1).hasOneUse());
8694 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
8695 // 1 and that the SETCC node has a single use.
8696 static bool isXor1OfSetCC(SDValue Op) {
8697 if (Op.getOpcode() != ISD::XOR)
8699 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
8700 if (N1C && N1C->getAPIntValue() == 1) {
8701 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
8702 Op.getOperand(0).hasOneUse();
8707 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
8708 bool addTest = true;
8709 SDValue Chain = Op.getOperand(0);
8710 SDValue Cond = Op.getOperand(1);
8711 SDValue Dest = Op.getOperand(2);
8712 DebugLoc dl = Op.getDebugLoc();
8715 if (Cond.getOpcode() == ISD::SETCC) {
8716 SDValue NewCond = LowerSETCC(Cond, DAG);
8717 if (NewCond.getNode())
8721 // FIXME: LowerXALUO doesn't handle these!!
8722 else if (Cond.getOpcode() == X86ISD::ADD ||
8723 Cond.getOpcode() == X86ISD::SUB ||
8724 Cond.getOpcode() == X86ISD::SMUL ||
8725 Cond.getOpcode() == X86ISD::UMUL)
8726 Cond = LowerXALUO(Cond, DAG);
8729 // Look pass (and (setcc_carry (cmp ...)), 1).
8730 if (Cond.getOpcode() == ISD::AND &&
8731 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
8732 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
8733 if (C && C->getAPIntValue() == 1)
8734 Cond = Cond.getOperand(0);
8737 // If condition flag is set by a X86ISD::CMP, then use it as the condition
8738 // setting operand in place of the X86ISD::SETCC.
8739 if (Cond.getOpcode() == X86ISD::SETCC ||
8740 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
8741 CC = Cond.getOperand(0);
8743 SDValue Cmp = Cond.getOperand(1);
8744 unsigned Opc = Cmp.getOpcode();
8745 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
8746 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
8750 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
8754 // These can only come from an arithmetic instruction with overflow,
8755 // e.g. SADDO, UADDO.
8756 Cond = Cond.getNode()->getOperand(1);
8763 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
8764 SDValue Cmp = Cond.getOperand(0).getOperand(1);
8765 if (CondOpc == ISD::OR) {
8766 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
8767 // two branches instead of an explicit OR instruction with a
8769 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8770 isX86LogicalCmp(Cmp)) {
8771 CC = Cond.getOperand(0).getOperand(0);
8772 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8773 Chain, Dest, CC, Cmp);
8774 CC = Cond.getOperand(1).getOperand(0);
8778 } else { // ISD::AND
8779 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
8780 // two branches instead of an explicit AND instruction with a
8781 // separate test. However, we only do this if this block doesn't
8782 // have a fall-through edge, because this requires an explicit
8783 // jmp when the condition is false.
8784 if (Cmp == Cond.getOperand(1).getOperand(1) &&
8785 isX86LogicalCmp(Cmp) &&
8786 Op.getNode()->hasOneUse()) {
8787 X86::CondCode CCode =
8788 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8789 CCode = X86::GetOppositeBranchCondition(CCode);
8790 CC = DAG.getConstant(CCode, MVT::i8);
8791 SDNode *User = *Op.getNode()->use_begin();
8792 // Look for an unconditional branch following this conditional branch.
8793 // We need this because we need to reverse the successors in order
8794 // to implement FCMP_OEQ.
8795 if (User->getOpcode() == ISD::BR) {
8796 SDValue FalseBB = User->getOperand(1);
8798 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
8799 assert(NewBR == User);
8803 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8804 Chain, Dest, CC, Cmp);
8805 X86::CondCode CCode =
8806 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
8807 CCode = X86::GetOppositeBranchCondition(CCode);
8808 CC = DAG.getConstant(CCode, MVT::i8);
8814 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
8815 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
8816 // It should be transformed during dag combiner except when the condition
8817 // is set by a arithmetics with overflow node.
8818 X86::CondCode CCode =
8819 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
8820 CCode = X86::GetOppositeBranchCondition(CCode);
8821 CC = DAG.getConstant(CCode, MVT::i8);
8822 Cond = Cond.getOperand(0).getOperand(1);
8828 // Look pass the truncate.
8829 if (Cond.getOpcode() == ISD::TRUNCATE)
8830 Cond = Cond.getOperand(0);
8832 // We know the result of AND is compared against zero. Try to match
8834 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
8835 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
8836 if (NewSetCC.getNode()) {
8837 CC = NewSetCC.getOperand(0);
8838 Cond = NewSetCC.getOperand(1);
8845 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
8846 Cond = EmitTest(Cond, X86::COND_NE, DAG);
8848 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
8849 Chain, Dest, CC, Cond);
8853 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
8854 // Calls to _alloca is needed to probe the stack when allocating more than 4k
8855 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
8856 // that the guard pages used by the OS virtual memory manager are allocated in
8857 // correct sequence.
8859 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
8860 SelectionDAG &DAG) const {
8861 assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows() ||
8862 EnableSegmentedStacks) &&
8863 "This should be used only on Windows targets or when segmented stacks "
8865 assert(!Subtarget->isTargetEnvMacho() && "Not implemented");
8866 DebugLoc dl = Op.getDebugLoc();
8869 SDValue Chain = Op.getOperand(0);
8870 SDValue Size = Op.getOperand(1);
8871 // FIXME: Ensure alignment here
8873 bool Is64Bit = Subtarget->is64Bit();
8874 EVT SPTy = Is64Bit ? MVT::i64 : MVT::i32;
8876 if (EnableSegmentedStacks) {
8877 MachineFunction &MF = DAG.getMachineFunction();
8878 MachineRegisterInfo &MRI = MF.getRegInfo();
8881 // The 64 bit implementation of segmented stacks needs to clobber both r10
8882 // r11. This makes it impossible to use it along with nested parameters.
8883 const Function *F = MF.getFunction();
8885 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
8887 if (I->hasNestAttr())
8888 report_fatal_error("Cannot use segmented stacks with functions that "
8889 "have nested arguments.");
8892 const TargetRegisterClass *AddrRegClass =
8893 getRegClassFor(Subtarget->is64Bit() ? MVT::i64:MVT::i32);
8894 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
8895 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
8896 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
8897 DAG.getRegister(Vreg, SPTy));
8898 SDValue Ops1[2] = { Value, Chain };
8899 return DAG.getMergeValues(Ops1, 2, dl);
8902 unsigned Reg = (Subtarget->is64Bit() ? X86::RAX : X86::EAX);
8904 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
8905 Flag = Chain.getValue(1);
8906 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
8908 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
8909 Flag = Chain.getValue(1);
8911 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
8913 SDValue Ops1[2] = { Chain.getValue(0), Chain };
8914 return DAG.getMergeValues(Ops1, 2, dl);
8918 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
8919 MachineFunction &MF = DAG.getMachineFunction();
8920 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
8922 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8923 DebugLoc DL = Op.getDebugLoc();
8925 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
8926 // vastart just stores the address of the VarArgsFrameIndex slot into the
8927 // memory location argument.
8928 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8930 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
8931 MachinePointerInfo(SV), false, false, 0);
8935 // gp_offset (0 - 6 * 8)
8936 // fp_offset (48 - 48 + 8 * 16)
8937 // overflow_arg_area (point to parameters coming in memory).
8939 SmallVector<SDValue, 8> MemOps;
8940 SDValue FIN = Op.getOperand(1);
8942 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
8943 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
8945 FIN, MachinePointerInfo(SV), false, false, 0);
8946 MemOps.push_back(Store);
8949 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8950 FIN, DAG.getIntPtrConstant(4));
8951 Store = DAG.getStore(Op.getOperand(0), DL,
8952 DAG.getConstant(FuncInfo->getVarArgsFPOffset(),
8954 FIN, MachinePointerInfo(SV, 4), false, false, 0);
8955 MemOps.push_back(Store);
8957 // Store ptr to overflow_arg_area
8958 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8959 FIN, DAG.getIntPtrConstant(4));
8960 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(),
8962 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
8963 MachinePointerInfo(SV, 8),
8965 MemOps.push_back(Store);
8967 // Store ptr to reg_save_area.
8968 FIN = DAG.getNode(ISD::ADD, DL, getPointerTy(),
8969 FIN, DAG.getIntPtrConstant(8));
8970 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
8972 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
8973 MachinePointerInfo(SV, 16), false, false, 0);
8974 MemOps.push_back(Store);
8975 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
8976 &MemOps[0], MemOps.size());
8979 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
8980 assert(Subtarget->is64Bit() &&
8981 "LowerVAARG only handles 64-bit va_arg!");
8982 assert((Subtarget->isTargetLinux() ||
8983 Subtarget->isTargetDarwin()) &&
8984 "Unhandled target in LowerVAARG");
8985 assert(Op.getNode()->getNumOperands() == 4);
8986 SDValue Chain = Op.getOperand(0);
8987 SDValue SrcPtr = Op.getOperand(1);
8988 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
8989 unsigned Align = Op.getConstantOperandVal(3);
8990 DebugLoc dl = Op.getDebugLoc();
8992 EVT ArgVT = Op.getNode()->getValueType(0);
8993 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
8994 uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
8997 // Decide which area this value should be read from.
8998 // TODO: Implement the AMD64 ABI in its entirety. This simple
8999 // selection mechanism works only for the basic types.
9000 if (ArgVT == MVT::f80) {
9001 llvm_unreachable("va_arg for f80 not yet implemented");
9002 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
9003 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
9004 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
9005 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
9007 llvm_unreachable("Unhandled argument type in LowerVAARG");
9011 // Sanity Check: Make sure using fp_offset makes sense.
9012 assert(!UseSoftFloat &&
9013 !(DAG.getMachineFunction()
9014 .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
9015 Subtarget->hasXMM());
9018 // Insert VAARG_64 node into the DAG
9019 // VAARG_64 returns two values: Variable Argument Address, Chain
9020 SmallVector<SDValue, 11> InstOps;
9021 InstOps.push_back(Chain);
9022 InstOps.push_back(SrcPtr);
9023 InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
9024 InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
9025 InstOps.push_back(DAG.getConstant(Align, MVT::i32));
9026 SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
9027 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
9028 VTs, &InstOps[0], InstOps.size(),
9030 MachinePointerInfo(SV),
9035 Chain = VAARG.getValue(1);
9037 // Load the next argument and return it
9038 return DAG.getLoad(ArgVT, dl,
9041 MachinePointerInfo(),
9045 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
9046 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
9047 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
9048 SDValue Chain = Op.getOperand(0);
9049 SDValue DstPtr = Op.getOperand(1);
9050 SDValue SrcPtr = Op.getOperand(2);
9051 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
9052 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
9053 DebugLoc DL = Op.getDebugLoc();
9055 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
9056 DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
9058 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
9062 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const {
9063 DebugLoc dl = Op.getDebugLoc();
9064 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9066 default: return SDValue(); // Don't custom lower most intrinsics.
9067 // Comparison intrinsics.
9068 case Intrinsic::x86_sse_comieq_ss:
9069 case Intrinsic::x86_sse_comilt_ss:
9070 case Intrinsic::x86_sse_comile_ss:
9071 case Intrinsic::x86_sse_comigt_ss:
9072 case Intrinsic::x86_sse_comige_ss:
9073 case Intrinsic::x86_sse_comineq_ss:
9074 case Intrinsic::x86_sse_ucomieq_ss:
9075 case Intrinsic::x86_sse_ucomilt_ss:
9076 case Intrinsic::x86_sse_ucomile_ss:
9077 case Intrinsic::x86_sse_ucomigt_ss:
9078 case Intrinsic::x86_sse_ucomige_ss:
9079 case Intrinsic::x86_sse_ucomineq_ss:
9080 case Intrinsic::x86_sse2_comieq_sd:
9081 case Intrinsic::x86_sse2_comilt_sd:
9082 case Intrinsic::x86_sse2_comile_sd:
9083 case Intrinsic::x86_sse2_comigt_sd:
9084 case Intrinsic::x86_sse2_comige_sd:
9085 case Intrinsic::x86_sse2_comineq_sd:
9086 case Intrinsic::x86_sse2_ucomieq_sd:
9087 case Intrinsic::x86_sse2_ucomilt_sd:
9088 case Intrinsic::x86_sse2_ucomile_sd:
9089 case Intrinsic::x86_sse2_ucomigt_sd:
9090 case Intrinsic::x86_sse2_ucomige_sd:
9091 case Intrinsic::x86_sse2_ucomineq_sd: {
9093 ISD::CondCode CC = ISD::SETCC_INVALID;
9096 case Intrinsic::x86_sse_comieq_ss:
9097 case Intrinsic::x86_sse2_comieq_sd:
9101 case Intrinsic::x86_sse_comilt_ss:
9102 case Intrinsic::x86_sse2_comilt_sd:
9106 case Intrinsic::x86_sse_comile_ss:
9107 case Intrinsic::x86_sse2_comile_sd:
9111 case Intrinsic::x86_sse_comigt_ss:
9112 case Intrinsic::x86_sse2_comigt_sd:
9116 case Intrinsic::x86_sse_comige_ss:
9117 case Intrinsic::x86_sse2_comige_sd:
9121 case Intrinsic::x86_sse_comineq_ss:
9122 case Intrinsic::x86_sse2_comineq_sd:
9126 case Intrinsic::x86_sse_ucomieq_ss:
9127 case Intrinsic::x86_sse2_ucomieq_sd:
9128 Opc = X86ISD::UCOMI;
9131 case Intrinsic::x86_sse_ucomilt_ss:
9132 case Intrinsic::x86_sse2_ucomilt_sd:
9133 Opc = X86ISD::UCOMI;
9136 case Intrinsic::x86_sse_ucomile_ss:
9137 case Intrinsic::x86_sse2_ucomile_sd:
9138 Opc = X86ISD::UCOMI;
9141 case Intrinsic::x86_sse_ucomigt_ss:
9142 case Intrinsic::x86_sse2_ucomigt_sd:
9143 Opc = X86ISD::UCOMI;
9146 case Intrinsic::x86_sse_ucomige_ss:
9147 case Intrinsic::x86_sse2_ucomige_sd:
9148 Opc = X86ISD::UCOMI;
9151 case Intrinsic::x86_sse_ucomineq_ss:
9152 case Intrinsic::x86_sse2_ucomineq_sd:
9153 Opc = X86ISD::UCOMI;
9158 SDValue LHS = Op.getOperand(1);
9159 SDValue RHS = Op.getOperand(2);
9160 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
9161 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
9162 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
9163 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
9164 DAG.getConstant(X86CC, MVT::i8), Cond);
9165 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
9167 // ptest and testp intrinsics. The intrinsic these come from are designed to
9168 // return an integer value, not just an instruction so lower it to the ptest
9169 // or testp pattern and a setcc for the result.
9170 case Intrinsic::x86_sse41_ptestz:
9171 case Intrinsic::x86_sse41_ptestc:
9172 case Intrinsic::x86_sse41_ptestnzc:
9173 case Intrinsic::x86_avx_ptestz_256:
9174 case Intrinsic::x86_avx_ptestc_256:
9175 case Intrinsic::x86_avx_ptestnzc_256:
9176 case Intrinsic::x86_avx_vtestz_ps:
9177 case Intrinsic::x86_avx_vtestc_ps:
9178 case Intrinsic::x86_avx_vtestnzc_ps:
9179 case Intrinsic::x86_avx_vtestz_pd:
9180 case Intrinsic::x86_avx_vtestc_pd:
9181 case Intrinsic::x86_avx_vtestnzc_pd:
9182 case Intrinsic::x86_avx_vtestz_ps_256:
9183 case Intrinsic::x86_avx_vtestc_ps_256:
9184 case Intrinsic::x86_avx_vtestnzc_ps_256:
9185 case Intrinsic::x86_avx_vtestz_pd_256:
9186 case Intrinsic::x86_avx_vtestc_pd_256:
9187 case Intrinsic::x86_avx_vtestnzc_pd_256: {
9188 bool IsTestPacked = false;
9191 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
9192 case Intrinsic::x86_avx_vtestz_ps:
9193 case Intrinsic::x86_avx_vtestz_pd:
9194 case Intrinsic::x86_avx_vtestz_ps_256:
9195 case Intrinsic::x86_avx_vtestz_pd_256:
9196 IsTestPacked = true; // Fallthrough
9197 case Intrinsic::x86_sse41_ptestz:
9198 case Intrinsic::x86_avx_ptestz_256:
9200 X86CC = X86::COND_E;
9202 case Intrinsic::x86_avx_vtestc_ps:
9203 case Intrinsic::x86_avx_vtestc_pd:
9204 case Intrinsic::x86_avx_vtestc_ps_256:
9205 case Intrinsic::x86_avx_vtestc_pd_256:
9206 IsTestPacked = true; // Fallthrough
9207 case Intrinsic::x86_sse41_ptestc:
9208 case Intrinsic::x86_avx_ptestc_256:
9210 X86CC = X86::COND_B;
9212 case Intrinsic::x86_avx_vtestnzc_ps:
9213 case Intrinsic::x86_avx_vtestnzc_pd:
9214 case Intrinsic::x86_avx_vtestnzc_ps_256:
9215 case Intrinsic::x86_avx_vtestnzc_pd_256:
9216 IsTestPacked = true; // Fallthrough
9217 case Intrinsic::x86_sse41_ptestnzc:
9218 case Intrinsic::x86_avx_ptestnzc_256:
9220 X86CC = X86::COND_A;
9224 SDValue LHS = Op.getOperand(1);
9225 SDValue RHS = Op.getOperand(2);
9226 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
9227 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
9228 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
9229 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
9230 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
9233 // Fix vector shift instructions where the last operand is a non-immediate
9235 case Intrinsic::x86_sse2_pslli_w:
9236 case Intrinsic::x86_sse2_pslli_d:
9237 case Intrinsic::x86_sse2_pslli_q:
9238 case Intrinsic::x86_sse2_psrli_w:
9239 case Intrinsic::x86_sse2_psrli_d:
9240 case Intrinsic::x86_sse2_psrli_q:
9241 case Intrinsic::x86_sse2_psrai_w:
9242 case Intrinsic::x86_sse2_psrai_d:
9243 case Intrinsic::x86_mmx_pslli_w:
9244 case Intrinsic::x86_mmx_pslli_d:
9245 case Intrinsic::x86_mmx_pslli_q:
9246 case Intrinsic::x86_mmx_psrli_w:
9247 case Intrinsic::x86_mmx_psrli_d:
9248 case Intrinsic::x86_mmx_psrli_q:
9249 case Intrinsic::x86_mmx_psrai_w:
9250 case Intrinsic::x86_mmx_psrai_d: {
9251 SDValue ShAmt = Op.getOperand(2);
9252 if (isa<ConstantSDNode>(ShAmt))
9255 unsigned NewIntNo = 0;
9256 EVT ShAmtVT = MVT::v4i32;
9258 case Intrinsic::x86_sse2_pslli_w:
9259 NewIntNo = Intrinsic::x86_sse2_psll_w;
9261 case Intrinsic::x86_sse2_pslli_d:
9262 NewIntNo = Intrinsic::x86_sse2_psll_d;
9264 case Intrinsic::x86_sse2_pslli_q:
9265 NewIntNo = Intrinsic::x86_sse2_psll_q;
9267 case Intrinsic::x86_sse2_psrli_w:
9268 NewIntNo = Intrinsic::x86_sse2_psrl_w;
9270 case Intrinsic::x86_sse2_psrli_d:
9271 NewIntNo = Intrinsic::x86_sse2_psrl_d;
9273 case Intrinsic::x86_sse2_psrli_q:
9274 NewIntNo = Intrinsic::x86_sse2_psrl_q;
9276 case Intrinsic::x86_sse2_psrai_w:
9277 NewIntNo = Intrinsic::x86_sse2_psra_w;
9279 case Intrinsic::x86_sse2_psrai_d:
9280 NewIntNo = Intrinsic::x86_sse2_psra_d;
9283 ShAmtVT = MVT::v2i32;
9285 case Intrinsic::x86_mmx_pslli_w:
9286 NewIntNo = Intrinsic::x86_mmx_psll_w;
9288 case Intrinsic::x86_mmx_pslli_d:
9289 NewIntNo = Intrinsic::x86_mmx_psll_d;
9291 case Intrinsic::x86_mmx_pslli_q:
9292 NewIntNo = Intrinsic::x86_mmx_psll_q;
9294 case Intrinsic::x86_mmx_psrli_w:
9295 NewIntNo = Intrinsic::x86_mmx_psrl_w;
9297 case Intrinsic::x86_mmx_psrli_d:
9298 NewIntNo = Intrinsic::x86_mmx_psrl_d;
9300 case Intrinsic::x86_mmx_psrli_q:
9301 NewIntNo = Intrinsic::x86_mmx_psrl_q;
9303 case Intrinsic::x86_mmx_psrai_w:
9304 NewIntNo = Intrinsic::x86_mmx_psra_w;
9306 case Intrinsic::x86_mmx_psrai_d:
9307 NewIntNo = Intrinsic::x86_mmx_psra_d;
9309 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
9315 // The vector shift intrinsics with scalars uses 32b shift amounts but
9316 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
9320 ShOps[1] = DAG.getConstant(0, MVT::i32);
9321 if (ShAmtVT == MVT::v4i32) {
9322 ShOps[2] = DAG.getUNDEF(MVT::i32);
9323 ShOps[3] = DAG.getUNDEF(MVT::i32);
9324 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
9326 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
9327 // FIXME this must be lowered to get rid of the invalid type.
