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 "X86MCTargetExpr.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.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/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineJumpTableInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/PseudoSourceValue.h"
38 #include "llvm/MC/MCAsmInfo.h"
39 #include "llvm/MC/MCContext.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/raw_ostream.h"
53 STATISTIC(NumTailCalls, "Number of tail calls");
56 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
58 // Disable16Bit - 16-bit operations typically have a larger encoding than
59 // corresponding 32-bit instructions, and 16-bit code is slow on some
60 // processors. This is an experimental flag to disable 16-bit operations
61 // (which forces them to be Legalized to 32-bit operations).
63 Disable16Bit("disable-16bit", cl::Hidden,
64 cl::desc("Disable use of 16-bit instructions"));
66 // Forward declarations.
67 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
70 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
71 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
72 default: llvm_unreachable("unknown subtarget type");
73 case X86Subtarget::isDarwin:
74 if (TM.getSubtarget<X86Subtarget>().is64Bit())
75 return new X8664_MachoTargetObjectFile();
76 return new TargetLoweringObjectFileMachO();
77 case X86Subtarget::isELF:
78 if (TM.getSubtarget<X86Subtarget>().is64Bit())
79 return new X8664_ELFTargetObjectFile(TM);
80 return new X8632_ELFTargetObjectFile(TM);
81 case X86Subtarget::isMingw:
82 case X86Subtarget::isCygwin:
83 case X86Subtarget::isWindows:
84 return new TargetLoweringObjectFileCOFF();
88 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
89 : TargetLowering(TM, createTLOF(TM)) {
90 Subtarget = &TM.getSubtarget<X86Subtarget>();
91 X86ScalarSSEf64 = Subtarget->hasSSE2();
92 X86ScalarSSEf32 = Subtarget->hasSSE1();
93 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
95 RegInfo = TM.getRegisterInfo();
98 // Set up the TargetLowering object.
100 // X86 is weird, it always uses i8 for shift amounts and setcc results.
101 setShiftAmountType(MVT::i8);
102 setBooleanContents(ZeroOrOneBooleanContent);
103 setSchedulingPreference(SchedulingForRegPressure);
104 setStackPointerRegisterToSaveRestore(X86StackPtr);
106 if (Subtarget->isTargetDarwin()) {
107 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
108 setUseUnderscoreSetJmp(false);
109 setUseUnderscoreLongJmp(false);
110 } else if (Subtarget->isTargetMingw()) {
111 // MS runtime is weird: it exports _setjmp, but longjmp!
112 setUseUnderscoreSetJmp(true);
113 setUseUnderscoreLongJmp(false);
115 setUseUnderscoreSetJmp(true);
116 setUseUnderscoreLongJmp(true);
119 // Set up the register classes.
120 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
122 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
123 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
124 if (Subtarget->is64Bit())
125 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
127 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
129 // We don't accept any truncstore of integer registers.
130 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
132 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
133 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
135 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
136 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
137 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
139 // SETOEQ and SETUNE require checking two conditions.
140 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
141 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
142 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
143 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
144 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
145 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
147 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
149 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
150 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
151 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
153 if (Subtarget->is64Bit()) {
154 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
155 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
156 } else if (!UseSoftFloat) {
157 if (X86ScalarSSEf64) {
158 // We have an impenetrably clever algorithm for ui64->double only.
159 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
161 // We have an algorithm for SSE2, and we turn this into a 64-bit
162 // FILD for other targets.
163 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
166 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
168 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
169 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
172 // SSE has no i16 to fp conversion, only i32
173 if (X86ScalarSSEf32) {
174 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
175 // f32 and f64 cases are Legal, f80 case is not
176 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
178 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
179 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
182 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
183 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
186 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
187 // are Legal, f80 is custom lowered.
188 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
189 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
191 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
193 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
194 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
196 if (X86ScalarSSEf32) {
197 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
198 // f32 and f64 cases are Legal, f80 case is not
199 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
201 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
202 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
205 // Handle FP_TO_UINT by promoting the destination to a larger signed
207 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
208 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
209 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
211 if (Subtarget->is64Bit()) {
212 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
213 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
214 } else if (!UseSoftFloat) {
215 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
216 // Expand FP_TO_UINT into a select.
217 // FIXME: We would like to use a Custom expander here eventually to do
218 // the optimal thing for SSE vs. the default expansion in the legalizer.
219 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
221 // With SSE3 we can use fisttpll to convert to a signed i64; without
222 // SSE, we're stuck with a fistpll.
223 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
226 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
227 if (!X86ScalarSSEf64) {
228 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
229 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
232 // Scalar integer divide and remainder are lowered to use operations that
233 // produce two results, to match the available instructions. This exposes
234 // the two-result form to trivial CSE, which is able to combine x/y and x%y
235 // into a single instruction.
237 // Scalar integer multiply-high is also lowered to use two-result
238 // operations, to match the available instructions. However, plain multiply
239 // (low) operations are left as Legal, as there are single-result
240 // instructions for this in x86. Using the two-result multiply instructions
241 // when both high and low results are needed must be arranged by dagcombine.