9330 EVT VT = Op.getValueType();
9331 ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
9332 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9333 DAG.getConstant(NewIntNo, MVT::i32),
9334 Op.getOperand(1), ShAmt);
9339 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
9340 SelectionDAG &DAG) const {
9341 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
9342 MFI->setReturnAddressIsTaken(true);
9344 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9345 DebugLoc dl = Op.getDebugLoc();
9348 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
9350 DAG.getConstant(TD->getPointerSize(),
9351 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
9352 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
9353 DAG.getNode(ISD::ADD, dl, getPointerTy(),
9355 MachinePointerInfo(), false, false, 0);
9358 // Just load the return address.
9359 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
9360 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
9361 RetAddrFI, MachinePointerInfo(), false, false, 0);
9364 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
9365 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
9366 MFI->setFrameAddressIsTaken(true);
9368 EVT VT = Op.getValueType();
9369 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
9370 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
9371 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
9372 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
9374 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
9375 MachinePointerInfo(),
9380 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
9381 SelectionDAG &DAG) const {
9382 return DAG.getIntPtrConstant(2*TD->getPointerSize());
9385 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
9386 MachineFunction &MF = DAG.getMachineFunction();
9387 SDValue Chain = Op.getOperand(0);
9388 SDValue Offset = Op.getOperand(1);
9389 SDValue Handler = Op.getOperand(2);
9390 DebugLoc dl = Op.getDebugLoc();
9392 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl,
9393 Subtarget->is64Bit() ? X86::RBP : X86::EBP,
9395 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
9397 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), Frame,
9398 DAG.getIntPtrConstant(TD->getPointerSize()));
9399 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
9400 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
9402 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
9403 MF.getRegInfo().addLiveOut(StoreAddrReg);
9405 return DAG.getNode(X86ISD::EH_RETURN, dl,
9407 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
9410 SDValue X86TargetLowering::LowerADJUST_TRAMPOLINE(SDValue Op,
9411 SelectionDAG &DAG) const {
9412 return Op.getOperand(0);
9415 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
9416 SelectionDAG &DAG) const {
9417 SDValue Root = Op.getOperand(0);
9418 SDValue Trmp = Op.getOperand(1); // trampoline
9419 SDValue FPtr = Op.getOperand(2); // nested function
9420 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
9421 DebugLoc dl = Op.getDebugLoc();
9423 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
9425 if (Subtarget->is64Bit()) {
9426 SDValue OutChains[6];
9428 // Large code-model.
9429 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
9430 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
9432 const unsigned char N86R10 = X86_MC::getX86RegNum(X86::R10);
9433 const unsigned char N86R11 = X86_MC::getX86RegNum(X86::R11);
9435 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
9437 // Load the pointer to the nested function into R11.
9438 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
9439 SDValue Addr = Trmp;
9440 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
9441 Addr, MachinePointerInfo(TrmpAddr),
9444 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9445 DAG.getConstant(2, MVT::i64));
9446 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
9447 MachinePointerInfo(TrmpAddr, 2),
9450 // Load the 'nest' parameter value into R10.
9451 // R10 is specified in X86CallingConv.td
9452 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
9453 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9454 DAG.getConstant(10, MVT::i64));
9455 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
9456 Addr, MachinePointerInfo(TrmpAddr, 10),
9459 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9460 DAG.getConstant(12, MVT::i64));
9461 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
9462 MachinePointerInfo(TrmpAddr, 12),
9465 // Jump to the nested function.
9466 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
9467 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9468 DAG.getConstant(20, MVT::i64));
9469 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
9470 Addr, MachinePointerInfo(TrmpAddr, 20),
9473 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
9474 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
9475 DAG.getConstant(22, MVT::i64));
9476 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
9477 MachinePointerInfo(TrmpAddr, 22),
9480 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6);
9482 const Function *Func =
9483 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
9484 CallingConv::ID CC = Func->getCallingConv();
9489 llvm_unreachable("Unsupported calling convention");
9490 case CallingConv::C:
9491 case CallingConv::X86_StdCall: {
9492 // Pass 'nest' parameter in ECX.
9493 // Must be kept in sync with X86CallingConv.td
9496 // Check that ECX wasn't needed by an 'inreg' parameter.
9497 FunctionType *FTy = Func->getFunctionType();
9498 const AttrListPtr &Attrs = Func->getAttributes();
9500 if (!Attrs.isEmpty() && !Func->isVarArg()) {
9501 unsigned InRegCount = 0;
9504 for (FunctionType::param_iterator I = FTy->param_begin(),
9505 E = FTy->param_end(); I != E; ++I, ++Idx)
9506 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
9507 // FIXME: should only count parameters that are lowered to integers.
9508 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
9510 if (InRegCount > 2) {
9511 report_fatal_error("Nest register in use - reduce number of inreg"
9517 case CallingConv::X86_FastCall:
9518 case CallingConv::X86_ThisCall:
9519 case CallingConv::Fast:
9520 // Pass 'nest' parameter in EAX.
9521 // Must be kept in sync with X86CallingConv.td
9526 SDValue OutChains[4];
9529 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9530 DAG.getConstant(10, MVT::i32));
9531 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
9533 // This is storing the opcode for MOV32ri.
9534 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
9535 const unsigned char N86Reg = X86_MC::getX86RegNum(NestReg);
9536 OutChains[0] = DAG.getStore(Root, dl,
9537 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
9538 Trmp, MachinePointerInfo(TrmpAddr),
9541 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9542 DAG.getConstant(1, MVT::i32));
9543 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
9544 MachinePointerInfo(TrmpAddr, 1),
9547 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
9548 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9549 DAG.getConstant(5, MVT::i32));
9550 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
9551 MachinePointerInfo(TrmpAddr, 5),
9554 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
9555 DAG.getConstant(6, MVT::i32));
9556 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
9557 MachinePointerInfo(TrmpAddr, 6),
9560 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4);
9564 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
9565 SelectionDAG &DAG) const {
9567 The rounding mode is in bits 11:10 of FPSR, and has the following
9574 FLT_ROUNDS, on the other hand, expects the following:
9581 To perform the conversion, we do:
9582 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
9585 MachineFunction &MF = DAG.getMachineFunction();
9586 const TargetMachine &TM = MF.getTarget();
9587 const TargetFrameLowering &TFI = *TM.getFrameLowering();
9588 unsigned StackAlignment = TFI.getStackAlignment();
9589 EVT VT = Op.getValueType();
9590 DebugLoc DL = Op.getDebugLoc();
9592 // Save FP Control Word to stack slot
9593 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
9594 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
9597 MachineMemOperand *MMO =
9598 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
9599 MachineMemOperand::MOStore, 2, 2);
9601 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
9602 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
9603 DAG.getVTList(MVT::Other),
9604 Ops, 2, MVT::i16, MMO);
9606 // Load FP Control Word from stack slot
9607 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
9608 MachinePointerInfo(), false, false, 0);
9610 // Transform as necessary
9612 DAG.getNode(ISD::SRL, DL, MVT::i16,
9613 DAG.getNode(ISD::AND, DL, MVT::i16,
9614 CWD, DAG.getConstant(0x800, MVT::i16)),
9615 DAG.getConstant(11, MVT::i8));
9617 DAG.getNode(ISD::SRL, DL, MVT::i16,
9618 DAG.getNode(ISD::AND, DL, MVT::i16,
9619 CWD, DAG.getConstant(0x400, MVT::i16)),
9620 DAG.getConstant(9, MVT::i8));
9623 DAG.getNode(ISD::AND, DL, MVT::i16,
9624 DAG.getNode(ISD::ADD, DL, MVT::i16,
9625 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
9626 DAG.getConstant(1, MVT::i16)),
9627 DAG.getConstant(3, MVT::i16));
9630 return DAG.getNode((VT.getSizeInBits() < 16 ?
9631 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
9634 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) const {
9635 EVT VT = Op.getValueType();
9637 unsigned NumBits = VT.getSizeInBits();
9638 DebugLoc dl = Op.getDebugLoc();
9640 Op = Op.getOperand(0);
9641 if (VT == MVT::i8) {
9642 // Zero extend to i32 since there is not an i8 bsr.
9644 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
9647 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
9648 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
9649 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
9651 // If src is zero (i.e. bsr sets ZF), returns NumBits.
9654 DAG.getConstant(NumBits+NumBits-1, OpVT),
9655 DAG.getConstant(X86::COND_E, MVT::i8),
9658 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
9660 // Finally xor with NumBits-1.
9661 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
9664 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
9668 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) const {
9669 EVT VT = Op.getValueType();
9671 unsigned NumBits = VT.getSizeInBits();
9672 DebugLoc dl = Op.getDebugLoc();
9674 Op = Op.getOperand(0);
9675 if (VT == MVT::i8) {
9677 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
9680 // Issue a bsf (scan bits forward) which also sets EFLAGS.
9681 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
9682 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
9684 // If src is zero (i.e. bsf sets ZF), returns NumBits.
9687 DAG.getConstant(NumBits, OpVT),
9688 DAG.getConstant(X86::COND_E, MVT::i8),
9691 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
9694 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
9698 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
9699 // ones, and then concatenate the result back.
9700 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
9701 EVT VT = Op.getValueType();
9703 assert(VT.getSizeInBits() == 256 && VT.isInteger() &&
9704 "Unsupported value type for operation");
9706 int NumElems = VT.getVectorNumElements();
9707 DebugLoc dl = Op.getDebugLoc();
9708 SDValue Idx0 = DAG.getConstant(0, MVT::i32);
9709 SDValue Idx1 = DAG.getConstant(NumElems/2, MVT::i32);
9711 // Extract the LHS vectors
9712 SDValue LHS = Op.getOperand(0);
9713 SDValue LHS1 = Extract128BitVector(LHS, Idx0, DAG, dl);
9714 SDValue LHS2 = Extract128BitVector(LHS, Idx1, DAG, dl);
9716 // Extract the RHS vectors
9717 SDValue RHS = Op.getOperand(1);
9718 SDValue RHS1 = Extract128BitVector(RHS, Idx0, DAG, dl);
9719 SDValue RHS2 = Extract128BitVector(RHS, Idx1, DAG, dl);
9721 MVT EltVT = VT.getVectorElementType().getSimpleVT();
9722 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
9724 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
9725 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
9726 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
9729 SDValue X86TargetLowering::LowerADD(SDValue Op, SelectionDAG &DAG) const {
9730 assert(Op.getValueType().getSizeInBits() == 256 &&
9731 Op.getValueType().isInteger() &&
9732 "Only handle AVX 256-bit vector integer operation");
9733 return Lower256IntArith(Op, DAG);
9736 SDValue X86TargetLowering::LowerSUB(SDValue Op, SelectionDAG &DAG) const {
9737 assert(Op.getValueType().getSizeInBits() == 256 &&
9738 Op.getValueType().isInteger() &&
9739 "Only handle AVX 256-bit vector integer operation");
9740 return Lower256IntArith(Op, DAG);
9743 SDValue X86TargetLowering::LowerMUL(SDValue Op, SelectionDAG &DAG) const {
9744 EVT VT = Op.getValueType();
9746 // Decompose 256-bit ops into smaller 128-bit ops.
9747 if (VT.getSizeInBits() == 256)
9748 return Lower256IntArith(Op, DAG);
9750 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
9751 DebugLoc dl = Op.getDebugLoc();
9753 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
9754 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
9755 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
9756 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
9757 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
9759 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
9760 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
9761 // return AloBlo + AloBhi + AhiBlo;
9763 SDValue A = Op.getOperand(0);
9764 SDValue B = Op.getOperand(1);
9766 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9767 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9768 A, DAG.getConstant(32, MVT::i32));
9769 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9770 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9771 B, DAG.getConstant(32, MVT::i32));
9772 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9773 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9775 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9776 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9778 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9779 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
9781 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9782 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9783 AloBhi, DAG.getConstant(32, MVT::i32));
9784 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9785 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9786 AhiBlo, DAG.getConstant(32, MVT::i32));
9787 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
9788 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
9792 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
9794 EVT VT = Op.getValueType();
9795 DebugLoc dl = Op.getDebugLoc();
9796 SDValue R = Op.getOperand(0);
9797 SDValue Amt = Op.getOperand(1);
9798 LLVMContext *Context = DAG.getContext();
9800 if (!(Subtarget->hasSSE2() || Subtarget->hasAVX()))
9803 // Decompose 256-bit shifts into smaller 128-bit shifts.
9804 if (VT.getSizeInBits() == 256) {
9805 int NumElems = VT.getVectorNumElements();
9806 MVT EltVT = VT.getVectorElementType().getSimpleVT();
9807 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
9809 // Extract the two vectors
9810 SDValue V1 = Extract128BitVector(R, DAG.getConstant(0, MVT::i32), DAG, dl);
9811 SDValue V2 = Extract128BitVector(R, DAG.getConstant(NumElems/2, MVT::i32),
9814 // Recreate the shift amount vectors
9816 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
9817 // Constant shift amount
9818 SmallVector<SDValue, 4> Amt1Csts;
9819 SmallVector<SDValue, 4> Amt2Csts;
9820 for (int i = 0; i < NumElems/2; ++i)
9821 Amt1Csts.push_back(Amt->getOperand(i));
9822 for (int i = NumElems/2; i < NumElems; ++i)
9823 Amt2Csts.push_back(Amt->getOperand(i));
9825 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
9826 &Amt1Csts[0], NumElems/2);
9827 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT,
9828 &Amt2Csts[0], NumElems/2);
9830 // Variable shift amount
9831 Amt1 = Extract128BitVector(Amt, DAG.getConstant(0, MVT::i32), DAG, dl);
9832 Amt2 = Extract128BitVector(Amt, DAG.getConstant(NumElems/2, MVT::i32),
9836 // Issue new vector shifts for the smaller types
9837 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
9838 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
9840 // Concatenate the result back
9841 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
9844 // Optimize shl/srl/sra with constant shift amount.
9845 if (isSplatVector(Amt.getNode())) {
9846 SDValue SclrAmt = Amt->getOperand(0);
9847 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(SclrAmt)) {
9848 uint64_t ShiftAmt = C->getZExtValue();
9850 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SHL)
9851 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9852 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9853 R, DAG.getConstant(ShiftAmt, MVT::i32));
9855 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SHL)
9856 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9857 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9858 R, DAG.getConstant(ShiftAmt, MVT::i32));
9860 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SHL)
9861 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9862 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9863 R, DAG.getConstant(ShiftAmt, MVT::i32));
9865 if (VT == MVT::v2i64 && Op.getOpcode() == ISD::SRL)
9866 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9867 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9868 R, DAG.getConstant(ShiftAmt, MVT::i32));
9870 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRL)
9871 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9872 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9873 R, DAG.getConstant(ShiftAmt, MVT::i32));
9875 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRL)
9876 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9877 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9878 R, DAG.getConstant(ShiftAmt, MVT::i32));
9880 if (VT == MVT::v4i32 && Op.getOpcode() == ISD::SRA)
9881 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9882 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9883 R, DAG.getConstant(ShiftAmt, MVT::i32));
9885 if (VT == MVT::v8i16 && Op.getOpcode() == ISD::SRA)
9886 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9887 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9888 R, DAG.getConstant(ShiftAmt, MVT::i32));
9892 // Lower SHL with variable shift amount.
9893 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
9894 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9895 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9896 Op.getOperand(1), DAG.getConstant(23, MVT::i32));
9898 ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
9900 std::vector<Constant*> CV(4, CI);
9901 Constant *C = ConstantVector::get(CV);
9902 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9903 SDValue Addend = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9904 MachinePointerInfo::getConstantPool(),
9907 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
9908 Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
9909 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
9910 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
9912 if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
9914 Op = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9915 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9916 Op.getOperand(1), DAG.getConstant(5, MVT::i32));
9918 ConstantInt *CM1 = ConstantInt::get(*Context, APInt(8, 15));
9919 ConstantInt *CM2 = ConstantInt::get(*Context, APInt(8, 63));
9921 std::vector<Constant*> CVM1(16, CM1);
9922 std::vector<Constant*> CVM2(16, CM2);
9923 Constant *C = ConstantVector::get(CVM1);
9924 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9925 SDValue M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9926 MachinePointerInfo::getConstantPool(),
9929 // r = pblendv(r, psllw(r & (char16)15, 4), a);
9930 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9931 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9932 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9933 DAG.getConstant(4, MVT::i32));
9934 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9936 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9938 C = ConstantVector::get(CVM2);
9939 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
9940 M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
9941 MachinePointerInfo::getConstantPool(),
9944 // r = pblendv(r, psllw(r & (char16)63, 2), a);
9945 M = DAG.getNode(ISD::AND, dl, VT, R, M);
9946 M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
9947 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
9948 DAG.getConstant(2, MVT::i32));
9949 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
9951 Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
9953 // return pblendv(r, r+r, a);
9954 R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
9955 R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
9961 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) const {
9962 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
9963 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
9964 // looks for this combo and may remove the "setcc" instruction if the "setcc"
9965 // has only one use.
9966 SDNode *N = Op.getNode();
9967 SDValue LHS = N->getOperand(0);
9968 SDValue RHS = N->getOperand(1);
9969 unsigned BaseOp = 0;
9971 DebugLoc DL = Op.getDebugLoc();
9972 switch (Op.getOpcode()) {
9973 default: llvm_unreachable("Unknown ovf instruction!");
9975 // A subtract of one will be selected as a INC. Note that INC doesn't
9976 // set CF, so we can't do this for UADDO.
9977 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9979 BaseOp = X86ISD::INC;
9983 BaseOp = X86ISD::ADD;
9987 BaseOp = X86ISD::ADD;
9991 // A subtract of one will be selected as a DEC. Note that DEC doesn't
9992 // set CF, so we can't do this for USUBO.
9993 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
9995 BaseOp = X86ISD::DEC;
9999 BaseOp = X86ISD::SUB;
10000 Cond = X86::COND_O;
10003 BaseOp = X86ISD::SUB;
10004 Cond = X86::COND_B;
10007 BaseOp = X86ISD::SMUL;
10008 Cond = X86::COND_O;
10010 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
10011 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
10013 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
10016 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
10017 DAG.getConstant(X86::COND_O, MVT::i32),
10018 SDValue(Sum.getNode(), 2));
10020 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
10024 // Also sets EFLAGS.
10025 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
10026 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
10029 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
10030 DAG.getConstant(Cond, MVT::i32),
10031 SDValue(Sum.getNode(), 1));
10033 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
10036 SDValue X86TargetLowering::LowerSIGN_EXTEND_INREG(SDValue Op, SelectionDAG &DAG) const{
10037 DebugLoc dl = Op.getDebugLoc();
10038 SDNode* Node = Op.getNode();
10039 EVT ExtraVT = cast<VTSDNode>(Node->getOperand(1))->getVT();
10040 EVT VT = Node->getValueType(0);
10042 if (Subtarget->hasSSE2() && VT.isVector()) {
10043 unsigned BitsDiff = VT.getScalarType().getSizeInBits() -
10044 ExtraVT.getScalarType().getSizeInBits();
10045 SDValue ShAmt = DAG.getConstant(BitsDiff, MVT::i32);
10047 unsigned SHLIntrinsicsID = 0;
10048 unsigned SRAIntrinsicsID = 0;
10049 switch (VT.getSimpleVT().SimpleTy) {
10053 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_q;
10054 SRAIntrinsicsID = 0;
10058 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_d;
10059 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_d;
10063 SHLIntrinsicsID = Intrinsic::x86_sse2_pslli_w;
10064 SRAIntrinsicsID = Intrinsic::x86_sse2_psrai_w;
10069 SDValue Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
10070 DAG.getConstant(SHLIntrinsicsID, MVT::i32),
10071 Node->getOperand(0), ShAmt);
10073 // In case of 1 bit sext, no need to shr
10074 if (ExtraVT.getScalarType().getSizeInBits() == 1) return Tmp1;
10076 if (SRAIntrinsicsID) {
10077 Tmp1 = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
10078 DAG.getConstant(SRAIntrinsicsID, MVT::i32),
10088 SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
10089 DebugLoc dl = Op.getDebugLoc();
10091 // Go ahead and emit the fence on x86-64 even if we asked for no-sse2.
10092 // There isn't any reason to disable it if the target processor supports it.
10093 if (!Subtarget->hasSSE2() && !Subtarget->is64Bit()) {
10094 SDValue Chain = Op.getOperand(0);
10095 SDValue Zero = DAG.getConstant(0, MVT::i32);
10097 DAG.getRegister(X86::ESP, MVT::i32), // Base
10098 DAG.getTargetConstant(1, MVT::i8), // Scale
10099 DAG.getRegister(0, MVT::i32), // Index
10100 DAG.getTargetConstant(0, MVT::i32), // Disp
10101 DAG.getRegister(0, MVT::i32), // Segment.
10106 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
10107 array_lengthof(Ops));
10108 return SDValue(Res, 0);
10111 unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
10113 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
10115 unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10116 unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
10117 unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
10118 unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
10120 // def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
10121 if (!Op1 && !Op2 && !Op3 && Op4)
10122 return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
10124 // def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
10125 if (Op1 && !Op2 && !Op3 && !Op4)
10126 return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
10128 // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
10130 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
10133 SDValue X86TargetLowering::LowerATOMIC_FENCE(SDValue Op,
10134 SelectionDAG &DAG) const {
10135 DebugLoc dl = Op.getDebugLoc();
10136 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
10137 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
10138 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
10139 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
10141 // The only fence that needs an instruction is a sequentially-consistent
10142 // cross-thread fence.
10143 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
10144 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
10145 // no-sse2). There isn't any reason to disable it if the target processor
10147 if (Subtarget->hasSSE2() || Subtarget->is64Bit())
10148 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
10150 SDValue Chain = Op.getOperand(0);
10151 SDValue Zero = DAG.getConstant(0, MVT::i32);
10153 DAG.getRegister(X86::ESP, MVT::i32), // Base
10154 DAG.getTargetConstant(1, MVT::i8), // Scale
10155 DAG.getRegister(0, MVT::i32), // Index
10156 DAG.getTargetConstant(0, MVT::i32), // Disp
10157 DAG.getRegister(0, MVT::i32), // Segment.
10162 DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
10163 array_lengthof(Ops));
10164 return SDValue(Res, 0);
10167 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
10168 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
10172 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) const {
10173 EVT T = Op.getValueType();
10174 DebugLoc DL = Op.getDebugLoc();
10177 switch(T.getSimpleVT().SimpleTy) {
10179 assert(false && "Invalid value type!");
10180 case MVT::i8: Reg = X86::AL; size = 1; break;
10181 case MVT::i16: Reg = X86::AX; size = 2; break;
10182 case MVT::i32: Reg = X86::EAX; size = 4; break;
10184 assert(Subtarget->is64Bit() && "Node not type legal!");
10185 Reg = X86::RAX; size = 8;
10188 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
10189 Op.getOperand(2), SDValue());
10190 SDValue Ops[] = { cpIn.getValue(0),
10193 DAG.getTargetConstant(size, MVT::i8),
10194 cpIn.getValue(1) };
10195 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
10196 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
10197 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
10200 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
10204 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
10205 SelectionDAG &DAG) const {
10206 assert(Subtarget->is64Bit() && "Result not type legalized?");
10207 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
10208 SDValue TheChain = Op.getOperand(0);
10209 DebugLoc dl = Op.getDebugLoc();
10210 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
10211 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
10212 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
10214 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
10215 DAG.getConstant(32, MVT::i8));
10217 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
10220 return DAG.getMergeValues(Ops, 2, dl);
10223 SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
10224 SelectionDAG &DAG) const {
10225 EVT SrcVT = Op.getOperand(0).getValueType();
10226 EVT DstVT = Op.getValueType();
10227 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
10228 Subtarget->hasMMX() && "Unexpected custom BITCAST");
10229 assert((DstVT == MVT::i64 ||
10230 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
10231 "Unexpected custom BITCAST");
10232 // i64 <=> MMX conversions are Legal.
10233 if (SrcVT==MVT::i64 && DstVT.isVector())
10235 if (DstVT==MVT::i64 && SrcVT.isVector())
10237 // MMX <=> MMX conversions are Legal.
10238 if (SrcVT.isVector() && DstVT.isVector())
10240 // All other conversions need to be expanded.
10244 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
10245 SDNode *Node = Op.getNode();
10246 DebugLoc dl = Node->getDebugLoc();
10247 EVT T = Node->getValueType(0);
10248 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
10249 DAG.getConstant(0, T), Node->getOperand(2));
10250 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
10251 cast<AtomicSDNode>(Node)->getMemoryVT(),
10252 Node->getOperand(0),
10253 Node->getOperand(1), negOp,
10254 cast<AtomicSDNode>(Node)->getSrcValue(),
10255 cast<AtomicSDNode>(Node)->getAlignment(),
10256 cast<AtomicSDNode>(Node)->getOrdering(),
10257 cast<AtomicSDNode>(Node)->getSynchScope());
10260 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
10261 SDNode *Node = Op.getNode();
10262 DebugLoc dl = Node->getDebugLoc();
10263 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
10265 // Convert seq_cst store -> xchg
10266 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
10267 // FIXME: On 32-bit, store -> fist or movq would be more efficient
10268 // (The only way to get a 16-byte store is cmpxchg16b)
10269 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
10270 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
10271 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
10272 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
10273 cast<AtomicSDNode>(Node)->getMemoryVT(),
10274 Node->getOperand(0),
10275 Node->getOperand(1), Node->getOperand(2),
10276 cast<AtomicSDNode>(Node)->getMemOperand(),
10277 cast<AtomicSDNode>(Node)->getOrdering(),
10278 cast<AtomicSDNode>(Node)->getSynchScope());
10279 return Swap.getValue(1);
10281 // Other atomic stores have a simple pattern.
10285 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
10286 EVT VT = Op.getNode()->getValueType(0);
10288 // Let legalize expand this if it isn't a legal type yet.
10289 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
10292 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
10295 bool ExtraOp = false;
10296 switch (Op.getOpcode()) {
10297 default: assert(0 && "Invalid code");
10298 case ISD::ADDC: Opc = X86ISD::ADD; break;
10299 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
10300 case ISD::SUBC: Opc = X86ISD::SUB; break;
10301 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
10305 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
10307 return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
10308 Op.getOperand(1), Op.getOperand(2));
10311 /// LowerOperation - Provide custom lowering hooks for some operations.
10313 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
10314 switch (Op.getOpcode()) {
10315 default: llvm_unreachable("Should not custom lower this!");
10316 case ISD::SIGN_EXTEND_INREG: return LowerSIGN_EXTEND_INREG(Op,DAG);
10317 case ISD::MEMBARRIER: return LowerMEMBARRIER(Op,DAG);
10318 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op,DAG);
10319 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
10320 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
10321 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
10322 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
10323 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
10324 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
10325 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
10326 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
10327 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
10328 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, DAG);
10329 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
10330 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
10331 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
10332 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
10333 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
10334 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
10335 case ISD::SHL_PARTS:
10336 case ISD::SRA_PARTS:
10337 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
10338 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
10339 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
10340 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
10341 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
10342 case ISD::FABS: return LowerFABS(Op, DAG);
10343 case ISD::FNEG: return LowerFNEG(Op, DAG);
10344 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
10345 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
10346 case ISD::SETCC: return LowerSETCC(Op, DAG);
10347 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
10348 case ISD::SELECT: return LowerSELECT(Op, DAG);
10349 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
10350 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
10351 case ISD::VASTART: return LowerVASTART(Op, DAG);
10352 case ISD::VAARG: return LowerVAARG(Op, DAG);
10353 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
10354 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
10355 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
10356 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
10357 case ISD::FRAME_TO_ARGS_OFFSET:
10358 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
10359 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
10360 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
10361 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
10362 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
10363 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
10364 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
10365 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
10366 case ISD::MUL: return LowerMUL(Op, DAG);
10369 case ISD::SHL: return LowerShift(Op, DAG);
10375 case ISD::UMULO: return LowerXALUO(Op, DAG);
10376 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
10377 case ISD::BITCAST: return LowerBITCAST(Op, DAG);
10381 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
10382 case ISD::ADD: return LowerADD(Op, DAG);
10383 case ISD::SUB: return LowerSUB(Op, DAG);
10387 static void ReplaceATOMIC_LOAD(SDNode *Node,
10388 SmallVectorImpl<SDValue> &Results,
10389 SelectionDAG &DAG) {
10390 DebugLoc dl = Node->getDebugLoc();
10391 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
10393 // Convert wide load -> cmpxchg8b/cmpxchg16b
10394 // FIXME: On 32-bit, load -> fild or movq would be more efficient
10395 // (The only way to get a 16-byte load is cmpxchg16b)
10396 // FIXME: 16-byte ATOMIC_CMP_SWAP isn't actually hooked up at the moment.
10397 SDValue Zero = DAG.getConstant(0, VT);
10398 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_CMP_SWAP, dl, VT,
10399 Node->getOperand(0),
10400 Node->getOperand(1), Zero, Zero,
10401 cast<AtomicSDNode>(Node)->getMemOperand(),
10402 cast<AtomicSDNode>(Node)->getOrdering(),
10403 cast<AtomicSDNode>(Node)->getSynchScope());
10404 Results.push_back(Swap.getValue(0));
10405 Results.push_back(Swap.getValue(1));
10408 void X86TargetLowering::
10409 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
10410 SelectionDAG &DAG, unsigned NewOp) const {
10411 EVT T = Node->getValueType(0);
10412 DebugLoc dl = Node->getDebugLoc();
10413 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
10415 SDValue Chain = Node->getOperand(0);
10416 SDValue In1 = Node->getOperand(1);
10417 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
10418 Node->getOperand(2), DAG.getIntPtrConstant(0));
10419 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
10420 Node->getOperand(2), DAG.getIntPtrConstant(1));
10421 SDValue Ops[] = { Chain, In1, In2L, In2H };
10422 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
10424 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
10425 cast<MemSDNode>(Node)->getMemOperand());
10426 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
10427 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
10428 Results.push_back(Result.getValue(2));
10431 /// ReplaceNodeResults - Replace a node with an illegal result type
10432 /// with a new node built out of custom code.
10433 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
10434 SmallVectorImpl<SDValue>&Results,
10435 SelectionDAG &DAG) const {
10436 DebugLoc dl = N->getDebugLoc();
10437 switch (N->getOpcode()) {
10439 assert(false && "Do not know how to custom type legalize this operation!");
10441 case ISD::SIGN_EXTEND_INREG:
10446 // We don't want to expand or promote these.
10448 case ISD::FP_TO_SINT: {
10449 std::pair<SDValue,SDValue> Vals =
10450 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
10451 SDValue FIST = Vals.first, StackSlot = Vals.second;
10452 if (FIST.getNode() != 0) {
10453 EVT VT = N->getValueType(0);
10454 // Return a load from the stack slot.
10455 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
10456 MachinePointerInfo(), false, false, 0));
10460 case ISD::READCYCLECOUNTER: {
10461 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
10462 SDValue TheChain = N->getOperand(0);
10463 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
10464 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
10466 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
10468 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
10469 SDValue Ops[] = { eax, edx };
10470 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
10471 Results.push_back(edx.getValue(1));
10474 case ISD::ATOMIC_CMP_SWAP: {
10475 EVT T = N->getValueType(0);
10476 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
10477 bool Regs64bit = T == MVT::i128;
10478 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
10479 SDValue cpInL, cpInH;
10480 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
10481 DAG.getConstant(0, HalfT));
10482 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
10483 DAG.getConstant(1, HalfT));
10484 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
10485 Regs64bit ? X86::RAX : X86::EAX,
10487 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
10488 Regs64bit ? X86::RDX : X86::EDX,
10489 cpInH, cpInL.getValue(1));
10490 SDValue swapInL, swapInH;
10491 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
10492 DAG.getConstant(0, HalfT));
10493 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
10494 DAG.getConstant(1, HalfT));
10495 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
10496 Regs64bit ? X86::RBX : X86::EBX,
10497 swapInL, cpInH.getValue(1));
10498 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
10499 Regs64bit ? X86::RCX : X86::ECX,
10500 swapInH, swapInL.getValue(1));
10501 SDValue Ops[] = { swapInH.getValue(0),
10503 swapInH.getValue(1) };
10504 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
10505 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
10506 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
10507 X86ISD::LCMPXCHG8_DAG;
10508 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys,
10510 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
10511 Regs64bit ? X86::RAX : X86::EAX,
10512 HalfT, Result.getValue(1));
10513 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
10514 Regs64bit ? X86::RDX : X86::EDX,
10515 HalfT, cpOutL.getValue(2));
10516 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
10517 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF, 2));
10518 Results.push_back(cpOutH.getValue(1));
10521 case ISD::ATOMIC_LOAD_ADD:
10522 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
10524 case ISD::ATOMIC_LOAD_AND:
10525 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
10527 case ISD::ATOMIC_LOAD_NAND:
10528 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
10530 case ISD::ATOMIC_LOAD_OR:
10531 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
10533 case ISD::ATOMIC_LOAD_SUB:
10534 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
10536 case ISD::ATOMIC_LOAD_XOR:
10537 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
10539 case ISD::ATOMIC_SWAP:
10540 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
10542 case ISD::ATOMIC_LOAD:
10543 ReplaceATOMIC_LOAD(N, Results, DAG);
10547 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
10549 default: return NULL;
10550 case X86ISD::BSF: return "X86ISD::BSF";
10551 case X86ISD::BSR: return "X86ISD::BSR";
10552 case X86ISD::SHLD: return "X86ISD::SHLD";
10553 case X86ISD::SHRD: return "X86ISD::SHRD";
10554 case X86ISD::FAND: return "X86ISD::FAND";
10555 case X86ISD::FOR: return "X86ISD::FOR";
10556 case X86ISD::FXOR: return "X86ISD::FXOR";
10557 case X86ISD::FSRL: return "X86ISD::FSRL";
10558 case X86ISD::FILD: return "X86ISD::FILD";
10559 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
10560 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
10561 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
10562 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
10563 case X86ISD::FLD: return "X86ISD::FLD";
10564 case X86ISD::FST: return "X86ISD::FST";
10565 case X86ISD::CALL: return "X86ISD::CALL";
10566 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
10567 case X86ISD::BT: return "X86ISD::BT";
10568 case X86ISD::CMP: return "X86ISD::CMP";
10569 case X86ISD::COMI: return "X86ISD::COMI";
10570 case X86ISD::UCOMI: return "X86ISD::UCOMI";
10571 case X86ISD::SETCC: return "X86ISD::SETCC";
10572 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
10573 case X86ISD::FSETCCsd: return "X86ISD::FSETCCsd";
10574 case X86ISD::FSETCCss: return "X86ISD::FSETCCss";
10575 case X86ISD::CMOV: return "X86ISD::CMOV";
10576 case X86ISD::BRCOND: return "X86ISD::BRCOND";
10577 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
10578 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
10579 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
10580 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
10581 case X86ISD::Wrapper: return "X86ISD::Wrapper";
10582 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
10583 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
10584 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
10585 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
10586 case X86ISD::PINSRB: return "X86ISD::PINSRB";
10587 case X86ISD::PINSRW: return "X86ISD::PINSRW";
10588 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
10589 case X86ISD::ANDNP: return "X86ISD::ANDNP";
10590 case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
10591 case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
10592 case X86ISD::PSIGND: return "X86ISD::PSIGND";
10593 case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
10594 case X86ISD::FMAX: return "X86ISD::FMAX";
10595 case X86ISD::FMIN: return "X86ISD::FMIN";
10596 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
10597 case X86ISD::FRCP: return "X86ISD::FRCP";
10598 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
10599 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
10600 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
10601 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
10602 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
10603 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
10604 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
10605 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
10606 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
10607 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
10608 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
10609 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
10610 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
10611 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
10612 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
10613 case X86ISD::VSHL: return "X86ISD::VSHL";
10614 case X86ISD::VSRL: return "X86ISD::VSRL";
10615 case X86ISD::CMPPD: return "X86ISD::CMPPD";
10616 case X86ISD::CMPPS: return "X86ISD::CMPPS";
10617 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
10618 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
10619 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
10620 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
10621 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
10622 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
10623 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
10624 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
10625 case X86ISD::ADD: return "X86ISD::ADD";
10626 case X86ISD::SUB: return "X86ISD::SUB";
10627 case X86ISD::ADC: return "X86ISD::ADC";
10628 case X86ISD::SBB: return "X86ISD::SBB";
10629 case X86ISD::SMUL: return "X86ISD::SMUL";
10630 case X86ISD::UMUL: return "X86ISD::UMUL";
10631 case X86ISD::INC: return "X86ISD::INC";
10632 case X86ISD::DEC: return "X86ISD::DEC";
10633 case X86ISD::OR: return "X86ISD::OR";
10634 case X86ISD::XOR: return "X86ISD::XOR";
10635 case X86ISD::AND: return "X86ISD::AND";
10636 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
10637 case X86ISD::PTEST: return "X86ISD::PTEST";
10638 case X86ISD::TESTP: return "X86ISD::TESTP";
10639 case X86ISD::PALIGN: return "X86ISD::PALIGN";
10640 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
10641 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
10642 case X86ISD::PSHUFHW_LD: return "X86ISD::PSHUFHW_LD";
10643 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
10644 case X86ISD::PSHUFLW_LD: return "X86ISD::PSHUFLW_LD";
10645 case X86ISD::SHUFPS: return "X86ISD::SHUFPS";
10646 case X86ISD::SHUFPD: return "X86ISD::SHUFPD";
10647 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
10648 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
10649 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
10650 case X86ISD::MOVHLPD: return "X86ISD::MOVHLPD";
10651 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
10652 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
10653 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
10654 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
10655 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
10656 case X86ISD::MOVSHDUP_LD: return "X86ISD::MOVSHDUP_LD";
10657 case X86ISD::MOVSLDUP_LD: return "X86ISD::MOVSLDUP_LD";
10658 case X86ISD::MOVSD: return "X86ISD::MOVSD";
10659 case X86ISD::MOVSS: return "X86ISD::MOVSS";
10660 case X86ISD::UNPCKLPS: return "X86ISD::UNPCKLPS";
10661 case X86ISD::UNPCKLPD: return "X86ISD::UNPCKLPD";
10662 case X86ISD::VUNPCKLPDY: return "X86ISD::VUNPCKLPDY";
10663 case X86ISD::UNPCKHPS: return "X86ISD::UNPCKHPS";
10664 case X86ISD::UNPCKHPD: return "X86ISD::UNPCKHPD";
10665 case X86ISD::PUNPCKLBW: return "X86ISD::PUNPCKLBW";
10666 case X86ISD::PUNPCKLWD: return "X86ISD::PUNPCKLWD";
10667 case X86ISD::PUNPCKLDQ: return "X86ISD::PUNPCKLDQ";
10668 case X86ISD::PUNPCKLQDQ: return "X86ISD::PUNPCKLQDQ";
10669 case X86ISD::PUNPCKHBW: return "X86ISD::PUNPCKHBW";
10670 case X86ISD::PUNPCKHWD: return "X86ISD::PUNPCKHWD";
10671 case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
10672 case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
10673 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
10674 case X86ISD::VPERMILPS: return "X86ISD::VPERMILPS";
10675 case X86ISD::VPERMILPSY: return "X86ISD::VPERMILPSY";
10676 case X86ISD::VPERMILPD: return "X86ISD::VPERMILPD";
10677 case X86ISD::VPERMILPDY: return "X86ISD::VPERMILPDY";
10678 case X86ISD::VPERM2F128: return "X86ISD::VPERM2F128";
10679 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
10680 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
10681 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
10682 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
10683 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
10687 // isLegalAddressingMode - Return true if the addressing mode represented
10688 // by AM is legal for this target, for a load/store of the specified type.
10689 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
10691 // X86 supports extremely general addressing modes.
10692 CodeModel::Model M = getTargetMachine().getCodeModel();
10693 Reloc::Model R = getTargetMachine().getRelocationModel();
10695 // X86 allows a sign-extended 32-bit immediate field as a displacement.
10696 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
10701 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
10703 // If a reference to this global requires an extra load, we can't fold it.
10704 if (isGlobalStubReference(GVFlags))
10707 // If BaseGV requires a register for the PIC base, we cannot also have a
10708 // BaseReg specified.
10709 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
10712 // If lower 4G is not available, then we must use rip-relative addressing.
10713 if ((M != CodeModel::Small || R != Reloc::Static) &&
10714 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
10718 switch (AM.Scale) {
10724 // These scales always work.
10729 // These scales are formed with basereg+scalereg. Only accept if there is
10734 default: // Other stuff never works.
10742 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
10743 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
10745 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
10746 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
10747 if (NumBits1 <= NumBits2)
10752 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
10753 if (!VT1.isInteger() || !VT2.isInteger())
10755 unsigned NumBits1 = VT1.getSizeInBits();
10756 unsigned NumBits2 = VT2.getSizeInBits();
10757 if (NumBits1 <= NumBits2)
10762 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
10763 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
10764 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
10767 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
10768 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
10769 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
10772 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
10773 // i16 instructions are longer (0x66 prefix) and potentially slower.
10774 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
10777 /// isShuffleMaskLegal - Targets can use this to indicate that they only
10778 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
10779 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
10780 /// are assumed to be legal.
10782 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
10784 // Very little shuffling can be done for 64-bit vectors right now.
10785 if (VT.getSizeInBits() == 64)
10786 return isPALIGNRMask(M, VT, Subtarget->hasSSSE3());
10788 // FIXME: pshufb, blends, shifts.
10789 return (VT.getVectorNumElements() == 2 ||
10790 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
10791 isMOVLMask(M, VT) ||
10792 isSHUFPMask(M, VT) ||
10793 isPSHUFDMask(M, VT) ||
10794 isPSHUFHWMask(M, VT) ||
10795 isPSHUFLWMask(M, VT) ||
10796 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
10797 isUNPCKLMask(M, VT) ||
10798 isUNPCKHMask(M, VT) ||
10799 isUNPCKL_v_undef_Mask(M, VT) ||
10800 isUNPCKH_v_undef_Mask(M, VT));
10804 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
10806 unsigned NumElts = VT.getVectorNumElements();
10807 // FIXME: This collection of masks seems suspect.