242 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
243 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
244 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
245 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
246 setOperationAction(ISD::SREM , MVT::i8 , Expand);
247 setOperationAction(ISD::UREM , MVT::i8 , Expand);
248 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
249 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
250 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
251 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
252 setOperationAction(ISD::SREM , MVT::i16 , Expand);
253 setOperationAction(ISD::UREM , MVT::i16 , Expand);
254 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
255 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
256 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
257 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
258 setOperationAction(ISD::SREM , MVT::i32 , Expand);
259 setOperationAction(ISD::UREM , MVT::i32 , Expand);
260 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
261 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
262 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
263 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
264 setOperationAction(ISD::SREM , MVT::i64 , Expand);
265 setOperationAction(ISD::UREM , MVT::i64 , Expand);
267 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
268 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
269 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
270 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
271 if (Subtarget->is64Bit())
272 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
273 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
274 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
275 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
276 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
277 setOperationAction(ISD::FREM , MVT::f32 , Expand);
278 setOperationAction(ISD::FREM , MVT::f64 , Expand);
279 setOperationAction(ISD::FREM , MVT::f80 , Expand);
280 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
282 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
283 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
284 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
285 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
287 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
288 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
290 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
291 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
293 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
294 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
295 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
296 if (Subtarget->is64Bit()) {
297 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
298 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
299 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
302 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
303 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
305 // These should be promoted to a larger select which is supported.
306 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
307 // X86 wants to expand cmov itself.
308 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
310 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
312 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
313 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
314 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
315 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
316 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
317 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
319 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
321 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
322 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
323 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
324 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
325 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
326 if (Subtarget->is64Bit()) {
327 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
328 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
330 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
333 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
334 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
335 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
336 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
337 if (Subtarget->is64Bit())
338 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
339 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
340 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
341 if (Subtarget->is64Bit()) {
342 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
343 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
344 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
345 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
346 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
348 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
349 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
350 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
351 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
352 if (Subtarget->is64Bit()) {
353 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
354 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
355 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
358 if (Subtarget->hasSSE1())
359 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
361 if (!Subtarget->hasSSE2())
362 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
364 // Expand certain atomics
365 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
366 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
367 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
368 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
373 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
375 if (!Subtarget->is64Bit()) {
376 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
377 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
378 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
379 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
380 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
381 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
382 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
385 // FIXME - use subtarget debug flags
386 if (!Subtarget->isTargetDarwin() &&
387 !Subtarget->isTargetELF() &&
388 !Subtarget->isTargetCygMing()) {
389 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
392 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
393 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
394 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
395 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
396 if (Subtarget->is64Bit()) {
397 setExceptionPointerRegister(X86::RAX);
398 setExceptionSelectorRegister(X86::RDX);
400 setExceptionPointerRegister(X86::EAX);
401 setExceptionSelectorRegister(X86::EDX);
403 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
404 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
406 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
408 setOperationAction(ISD::TRAP, MVT::Other, Legal);
410 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
411 setOperationAction(ISD::VASTART , MVT::Other, Custom);
412 setOperationAction(ISD::VAEND , MVT::Other, Expand);
413 if (Subtarget->is64Bit()) {
414 setOperationAction(ISD::VAARG , MVT::Other, Custom);
415 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
417 setOperationAction(ISD::VAARG , MVT::Other, Expand);
418 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
421 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
422 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
423 if (Subtarget->is64Bit())
424 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
425 if (Subtarget->isTargetCygMing())
426 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
428 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
430 if (!UseSoftFloat && X86ScalarSSEf64) {
431 // f32 and f64 use SSE.
432 // Set up the FP register classes.
433 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
434 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
436 // Use ANDPD to simulate FABS.
437 setOperationAction(ISD::FABS , MVT::f64, Custom);
438 setOperationAction(ISD::FABS , MVT::f32, Custom);
440 // Use XORP to simulate FNEG.
441 setOperationAction(ISD::FNEG , MVT::f64, Custom);
442 setOperationAction(ISD::FNEG , MVT::f32, Custom);
444 // Use ANDPD and ORPD to simulate FCOPYSIGN.
445 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
446 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
448 // We don't support sin/cos/fmod
449 setOperationAction(ISD::FSIN , MVT::f64, Expand);
450 setOperationAction(ISD::FCOS , MVT::f64, Expand);
451 setOperationAction(ISD::FSIN , MVT::f32, Expand);
452 setOperationAction(ISD::FCOS , MVT::f32, Expand);
454 // Expand FP immediates into loads from the stack, except for the special
456 addLegalFPImmediate(APFloat(+0.0)); // xorpd
457 addLegalFPImmediate(APFloat(+0.0f)); // xorps
458 } else if (!UseSoftFloat && X86ScalarSSEf32) {
459 // Use SSE for f32, x87 for f64.
460 // Set up the FP register classes.
461 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
462 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
464 // Use ANDPS to simulate FABS.
465 setOperationAction(ISD::FABS , MVT::f32, Custom);
467 // Use XORP to simulate FNEG.
468 setOperationAction(ISD::FNEG , MVT::f32, Custom);
470 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
472 // Use ANDPS and ORPS to simulate FCOPYSIGN.
473 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
474 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
476 // We don't support sin/cos/fmod
477 setOperationAction(ISD::FSIN , MVT::f32, Expand);
478 setOperationAction(ISD::FCOS , MVT::f32, Expand);
480 // Special cases we handle for FP constants.