10810 if (NumElts == 4 && VT.getSizeInBits() == 128) {
10811 return (isMOVLMask(Mask, VT) ||
10812 isCommutedMOVLMask(Mask, VT, true) ||
10813 isSHUFPMask(Mask, VT) ||
10814 isCommutedSHUFPMask(Mask, VT));
10819 //===----------------------------------------------------------------------===//
10820 // X86 Scheduler Hooks
10821 //===----------------------------------------------------------------------===//
10823 // private utility function
10824 MachineBasicBlock *
10825 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
10826 MachineBasicBlock *MBB,
10833 TargetRegisterClass *RC,
10834 bool invSrc) const {
10835 // For the atomic bitwise operator, we generate
10838 // ld t1 = [bitinstr.addr]
10839 // op t2 = t1, [bitinstr.val]
10841 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
10843 // fallthrough -->nextMBB
10844 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10845 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10846 MachineFunction::iterator MBBIter = MBB;
10849 /// First build the CFG
10850 MachineFunction *F = MBB->getParent();
10851 MachineBasicBlock *thisMBB = MBB;
10852 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10853 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10854 F->insert(MBBIter, newMBB);
10855 F->insert(MBBIter, nextMBB);
10857 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10858 nextMBB->splice(nextMBB->begin(), thisMBB,
10859 llvm::next(MachineBasicBlock::iterator(bInstr)),
10861 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10863 // Update thisMBB to fall through to newMBB
10864 thisMBB->addSuccessor(newMBB);
10866 // newMBB jumps to itself and fall through to nextMBB
10867 newMBB->addSuccessor(nextMBB);
10868 newMBB->addSuccessor(newMBB);
10870 // Insert instructions into newMBB based on incoming instruction
10871 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
10872 "unexpected number of operands");
10873 DebugLoc dl = bInstr->getDebugLoc();
10874 MachineOperand& destOper = bInstr->getOperand(0);
10875 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10876 int numArgs = bInstr->getNumOperands() - 1;
10877 for (int i=0; i < numArgs; ++i)
10878 argOpers[i] = &bInstr->getOperand(i+1);
10880 // x86 address has 4 operands: base, index, scale, and displacement
10881 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
10882 int valArgIndx = lastAddrIndx + 1;
10884 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
10885 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
10886 for (int i=0; i <= lastAddrIndx; ++i)
10887 (*MIB).addOperand(*argOpers[i]);
10889 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
10891 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
10896 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
10897 assert((argOpers[valArgIndx]->isReg() ||
10898 argOpers[valArgIndx]->isImm()) &&
10899 "invalid operand");
10900 if (argOpers[valArgIndx]->isReg())
10901 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
10903 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
10905 (*MIB).addOperand(*argOpers[valArgIndx]);
10907 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), EAXreg);
10910 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
10911 for (int i=0; i <= lastAddrIndx; ++i)
10912 (*MIB).addOperand(*argOpers[i]);
10914 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
10915 (*MIB).setMemRefs(bInstr->memoperands_begin(),
10916 bInstr->memoperands_end());
10918 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
10919 MIB.addReg(EAXreg);
10922 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
10924 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
10928 // private utility function: 64 bit atomics on 32 bit host.
10929 MachineBasicBlock *
10930 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
10931 MachineBasicBlock *MBB,
10936 bool invSrc) const {
10937 // For the atomic bitwise operator, we generate
10938 // thisMBB (instructions are in pairs, except cmpxchg8b)
10939 // ld t1,t2 = [bitinstr.addr]
10941 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
10942 // op t5, t6 <- out1, out2, [bitinstr.val]
10943 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
10944 // mov ECX, EBX <- t5, t6
10945 // mov EAX, EDX <- t1, t2
10946 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
10947 // mov t3, t4 <- EAX, EDX
10949 // result in out1, out2
10950 // fallthrough -->nextMBB
10952 const TargetRegisterClass *RC = X86::GR32RegisterClass;
10953 const unsigned LoadOpc = X86::MOV32rm;
10954 const unsigned NotOpc = X86::NOT32r;
10955 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
10956 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
10957 MachineFunction::iterator MBBIter = MBB;
10960 /// First build the CFG
10961 MachineFunction *F = MBB->getParent();
10962 MachineBasicBlock *thisMBB = MBB;
10963 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
10964 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
10965 F->insert(MBBIter, newMBB);
10966 F->insert(MBBIter, nextMBB);
10968 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
10969 nextMBB->splice(nextMBB->begin(), thisMBB,
10970 llvm::next(MachineBasicBlock::iterator(bInstr)),
10972 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
10974 // Update thisMBB to fall through to newMBB
10975 thisMBB->addSuccessor(newMBB);
10977 // newMBB jumps to itself and fall through to nextMBB
10978 newMBB->addSuccessor(nextMBB);
10979 newMBB->addSuccessor(newMBB);
10981 DebugLoc dl = bInstr->getDebugLoc();
10982 // Insert instructions into newMBB based on incoming instruction
10983 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
10984 assert(bInstr->getNumOperands() < X86::AddrNumOperands + 14 &&
10985 "unexpected number of operands");
10986 MachineOperand& dest1Oper = bInstr->getOperand(0);
10987 MachineOperand& dest2Oper = bInstr->getOperand(1);
10988 MachineOperand* argOpers[2 + X86::AddrNumOperands];
10989 for (int i=0; i < 2 + X86::AddrNumOperands; ++i) {
10990 argOpers[i] = &bInstr->getOperand(i+2);
10992 // We use some of the operands multiple times, so conservatively just
10993 // clear any kill flags that might be present.
10994 if (argOpers[i]->isReg() && argOpers[i]->isUse())
10995 argOpers[i]->setIsKill(false);
10998 // x86 address has 5 operands: base, index, scale, displacement, and segment.
10999 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
11001 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
11002 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
11003 for (int i=0; i <= lastAddrIndx; ++i)
11004 (*MIB).addOperand(*argOpers[i]);
11005 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
11006 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
11007 // add 4 to displacement.
11008 for (int i=0; i <= lastAddrIndx-2; ++i)
11009 (*MIB).addOperand(*argOpers[i]);
11010 MachineOperand newOp3 = *(argOpers[3]);
11011 if (newOp3.isImm())
11012 newOp3.setImm(newOp3.getImm()+4);
11014 newOp3.setOffset(newOp3.getOffset()+4);
11015 (*MIB).addOperand(newOp3);
11016 (*MIB).addOperand(*argOpers[lastAddrIndx]);
11018 // t3/4 are defined later, at the bottom of the loop
11019 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
11020 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
11021 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
11022 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
11023 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
11024 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
11026 // The subsequent operations should be using the destination registers of
11027 //the PHI instructions.
11029 t1 = F->getRegInfo().createVirtualRegister(RC);
11030 t2 = F->getRegInfo().createVirtualRegister(RC);
11031 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
11032 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
11034 t1 = dest1Oper.getReg();
11035 t2 = dest2Oper.getReg();
11038 int valArgIndx = lastAddrIndx + 1;
11039 assert((argOpers[valArgIndx]->isReg() ||
11040 argOpers[valArgIndx]->isImm()) &&
11041 "invalid operand");
11042 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
11043 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
11044 if (argOpers[valArgIndx]->isReg())
11045 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
11047 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
11048 if (regOpcL != X86::MOV32rr)
11050 (*MIB).addOperand(*argOpers[valArgIndx]);
11051 assert(argOpers[valArgIndx + 1]->isReg() ==
11052 argOpers[valArgIndx]->isReg());
11053 assert(argOpers[valArgIndx + 1]->isImm() ==
11054 argOpers[valArgIndx]->isImm());
11055 if (argOpers[valArgIndx + 1]->isReg())
11056 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
11058 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
11059 if (regOpcH != X86::MOV32rr)
11061 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
11063 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
11065 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EDX);
11068 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EBX);
11070 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::ECX);
11073 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
11074 for (int i=0; i <= lastAddrIndx; ++i)
11075 (*MIB).addOperand(*argOpers[i]);
11077 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
11078 (*MIB).setMemRefs(bInstr->memoperands_begin(),
11079 bInstr->memoperands_end());
11081 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t3);
11082 MIB.addReg(X86::EAX);
11083 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t4);
11084 MIB.addReg(X86::EDX);
11087 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
11089 bInstr->eraseFromParent(); // The pseudo instruction is gone now.
11093 // private utility function
11094 MachineBasicBlock *
11095 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
11096 MachineBasicBlock *MBB,
11097 unsigned cmovOpc) const {
11098 // For the atomic min/max operator, we generate
11101 // ld t1 = [min/max.addr]
11102 // mov t2 = [min/max.val]
11104 // cmov[cond] t2 = t1
11106 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
11108 // fallthrough -->nextMBB
11110 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11111 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
11112 MachineFunction::iterator MBBIter = MBB;
11115 /// First build the CFG
11116 MachineFunction *F = MBB->getParent();
11117 MachineBasicBlock *thisMBB = MBB;
11118 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
11119 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
11120 F->insert(MBBIter, newMBB);
11121 F->insert(MBBIter, nextMBB);
11123 // Transfer the remainder of thisMBB and its successor edges to nextMBB.
11124 nextMBB->splice(nextMBB->begin(), thisMBB,
11125 llvm::next(MachineBasicBlock::iterator(mInstr)),
11127 nextMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
11129 // Update thisMBB to fall through to newMBB
11130 thisMBB->addSuccessor(newMBB);
11132 // newMBB jumps to newMBB and fall through to nextMBB
11133 newMBB->addSuccessor(nextMBB);
11134 newMBB->addSuccessor(newMBB);
11136 DebugLoc dl = mInstr->getDebugLoc();
11137 // Insert instructions into newMBB based on incoming instruction
11138 assert(mInstr->getNumOperands() < X86::AddrNumOperands + 4 &&
11139 "unexpected number of operands");
11140 MachineOperand& destOper = mInstr->getOperand(0);
11141 MachineOperand* argOpers[2 + X86::AddrNumOperands];
11142 int numArgs = mInstr->getNumOperands() - 1;
11143 for (int i=0; i < numArgs; ++i)
11144 argOpers[i] = &mInstr->getOperand(i+1);
11146 // x86 address has 4 operands: base, index, scale, and displacement
11147 int lastAddrIndx = X86::AddrNumOperands - 1; // [0,3]
11148 int valArgIndx = lastAddrIndx + 1;
11150 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
11151 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
11152 for (int i=0; i <= lastAddrIndx; ++i)
11153 (*MIB).addOperand(*argOpers[i]);
11155 // We only support register and immediate values
11156 assert((argOpers[valArgIndx]->isReg() ||
11157 argOpers[valArgIndx]->isImm()) &&
11158 "invalid operand");
11160 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
11161 if (argOpers[valArgIndx]->isReg())
11162 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), t2);
11164 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
11165 (*MIB).addOperand(*argOpers[valArgIndx]);
11167 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), X86::EAX);
11170 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
11175 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
11176 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
11180 // Cmp and exchange if none has modified the memory location
11181 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
11182 for (int i=0; i <= lastAddrIndx; ++i)
11183 (*MIB).addOperand(*argOpers[i]);
11185 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
11186 (*MIB).setMemRefs(mInstr->memoperands_begin(),
11187 mInstr->memoperands_end());
11189 MIB = BuildMI(newMBB, dl, TII->get(TargetOpcode::COPY), destOper.getReg());
11190 MIB.addReg(X86::EAX);
11193 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
11195 mInstr->eraseFromParent(); // The pseudo instruction is gone now.
11199 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
11200 // or XMM0_V32I8 in AVX all of this code can be replaced with that
11201 // in the .td file.
11202 MachineBasicBlock *
11203 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
11204 unsigned numArgs, bool memArg) const {
11205 assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
11206 "Target must have SSE4.2 or AVX features enabled");
11208 DebugLoc dl = MI->getDebugLoc();
11209 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11211 if (!Subtarget->hasAVX()) {
11213 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
11215 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
11218 Opc = numArgs == 3 ? X86::VPCMPISTRM128rm : X86::VPCMPESTRM128rm;
11220 Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
11223 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
11224 for (unsigned i = 0; i < numArgs; ++i) {
11225 MachineOperand &Op = MI->getOperand(i+1);
11226 if (!(Op.isReg() && Op.isImplicit()))
11227 MIB.addOperand(Op);
11229 BuildMI(*BB, MI, dl,
11230 TII->get(Subtarget->hasAVX() ? X86::VMOVAPSrr : X86::MOVAPSrr),
11231 MI->getOperand(0).getReg())
11232 .addReg(X86::XMM0);
11234 MI->eraseFromParent();
11238 MachineBasicBlock *
11239 X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
11240 DebugLoc dl = MI->getDebugLoc();
11241 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11243 // Address into RAX/EAX, other two args into ECX, EDX.
11244 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
11245 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11246 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
11247 for (int i = 0; i < X86::AddrNumOperands; ++i)
11248 MIB.addOperand(MI->getOperand(i));
11250 unsigned ValOps = X86::AddrNumOperands;
11251 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
11252 .addReg(MI->getOperand(ValOps).getReg());
11253 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
11254 .addReg(MI->getOperand(ValOps+1).getReg());
11256 // The instruction doesn't actually take any operands though.
11257 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
11259 MI->eraseFromParent(); // The pseudo is gone now.
11263 MachineBasicBlock *
11264 X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
11265 DebugLoc dl = MI->getDebugLoc();
11266 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11268 // First arg in ECX, the second in EAX.
11269 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
11270 .addReg(MI->getOperand(0).getReg());
11271 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
11272 .addReg(MI->getOperand(1).getReg());
11274 // The instruction doesn't actually take any operands though.
11275 BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
11277 MI->eraseFromParent(); // The pseudo is gone now.
11281 MachineBasicBlock *
11282 X86TargetLowering::EmitVAARG64WithCustomInserter(
11284 MachineBasicBlock *MBB) const {
11285 // Emit va_arg instruction on X86-64.
11287 // Operands to this pseudo-instruction:
11288 // 0 ) Output : destination address (reg)
11289 // 1-5) Input : va_list address (addr, i64mem)
11290 // 6 ) ArgSize : Size (in bytes) of vararg type
11291 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
11292 // 8 ) Align : Alignment of type
11293 // 9 ) EFLAGS (implicit-def)
11295 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
11296 assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
11298 unsigned DestReg = MI->getOperand(0).getReg();
11299 MachineOperand &Base = MI->getOperand(1);
11300 MachineOperand &Scale = MI->getOperand(2);
11301 MachineOperand &Index = MI->getOperand(3);
11302 MachineOperand &Disp = MI->getOperand(4);
11303 MachineOperand &Segment = MI->getOperand(5);
11304 unsigned ArgSize = MI->getOperand(6).getImm();
11305 unsigned ArgMode = MI->getOperand(7).getImm();
11306 unsigned Align = MI->getOperand(8).getImm();
11308 // Memory Reference
11309 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
11310 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
11311 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
11313 // Machine Information
11314 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11315 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
11316 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
11317 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
11318 DebugLoc DL = MI->getDebugLoc();
11320 // struct va_list {
11323 // i64 overflow_area (address)
11324 // i64 reg_save_area (address)
11326 // sizeof(va_list) = 24
11327 // alignment(va_list) = 8
11329 unsigned TotalNumIntRegs = 6;
11330 unsigned TotalNumXMMRegs = 8;
11331 bool UseGPOffset = (ArgMode == 1);
11332 bool UseFPOffset = (ArgMode == 2);
11333 unsigned MaxOffset = TotalNumIntRegs * 8 +
11334 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
11336 /* Align ArgSize to a multiple of 8 */
11337 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
11338 bool NeedsAlign = (Align > 8);
11340 MachineBasicBlock *thisMBB = MBB;
11341 MachineBasicBlock *overflowMBB;
11342 MachineBasicBlock *offsetMBB;
11343 MachineBasicBlock *endMBB;
11345 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
11346 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
11347 unsigned OffsetReg = 0;
11349 if (!UseGPOffset && !UseFPOffset) {
11350 // If we only pull from the overflow region, we don't create a branch.
11351 // We don't need to alter control flow.
11352 OffsetDestReg = 0; // unused
11353 OverflowDestReg = DestReg;
11356 overflowMBB = thisMBB;
11359 // First emit code to check if gp_offset (or fp_offset) is below the bound.
11360 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
11361 // If not, pull from overflow_area. (branch to overflowMBB)
11366 // offsetMBB overflowMBB
11371 // Registers for the PHI in endMBB
11372 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
11373 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
11375 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
11376 MachineFunction *MF = MBB->getParent();
11377 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
11378 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
11379 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
11381 MachineFunction::iterator MBBIter = MBB;
11384 // Insert the new basic blocks
11385 MF->insert(MBBIter, offsetMBB);
11386 MF->insert(MBBIter, overflowMBB);
11387 MF->insert(MBBIter, endMBB);
11389 // Transfer the remainder of MBB and its successor edges to endMBB.
11390 endMBB->splice(endMBB->begin(), thisMBB,
11391 llvm::next(MachineBasicBlock::iterator(MI)),
11393 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
11395 // Make offsetMBB and overflowMBB successors of thisMBB
11396 thisMBB->addSuccessor(offsetMBB);
11397 thisMBB->addSuccessor(overflowMBB);
11399 // endMBB is a successor of both offsetMBB and overflowMBB
11400 offsetMBB->addSuccessor(endMBB);
11401 overflowMBB->addSuccessor(endMBB);
11403 // Load the offset value into a register
11404 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
11405 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
11409 .addDisp(Disp, UseFPOffset ? 4 : 0)
11410 .addOperand(Segment)
11411 .setMemRefs(MMOBegin, MMOEnd);
11413 // Check if there is enough room left to pull this argument.
11414 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
11416 .addImm(MaxOffset + 8 - ArgSizeA8);
11418 // Branch to "overflowMBB" if offset >= max
11419 // Fall through to "offsetMBB" otherwise
11420 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
11421 .addMBB(overflowMBB);
11424 // In offsetMBB, emit code to use the reg_save_area.
11426 assert(OffsetReg != 0);
11428 // Read the reg_save_area address.
11429 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
11430 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
11435 .addOperand(Segment)
11436 .setMemRefs(MMOBegin, MMOEnd);
11438 // Zero-extend the offset
11439 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
11440 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
11443 .addImm(X86::sub_32bit);
11445 // Add the offset to the reg_save_area to get the final address.
11446 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
11447 .addReg(OffsetReg64)
11448 .addReg(RegSaveReg);
11450 // Compute the offset for the next argument
11451 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
11452 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
11454 .addImm(UseFPOffset ? 16 : 8);
11456 // Store it back into the va_list.
11457 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
11461 .addDisp(Disp, UseFPOffset ? 4 : 0)
11462 .addOperand(Segment)
11463 .addReg(NextOffsetReg)
11464 .setMemRefs(MMOBegin, MMOEnd);
11467 BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
11472 // Emit code to use overflow area
11475 // Load the overflow_area address into a register.
11476 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
11477 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
11482 .addOperand(Segment)
11483 .setMemRefs(MMOBegin, MMOEnd);
11485 // If we need to align it, do so. Otherwise, just copy the address
11486 // to OverflowDestReg.
11488 // Align the overflow address
11489 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
11490 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
11492 // aligned_addr = (addr + (align-1)) & ~(align-1)
11493 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
11494 .addReg(OverflowAddrReg)
11497 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
11499 .addImm(~(uint64_t)(Align-1));
11501 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
11502 .addReg(OverflowAddrReg);
11505 // Compute the next overflow address after this argument.
11506 // (the overflow address should be kept 8-byte aligned)
11507 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
11508 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
11509 .addReg(OverflowDestReg)
11510 .addImm(ArgSizeA8);
11512 // Store the new overflow address.
11513 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
11518 .addOperand(Segment)
11519 .addReg(NextAddrReg)
11520 .setMemRefs(MMOBegin, MMOEnd);
11522 // If we branched, emit the PHI to the front of endMBB.
11524 BuildMI(*endMBB, endMBB->begin(), DL,
11525 TII->get(X86::PHI), DestReg)
11526 .addReg(OffsetDestReg).addMBB(offsetMBB)
11527 .addReg(OverflowDestReg).addMBB(overflowMBB);
11530 // Erase the pseudo instruction
11531 MI->eraseFromParent();
11536 MachineBasicBlock *
11537 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
11539 MachineBasicBlock *MBB) const {
11540 // Emit code to save XMM registers to the stack. The ABI says that the
11541 // number of registers to save is given in %al, so it's theoretically
11542 // possible to do an indirect jump trick to avoid saving all of them,
11543 // however this code takes a simpler approach and just executes all
11544 // of the stores if %al is non-zero. It's less code, and it's probably
11545 // easier on the hardware branch predictor, and stores aren't all that
11546 // expensive anyway.
11548 // Create the new basic blocks. One block contains all the XMM stores,
11549 // and one block is the final destination regardless of whether any
11550 // stores were performed.
11551 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
11552 MachineFunction *F = MBB->getParent();
11553 MachineFunction::iterator MBBIter = MBB;
11555 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
11556 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
11557 F->insert(MBBIter, XMMSaveMBB);
11558 F->insert(MBBIter, EndMBB);
11560 // Transfer the remainder of MBB and its successor edges to EndMBB.
11561 EndMBB->splice(EndMBB->begin(), MBB,
11562 llvm::next(MachineBasicBlock::iterator(MI)),
11564 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
11566 // The original block will now fall through to the XMM save block.
11567 MBB->addSuccessor(XMMSaveMBB);
11568 // The XMMSaveMBB will fall through to the end block.
11569 XMMSaveMBB->addSuccessor(EndMBB);
11571 // Now add the instructions.
11572 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11573 DebugLoc DL = MI->getDebugLoc();
11575 unsigned CountReg = MI->getOperand(0).getReg();
11576 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
11577 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
11579 if (!Subtarget->isTargetWin64()) {
11580 // If %al is 0, branch around the XMM save block.
11581 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
11582 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
11583 MBB->addSuccessor(EndMBB);
11586 unsigned MOVOpc = Subtarget->hasAVX() ? X86::VMOVAPSmr : X86::MOVAPSmr;
11587 // In the XMM save block, save all the XMM argument registers.
11588 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
11589 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
11590 MachineMemOperand *MMO =
11591 F->getMachineMemOperand(
11592 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
11593 MachineMemOperand::MOStore,
11594 /*Size=*/16, /*Align=*/16);
11595 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
11596 .addFrameIndex(RegSaveFrameIndex)
11597 .addImm(/*Scale=*/1)
11598 .addReg(/*IndexReg=*/0)
11599 .addImm(/*Disp=*/Offset)
11600 .addReg(/*Segment=*/0)
11601 .addReg(MI->getOperand(i).getReg())
11602 .addMemOperand(MMO);
11605 MI->eraseFromParent(); // The pseudo instruction is gone now.
11610 MachineBasicBlock *
11611 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
11612 MachineBasicBlock *BB) const {
11613 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11614 DebugLoc DL = MI->getDebugLoc();
11616 // To "insert" a SELECT_CC instruction, we actually have to insert the
11617 // diamond control-flow pattern. The incoming instruction knows the
11618 // destination vreg to set, the condition code register to branch on, the
11619 // true/false values to select between, and a branch opcode to use.
11620 const BasicBlock *LLVM_BB = BB->getBasicBlock();
11621 MachineFunction::iterator It = BB;
11627 // cmpTY ccX, r1, r2
11629 // fallthrough --> copy0MBB
11630 MachineBasicBlock *thisMBB = BB;
11631 MachineFunction *F = BB->getParent();
11632 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
11633 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
11634 F->insert(It, copy0MBB);
11635 F->insert(It, sinkMBB);
11637 // If the EFLAGS register isn't dead in the terminator, then claim that it's
11638 // live into the sink and copy blocks.
11639 if (!MI->killsRegister(X86::EFLAGS)) {
11640 copy0MBB->addLiveIn(X86::EFLAGS);
11641 sinkMBB->addLiveIn(X86::EFLAGS);
11644 // Transfer the remainder of BB and its successor edges to sinkMBB.
11645 sinkMBB->splice(sinkMBB->begin(), BB,
11646 llvm::next(MachineBasicBlock::iterator(MI)),
11648 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
11650 // Add the true and fallthrough blocks as its successors.
11651 BB->addSuccessor(copy0MBB);
11652 BB->addSuccessor(sinkMBB);
11654 // Create the conditional branch instruction.
11656 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
11657 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
11660 // %FalseValue = ...
11661 // # fallthrough to sinkMBB
11662 copy0MBB->addSuccessor(sinkMBB);
11665 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
11667 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
11668 TII->get(X86::PHI), MI->getOperand(0).getReg())
11669 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
11670 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
11672 MI->eraseFromParent(); // The pseudo instruction is gone now.
11676 MachineBasicBlock *
11677 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI, MachineBasicBlock *BB,
11678 bool Is64Bit) const {
11679 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11680 DebugLoc DL = MI->getDebugLoc();
11681 MachineFunction *MF = BB->getParent();
11682 const BasicBlock *LLVM_BB = BB->getBasicBlock();
11684 assert(EnableSegmentedStacks);
11686 unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
11687 unsigned TlsOffset = Is64Bit ? 0x70 : 0x30;
11690 // ... [Till the alloca]
11691 // If stacklet is not large enough, jump to mallocMBB
11694 // Allocate by subtracting from RSP
11695 // Jump to continueMBB
11698 // Allocate by call to runtime
11702 // [rest of original BB]
11705 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
11706 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
11707 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
11709 MachineRegisterInfo &MRI = MF->getRegInfo();
11710 const TargetRegisterClass *AddrRegClass =
11711 getRegClassFor(Is64Bit ? MVT::i64:MVT::i32);
11713 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
11714 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
11715 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
11716 sizeVReg = MI->getOperand(1).getReg(),
11717 physSPReg = Is64Bit ? X86::RSP : X86::ESP;
11719 MachineFunction::iterator MBBIter = BB;
11722 MF->insert(MBBIter, bumpMBB);
11723 MF->insert(MBBIter, mallocMBB);
11724 MF->insert(MBBIter, continueMBB);
11726 continueMBB->splice(continueMBB->begin(), BB, llvm::next
11727 (MachineBasicBlock::iterator(MI)), BB->end());
11728 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
11730 // Add code to the main basic block to check if the stack limit has been hit,
11731 // and if so, jump to mallocMBB otherwise to bumpMBB.
11732 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
11733 BuildMI(BB, DL, TII->get(Is64Bit ? X86::SUB64rr:X86::SUB32rr), tmpSPVReg)
11734 .addReg(tmpSPVReg).addReg(sizeVReg);
11735 BuildMI(BB, DL, TII->get(Is64Bit ? X86::CMP64mr:X86::CMP32mr))
11736 .addReg(0).addImm(0).addReg(0).addImm(TlsOffset).addReg(TlsReg)
11737 .addReg(tmpSPVReg);
11738 BuildMI(BB, DL, TII->get(X86::JG_4)).addMBB(mallocMBB);
11740 // bumpMBB simply decreases the stack pointer, since we know the current
11741 // stacklet has enough space.
11742 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
11743 .addReg(tmpSPVReg);
11744 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
11745 .addReg(tmpSPVReg);
11746 BuildMI(bumpMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
11748 // Calls into a routine in libgcc to allocate more space from the heap.
11750 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
11752 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
11753 .addExternalSymbol("__morestack_allocate_stack_space").addReg(X86::RDI);
11755 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
11757 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
11758 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
11759 .addExternalSymbol("__morestack_allocate_stack_space");
11763 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
11766 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
11767 .addReg(Is64Bit ? X86::RAX : X86::EAX);
11768 BuildMI(mallocMBB, DL, TII->get(X86::JMP_4)).addMBB(continueMBB);
11770 // Set up the CFG correctly.
11771 BB->addSuccessor(bumpMBB);
11772 BB->addSuccessor(mallocMBB);
11773 mallocMBB->addSuccessor(continueMBB);
11774 bumpMBB->addSuccessor(continueMBB);
11776 // Take care of the PHI nodes.
11777 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
11778 MI->getOperand(0).getReg())
11779 .addReg(mallocPtrVReg).addMBB(mallocMBB)
11780 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
11782 // Delete the original pseudo instruction.
11783 MI->eraseFromParent();
11786 return continueMBB;
11789 MachineBasicBlock *
11790 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
11791 MachineBasicBlock *BB) const {
11792 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11793 DebugLoc DL = MI->getDebugLoc();
11795 assert(!Subtarget->isTargetEnvMacho());
11797 // The lowering is pretty easy: we're just emitting the call to _alloca. The
11798 // non-trivial part is impdef of ESP.
11800 if (Subtarget->isTargetWin64()) {
11801 if (Subtarget->isTargetCygMing()) {
11802 // ___chkstk(Mingw64):
11803 // Clobbers R10, R11, RAX and EFLAGS.
11805 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
11806 .addExternalSymbol("___chkstk")
11807 .addReg(X86::RAX, RegState::Implicit)
11808 .addReg(X86::RSP, RegState::Implicit)
11809 .addReg(X86::RAX, RegState::Define | RegState::Implicit)
11810 .addReg(X86::RSP, RegState::Define | RegState::Implicit)
11811 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11813 // __chkstk(MSVCRT): does not update stack pointer.
11814 // Clobbers R10, R11 and EFLAGS.
11815 // FIXME: RAX(allocated size) might be reused and not killed.
11816 BuildMI(*BB, MI, DL, TII->get(X86::W64ALLOCA))
11817 .addExternalSymbol("__chkstk")
11818 .addReg(X86::RAX, RegState::Implicit)
11819 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11820 // RAX has the offset to subtracted from RSP.
11821 BuildMI(*BB, MI, DL, TII->get(X86::SUB64rr), X86::RSP)
11826 const char *StackProbeSymbol =
11827 Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
11829 BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
11830 .addExternalSymbol(StackProbeSymbol)
11831 .addReg(X86::EAX, RegState::Implicit)
11832 .addReg(X86::ESP, RegState::Implicit)
11833 .addReg(X86::EAX, RegState::Define | RegState::Implicit)
11834 .addReg(X86::ESP, RegState::Define | RegState::Implicit)
11835 .addReg(X86::EFLAGS, RegState::Define | RegState::Implicit);
11838 MI->eraseFromParent(); // The pseudo instruction is gone now.
11842 MachineBasicBlock *
11843 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
11844 MachineBasicBlock *BB) const {
11845 // This is pretty easy. We're taking the value that we received from
11846 // our load from the relocation, sticking it in either RDI (x86-64)
11847 // or EAX and doing an indirect call. The return value will then
11848 // be in the normal return register.
11849 const X86InstrInfo *TII
11850 = static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
11851 DebugLoc DL = MI->getDebugLoc();
11852 MachineFunction *F = BB->getParent();
11854 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
11855 assert(MI->getOperand(3).isGlobal() && "This should be a global");
11857 if (Subtarget->is64Bit()) {
11858 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11859 TII->get(X86::MOV64rm), X86::RDI)
11861 .addImm(0).addReg(0)
11862 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11863 MI->getOperand(3).getTargetFlags())
11865 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
11866 addDirectMem(MIB, X86::RDI);
11867 } else if (getTargetMachine().getRelocationModel() != Reloc::PIC_) {
11868 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11869 TII->get(X86::MOV32rm), X86::EAX)
11871 .addImm(0).addReg(0)
11872 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11873 MI->getOperand(3).getTargetFlags())
11875 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
11876 addDirectMem(MIB, X86::EAX);
11878 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
11879 TII->get(X86::MOV32rm), X86::EAX)
11880 .addReg(TII->getGlobalBaseReg(F))
11881 .addImm(0).addReg(0)
11882 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
11883 MI->getOperand(3).getTargetFlags())
11885 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
11886 addDirectMem(MIB, X86::EAX);
11889 MI->eraseFromParent(); // The pseudo instruction is gone now.
11893 MachineBasicBlock *
11894 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
11895 MachineBasicBlock *BB) const {
11896 switch (MI->getOpcode()) {
11897 default: assert(false && "Unexpected instr type to insert");
11898 case X86::TAILJMPd64:
11899 case X86::TAILJMPr64:
11900 case X86::TAILJMPm64:
11901 assert(!"TAILJMP64 would not be touched here.");
11902 case X86::TCRETURNdi64:
11903 case X86::TCRETURNri64:
11904 case X86::TCRETURNmi64:
11905 // Defs of TCRETURNxx64 has Win64's callee-saved registers, as subset.
11906 // On AMD64, additional defs should be added before register allocation.
11907 if (!Subtarget->isTargetWin64()) {
11908 MI->addRegisterDefined(X86::RSI);
11909 MI->addRegisterDefined(X86::RDI);
11910 MI->addRegisterDefined(X86::XMM6);
11911 MI->addRegisterDefined(X86::XMM7);
11912 MI->addRegisterDefined(X86::XMM8);
11913 MI->addRegisterDefined(X86::XMM9);
11914 MI->addRegisterDefined(X86::XMM10);
11915 MI->addRegisterDefined(X86::XMM11);
11916 MI->addRegisterDefined(X86::XMM12);
11917 MI->addRegisterDefined(X86::XMM13);
11918 MI->addRegisterDefined(X86::XMM14);
11919 MI->addRegisterDefined(X86::XMM15);
11922 case X86::WIN_ALLOCA:
11923 return EmitLoweredWinAlloca(MI, BB);
11924 case X86::SEG_ALLOCA_32:
11925 return EmitLoweredSegAlloca(MI, BB, false);
11926 case X86::SEG_ALLOCA_64:
11927 return EmitLoweredSegAlloca(MI, BB, true);
11928 case X86::TLSCall_32:
11929 case X86::TLSCall_64:
11930 return EmitLoweredTLSCall(MI, BB);
11931 case X86::CMOV_GR8:
11932 case X86::CMOV_FR32:
11933 case X86::CMOV_FR64:
11934 case X86::CMOV_V4F32:
11935 case X86::CMOV_V2F64:
11936 case X86::CMOV_V2I64:
11937 case X86::CMOV_V8F32:
11938 case X86::CMOV_V4F64:
11939 case X86::CMOV_V4I64:
11940 case X86::CMOV_GR16:
11941 case X86::CMOV_GR32:
11942 case X86::CMOV_RFP32:
11943 case X86::CMOV_RFP64:
11944 case X86::CMOV_RFP80:
11945 return EmitLoweredSelect(MI, BB);
11947 case X86::FP32_TO_INT16_IN_MEM:
11948 case X86::FP32_TO_INT32_IN_MEM:
11949 case X86::FP32_TO_INT64_IN_MEM:
11950 case X86::FP64_TO_INT16_IN_MEM:
11951 case X86::FP64_TO_INT32_IN_MEM:
11952 case X86::FP64_TO_INT64_IN_MEM:
11953 case X86::FP80_TO_INT16_IN_MEM:
11954 case X86::FP80_TO_INT32_IN_MEM:
11955 case X86::FP80_TO_INT64_IN_MEM: {
11956 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
11957 DebugLoc DL = MI->getDebugLoc();
11959 // Change the floating point control register to use "round towards zero"
11960 // mode when truncating to an integer value.
11961 MachineFunction *F = BB->getParent();
11962 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
11963 addFrameReference(BuildMI(*BB, MI, DL,
11964 TII->get(X86::FNSTCW16m)), CWFrameIdx);
11966 // Load the old value of the high byte of the control word...
11968 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
11969 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
11972 // Set the high part to be round to zero...
11973 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
11976 // Reload the modified control word now...
11977 addFrameReference(BuildMI(*BB, MI, DL,
11978 TII->get(X86::FLDCW16m)), CWFrameIdx);
11980 // Restore the memory image of control word to original value
11981 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
11984 // Get the X86 opcode to use.
11986 switch (MI->getOpcode()) {
11987 default: llvm_unreachable("illegal opcode!");
11988 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
11989 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
11990 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
11991 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
11992 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
11993 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
11994 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
11995 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
11996 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
12000 MachineOperand &Op = MI->getOperand(0);
12002 AM.BaseType = X86AddressMode::RegBase;
12003 AM.Base.Reg = Op.getReg();
12005 AM.BaseType = X86AddressMode::FrameIndexBase;
12006 AM.Base.FrameIndex = Op.getIndex();
12008 Op = MI->getOperand(1);
12010 AM.Scale = Op.getImm();
12011 Op = MI->getOperand(2);
12013 AM.IndexReg = Op.getImm();
12014 Op = MI->getOperand(3);
12015 if (Op.isGlobal()) {
12016 AM.GV = Op.getGlobal();
12018 AM.Disp = Op.getImm();
12020 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
12021 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
12023 // Reload the original control word now.
12024 addFrameReference(BuildMI(*BB, MI, DL,
12025 TII->get(X86::FLDCW16m)), CWFrameIdx);
12027 MI->eraseFromParent(); // The pseudo instruction is gone now.
12030 // String/text processing lowering.
12031 case X86::PCMPISTRM128REG:
12032 case X86::VPCMPISTRM128REG:
12033 return EmitPCMP(MI, BB, 3, false /* in-mem */);
12034 case X86::PCMPISTRM128MEM:
12035 case X86::VPCMPISTRM128MEM:
12036 return EmitPCMP(MI, BB, 3, true /* in-mem */);
12037 case X86::PCMPESTRM128REG:
12038 case X86::VPCMPESTRM128REG:
12039 return EmitPCMP(MI, BB, 5, false /* in mem */);
12040 case X86::PCMPESTRM128MEM:
12041 case X86::VPCMPESTRM128MEM:
12042 return EmitPCMP(MI, BB, 5, true /* in mem */);
12044 // Thread synchronization.
12046 return EmitMonitor(MI, BB);
12048 return EmitMwait(MI, BB);
12050 // Atomic Lowering.
12051 case X86::ATOMAND32:
12052 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
12053 X86::AND32ri, X86::MOV32rm,
12055 X86::NOT32r, X86::EAX,
12056 X86::GR32RegisterClass);
12057 case X86::ATOMOR32:
12058 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
12059 X86::OR32ri, X86::MOV32rm,
12061 X86::NOT32r, X86::EAX,
12062 X86::GR32RegisterClass);
12063 case X86::ATOMXOR32:
12064 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
12065 X86::XOR32ri, X86::MOV32rm,
12067 X86::NOT32r, X86::EAX,
12068 X86::GR32RegisterClass);
12069 case X86::ATOMNAND32:
12070 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
12071 X86::AND32ri, X86::MOV32rm,
12073 X86::NOT32r, X86::EAX,
12074 X86::GR32RegisterClass, true);
12075 case X86::ATOMMIN32:
12076 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
12077 case X86::ATOMMAX32:
12078 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
12079 case X86::ATOMUMIN32:
12080 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
12081 case X86::ATOMUMAX32:
12082 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
12084 case X86::ATOMAND16:
12085 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
12086 X86::AND16ri, X86::MOV16rm,
12088 X86::NOT16r, X86::AX,
12089 X86::GR16RegisterClass);
12090 case X86::ATOMOR16:
12091 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
12092 X86::OR16ri, X86::MOV16rm,
12094 X86::NOT16r, X86::AX,
12095 X86::GR16RegisterClass);
12096 case X86::ATOMXOR16:
12097 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
12098 X86::XOR16ri, X86::MOV16rm,
12100 X86::NOT16r, X86::AX,
12101 X86::GR16RegisterClass);
12102 case X86::ATOMNAND16:
12103 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
12104 X86::AND16ri, X86::MOV16rm,
12106 X86::NOT16r, X86::AX,
12107 X86::GR16RegisterClass, true);
12108 case X86::ATOMMIN16:
12109 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
12110 case X86::ATOMMAX16:
12111 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
12112 case X86::ATOMUMIN16:
12113 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
12114 case X86::ATOMUMAX16:
12115 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
12117 case X86::ATOMAND8:
12118 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
12119 X86::AND8ri, X86::MOV8rm,
12121 X86::NOT8r, X86::AL,
12122 X86::GR8RegisterClass);
12124 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
12125 X86::OR8ri, X86::MOV8rm,
12127 X86::NOT8r, X86::AL,
12128 X86::GR8RegisterClass);
12129 case X86::ATOMXOR8:
12130 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
12131 X86::XOR8ri, X86::MOV8rm,
12133 X86::NOT8r, X86::AL,
12134 X86::GR8RegisterClass);
12135 case X86::ATOMNAND8:
12136 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
12137 X86::AND8ri, X86::MOV8rm,
12139 X86::NOT8r, X86::AL,
12140 X86::GR8RegisterClass, true);
12141 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
12142 // This group is for 64-bit host.
12143 case X86::ATOMAND64:
12144 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
12145 X86::AND64ri32, X86::MOV64rm,
12147 X86::NOT64r, X86::RAX,
12148 X86::GR64RegisterClass);
12149 case X86::ATOMOR64:
12150 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
12151 X86::OR64ri32, X86::MOV64rm,
12153 X86::NOT64r, X86::RAX,
12154 X86::GR64RegisterClass);
12155 case X86::ATOMXOR64:
12156 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
12157 X86::XOR64ri32, X86::MOV64rm,
12159 X86::NOT64r, X86::RAX,
12160 X86::GR64RegisterClass);
12161 case X86::ATOMNAND64:
12162 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
12163 X86::AND64ri32, X86::MOV64rm,
12165 X86::NOT64r, X86::RAX,
12166 X86::GR64RegisterClass, true);
12167 case X86::ATOMMIN64:
12168 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
12169 case X86::ATOMMAX64:
12170 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
12171 case X86::ATOMUMIN64:
12172 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
12173 case X86::ATOMUMAX64:
12174 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
12176 // This group does 64-bit operations on a 32-bit host.
12177 case X86::ATOMAND6432:
12178 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12179 X86::AND32rr, X86::AND32rr,
12180 X86::AND32ri, X86::AND32ri,
12182 case X86::ATOMOR6432:
12183 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12184 X86::OR32rr, X86::OR32rr,
12185 X86::OR32ri, X86::OR32ri,
12187 case X86::ATOMXOR6432:
12188 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12189 X86::XOR32rr, X86::XOR32rr,
12190 X86::XOR32ri, X86::XOR32ri,
12192 case X86::ATOMNAND6432:
12193 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12194 X86::AND32rr, X86::AND32rr,
12195 X86::AND32ri, X86::AND32ri,
12197 case X86::ATOMADD6432:
12198 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12199 X86::ADD32rr, X86::ADC32rr,
12200 X86::ADD32ri, X86::ADC32ri,
12202 case X86::ATOMSUB6432:
12203 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12204 X86::SUB32rr, X86::SBB32rr,
12205 X86::SUB32ri, X86::SBB32ri,
12207 case X86::ATOMSWAP6432:
12208 return EmitAtomicBit6432WithCustomInserter(MI, BB,
12209 X86::MOV32rr, X86::MOV32rr,
12210 X86::MOV32ri, X86::MOV32ri,
12212 case X86::VASTART_SAVE_XMM_REGS:
12213 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
12215 case X86::VAARG_64:
12216 return EmitVAARG64WithCustomInserter(MI, BB);
12220 //===----------------------------------------------------------------------===//
12221 // X86 Optimization Hooks
12222 //===----------------------------------------------------------------------===//
12224 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
12228 const SelectionDAG &DAG,
12229 unsigned Depth) const {
12230 unsigned Opc = Op.getOpcode();
12231 assert((Opc >= ISD::BUILTIN_OP_END ||
12232 Opc == ISD::INTRINSIC_WO_CHAIN ||
12233 Opc == ISD::INTRINSIC_W_CHAIN ||
12234 Opc == ISD::INTRINSIC_VOID) &&
12235 "Should use MaskedValueIsZero if you don't know whether Op"
12236 " is a target node!");
12238 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
12252 // These nodes' second result is a boolean.