481 addLegalFPImmediate(APFloat(+0.0f)); // xorps
482 addLegalFPImmediate(APFloat(+0.0)); // FLD0
483 addLegalFPImmediate(APFloat(+1.0)); // FLD1
484 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
485 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
488 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
489 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
491 } else if (!UseSoftFloat) {
492 // f32 and f64 in x87.
493 // Set up the FP register classes.
494 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
495 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
497 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
498 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
499 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
500 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
503 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
504 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
506 addLegalFPImmediate(APFloat(+0.0)); // FLD0
507 addLegalFPImmediate(APFloat(+1.0)); // FLD1
508 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
509 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
510 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
511 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
512 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
513 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
516 // Long double always uses X87.
518 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
519 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
520 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
523 APFloat TmpFlt(+0.0);
524 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
526 addLegalFPImmediate(TmpFlt); // FLD0
528 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
529 APFloat TmpFlt2(+1.0);
530 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
532 addLegalFPImmediate(TmpFlt2); // FLD1
533 TmpFlt2.changeSign();
534 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
538 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
539 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
543 // Always use a library call for pow.
544 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
545 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
546 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
548 setOperationAction(ISD::FLOG, MVT::f80, Expand);
549 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
550 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
551 setOperationAction(ISD::FEXP, MVT::f80, Expand);
552 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
554 // First set operation action for all vector types to either promote
555 // (for widening) or expand (for scalarization). Then we will selectively
556 // turn on ones that can be effectively codegen'd.
557 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
558 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
559 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
574 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
575 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
605 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
606 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
607 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
608 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
609 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
610 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
611 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
612 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
613 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
614 setTruncStoreAction((MVT::SimpleValueType)VT,
615 (MVT::SimpleValueType)InnerVT, Expand);
616 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
617 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
618 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
621 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
622 // with -msoft-float, disable use of MMX as well.
623 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
624 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
625 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
626 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
627 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
628 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
630 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
631 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
632 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
633 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
635 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
636 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
637 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
638 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
640 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
641 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
643 setOperationAction(ISD::AND, MVT::v8i8, Promote);
644 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
645 setOperationAction(ISD::AND, MVT::v4i16, Promote);
646 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
647 setOperationAction(ISD::AND, MVT::v2i32, Promote);
648 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
649 setOperationAction(ISD::AND, MVT::v1i64, Legal);
651 setOperationAction(ISD::OR, MVT::v8i8, Promote);
652 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
653 setOperationAction(ISD::OR, MVT::v4i16, Promote);
654 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
655 setOperationAction(ISD::OR, MVT::v2i32, Promote);
656 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
657 setOperationAction(ISD::OR, MVT::v1i64, Legal);
659 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
660 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
661 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
662 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
663 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
664 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
665 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
667 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
668 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
669 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
670 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
671 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
672 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
673 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
674 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
675 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
677 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
678 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
679 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
680 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
681 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
683 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
684 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
685 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
686 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
688 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
689 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
690 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
691 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
693 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
695 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
696 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
697 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
698 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
699 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
700 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
701 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
704 if (!UseSoftFloat && Subtarget->hasSSE1()) {
705 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
707 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
708 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
709 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
710 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
711 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
712 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
713 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
714 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
715 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
716 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
717 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
718 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
721 if (!UseSoftFloat && Subtarget->hasSSE2()) {
722 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
724 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
725 // registers cannot be used even for integer operations.
726 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
727 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
728 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
729 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
731 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
732 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
733 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
734 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
735 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
736 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
737 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
738 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
739 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
740 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
741 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
742 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
743 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
744 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
745 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
746 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
748 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
749 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
750 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
751 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
753 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
754 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
755 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
756 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
757 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
759 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
760 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
761 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
762 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
763 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
765 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
766 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
767 EVT VT = (MVT::SimpleValueType)i;
768 // Do not attempt to custom lower non-power-of-2 vectors
769 if (!isPowerOf2_32(VT.getVectorNumElements()))
771 // Do not attempt to custom lower non-128-bit vectors
772 if (!VT.is128BitVector())
774 setOperationAction(ISD::BUILD_VECTOR,
775 VT.getSimpleVT().SimpleTy, Custom);
776 setOperationAction(ISD::VECTOR_SHUFFLE,
777 VT.getSimpleVT().SimpleTy, Custom);
778 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
779 VT.getSimpleVT().SimpleTy, Custom);
782 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
783 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
784 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
785 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
786 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
787 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
789 if (Subtarget->is64Bit()) {
790 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
791 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
794 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
795 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
796 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
799 // Do not attempt to promote non-128-bit vectors
800 if (!VT.is128BitVector()) {
803 setOperationAction(ISD::AND, SVT, Promote);
804 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
805 setOperationAction(ISD::OR, SVT, Promote);
806 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
807 setOperationAction(ISD::XOR, SVT, Promote);
808 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
809 setOperationAction(ISD::LOAD, SVT, Promote);
810 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
811 setOperationAction(ISD::SELECT, SVT, Promote);
812 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
815 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
817 // Custom lower v2i64 and v2f64 selects.
818 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
819 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
820 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
821 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
823 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
824 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
825 if (!DisableMMX && Subtarget->hasMMX()) {
826 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
827 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
831 if (Subtarget->hasSSE41()) {
832 // FIXME: Do we need to handle scalar-to-vector here?