12253 if (Op.getResNo() == 0)
12256 case X86ISD::SETCC:
12257 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
12258 Mask.getBitWidth() - 1);
12263 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
12264 unsigned Depth) const {
12265 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
12266 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
12267 return Op.getValueType().getScalarType().getSizeInBits();
12273 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
12274 /// node is a GlobalAddress + offset.
12275 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
12276 const GlobalValue* &GA,
12277 int64_t &Offset) const {
12278 if (N->getOpcode() == X86ISD::Wrapper) {
12279 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
12280 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
12281 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
12285 return TargetLowering::isGAPlusOffset(N, GA, Offset);
12288 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
12289 /// same as extracting the high 128-bit part of 256-bit vector and then
12290 /// inserting the result into the low part of a new 256-bit vector
12291 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
12292 EVT VT = SVOp->getValueType(0);
12293 int NumElems = VT.getVectorNumElements();
12295 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
12296 for (int i = 0, j = NumElems/2; i < NumElems/2; ++i, ++j)
12297 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
12298 SVOp->getMaskElt(j) >= 0)
12304 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
12305 /// same as extracting the low 128-bit part of 256-bit vector and then
12306 /// inserting the result into the high part of a new 256-bit vector
12307 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
12308 EVT VT = SVOp->getValueType(0);
12309 int NumElems = VT.getVectorNumElements();
12311 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
12312 for (int i = NumElems/2, j = 0; i < NumElems; ++i, ++j)
12313 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
12314 SVOp->getMaskElt(j) >= 0)
12320 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
12321 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
12322 TargetLowering::DAGCombinerInfo &DCI) {
12323 DebugLoc dl = N->getDebugLoc();
12324 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
12325 SDValue V1 = SVOp->getOperand(0);
12326 SDValue V2 = SVOp->getOperand(1);
12327 EVT VT = SVOp->getValueType(0);
12328 int NumElems = VT.getVectorNumElements();
12330 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
12331 V2.getOpcode() == ISD::CONCAT_VECTORS) {
12335 // V UNDEF BUILD_VECTOR UNDEF
12337 // CONCAT_VECTOR CONCAT_VECTOR
12340 // RESULT: V + zero extended
12342 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
12343 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
12344 V1.getOperand(1).getOpcode() != ISD::UNDEF)
12347 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
12350 // To match the shuffle mask, the first half of the mask should
12351 // be exactly the first vector, and all the rest a splat with the
12352 // first element of the second one.
12353 for (int i = 0; i < NumElems/2; ++i)
12354 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
12355 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
12358 // Emit a zeroed vector and insert the desired subvector on its
12360 SDValue Zeros = getZeroVector(VT, true /* HasSSE2 */, DAG, dl);
12361 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0),
12362 DAG.getConstant(0, MVT::i32), DAG, dl);
12363 return DCI.CombineTo(N, InsV);
12366 //===--------------------------------------------------------------------===//
12367 // Combine some shuffles into subvector extracts and inserts:
12370 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
12371 if (isShuffleHigh128VectorInsertLow(SVOp)) {
12372 SDValue V = Extract128BitVector(V1, DAG.getConstant(NumElems/2, MVT::i32),
12374 SDValue InsV = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT),
12375 V, DAG.getConstant(0, MVT::i32), DAG, dl);
12376 return DCI.CombineTo(N, InsV);
12379 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
12380 if (isShuffleLow128VectorInsertHigh(SVOp)) {
12381 SDValue V = Extract128BitVector(V1, DAG.getConstant(0, MVT::i32), DAG, dl);
12382 SDValue InsV = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT),
12383 V, DAG.getConstant(NumElems/2, MVT::i32), DAG, dl);
12384 return DCI.CombineTo(N, InsV);
12390 /// PerformShuffleCombine - Performs several different shuffle combines.
12391 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
12392 TargetLowering::DAGCombinerInfo &DCI,
12393 const X86Subtarget *Subtarget) {
12394 DebugLoc dl = N->getDebugLoc();
12395 EVT VT = N->getValueType(0);
12397 // Don't create instructions with illegal types after legalize types has run.
12398 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12399 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
12402 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
12403 if (Subtarget->hasAVX() && VT.getSizeInBits() == 256 &&
12404 N->getOpcode() == ISD::VECTOR_SHUFFLE)
12405 return PerformShuffleCombine256(N, DAG, DCI);
12407 // Only handle 128 wide vector from here on.
12408 if (VT.getSizeInBits() != 128)
12411 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
12412 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
12413 // consecutive, non-overlapping, and in the right order.
12414 SmallVector<SDValue, 16> Elts;
12415 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
12416 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
12418 return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
12421 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
12422 /// generation and convert it from being a bunch of shuffles and extracts
12423 /// to a simple store and scalar loads to extract the elements.
12424 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
12425 const TargetLowering &TLI) {
12426 SDValue InputVector = N->getOperand(0);
12428 // Only operate on vectors of 4 elements, where the alternative shuffling
12429 // gets to be more expensive.
12430 if (InputVector.getValueType() != MVT::v4i32)
12433 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
12434 // single use which is a sign-extend or zero-extend, and all elements are
12436 SmallVector<SDNode *, 4> Uses;
12437 unsigned ExtractedElements = 0;
12438 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
12439 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
12440 if (UI.getUse().getResNo() != InputVector.getResNo())
12443 SDNode *Extract = *UI;
12444 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12447 if (Extract->getValueType(0) != MVT::i32)
12449 if (!Extract->hasOneUse())
12451 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
12452 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
12454 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
12457 // Record which element was extracted.
12458 ExtractedElements |=
12459 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
12461 Uses.push_back(Extract);
12464 // If not all the elements were used, this may not be worthwhile.
12465 if (ExtractedElements != 15)
12468 // Ok, we've now decided to do the transformation.
12469 DebugLoc dl = InputVector.getDebugLoc();
12471 // Store the value to a temporary stack slot.
12472 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
12473 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
12474 MachinePointerInfo(), false, false, 0);
12476 // Replace each use (extract) with a load of the appropriate element.
12477 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
12478 UE = Uses.end(); UI != UE; ++UI) {
12479 SDNode *Extract = *UI;
12481 // cOMpute the element's address.
12482 SDValue Idx = Extract->getOperand(1);
12484 InputVector.getValueType().getVectorElementType().getSizeInBits()/8;
12485 uint64_t Offset = EltSize * cast<ConstantSDNode>(Idx)->getZExtValue();
12486 SDValue OffsetVal = DAG.getConstant(Offset, TLI.getPointerTy());
12488 SDValue ScalarAddr = DAG.getNode(ISD::ADD, dl, TLI.getPointerTy(),
12489 StackPtr, OffsetVal);
12491 // Load the scalar.
12492 SDValue LoadScalar = DAG.getLoad(Extract->getValueType(0), dl, Ch,
12493 ScalarAddr, MachinePointerInfo(),
12496 // Replace the exact with the load.
12497 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), LoadScalar);
12500 // The replacement was made in place; don't return anything.
12504 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
12505 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
12506 const X86Subtarget *Subtarget) {
12507 DebugLoc DL = N->getDebugLoc();
12508 SDValue Cond = N->getOperand(0);
12509 // Get the LHS/RHS of the select.
12510 SDValue LHS = N->getOperand(1);
12511 SDValue RHS = N->getOperand(2);
12513 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
12514 // instructions match the semantics of the common C idiom x<y?x:y but not
12515 // x<=y?x:y, because of how they handle negative zero (which can be
12516 // ignored in unsafe-math mode).
12517 if (Subtarget->hasSSE2() &&
12518 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
12519 Cond.getOpcode() == ISD::SETCC) {
12520 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
12522 unsigned Opcode = 0;
12523 // Check for x CC y ? x : y.
12524 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
12525 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
12529 // Converting this to a min would handle NaNs incorrectly, and swapping
12530 // the operands would cause it to handle comparisons between positive
12531 // and negative zero incorrectly.
12532 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
12533 if (!UnsafeFPMath &&
12534 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
12536 std::swap(LHS, RHS);
12538 Opcode = X86ISD::FMIN;
12541 // Converting this to a min would handle comparisons between positive
12542 // and negative zero incorrectly.
12543 if (!UnsafeFPMath &&
12544 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
12546 Opcode = X86ISD::FMIN;
12549 // Converting this to a min would handle both negative zeros and NaNs
12550 // incorrectly, but we can swap the operands to fix both.
12551 std::swap(LHS, RHS);
12555 Opcode = X86ISD::FMIN;
12559 // Converting this to a max would handle comparisons between positive
12560 // and negative zero incorrectly.
12561 if (!UnsafeFPMath &&
12562 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
12564 Opcode = X86ISD::FMAX;
12567 // Converting this to a max would handle NaNs incorrectly, and swapping
12568 // the operands would cause it to handle comparisons between positive
12569 // and negative zero incorrectly.
12570 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
12571 if (!UnsafeFPMath &&
12572 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
12574 std::swap(LHS, RHS);
12576 Opcode = X86ISD::FMAX;
12579 // Converting this to a max would handle both negative zeros and NaNs
12580 // incorrectly, but we can swap the operands to fix both.
12581 std::swap(LHS, RHS);
12585 Opcode = X86ISD::FMAX;
12588 // Check for x CC y ? y : x -- a min/max with reversed arms.
12589 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
12590 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
12594 // Converting this to a min would handle comparisons between positive
12595 // and negative zero incorrectly, and swapping the operands would
12596 // cause it to handle NaNs incorrectly.
12597 if (!UnsafeFPMath &&
12598 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
12599 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
12601 std::swap(LHS, RHS);
12603 Opcode = X86ISD::FMIN;
12606 // Converting this to a min would handle NaNs incorrectly.
12607 if (!UnsafeFPMath &&
12608 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
12610 Opcode = X86ISD::FMIN;
12613 // Converting this to a min would handle both negative zeros and NaNs
12614 // incorrectly, but we can swap the operands to fix both.
12615 std::swap(LHS, RHS);
12619 Opcode = X86ISD::FMIN;
12623 // Converting this to a max would handle NaNs incorrectly.
12624 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
12626 Opcode = X86ISD::FMAX;
12629 // Converting this to a max would handle comparisons between positive
12630 // and negative zero incorrectly, and swapping the operands would
12631 // cause it to handle NaNs incorrectly.
12632 if (!UnsafeFPMath &&
12633 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
12634 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
12636 std::swap(LHS, RHS);
12638 Opcode = X86ISD::FMAX;
12641 // Converting this to a max would handle both negative zeros and NaNs
12642 // incorrectly, but we can swap the operands to fix both.
12643 std::swap(LHS, RHS);
12647 Opcode = X86ISD::FMAX;
12653 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
12656 // If this is a select between two integer constants, try to do some
12658 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
12659 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
12660 // Don't do this for crazy integer types.
12661 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
12662 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
12663 // so that TrueC (the true value) is larger than FalseC.
12664 bool NeedsCondInvert = false;
12666 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
12667 // Efficiently invertible.
12668 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
12669 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
12670 isa<ConstantSDNode>(Cond.getOperand(1))))) {
12671 NeedsCondInvert = true;
12672 std::swap(TrueC, FalseC);
12675 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
12676 if (FalseC->getAPIntValue() == 0 &&
12677 TrueC->getAPIntValue().isPowerOf2()) {
12678 if (NeedsCondInvert) // Invert the condition if needed.
12679 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
12680 DAG.getConstant(1, Cond.getValueType()));
12682 // Zero extend the condition if needed.
12683 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
12685 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
12686 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
12687 DAG.getConstant(ShAmt, MVT::i8));
12690 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
12691 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
12692 if (NeedsCondInvert) // Invert the condition if needed.
12693 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
12694 DAG.getConstant(1, Cond.getValueType()));
12696 // Zero extend the condition if needed.
12697 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
12698 FalseC->getValueType(0), Cond);
12699 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12700 SDValue(FalseC, 0));
12703 // Optimize cases that will turn into an LEA instruction. This requires
12704 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
12705 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
12706 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
12707 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
12709 bool isFastMultiplier = false;
12711 switch ((unsigned char)Diff) {
12713 case 1: // result = add base, cond
12714 case 2: // result = lea base( , cond*2)
12715 case 3: // result = lea base(cond, cond*2)
12716 case 4: // result = lea base( , cond*4)
12717 case 5: // result = lea base(cond, cond*4)
12718 case 8: // result = lea base( , cond*8)
12719 case 9: // result = lea base(cond, cond*8)
12720 isFastMultiplier = true;
12725 if (isFastMultiplier) {
12726 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
12727 if (NeedsCondInvert) // Invert the condition if needed.
12728 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
12729 DAG.getConstant(1, Cond.getValueType()));
12731 // Zero extend the condition if needed.
12732 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
12734 // Scale the condition by the difference.
12736 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
12737 DAG.getConstant(Diff, Cond.getValueType()));
12739 // Add the base if non-zero.
12740 if (FalseC->getAPIntValue() != 0)
12741 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12742 SDValue(FalseC, 0));
12752 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
12753 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
12754 TargetLowering::DAGCombinerInfo &DCI) {
12755 DebugLoc DL = N->getDebugLoc();
12757 // If the flag operand isn't dead, don't touch this CMOV.
12758 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
12761 SDValue FalseOp = N->getOperand(0);
12762 SDValue TrueOp = N->getOperand(1);
12763 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
12764 SDValue Cond = N->getOperand(3);
12765 if (CC == X86::COND_E || CC == X86::COND_NE) {
12766 switch (Cond.getOpcode()) {
12770 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
12771 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
12772 return (CC == X86::COND_E) ? FalseOp : TrueOp;
12776 // If this is a select between two integer constants, try to do some
12777 // optimizations. Note that the operands are ordered the opposite of SELECT
12779 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
12780 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
12781 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
12782 // larger than FalseC (the false value).
12783 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
12784 CC = X86::GetOppositeBranchCondition(CC);
12785 std::swap(TrueC, FalseC);
12788 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
12789 // This is efficient for any integer data type (including i8/i16) and
12791 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
12792 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
12793 DAG.getConstant(CC, MVT::i8), Cond);
12795 // Zero extend the condition if needed.
12796 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
12798 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
12799 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
12800 DAG.getConstant(ShAmt, MVT::i8));
12801 if (N->getNumValues() == 2) // Dead flag value?
12802 return DCI.CombineTo(N, Cond, SDValue());
12806 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
12807 // for any integer data type, including i8/i16.
12808 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
12809 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
12810 DAG.getConstant(CC, MVT::i8), Cond);
12812 // Zero extend the condition if needed.
12813 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
12814 FalseC->getValueType(0), Cond);
12815 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12816 SDValue(FalseC, 0));
12818 if (N->getNumValues() == 2) // Dead flag value?
12819 return DCI.CombineTo(N, Cond, SDValue());
12823 // Optimize cases that will turn into an LEA instruction. This requires
12824 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
12825 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
12826 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
12827 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
12829 bool isFastMultiplier = false;
12831 switch ((unsigned char)Diff) {
12833 case 1: // result = add base, cond
12834 case 2: // result = lea base( , cond*2)
12835 case 3: // result = lea base(cond, cond*2)
12836 case 4: // result = lea base( , cond*4)
12837 case 5: // result = lea base(cond, cond*4)
12838 case 8: // result = lea base( , cond*8)
12839 case 9: // result = lea base(cond, cond*8)
12840 isFastMultiplier = true;
12845 if (isFastMultiplier) {
12846 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
12847 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
12848 DAG.getConstant(CC, MVT::i8), Cond);
12849 // Zero extend the condition if needed.
12850 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
12852 // Scale the condition by the difference.
12854 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
12855 DAG.getConstant(Diff, Cond.getValueType()));
12857 // Add the base if non-zero.
12858 if (FalseC->getAPIntValue() != 0)
12859 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
12860 SDValue(FalseC, 0));
12861 if (N->getNumValues() == 2) // Dead flag value?
12862 return DCI.CombineTo(N, Cond, SDValue());
12872 /// PerformMulCombine - Optimize a single multiply with constant into two
12873 /// in order to implement it with two cheaper instructions, e.g.
12874 /// LEA + SHL, LEA + LEA.
12875 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
12876 TargetLowering::DAGCombinerInfo &DCI) {
12877 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
12880 EVT VT = N->getValueType(0);
12881 if (VT != MVT::i64)
12884 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
12887 uint64_t MulAmt = C->getZExtValue();
12888 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
12891 uint64_t MulAmt1 = 0;
12892 uint64_t MulAmt2 = 0;
12893 if ((MulAmt % 9) == 0) {
12895 MulAmt2 = MulAmt / 9;
12896 } else if ((MulAmt % 5) == 0) {
12898 MulAmt2 = MulAmt / 5;
12899 } else if ((MulAmt % 3) == 0) {
12901 MulAmt2 = MulAmt / 3;
12904 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
12905 DebugLoc DL = N->getDebugLoc();
12907 if (isPowerOf2_64(MulAmt2) &&
12908 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
12909 // If second multiplifer is pow2, issue it first. We want the multiply by
12910 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
12912 std::swap(MulAmt1, MulAmt2);
12915 if (isPowerOf2_64(MulAmt1))
12916 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
12917 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
12919 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
12920 DAG.getConstant(MulAmt1, VT));
12922 if (isPowerOf2_64(MulAmt2))
12923 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
12924 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
12926 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
12927 DAG.getConstant(MulAmt2, VT));
12929 // Do not add new nodes to DAG combiner worklist.
12930 DCI.CombineTo(N, NewMul, false);
12935 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
12936 SDValue N0 = N->getOperand(0);
12937 SDValue N1 = N->getOperand(1);
12938 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
12939 EVT VT = N0.getValueType();
12941 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
12942 // since the result of setcc_c is all zero's or all ones.
12943 if (N1C && N0.getOpcode() == ISD::AND &&
12944 N0.getOperand(1).getOpcode() == ISD::Constant) {
12945 SDValue N00 = N0.getOperand(0);
12946 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
12947 ((N00.getOpcode() == ISD::ANY_EXTEND ||
12948 N00.getOpcode() == ISD::ZERO_EXTEND) &&
12949 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
12950 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
12951 APInt ShAmt = N1C->getAPIntValue();
12952 Mask = Mask.shl(ShAmt);
12954 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
12955 N00, DAG.getConstant(Mask, VT));
12962 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
12964 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
12965 const X86Subtarget *Subtarget) {
12966 EVT VT = N->getValueType(0);
12967 if (!VT.isVector() && VT.isInteger() &&
12968 N->getOpcode() == ISD::SHL)
12969 return PerformSHLCombine(N, DAG);
12971 // On X86 with SSE2 support, we can transform this to a vector shift if
12972 // all elements are shifted by the same amount. We can't do this in legalize
12973 // because the a constant vector is typically transformed to a constant pool
12974 // so we have no knowledge of the shift amount.
12975 if (!(Subtarget->hasSSE2() || Subtarget->hasAVX()))
12978 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
12981 SDValue ShAmtOp = N->getOperand(1);
12982 EVT EltVT = VT.getVectorElementType();
12983 DebugLoc DL = N->getDebugLoc();
12984 SDValue BaseShAmt = SDValue();
12985 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
12986 unsigned NumElts = VT.getVectorNumElements();
12988 for (; i != NumElts; ++i) {
12989 SDValue Arg = ShAmtOp.getOperand(i);
12990 if (Arg.getOpcode() == ISD::UNDEF) continue;
12994 for (; i != NumElts; ++i) {
12995 SDValue Arg = ShAmtOp.getOperand(i);
12996 if (Arg.getOpcode() == ISD::UNDEF) continue;
12997 if (Arg != BaseShAmt) {
13001 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
13002 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
13003 SDValue InVec = ShAmtOp.getOperand(0);
13004 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
13005 unsigned NumElts = InVec.getValueType().getVectorNumElements();
13007 for (; i != NumElts; ++i) {
13008 SDValue Arg = InVec.getOperand(i);
13009 if (Arg.getOpcode() == ISD::UNDEF) continue;
13013 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
13014 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
13015 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
13016 if (C->getZExtValue() == SplatIdx)
13017 BaseShAmt = InVec.getOperand(1);
13020 if (BaseShAmt.getNode() == 0)
13021 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
13022 DAG.getIntPtrConstant(0));
13026 // The shift amount is an i32.
13027 if (EltVT.bitsGT(MVT::i32))
13028 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
13029 else if (EltVT.bitsLT(MVT::i32))
13030 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
13032 // The shift amount is identical so we can do a vector shift.
13033 SDValue ValOp = N->getOperand(0);
13034 switch (N->getOpcode()) {
13036 llvm_unreachable("Unknown shift opcode!");
13039 if (VT == MVT::v2i64)
13040 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13041 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
13043 if (VT == MVT::v4i32)
13044 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13045 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
13047 if (VT == MVT::v8i16)
13048 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13049 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
13053 if (VT == MVT::v4i32)
13054 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13055 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
13057 if (VT == MVT::v8i16)
13058 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13059 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
13063 if (VT == MVT::v2i64)
13064 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13065 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
13067 if (VT == MVT::v4i32)
13068 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13069 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
13071 if (VT == MVT::v8i16)
13072 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
13073 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
13081 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
13082 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
13083 // and friends. Likewise for OR -> CMPNEQSS.
13084 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
13085 TargetLowering::DAGCombinerInfo &DCI,
13086 const X86Subtarget *Subtarget) {
13089 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
13090 // we're requiring SSE2 for both.