833 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
835 // i8 and i16 vectors are custom , because the source register and source
836 // source memory operand types are not the same width. f32 vectors are
837 // custom since the immediate controlling the insert encodes additional
839 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
840 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
841 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
842 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
844 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
845 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
846 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
847 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
849 if (Subtarget->is64Bit()) {
850 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
851 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
855 if (Subtarget->hasSSE42()) {
856 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
859 if (!UseSoftFloat && Subtarget->hasAVX()) {
860 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
861 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
862 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
863 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
865 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
866 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
867 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
868 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
869 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
870 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
871 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
872 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
873 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
874 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
875 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
876 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
877 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
878 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
879 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
881 // Operations to consider commented out -v16i16 v32i8
882 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
883 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
884 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
885 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
886 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
887 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
888 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
889 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
890 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
891 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
892 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
893 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
894 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
895 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
897 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
898 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
899 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
900 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
902 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
903 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
904 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
905 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
906 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
908 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
909 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
910 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
911 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
912 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
913 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
916 // Not sure we want to do this since there are no 256-bit integer
919 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
920 // This includes 256-bit vectors
921 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
922 EVT VT = (MVT::SimpleValueType)i;
924 // Do not attempt to custom lower non-power-of-2 vectors
925 if (!isPowerOf2_32(VT.getVectorNumElements()))
928 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
929 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
930 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
933 if (Subtarget->is64Bit()) {
934 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
935 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
940 // Not sure we want to do this since there are no 256-bit integer
943 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
944 // Including 256-bit vectors
945 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
946 EVT VT = (MVT::SimpleValueType)i;
948 if (!VT.is256BitVector()) {
951 setOperationAction(ISD::AND, VT, Promote);
952 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
953 setOperationAction(ISD::OR, VT, Promote);
954 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
955 setOperationAction(ISD::XOR, VT, Promote);
956 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
957 setOperationAction(ISD::LOAD, VT, Promote);
958 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
959 setOperationAction(ISD::SELECT, VT, Promote);
960 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
963 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
967 // We want to custom lower some of our intrinsics.
968 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
970 // Add/Sub/Mul with overflow operations are custom lowered.
971 setOperationAction(ISD::SADDO, MVT::i32, Custom);
972 setOperationAction(ISD::SADDO, MVT::i64, Custom);
973 setOperationAction(ISD::UADDO, MVT::i32, Custom);
974 setOperationAction(ISD::UADDO, MVT::i64, Custom);
975 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
976 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
977 setOperationAction(ISD::USUBO, MVT::i32, Custom);
978 setOperationAction(ISD::USUBO, MVT::i64, Custom);
979 setOperationAction(ISD::SMULO, MVT::i32, Custom);
980 setOperationAction(ISD::SMULO, MVT::i64, Custom);
982 if (!Subtarget->is64Bit()) {
983 // These libcalls are not available in 32-bit.
984 setLibcallName(RTLIB::SHL_I128, 0);
985 setLibcallName(RTLIB::SRL_I128, 0);
986 setLibcallName(RTLIB::SRA_I128, 0);
989 // We have target-specific dag combine patterns for the following nodes:
990 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
991 setTargetDAGCombine(ISD::BUILD_VECTOR);
992 setTargetDAGCombine(ISD::SELECT);
993 setTargetDAGCombine(ISD::AND);
994 setTargetDAGCombine(ISD::SHL);
995 setTargetDAGCombine(ISD::SRA);
996 setTargetDAGCombine(ISD::SRL);
997 setTargetDAGCombine(ISD::OR);
998 setTargetDAGCombine(ISD::STORE);
999 setTargetDAGCombine(ISD::MEMBARRIER);
1000 setTargetDAGCombine(ISD::ZERO_EXTEND);
1001 if (Subtarget->is64Bit())
1002 setTargetDAGCombine(ISD::MUL);
1004 computeRegisterProperties();
1006 // FIXME: These should be based on subtarget info. Plus, the values should
1007 // be smaller when we are in optimizing for size mode.
1008 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1009 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1010 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1011 setPrefLoopAlignment(16);
1012 benefitFromCodePlacementOpt = true;
1016 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1021 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1022 /// the desired ByVal argument alignment.
1023 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1026 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1027 if (VTy->getBitWidth() == 128)
1029 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1030 unsigned EltAlign = 0;
1031 getMaxByValAlign(ATy->getElementType(), EltAlign);
1032 if (EltAlign > MaxAlign)
1033 MaxAlign = EltAlign;
1034 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1035 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1036 unsigned EltAlign = 0;
1037 getMaxByValAlign(STy->getElementType(i), EltAlign);
1038 if (EltAlign > MaxAlign)
1039 MaxAlign = EltAlign;
1047 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1048 /// function arguments in the caller parameter area. For X86, aggregates
1049 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1050 /// are at 4-byte boundaries.
1051 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1052 if (Subtarget->is64Bit()) {
1053 // Max of 8 and alignment of type.
1054 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1061 if (Subtarget->hasSSE1())
1062 getMaxByValAlign(Ty, Align);
1066 /// getOptimalMemOpType - Returns the target specific optimal type for load
1067 /// and store operations as a result of memset, memcpy, and memmove
1068 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1071 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1072 bool isSrcConst, bool isSrcStr,
1073 SelectionDAG &DAG) const {
1074 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1075 // linux. This is because the stack realignment code can't handle certain
1076 // cases like PR2962. This should be removed when PR2962 is fixed.