13091 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
13092 SDValue N0 = N->getOperand(0);
13093 SDValue N1 = N->getOperand(1);
13094 SDValue CMP0 = N0->getOperand(1);
13095 SDValue CMP1 = N1->getOperand(1);
13096 DebugLoc DL = N->getDebugLoc();
13098 // The SETCCs should both refer to the same CMP.
13099 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
13102 SDValue CMP00 = CMP0->getOperand(0);
13103 SDValue CMP01 = CMP0->getOperand(1);
13104 EVT VT = CMP00.getValueType();
13106 if (VT == MVT::f32 || VT == MVT::f64) {
13107 bool ExpectingFlags = false;
13108 // Check for any users that want flags:
13109 for (SDNode::use_iterator UI = N->use_begin(),
13111 !ExpectingFlags && UI != UE; ++UI)
13112 switch (UI->getOpcode()) {
13117 ExpectingFlags = true;
13119 case ISD::CopyToReg:
13120 case ISD::SIGN_EXTEND:
13121 case ISD::ZERO_EXTEND:
13122 case ISD::ANY_EXTEND:
13126 if (!ExpectingFlags) {
13127 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
13128 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
13130 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
13131 X86::CondCode tmp = cc0;
13136 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
13137 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
13138 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
13139 X86ISD::NodeType NTOperator = is64BitFP ?
13140 X86ISD::FSETCCsd : X86ISD::FSETCCss;
13141 // FIXME: need symbolic constants for these magic numbers.
13142 // See X86ATTInstPrinter.cpp:printSSECC().
13143 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
13144 SDValue OnesOrZeroesF = DAG.getNode(NTOperator, DL, MVT::f32, CMP00, CMP01,
13145 DAG.getConstant(x86cc, MVT::i8));
13146 SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, MVT::i32,
13148 SDValue ANDed = DAG.getNode(ISD::AND, DL, MVT::i32, OnesOrZeroesI,
13149 DAG.getConstant(1, MVT::i32));
13150 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
13151 return OneBitOfTruth;
13159 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
13160 /// so it can be folded inside ANDNP.
13161 static bool CanFoldXORWithAllOnes(const SDNode *N) {
13162 EVT VT = N->getValueType(0);
13164 // Match direct AllOnes for 128 and 256-bit vectors
13165 if (ISD::isBuildVectorAllOnes(N))
13168 // Look through a bit convert.
13169 if (N->getOpcode() == ISD::BITCAST)
13170 N = N->getOperand(0).getNode();
13172 // Sometimes the operand may come from a insert_subvector building a 256-bit
13174 if (VT.getSizeInBits() == 256 &&
13175 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
13176 SDValue V1 = N->getOperand(0);
13177 SDValue V2 = N->getOperand(1);
13179 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
13180 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
13181 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
13182 ISD::isBuildVectorAllOnes(V2.getNode()))
13189 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
13190 TargetLowering::DAGCombinerInfo &DCI,
13191 const X86Subtarget *Subtarget) {
13192 if (DCI.isBeforeLegalizeOps())
13195 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
13199 // Want to form ANDNP nodes:
13200 // 1) In the hopes of then easily combining them with OR and AND nodes
13201 // to form PBLEND/PSIGN.
13202 // 2) To match ANDN packed intrinsics
13203 EVT VT = N->getValueType(0);
13204 if (VT != MVT::v2i64 && VT != MVT::v4i64)
13207 SDValue N0 = N->getOperand(0);
13208 SDValue N1 = N->getOperand(1);
13209 DebugLoc DL = N->getDebugLoc();
13211 // Check LHS for vnot
13212 if (N0.getOpcode() == ISD::XOR &&
13213 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
13214 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
13215 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
13217 // Check RHS for vnot
13218 if (N1.getOpcode() == ISD::XOR &&
13219 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
13220 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
13221 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
13226 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
13227 TargetLowering::DAGCombinerInfo &DCI,
13228 const X86Subtarget *Subtarget) {
13229 if (DCI.isBeforeLegalizeOps())
13232 SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget);
13236 EVT VT = N->getValueType(0);
13237 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
13240 SDValue N0 = N->getOperand(0);
13241 SDValue N1 = N->getOperand(1);
13243 // look for psign/blend
13244 if (Subtarget->hasSSSE3()) {
13245 if (VT == MVT::v2i64) {
13246 // Canonicalize pandn to RHS
13247 if (N0.getOpcode() == X86ISD::ANDNP)
13249 // or (and (m, x), (pandn m, y))
13250 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
13251 SDValue Mask = N1.getOperand(0);
13252 SDValue X = N1.getOperand(1);
13254 if (N0.getOperand(0) == Mask)
13255 Y = N0.getOperand(1);
13256 if (N0.getOperand(1) == Mask)
13257 Y = N0.getOperand(0);
13259 // Check to see if the mask appeared in both the AND and ANDNP and
13263 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
13264 if (Mask.getOpcode() != ISD::BITCAST ||
13265 X.getOpcode() != ISD::BITCAST ||
13266 Y.getOpcode() != ISD::BITCAST)
13269 // Look through mask bitcast.
13270 Mask = Mask.getOperand(0);
13271 EVT MaskVT = Mask.getValueType();
13273 // Validate that the Mask operand is a vector sra node. The sra node
13274 // will be an intrinsic.
13275 if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
13278 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
13279 // there is no psrai.b
13280 switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
13281 case Intrinsic::x86_sse2_psrai_w:
13282 case Intrinsic::x86_sse2_psrai_d:
13284 default: return SDValue();
13287 // Check that the SRA is all signbits.
13288 SDValue SraC = Mask.getOperand(2);
13289 unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
13290 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
13291 if ((SraAmt + 1) != EltBits)
13294 DebugLoc DL = N->getDebugLoc();
13296 // Now we know we at least have a plendvb with the mask val. See if
13297 // we can form a psignb/w/d.
13298 // psign = x.type == y.type == mask.type && y = sub(0, x);
13299 X = X.getOperand(0);
13300 Y = Y.getOperand(0);
13301 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
13302 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
13303 X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
13306 case 8: Opc = X86ISD::PSIGNB; break;
13307 case 16: Opc = X86ISD::PSIGNW; break;
13308 case 32: Opc = X86ISD::PSIGND; break;
13312 SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
13313 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
13316 // PBLENDVB only available on SSE 4.1
13317 if (!Subtarget->hasSSE41())
13320 X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
13321 Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
13322 Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
13323 Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
13324 return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
13329 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
13330 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
13332 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
13334 if (!N0.hasOneUse() || !N1.hasOneUse())
13337 SDValue ShAmt0 = N0.getOperand(1);
13338 if (ShAmt0.getValueType() != MVT::i8)
13340 SDValue ShAmt1 = N1.getOperand(1);
13341 if (ShAmt1.getValueType() != MVT::i8)
13343 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
13344 ShAmt0 = ShAmt0.getOperand(0);
13345 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
13346 ShAmt1 = ShAmt1.getOperand(0);
13348 DebugLoc DL = N->getDebugLoc();
13349 unsigned Opc = X86ISD::SHLD;
13350 SDValue Op0 = N0.getOperand(0);
13351 SDValue Op1 = N1.getOperand(0);
13352 if (ShAmt0.getOpcode() == ISD::SUB) {
13353 Opc = X86ISD::SHRD;
13354 std::swap(Op0, Op1);
13355 std::swap(ShAmt0, ShAmt1);
13358 unsigned Bits = VT.getSizeInBits();
13359 if (ShAmt1.getOpcode() == ISD::SUB) {
13360 SDValue Sum = ShAmt1.getOperand(0);
13361 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
13362 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
13363 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
13364 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
13365 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
13366 return DAG.getNode(Opc, DL, VT,
13368 DAG.getNode(ISD::TRUNCATE, DL,
13371 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
13372 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
13374 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
13375 return DAG.getNode(Opc, DL, VT,
13376 N0.getOperand(0), N1.getOperand(0),
13377 DAG.getNode(ISD::TRUNCATE, DL,
13384 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
13385 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
13386 const X86Subtarget *Subtarget) {
13387 StoreSDNode *St = cast<StoreSDNode>(N);
13388 EVT VT = St->getValue().getValueType();
13389 EVT StVT = St->getMemoryVT();
13390 DebugLoc dl = St->getDebugLoc();
13391 SDValue StoredVal = St->getOperand(1);
13392 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13394 // If we are saving a concatination of two XMM registers, perform two stores.
13395 // This is better in Sandy Bridge cause one 256-bit mem op is done via two
13396 // 128-bit ones. If in the future the cost becomes only one memory access the
13397 // first version would be better.
13398 if (VT.getSizeInBits() == 256 &&
13399 StoredVal.getNode()->getOpcode() == ISD::CONCAT_VECTORS &&
13400 StoredVal.getNumOperands() == 2) {
13402 SDValue Value0 = StoredVal.getOperand(0);
13403 SDValue Value1 = StoredVal.getOperand(1);
13405 SDValue Stride = DAG.getConstant(16, TLI.getPointerTy());
13406 SDValue Ptr0 = St->getBasePtr();
13407 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
13409 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
13410 St->getPointerInfo(), St->isVolatile(),
13411 St->isNonTemporal(), St->getAlignment());
13412 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
13413 St->getPointerInfo(), St->isVolatile(),
13414 St->isNonTemporal(), St->getAlignment());
13415 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
13418 // Optimize trunc store (of multiple scalars) to shuffle and store.
13419 // First, pack all of the elements in one place. Next, store to memory
13420 // in fewer chunks.
13421 if (St->isTruncatingStore() && VT.isVector()) {
13422 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13423 unsigned NumElems = VT.getVectorNumElements();
13424 assert(StVT != VT && "Cannot truncate to the same type");
13425 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
13426 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
13428 // From, To sizes and ElemCount must be pow of two
13429 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
13430 // We are going to use the original vector elt for storing.
13431 // accumulated smaller vector elements must be a multiple of bigger size.
13432 if (0 != (NumElems * ToSz) % FromSz) return SDValue();
13433 unsigned SizeRatio = FromSz / ToSz;
13435 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
13437 // Create a type on which we perform the shuffle
13438 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
13439 StVT.getScalarType(), NumElems*SizeRatio);
13441 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
13443 SDValue WideVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, St->getValue());
13444 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
13445 for (unsigned i = 0; i < NumElems; i++ ) ShuffleVec[i] = i * SizeRatio;
13447 // Can't shuffle using an illegal type
13448 if (!TLI.isTypeLegal(WideVecVT)) return SDValue();
13450 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
13451 DAG.getUNDEF(WideVec.getValueType()),
13452 ShuffleVec.data());
13453 // At this point all of the data is stored at the bottom of the
13454 // register. We now need to save it to mem.
13456 // Find the largest store unit
13457 MVT StoreType = MVT::i8;
13458 for (unsigned tp = MVT::FIRST_INTEGER_VALUETYPE;
13459 tp < MVT::LAST_INTEGER_VALUETYPE; ++tp) {
13460 MVT Tp = (MVT::SimpleValueType)tp;
13461 if (TLI.isTypeLegal(Tp) && StoreType.getSizeInBits() < NumElems * ToSz)
13465 // Bitcast the original vector into a vector of store-size units
13466 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
13467 StoreType, VT.getSizeInBits()/EVT(StoreType).getSizeInBits());
13468 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
13469 SDValue ShuffWide = DAG.getNode(ISD::BITCAST, dl, StoreVecVT, Shuff);
13470 SmallVector<SDValue, 8> Chains;
13471 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits()/8,
13472 TLI.getPointerTy());
13473 SDValue Ptr = St->getBasePtr();
13475 // Perform one or more big stores into memory.
13476 for (unsigned i = 0; i < (ToSz*NumElems)/StoreType.getSizeInBits() ; i++) {
13477 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
13478 StoreType, ShuffWide,
13479 DAG.getIntPtrConstant(i));
13480 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
13481 St->getPointerInfo(), St->isVolatile(),
13482 St->isNonTemporal(), St->getAlignment());
13483 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
13484 Chains.push_back(Ch);
13487 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0],
13492 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
13493 // the FP state in cases where an emms may be missing.
13494 // A preferable solution to the general problem is to figure out the right
13495 // places to insert EMMS. This qualifies as a quick hack.
13497 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
13498 if (VT.getSizeInBits() != 64)
13501 const Function *F = DAG.getMachineFunction().getFunction();
13502 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
13503 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
13504 && Subtarget->hasSSE2();
13505 if ((VT.isVector() ||
13506 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
13507 isa<LoadSDNode>(St->getValue()) &&
13508 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
13509 St->getChain().hasOneUse() && !St->isVolatile()) {
13510 SDNode* LdVal = St->getValue().getNode();
13511 LoadSDNode *Ld = 0;
13512 int TokenFactorIndex = -1;
13513 SmallVector<SDValue, 8> Ops;
13514 SDNode* ChainVal = St->getChain().getNode();
13515 // Must be a store of a load. We currently handle two cases: the load
13516 // is a direct child, and it's under an intervening TokenFactor. It is
13517 // possible to dig deeper under nested TokenFactors.
13518 if (ChainVal == LdVal)
13519 Ld = cast<LoadSDNode>(St->getChain());
13520 else if (St->getValue().hasOneUse() &&
13521 ChainVal->getOpcode() == ISD::TokenFactor) {
13522 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
13523 if (ChainVal->getOperand(i).getNode() == LdVal) {
13524 TokenFactorIndex = i;
13525 Ld = cast<LoadSDNode>(St->getValue());
13527 Ops.push_back(ChainVal->getOperand(i));
13531 if (!Ld || !ISD::isNormalLoad(Ld))
13534 // If this is not the MMX case, i.e. we are just turning i64 load/store
13535 // into f64 load/store, avoid the transformation if there are multiple
13536 // uses of the loaded value.
13537 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
13540 DebugLoc LdDL = Ld->getDebugLoc();
13541 DebugLoc StDL = N->getDebugLoc();
13542 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
13543 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
13545 if (Subtarget->is64Bit() || F64IsLegal) {
13546 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
13547 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
13548 Ld->getPointerInfo(), Ld->isVolatile(),
13549 Ld->isNonTemporal(), Ld->getAlignment());
13550 SDValue NewChain = NewLd.getValue(1);
13551 if (TokenFactorIndex != -1) {
13552 Ops.push_back(NewChain);
13553 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
13556 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
13557 St->getPointerInfo(),
13558 St->isVolatile(), St->isNonTemporal(),
13559 St->getAlignment());
13562 // Otherwise, lower to two pairs of 32-bit loads / stores.
13563 SDValue LoAddr = Ld->getBasePtr();
13564 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
13565 DAG.getConstant(4, MVT::i32));
13567 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
13568 Ld->getPointerInfo(),
13569 Ld->isVolatile(), Ld->isNonTemporal(),
13570 Ld->getAlignment());
13571 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
13572 Ld->getPointerInfo().getWithOffset(4),
13573 Ld->isVolatile(), Ld->isNonTemporal(),
13574 MinAlign(Ld->getAlignment(), 4));
13576 SDValue NewChain = LoLd.getValue(1);
13577 if (TokenFactorIndex != -1) {
13578 Ops.push_back(LoLd);
13579 Ops.push_back(HiLd);
13580 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
13584 LoAddr = St->getBasePtr();
13585 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
13586 DAG.getConstant(4, MVT::i32));
13588 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
13589 St->getPointerInfo(),
13590 St->isVolatile(), St->isNonTemporal(),
13591 St->getAlignment());
13592 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
13593 St->getPointerInfo().getWithOffset(4),
13595 St->isNonTemporal(),
13596 MinAlign(St->getAlignment(), 4));
13597 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
13602 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
13603 /// X86ISD::FXOR nodes.
13604 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
13605 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
13606 // F[X]OR(0.0, x) -> x
13607 // F[X]OR(x, 0.0) -> x
13608 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
13609 if (C->getValueAPF().isPosZero())
13610 return N->getOperand(1);
13611 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
13612 if (C->getValueAPF().isPosZero())
13613 return N->getOperand(0);
13617 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
13618 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
13619 // FAND(0.0, x) -> 0.0
13620 // FAND(x, 0.0) -> 0.0
13621 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
13622 if (C->getValueAPF().isPosZero())
13623 return N->getOperand(0);
13624 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
13625 if (C->getValueAPF().isPosZero())
13626 return N->getOperand(1);
13630 static SDValue PerformBTCombine(SDNode *N,
13632 TargetLowering::DAGCombinerInfo &DCI) {
13633 // BT ignores high bits in the bit index operand.
13634 SDValue Op1 = N->getOperand(1);
13635 if (Op1.hasOneUse()) {
13636 unsigned BitWidth = Op1.getValueSizeInBits();
13637 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
13638 APInt KnownZero, KnownOne;
13639 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
13640 !DCI.isBeforeLegalizeOps());
13641 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13642 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
13643 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
13644 DCI.CommitTargetLoweringOpt(TLO);
13649 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
13650 SDValue Op = N->getOperand(0);
13651 if (Op.getOpcode() == ISD::BITCAST)
13652 Op = Op.getOperand(0);
13653 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
13654 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
13655 VT.getVectorElementType().getSizeInBits() ==
13656 OpVT.getVectorElementType().getSizeInBits()) {
13657 return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
13662 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
13663 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
13664 // (and (i32 x86isd::setcc_carry), 1)
13665 // This eliminates the zext. This transformation is necessary because
13666 // ISD::SETCC is always legalized to i8.
13667 DebugLoc dl = N->getDebugLoc();
13668 SDValue N0 = N->getOperand(0);
13669 EVT VT = N->getValueType(0);
13670 if (N0.getOpcode() == ISD::AND &&
13672 N0.getOperand(0).hasOneUse()) {
13673 SDValue N00 = N0.getOperand(0);
13674 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
13676 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
13677 if (!C || C->getZExtValue() != 1)
13679 return DAG.getNode(ISD::AND, dl, VT,
13680 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
13681 N00.getOperand(0), N00.getOperand(1)),
13682 DAG.getConstant(1, VT));
13688 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
13689 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
13690 unsigned X86CC = N->getConstantOperandVal(0);
13691 SDValue EFLAG = N->getOperand(1);
13692 DebugLoc DL = N->getDebugLoc();
13694 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
13695 // a zext and produces an all-ones bit which is more useful than 0/1 in some
13697 if (X86CC == X86::COND_B)
13698 return DAG.getNode(ISD::AND, DL, MVT::i8,
13699 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
13700 DAG.getConstant(X86CC, MVT::i8), EFLAG),
13701 DAG.getConstant(1, MVT::i8));
13706 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
13707 const X86TargetLowering *XTLI) {
13708 SDValue Op0 = N->getOperand(0);
13709 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
13710 // a 32-bit target where SSE doesn't support i64->FP operations.
13711 if (Op0.getOpcode() == ISD::LOAD) {
13712 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
13713 EVT VT = Ld->getValueType(0);
13714 if (!Ld->isVolatile() && !N->getValueType(0).isVector() &&
13715 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
13716 !XTLI->getSubtarget()->is64Bit() &&
13717 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
13718 SDValue FILDChain = XTLI->BuildFILD(SDValue(N, 0), Ld->getValueType(0),
13719 Ld->getChain(), Op0, DAG);
13720 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
13727 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
13728 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
13729 X86TargetLowering::DAGCombinerInfo &DCI) {
13730 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
13731 // the result is either zero or one (depending on the input carry bit).
13732 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
13733 if (X86::isZeroNode(N->getOperand(0)) &&
13734 X86::isZeroNode(N->getOperand(1)) &&
13735 // We don't have a good way to replace an EFLAGS use, so only do this when
13737 SDValue(N, 1).use_empty()) {
13738 DebugLoc DL = N->getDebugLoc();
13739 EVT VT = N->getValueType(0);
13740 SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
13741 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
13742 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
13743 DAG.getConstant(X86::COND_B,MVT::i8),
13745 DAG.getConstant(1, VT));
13746 return DCI.CombineTo(N, Res1, CarryOut);
13752 // fold (add Y, (sete X, 0)) -> adc 0, Y
13753 // (add Y, (setne X, 0)) -> sbb -1, Y
13754 // (sub (sete X, 0), Y) -> sbb 0, Y
13755 // (sub (setne X, 0), Y) -> adc -1, Y
13756 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
13757 DebugLoc DL = N->getDebugLoc();
13759 // Look through ZExts.
13760 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
13761 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
13764 SDValue SetCC = Ext.getOperand(0);
13765 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
13768 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
13769 if (CC != X86::COND_E && CC != X86::COND_NE)
13772 SDValue Cmp = SetCC.getOperand(1);
13773 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
13774 !X86::isZeroNode(Cmp.getOperand(1)) ||
13775 !Cmp.getOperand(0).getValueType().isInteger())
13778 SDValue CmpOp0 = Cmp.getOperand(0);
13779 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
13780 DAG.getConstant(1, CmpOp0.getValueType()));
13782 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
13783 if (CC == X86::COND_NE)
13784 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
13785 DL, OtherVal.getValueType(), OtherVal,
13786 DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
13787 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
13788 DL, OtherVal.getValueType(), OtherVal,
13789 DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
13792 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG) {
13793 SDValue Op0 = N->getOperand(0);
13794 SDValue Op1 = N->getOperand(1);
13796 // X86 can't encode an immediate LHS of a sub. See if we can push the
13797 // negation into a preceding instruction.
13798 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
13799 // If the RHS of the sub is a XOR with one use and a constant, invert the
13800 // immediate. Then add one to the LHS of the sub so we can turn
13801 // X-Y -> X+~Y+1, saving one register.
13802 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
13803 isa<ConstantSDNode>(Op1.getOperand(1))) {
13804 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
13805 EVT VT = Op0.getValueType();
13806 SDValue NewXor = DAG.getNode(ISD::XOR, Op1.getDebugLoc(), VT,
13808 DAG.getConstant(~XorC, VT));
13809 return DAG.getNode(ISD::ADD, N->getDebugLoc(), VT, NewXor,
13810 DAG.getConstant(C->getAPIntValue()+1, VT));
13814 return OptimizeConditionalInDecrement(N, DAG);
13817 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
13818 DAGCombinerInfo &DCI) const {
13819 SelectionDAG &DAG = DCI.DAG;
13820 switch (N->getOpcode()) {
13822 case ISD::EXTRACT_VECTOR_ELT:
13823 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
13824 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
13825 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
13826 case ISD::ADD: return OptimizeConditionalInDecrement(N, DAG);
13827 case ISD::SUB: return PerformSubCombine(N, DAG);
13828 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
13829 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
13832 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
13833 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
13834 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
13835 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
13836 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, this);
13838 case X86ISD::FOR: return PerformFORCombine(N, DAG);
13839 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
13840 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
13841 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
13842 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
13843 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
13844 case X86ISD::SHUFPS: // Handle all target specific shuffles
13845 case X86ISD::SHUFPD:
13846 case X86ISD::PALIGN:
13847 case X86ISD::PUNPCKHBW:
13848 case X86ISD::PUNPCKHWD:
13849 case X86ISD::PUNPCKHDQ:
13850 case X86ISD::PUNPCKHQDQ:
13851 case X86ISD::UNPCKHPS:
13852 case X86ISD::UNPCKHPD:
13853 case X86ISD::VUNPCKHPSY:
13854 case X86ISD::VUNPCKHPDY:
13855 case X86ISD::PUNPCKLBW:
13856 case X86ISD::PUNPCKLWD:
13857 case X86ISD::PUNPCKLDQ:
13858 case X86ISD::PUNPCKLQDQ:
13859 case X86ISD::UNPCKLPS:
13860 case X86ISD::UNPCKLPD:
13861 case X86ISD::VUNPCKLPSY:
13862 case X86ISD::VUNPCKLPDY:
13863 case X86ISD::MOVHLPS:
13864 case X86ISD::MOVLHPS:
13865 case X86ISD::PSHUFD:
13866 case X86ISD::PSHUFHW:
13867 case X86ISD::PSHUFLW:
13868 case X86ISD::MOVSS:
13869 case X86ISD::MOVSD:
13870 case X86ISD::VPERMILPS:
13871 case X86ISD::VPERMILPSY:
13872 case X86ISD::VPERMILPD:
13873 case X86ISD::VPERMILPDY:
13874 case X86ISD::VPERM2F128:
13875 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
13881 /// isTypeDesirableForOp - Return true if the target has native support for
13882 /// the specified value type and it is 'desirable' to use the type for the
13883 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
13884 /// instruction encodings are longer and some i16 instructions are slow.
13885 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
13886 if (!isTypeLegal(VT))
13888 if (VT != MVT::i16)
13895 case ISD::SIGN_EXTEND:
13896 case ISD::ZERO_EXTEND:
13897 case ISD::ANY_EXTEND:
13910 /// IsDesirableToPromoteOp - This method query the target whether it is
13911 /// beneficial for dag combiner to promote the specified node. If true, it
13912 /// should return the desired promotion type by reference.
13913 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
13914 EVT VT = Op.getValueType();
13915 if (VT != MVT::i16)
13918 bool Promote = false;
13919 bool Commute = false;
13920 switch (Op.getOpcode()) {
13923 LoadSDNode *LD = cast<LoadSDNode>(Op);
13924 // If the non-extending load has a single use and it's not live out, then it
13925 // might be folded.
13926 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
13927 Op.hasOneUse()*/) {
13928 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13929 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
13930 // The only case where we'd want to promote LOAD (rather then it being
13931 // promoted as an operand is when it's only use is liveout.
13932 if (UI->getOpcode() != ISD::CopyToReg)
13939 case ISD::SIGN_EXTEND:
13940 case ISD::ZERO_EXTEND:
13941 case ISD::ANY_EXTEND:
13946 SDValue N0 = Op.getOperand(0);
13947 // Look out for (store (shl (load), x)).
13948 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
13961 SDValue N0 = Op.getOperand(0);
13962 SDValue N1 = Op.getOperand(1);
13963 if (!Commute && MayFoldLoad(N1))
13965 // Avoid disabling potential load folding opportunities.
13966 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
13968 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
13978 //===----------------------------------------------------------------------===//
13979 // X86 Inline Assembly Support
13980 //===----------------------------------------------------------------------===//
13982 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
13983 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
13985 std::string AsmStr = IA->getAsmString();
13987 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
13988 SmallVector<StringRef, 4> AsmPieces;
13989 SplitString(AsmStr, AsmPieces, ";\n");
13991 switch (AsmPieces.size()) {
13992 default: return false;
13994 AsmStr = AsmPieces[0];
13996 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
13998 // FIXME: this should verify that we are targeting a 486 or better. If not,
13999 // we will turn this bswap into something that will be lowered to logical ops
14000 // instead of emitting the bswap asm. For now, we don't support 486 or lower
14001 // so don't worry about this.
14003 if (AsmPieces.size() == 2 &&
14004 (AsmPieces[0] == "bswap" ||
14005 AsmPieces[0] == "bswapq" ||
14006 AsmPieces[0] == "bswapl") &&
14007 (AsmPieces[1] == "$0" ||
14008 AsmPieces[1] == "${0:q}")) {
14009 // No need to check constraints, nothing other than the equivalent of
14010 // "=r,0" would be valid here.
14011 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
14012 if (!Ty || Ty->getBitWidth() % 16 != 0)
14014 return IntrinsicLowering::LowerToByteSwap(CI);
14016 // rorw $$8, ${0:w} --> llvm.bswap.i16
14017 if (CI->getType()->isIntegerTy(16) &&
14018 AsmPieces.size() == 3 &&
14019 (AsmPieces[0] == "rorw" || AsmPieces[0] == "rolw") &&
14020 AsmPieces[1] == "$$8," &&
14021 AsmPieces[2] == "${0:w}" &&
14022 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
14024 const std::string &ConstraintsStr = IA->getConstraintString();
14025 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
14026 std::sort(AsmPieces.begin(), AsmPieces.end());
14027 if (AsmPieces.size() == 4 &&
14028 AsmPieces[0] == "~{cc}" &&
14029 AsmPieces[1] == "~{dirflag}" &&
14030 AsmPieces[2] == "~{flags}" &&
14031 AsmPieces[3] == "~{fpsr}") {
14032 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
14033 if (!Ty || Ty->getBitWidth() % 16 != 0)
14035 return IntrinsicLowering::LowerToByteSwap(CI);
14040 if (CI->getType()->isIntegerTy(32) &&
14041 IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
14042 SmallVector<StringRef, 4> Words;
14043 SplitString(AsmPieces[0], Words, " \t,");
14044 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
14045 Words[2] == "${0:w}") {
14047 SplitString(AsmPieces[1], Words, " \t,");
14048 if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
14049 Words[2] == "$0") {
14051 SplitString(AsmPieces[2], Words, " \t,");
14052 if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
14053 Words[2] == "${0:w}") {
14055 const std::string &ConstraintsStr = IA->getConstraintString();
14056 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
14057 std::sort(AsmPieces.begin(), AsmPieces.end());
14058 if (AsmPieces.size() == 4 &&
14059 AsmPieces[0] == "~{cc}" &&
14060 AsmPieces[1] == "~{dirflag}" &&
14061 AsmPieces[2] == "~{flags}" &&
14062 AsmPieces[3] == "~{fpsr}") {
14063 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
14064 if (!Ty || Ty->getBitWidth() % 16 != 0)
14066 return IntrinsicLowering::LowerToByteSwap(CI);
14073 if (CI->getType()->isIntegerTy(64)) {
14074 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
14075 if (Constraints.size() >= 2 &&
14076 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
14077 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
14078 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
14079 SmallVector<StringRef, 4> Words;
14080 SplitString(AsmPieces[0], Words, " \t");
14081 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
14083 SplitString(AsmPieces[1], Words, " \t");
14084 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
14086 SplitString(AsmPieces[2], Words, " \t,");
14087 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
14088 Words[2] == "%edx") {
14089 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
14090 if (!Ty || Ty->getBitWidth() % 16 != 0)
14092 return IntrinsicLowering::LowerToByteSwap(CI);
14105 /// getConstraintType - Given a constraint letter, return the type of
14106 /// constraint it is for this target.
14107 X86TargetLowering::ConstraintType
14108 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
14109 if (Constraint.size() == 1) {
14110 switch (Constraint[0]) {
14121 return C_RegisterClass;
14145 return TargetLowering::getConstraintType(Constraint);
14148 /// Examine constraint type and operand type and determine a weight value.
14149 /// This object must already have been set up with the operand type
14150 /// and the current alternative constraint selected.
14151 TargetLowering::ConstraintWeight
14152 X86TargetLowering::getSingleConstraintMatchWeight(
14153 AsmOperandInfo &info, const char *constraint) const {
14154 ConstraintWeight weight = CW_Invalid;
14155 Value *CallOperandVal = info.CallOperandVal;
14156 // If we don't have a value, we can't do a match,
14157 // but allow it at the lowest weight.
14158 if (CallOperandVal == NULL)
14160 Type *type = CallOperandVal->getType();
14161 // Look at the constraint type.
14162 switch (*constraint) {
14164 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
14175 if (CallOperandVal->getType()->isIntegerTy())
14176 weight = CW_SpecificReg;
14181 if (type->isFloatingPointTy())
14182 weight = CW_SpecificReg;
14185 if (type->isX86_MMXTy() && Subtarget->hasMMX())
14186 weight = CW_SpecificReg;
14190 if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
14191 weight = CW_Register;
14194 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
14195 if (C->getZExtValue() <= 31)
14196 weight = CW_Constant;
14200 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14201 if (C->getZExtValue() <= 63)
14202 weight = CW_Constant;
14206 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14207 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
14208 weight = CW_Constant;
14212 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14213 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
14214 weight = CW_Constant;
14218 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14219 if (C->getZExtValue() <= 3)
14220 weight = CW_Constant;
14224 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14225 if (C->getZExtValue() <= 0xff)
14226 weight = CW_Constant;
14231 if (dyn_cast<ConstantFP>(CallOperandVal)) {
14232 weight = CW_Constant;
14236 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14237 if ((C->getSExtValue() >= -0x80000000LL) &&
14238 (C->getSExtValue() <= 0x7fffffffLL))
14239 weight = CW_Constant;
14243 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
14244 if (C->getZExtValue() <= 0xffffffff)
14245 weight = CW_Constant;
14252 /// LowerXConstraint - try to replace an X constraint, which matches anything,
14253 /// with another that has more specific requirements based on the type of the
14254 /// corresponding operand.
14255 const char *X86TargetLowering::
14256 LowerXConstraint(EVT ConstraintVT) const {
14257 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
14258 // 'f' like normal targets.
14259 if (ConstraintVT.isFloatingPoint()) {
14260 if (Subtarget->hasXMMInt())
14262 if (Subtarget->hasXMM())
14266 return TargetLowering::LowerXConstraint(ConstraintVT);
14269 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
14270 /// vector. If it is invalid, don't add anything to Ops.
14271 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
14272 std::string &Constraint,
14273 std::vector<SDValue>&Ops,
14274 SelectionDAG &DAG) const {
14275 SDValue Result(0, 0);
14277 // Only support length 1 constraints for now.
14278 if (Constraint.length() > 1) return;
14280 char ConstraintLetter = Constraint[0];
14281 switch (ConstraintLetter) {
14284 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
14285 if (C->getZExtValue() <= 31) {
14286 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
14292 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
14293 if (C->getZExtValue() <= 63) {
14294 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
14300 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
14301 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
14302 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
14308 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
14309 if (C->getZExtValue() <= 255) {
14310 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
14316 // 32-bit signed value
14317 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
14318 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
14319 C->getSExtValue())) {
14320 // Widen to 64 bits here to get it sign extended.
14321 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
14324 // FIXME gcc accepts some relocatable values here too, but only in certain
14325 // memory models; it's complicated.
14330 // 32-bit unsigned value
14331 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
14332 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
14333 C->getZExtValue())) {
14334 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
14338 // FIXME gcc accepts some relocatable values here too, but only in certain
14339 // memory models; it's complicated.
14343 // Literal immediates are always ok.
14344 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
14345 // Widen to 64 bits here to get it sign extended.
14346 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
14350 // In any sort of PIC mode addresses need to be computed at runtime by
14351 // adding in a register or some sort of table lookup. These can't
14352 // be used as immediates.
14353 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
14356 // If we are in non-pic codegen mode, we allow the address of a global (with
14357 // an optional displacement) to be used with 'i'.
14358 GlobalAddressSDNode *GA = 0;
14359 int64_t Offset = 0;
14361 // Match either (GA), (GA+C), (GA+C1+C2), etc.
14363 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
14364 Offset += GA->getOffset();
14366 } else if (Op.getOpcode() == ISD::ADD) {
14367 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
14368 Offset += C->getZExtValue();
14369 Op = Op.getOperand(0);
14372 } else if (Op.getOpcode() == ISD::SUB) {
14373 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
14374 Offset += -C->getZExtValue();
14375 Op = Op.getOperand(0);
14380 // Otherwise, this isn't something we can handle, reject it.
14384 const GlobalValue *GV = GA->getGlobal();
14385 // If we require an extra load to get this address, as in PIC mode, we
14386 // can't accept it.
14387 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
14388 getTargetMachine())))
14391 Result = DAG.getTargetGlobalAddress(GV, Op.getDebugLoc(),
14392 GA->getValueType(0), Offset);
14397 if (Result.getNode()) {
14398 Ops.push_back(Result);
14401 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
14404 std::pair<unsigned, const TargetRegisterClass*>
14405 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
14407 // First, see if this is a constraint that directly corresponds to an LLVM
14409 if (Constraint.size() == 1) {
14410 // GCC Constraint Letters
14411 switch (Constraint[0]) {
14413 // TODO: Slight differences here in allocation order and leaving
14414 // RIP in the class. Do they matter any more here than they do
14415 // in the normal allocation?
14416 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
14417 if (Subtarget->is64Bit()) {
14418 if (VT == MVT::i32 || VT == MVT::f32)
14419 return std::make_pair(0U, X86::GR32RegisterClass);
14420 else if (VT == MVT::i16)
14421 return std::make_pair(0U, X86::GR16RegisterClass);
14422 else if (VT == MVT::i8 || VT == MVT::i1)
14423 return std::make_pair(0U, X86::GR8RegisterClass);
14424 else if (VT == MVT::i64 || VT == MVT::f64)
14425 return std::make_pair(0U, X86::GR64RegisterClass);
14428 // 32-bit fallthrough
14429 case 'Q': // Q_REGS
14430 if (VT == MVT::i32 || VT == MVT::f32)
14431 return std::make_pair(0U, X86::GR32_ABCDRegisterClass);
14432 else if (VT == MVT::i16)
14433 return std::make_pair(0U, X86::GR16_ABCDRegisterClass);
14434 else if (VT == MVT::i8 || VT == MVT::i1)
14435 return std::make_pair(0U, X86::GR8_ABCD_LRegisterClass);
14436 else if (VT == MVT::i64)
14437 return std::make_pair(0U, X86::GR64_ABCDRegisterClass);
14439 case 'r': // GENERAL_REGS
14440 case 'l': // INDEX_REGS
14441 if (VT == MVT::i8 || VT == MVT::i1)
14442 return std::make_pair(0U, X86::GR8RegisterClass);
14443 if (VT == MVT::i16)
14444 return std::make_pair(0U, X86::GR16RegisterClass);
14445 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
14446 return std::make_pair(0U, X86::GR32RegisterClass);
14447 return std::make_pair(0U, X86::GR64RegisterClass);
14448 case 'R': // LEGACY_REGS
14449 if (VT == MVT::i8 || VT == MVT::i1)
14450 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
14451 if (VT == MVT::i16)
14452 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
14453 if (VT == MVT::i32 || !Subtarget->is64Bit())
14454 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
14455 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
14456 case 'f': // FP Stack registers.
14457 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
14458 // value to the correct fpstack register class.
14459 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
14460 return std::make_pair(0U, X86::RFP32RegisterClass);
14461 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
14462 return std::make_pair(0U, X86::RFP64RegisterClass);
14463 return std::make_pair(0U, X86::RFP80RegisterClass);
14464 case 'y': // MMX_REGS if MMX allowed.
14465 if (!Subtarget->hasMMX()) break;
14466 return std::make_pair(0U, X86::VR64RegisterClass);
14467 case 'Y': // SSE_REGS if SSE2 allowed
14468 if (!Subtarget->hasXMMInt()) break;
14470 case 'x': // SSE_REGS if SSE1 allowed
14471 if (!Subtarget->hasXMM()) break;
14473 switch (VT.getSimpleVT().SimpleTy) {
14475 // Scalar SSE types.
14478 return std::make_pair(0U, X86::FR32RegisterClass);
14481 return std::make_pair(0U, X86::FR64RegisterClass);
14489 return std::make_pair(0U, X86::VR128RegisterClass);
14495 // Use the default implementation in TargetLowering to convert the register
14496 // constraint into a member of a register class.
14497 std::pair<unsigned, const TargetRegisterClass*> Res;
14498 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
14500 // Not found as a standard register?
14501 if (Res.second == 0) {
14502 // Map st(0) -> st(7) -> ST0
14503 if (Constraint.size() == 7 && Constraint[0] == '{' &&
14504 tolower(Constraint[1]) == 's' &&
14505 tolower(Constraint[2]) == 't' &&
14506 Constraint[3] == '(' &&
14507 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
14508 Constraint[5] == ')' &&
14509 Constraint[6] == '}') {
14511 Res.first = X86::ST0+Constraint[4]-'0';
14512 Res.second = X86::RFP80RegisterClass;
14516 // GCC allows "st(0)" to be called just plain "st".
14517 if (StringRef("{st}").equals_lower(Constraint)) {
14518 Res.first = X86::ST0;
14519 Res.second = X86::RFP80RegisterClass;
14524 if (StringRef("{flags}").equals_lower(Constraint)) {
14525 Res.first = X86::EFLAGS;
14526 Res.second = X86::CCRRegisterClass;
14530 // 'A' means EAX + EDX.
14531 if (Constraint == "A") {
14532 Res.first = X86::EAX;
14533 Res.second = X86::GR32_ADRegisterClass;
14539 // Otherwise, check to see if this is a register class of the wrong value
14540 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
14541 // turn into {ax},{dx}.
14542 if (Res.second->hasType(VT))
14543 return Res; // Correct type already, nothing to do.
14545 // All of the single-register GCC register classes map their values onto
14546 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
14547 // really want an 8-bit or 32-bit register, map to the appropriate register
14548 // class and return the appropriate register.
14549 if (Res.second == X86::GR16RegisterClass) {
14550 if (VT == MVT::i8) {
14551 unsigned DestReg = 0;
14552 switch (Res.first) {
14554 case X86::AX: DestReg = X86::AL; break;
14555 case X86::DX: DestReg = X86::DL; break;
14556 case X86::CX: DestReg = X86::CL; break;
14557 case X86::BX: DestReg = X86::BL; break;
14560 Res.first = DestReg;
14561 Res.second = X86::GR8RegisterClass;
14563 } else if (VT == MVT::i32) {
14564 unsigned DestReg = 0;
14565 switch (Res.first) {
14567 case X86::AX: DestReg = X86::EAX; break;
14568 case X86::DX: DestReg = X86::EDX; break;
14569 case X86::CX: DestReg = X86::ECX; break;
14570 case X86::BX: DestReg = X86::EBX; break;
14571 case X86::SI: DestReg = X86::ESI; break;
14572 case X86::DI: DestReg = X86::EDI; break;
14573 case X86::BP: DestReg = X86::EBP; break;
14574 case X86::SP: DestReg = X86::ESP; break;
14577 Res.first = DestReg;
14578 Res.second = X86::GR32RegisterClass;
14580 } else if (VT == MVT::i64) {
14581 unsigned DestReg = 0;
14582 switch (Res.first) {
14584 case X86::AX: DestReg = X86::RAX; break;
14585 case X86::DX: DestReg = X86::RDX; break;
14586 case X86::CX: DestReg = X86::RCX; break;
14587 case X86::BX: DestReg = X86::RBX; break;
14588 case X86::SI: DestReg = X86::RSI; break;
14589 case X86::DI: DestReg = X86::RDI; break;
14590 case X86::BP: DestReg = X86::RBP; break;
14591 case X86::SP: DestReg = X86::RSP; break;
14594 Res.first = DestReg;
14595 Res.second = X86::GR64RegisterClass;
14598 } else if (Res.second == X86::FR32RegisterClass ||
14599 Res.second == X86::FR64RegisterClass ||
14600 Res.second == X86::VR128RegisterClass) {
14601 // Handle references to XMM physical registers that got mapped into the
14602 // wrong class. This can happen with constraints like {xmm0} where the
14603 // target independent register mapper will just pick the first match it can
14604 // find, ignoring the required type.
14605 if (VT == MVT::f32)
14606 Res.second = X86::FR32RegisterClass;
14607 else if (VT == MVT::f64)
14608 Res.second = X86::FR64RegisterClass;
14609 else if (X86::VR128RegisterClass->hasType(VT))
14610 Res.second = X86::VR128RegisterClass;