1077 const Function *F = DAG.getMachineFunction().getFunction();
1078 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1079 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1080 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1082 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1085 if (Subtarget->is64Bit() && Size >= 8)
1090 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1091 /// current function. The returned value is a member of the
1092 /// MachineJumpTableInfo::JTEntryKind enum.
1093 unsigned X86TargetLowering::getJumpTableEncoding() const {
1094 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1096 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1097 Subtarget->isPICStyleGOT())
1098 return MachineJumpTableInfo::EK_Custom32;
1100 // Otherwise, use the normal jump table encoding heuristics.
1101 return TargetLowering::getJumpTableEncoding();
1104 /// getPICBaseSymbol - Return the X86-32 PIC base.
1106 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1107 MCContext &Ctx) const {
1108 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1109 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1110 Twine(MF->getFunctionNumber())+"$pb");
1115 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1116 const MachineBasicBlock *MBB,
1117 unsigned uid,MCContext &Ctx) const{
1118 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1119 Subtarget->isPICStyleGOT());
1120 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1122 return X86MCTargetExpr::Create(MBB->getSymbol(Ctx),
1123 X86MCTargetExpr::GOTOFF, Ctx);
1126 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1128 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1129 SelectionDAG &DAG) const {
1130 if (!Subtarget->is64Bit())
1131 // This doesn't have DebugLoc associated with it, but is not really the
1132 // same as a Register.
1133 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1138 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1139 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1141 const MCExpr *X86TargetLowering::
1142 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1143 MCContext &Ctx) const {
1144 // X86-64 uses RIP relative addressing based on the jump table label.
1145 if (Subtarget->isPICStyleRIPRel())
1146 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1148 // Otherwise, the reference is relative to the PIC base.
1149 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1152 /// getFunctionAlignment - Return the Log2 alignment of this function.
1153 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1154 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1157 //===----------------------------------------------------------------------===//
1158 // Return Value Calling Convention Implementation
1159 //===----------------------------------------------------------------------===//
1161 #include "X86GenCallingConv.inc"
1164 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1165 const SmallVectorImpl<EVT> &OutTys,
1166 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1167 SelectionDAG &DAG) {
1168 SmallVector<CCValAssign, 16> RVLocs;
1169 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1170 RVLocs, *DAG.getContext());
1171 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1175 X86TargetLowering::LowerReturn(SDValue Chain,
1176 CallingConv::ID CallConv, bool isVarArg,
1177 const SmallVectorImpl<ISD::OutputArg> &Outs,
1178 DebugLoc dl, SelectionDAG &DAG) {
1180 SmallVector<CCValAssign, 16> RVLocs;
1181 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1182 RVLocs, *DAG.getContext());
1183 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1185 // Add the regs to the liveout set for the function.
1186 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1187 for (unsigned i = 0; i != RVLocs.size(); ++i)
1188 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1189 MRI.addLiveOut(RVLocs[i].getLocReg());
1193 SmallVector<SDValue, 6> RetOps;
1194 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1195 // Operand #1 = Bytes To Pop
1196 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1198 // Copy the result values into the output registers.
1199 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1200 CCValAssign &VA = RVLocs[i];
1201 assert(VA.isRegLoc() && "Can only return in registers!");
1202 SDValue ValToCopy = Outs[i].Val;
1204 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1205 // the RET instruction and handled by the FP Stackifier.
1206 if (VA.getLocReg() == X86::ST0 ||
1207 VA.getLocReg() == X86::ST1) {
1208 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1209 // change the value to the FP stack register class.
1210 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1211 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1212 RetOps.push_back(ValToCopy);
1213 // Don't emit a copytoreg.
1217 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1218 // which is returned in RAX / RDX.
1219 if (Subtarget->is64Bit()) {
1220 EVT ValVT = ValToCopy.getValueType();
1221 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1222 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1223 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1224 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1228 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1229 Flag = Chain.getValue(1);
1232 // The x86-64 ABI for returning structs by value requires that we copy
1233 // the sret argument into %rax for the return. We saved the argument into
1234 // a virtual register in the entry block, so now we copy the value out
1236 if (Subtarget->is64Bit() &&
1237 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1238 MachineFunction &MF = DAG.getMachineFunction();
1239 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1240 unsigned Reg = FuncInfo->getSRetReturnReg();
1242 Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
1243 FuncInfo->setSRetReturnReg(Reg);
1245 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1247 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1248 Flag = Chain.getValue(1);
1250 // RAX now acts like a return value.
1251 MRI.addLiveOut(X86::RAX);
1254 RetOps[0] = Chain; // Update chain.
1256 // Add the flag if we have it.
1258 RetOps.push_back(Flag);
1260 return DAG.getNode(X86ISD::RET_FLAG, dl,
1261 MVT::Other, &RetOps[0], RetOps.size());
1264 /// LowerCallResult - Lower the result values of a call into the
1265 /// appropriate copies out of appropriate physical registers.
1268 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1269 CallingConv::ID CallConv, bool isVarArg,
1270 const SmallVectorImpl<ISD::InputArg> &Ins,
1271 DebugLoc dl, SelectionDAG &DAG,
1272 SmallVectorImpl<SDValue> &InVals) {
1274 // Assign locations to each value returned by this call.
1275 SmallVector<CCValAssign, 16> RVLocs;
1276 bool Is64Bit = Subtarget->is64Bit();
1277 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1278 RVLocs, *DAG.getContext());
1279 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1281 // Copy all of the result registers out of their specified physreg.
1282 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1283 CCValAssign &VA = RVLocs[i];
1284 EVT CopyVT = VA.getValVT();
1286 // If this is x86-64, and we disabled SSE, we can't return FP values
1287 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1288 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1289 llvm_report_error("SSE register return with SSE disabled");
1292 // If this is a call to a function that returns an fp value on the floating
1293 // point stack, but where we prefer to use the value in xmm registers, copy
1294 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1295 if ((VA.getLocReg() == X86::ST0 ||
1296 VA.getLocReg() == X86::ST1) &&
1297 isScalarFPTypeInSSEReg(VA.getValVT())) {
1302 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1303 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1304 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1305 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1306 MVT::v2i64, InFlag).getValue(1);
1307 Val = Chain.getValue(0);
1308 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1309 Val, DAG.getConstant(0, MVT::i64));
1311 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1312 MVT::i64, InFlag).getValue(1);
1313 Val = Chain.getValue(0);
1315 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1317 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1318 CopyVT, InFlag).getValue(1);
1319 Val = Chain.getValue(0);
1321 InFlag = Chain.getValue(2);
1323 if (CopyVT != VA.getValVT()) {
1324 // Round the F80 the right size, which also moves to the appropriate xmm
1326 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1327 // This truncation won't change the value.
1328 DAG.getIntPtrConstant(1));
1331 InVals.push_back(Val);
1338 //===----------------------------------------------------------------------===//
1339 // C & StdCall & Fast Calling Convention implementation
1340 //===----------------------------------------------------------------------===//
1341 // StdCall calling convention seems to be standard for many Windows' API
1342 // routines and around. It differs from C calling convention just a little:
1343 // callee should clean up the stack, not caller. Symbols should be also
1344 // decorated in some fancy way :) It doesn't support any vector arguments.
1345 // For info on fast calling convention see Fast Calling Convention (tail call)
1346 // implementation LowerX86_32FastCCCallTo.
1348 /// CallIsStructReturn - Determines whether a call uses struct return
1350 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1354 return Outs[0].Flags.isSRet();
1357 /// ArgsAreStructReturn - Determines whether a function uses struct
1358 /// return semantics.
1360 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1364 return Ins[0].Flags.isSRet();
1367 /// IsCalleePop - Determines whether the callee is required to pop its
1368 /// own arguments. Callee pop is necessary to support tail calls.
1369 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1373 switch (CallingConv) {
1376 case CallingConv::X86_StdCall:
1377 return !Subtarget->is64Bit();
1378 case CallingConv::X86_FastCall:
1379 return !Subtarget->is64Bit();
1380 case CallingConv::Fast:
1381 return GuaranteedTailCallOpt;
1385 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1386 /// given CallingConvention value.
1387 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1388 if (Subtarget->is64Bit()) {
1389 if (Subtarget->isTargetWin64())
1390 return CC_X86_Win64_C;
1395 if (CC == CallingConv::X86_FastCall)
1396 return CC_X86_32_FastCall;
1397 else if (CC == CallingConv::Fast)
1398 return CC_X86_32_FastCC;
1403 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1404 /// by "Src" to address "Dst" with size and alignment information specified by
1405 /// the specific parameter attribute. The copy will be passed as a byval
1406 /// function parameter.
1408 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1409 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1411 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1412 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1413 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1416 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1417 /// a tailcall target by changing its ABI.
1418 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1419 return GuaranteedTailCallOpt && CC == CallingConv::Fast;
1423 X86TargetLowering::LowerMemArgument(SDValue Chain,
1424 CallingConv::ID CallConv,
1425 const SmallVectorImpl<ISD::InputArg> &Ins,
1426 DebugLoc dl, SelectionDAG &DAG,
1427 const CCValAssign &VA,
1428 MachineFrameInfo *MFI,
1430 // Create the nodes corresponding to a load from this parameter slot.
1431 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1432 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1433 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1436 // If value is passed by pointer we have address passed instead of the value
1438 if (VA.getLocInfo() == CCValAssign::Indirect)
1439 ValVT = VA.getLocVT();
1441 ValVT = VA.getValVT();
1443 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1444 // changed with more analysis.
1445 // In case of tail call optimization mark all arguments mutable. Since they
1446 // could be overwritten by lowering of arguments in case of a tail call.
1447 if (Flags.isByVal()) {
1448 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1449 VA.getLocMemOffset(), isImmutable, false);
1450 return DAG.getFrameIndex(FI, getPointerTy());
1452 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1453 VA.getLocMemOffset(), isImmutable, false);
1454 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1455 return DAG.getLoad(ValVT, dl, Chain, FIN,
1456 PseudoSourceValue::getFixedStack(FI), 0,
1462 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1463 CallingConv::ID CallConv,
1465 const SmallVectorImpl<ISD::InputArg> &Ins,
1468 SmallVectorImpl<SDValue> &InVals) {
1470 MachineFunction &MF = DAG.getMachineFunction();
1471 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1473 const Function* Fn = MF.getFunction();
1474 if (Fn->hasExternalLinkage() &&
1475 Subtarget->isTargetCygMing() &&
1476 Fn->getName() == "main")
1477 FuncInfo->setForceFramePointer(true);
1479 MachineFrameInfo *MFI = MF.getFrameInfo();
1480 bool Is64Bit = Subtarget->is64Bit();
1481 bool IsWin64 = Subtarget->isTargetWin64();
1483 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1484 "Var args not supported with calling convention fastcc");
1486 // Assign locations to all of the incoming arguments.
1487 SmallVector<CCValAssign, 16> ArgLocs;
1488 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1489 ArgLocs, *DAG.getContext());
1490 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1492 unsigned LastVal = ~0U;
1494 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1495 CCValAssign &VA = ArgLocs[i];
1496 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1498 assert(VA.getValNo() != LastVal &&
1499 "Don't support value assigned to multiple locs yet");
1500 LastVal = VA.getValNo();
1502 if (VA.isRegLoc()) {
1503 EVT RegVT = VA.getLocVT();
1504 TargetRegisterClass *RC = NULL;
1505 if (RegVT == MVT::i32)
1506 RC = X86::GR32RegisterClass;
1507 else if (Is64Bit && RegVT == MVT::i64)
1508 RC = X86::GR64RegisterClass;
1509 else if (RegVT == MVT::f32)
1510 RC = X86::FR32RegisterClass;
1511 else if (RegVT == MVT::f64)
1512 RC = X86::FR64RegisterClass;
1513 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1514 RC = X86::VR128RegisterClass;
1515 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1516 RC = X86::VR64RegisterClass;
1518 llvm_unreachable("Unknown argument type!");
1520 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1521 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1523 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1524 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1526 if (VA.getLocInfo() == CCValAssign::SExt)
1527 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1528 DAG.getValueType(VA.getValVT()));
1529 else if (VA.getLocInfo() == CCValAssign::ZExt)
1530 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1531 DAG.getValueType(VA.getValVT()));
1532 else if (VA.getLocInfo() == CCValAssign::BCvt)
1533 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1535 if (VA.isExtInLoc()) {
1536 // Handle MMX values passed in XMM regs.
1537 if (RegVT.isVector()) {
1538 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1539 ArgValue, DAG.getConstant(0, MVT::i64));
1540 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1542 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1545 assert(VA.isMemLoc());
1546 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1549 // If value is passed via pointer - do a load.
1550 if (VA.getLocInfo() == CCValAssign::Indirect)
1551 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1554 InVals.push_back(ArgValue);
1557 // The x86-64 ABI for returning structs by value requires that we copy
1558 // the sret argument into %rax for the return. Save the argument into
1559 // a virtual register so that we can access it from the return points.
1560 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1561 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1562 unsigned Reg = FuncInfo->getSRetReturnReg();
1564 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1565 FuncInfo->setSRetReturnReg(Reg);
1567 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1568 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1571 unsigned StackSize = CCInfo.getNextStackOffset();
1572 // Align stack specially for tail calls.
1573 if (FuncIsMadeTailCallSafe(CallConv))
1574 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1576 // If the function takes variable number of arguments, make a frame index for
1577 // the start of the first vararg value... for expansion of llvm.va_start.
1579 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1580 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1583 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1585 // FIXME: We should really autogenerate these arrays
1586 static const unsigned GPR64ArgRegsWin64[] = {
1587 X86::RCX, X86::RDX, X86::R8, X86::R9
1589 static const unsigned XMMArgRegsWin64[] = {
1590 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1592 static const unsigned GPR64ArgRegs64Bit[] = {
1593 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1595 static const unsigned XMMArgRegs64Bit[] = {
1596 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1597 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1599 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1602 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1603 GPR64ArgRegs = GPR64ArgRegsWin64;
1604 XMMArgRegs = XMMArgRegsWin64;
1606 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1607 GPR64ArgRegs = GPR64ArgRegs64Bit;
1608 XMMArgRegs = XMMArgRegs64Bit;
1610 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1612 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1615 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1616 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1617 "SSE register cannot be used when SSE is disabled!");
1618 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1619 "SSE register cannot be used when SSE is disabled!");
1620 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1621 // Kernel mode asks for SSE to be disabled, so don't push them
1623 TotalNumXMMRegs = 0;
1625 // For X86-64, if there are vararg parameters that are passed via
1626 // registers, then we must store them to their spots on the stack so they
1627 // may be loaded by deferencing the result of va_next.
1628 VarArgsGPOffset = NumIntRegs * 8;
1629 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1630 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1631 TotalNumXMMRegs * 16, 16,
1634 // Store the integer parameter registers.
1635 SmallVector<SDValue, 8> MemOps;
1636 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1637 unsigned Offset = VarArgsGPOffset;
1638 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1639 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1640 DAG.getIntPtrConstant(Offset));
1641 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1642 X86::GR64RegisterClass);
1643 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1645 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1646 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1647 Offset, false, false, 0);
1648 MemOps.push_back(Store);
1652 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1653 // Now store the XMM (fp + vector) parameter registers.
1654 SmallVector<SDValue, 11> SaveXMMOps;
1655 SaveXMMOps.push_back(Chain);
1657 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1658 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1659 SaveXMMOps.push_back(ALVal);
1661 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1662 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1664 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1665 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1666 X86::VR128RegisterClass);
1667 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1668 SaveXMMOps.push_back(Val);
1670 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1672 &SaveXMMOps[0], SaveXMMOps.size()));
1675 if (!MemOps.empty())
1676 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1677 &MemOps[0], MemOps.size());
1681 // Some CCs need callee pop.
1682 if (IsCalleePop(isVarArg, CallConv)) {
1683 BytesToPopOnReturn = StackSize; // Callee pops everything.
1685 BytesToPopOnReturn = 0; // Callee pops nothing.
1686 // If this is an sret function, the return should pop the hidden pointer.
1687 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1688 BytesToPopOnReturn = 4;
1692 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1693 if (CallConv == CallingConv::X86_FastCall)
1694 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1697 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1703 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1704 SDValue StackPtr, SDValue Arg,
1705 DebugLoc dl, SelectionDAG &DAG,
1706 const CCValAssign &VA,
1707 ISD::ArgFlagsTy Flags) {
1708 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1709 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1710 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1711 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1712 if (Flags.isByVal()) {
1713 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1715 return DAG.getStore(Chain, dl, Arg, PtrOff,
1716 PseudoSourceValue::getStack(), LocMemOffset,
1720 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1721 /// optimization is performed and it is required.
1723 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1724 SDValue &OutRetAddr, SDValue Chain,
1725 bool IsTailCall, bool Is64Bit,
1726 int FPDiff, DebugLoc dl) {
1727 // Adjust the Return address stack slot.
1728 EVT VT = getPointerTy();
1729 OutRetAddr = getReturnAddressFrameIndex(DAG);
1731 // Load the "old" Return address.
1732 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1733 return SDValue(OutRetAddr.getNode(), 1);
1736 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1737 /// optimization is performed and it is required (FPDiff!=0).
1739 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1740 SDValue Chain, SDValue RetAddrFrIdx,
1741 bool Is64Bit, int FPDiff, DebugLoc dl) {
1742 // Store the return address to the appropriate stack slot.
1743 if (!FPDiff) return Chain;
1744 // Calculate the new stack slot for the return address.
1745 int SlotSize = Is64Bit ? 8 : 4;
1746 int NewReturnAddrFI =
1747 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, false, false);
1748 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1749 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1750 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1751 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1757 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1758 CallingConv::ID CallConv, bool isVarArg,
1760 const SmallVectorImpl<ISD::OutputArg> &Outs,
1761 const SmallVectorImpl<ISD::InputArg> &Ins,
1762 DebugLoc dl, SelectionDAG &DAG,
1763 SmallVectorImpl<SDValue> &InVals) {
1764 MachineFunction &MF = DAG.getMachineFunction();
1765 bool Is64Bit = Subtarget->is64Bit();
1766 bool IsStructRet = CallIsStructReturn(Outs);
1767 bool IsSibcall = false;
1770 // Check if it's really possible to do a tail call.
1771 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
1774 // Sibcalls are automatically detected tailcalls which do not require
1776 if (!GuaranteedTailCallOpt && isTailCall)
1783 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1784 "Var args not supported with calling convention fastcc");
1786 // Analyze operands of the call, assigning locations to each operand.
1787 SmallVector<CCValAssign, 16> ArgLocs;
1788 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1789 ArgLocs, *DAG.getContext());
1790 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1792 // Get a count of how many bytes are to be pushed on the stack.
1793 unsigned NumBytes = CCInfo.getNextStackOffset();
1795 // This is a sibcall. The memory operands are available in caller's
1796 // own caller's stack.
1798 else if (GuaranteedTailCallOpt && CallConv == CallingConv::Fast)
1799 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1802 if (isTailCall && !IsSibcall) {
1803 // Lower arguments at fp - stackoffset + fpdiff.
1804 unsigned NumBytesCallerPushed =
1805 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1806 FPDiff = NumBytesCallerPushed - NumBytes;
1808 // Set the delta of movement of the returnaddr stackslot.
1809 // But only set if delta is greater than previous delta.
1810 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1811 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1815 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1817 SDValue RetAddrFrIdx;
1818 // Load return adress for tail calls.
1819 if (isTailCall && FPDiff)
1820 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1821 Is64Bit, FPDiff, dl);
1823 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1824 SmallVector<SDValue, 8> MemOpChains;
1827 // Walk the register/memloc assignments, inserting copies/loads. In the case
1828 // of tail call optimization arguments are handle later.
1829 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1830 CCValAssign &VA = ArgLocs[i];
1831 EVT RegVT = VA.getLocVT();
1832 SDValue Arg = Outs[i].Val;
1833 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1834 bool isByVal = Flags.isByVal();
1836 // Promote the value if needed.
1837 switch (VA.getLocInfo()) {
1838 default: llvm_unreachable("Unknown loc info!");
1839 case CCValAssign::Full: break;
1840 case CCValAssign::SExt:
1841 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1843 case CCValAssign::ZExt:
1844 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1846 case CCValAssign::AExt:
1847 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1848 // Special case: passing MMX values in XMM registers.
1849 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1850 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1851 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1853 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1855 case CCValAssign::BCvt:
1856 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1858 case CCValAssign::Indirect: {
1859 // Store the argument.
1860 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1861 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1862 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1863 PseudoSourceValue::getFixedStack(FI), 0,
1870 if (VA.isRegLoc()) {
1871 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1872 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1873 assert(VA.isMemLoc());
1874 if (StackPtr.getNode() == 0)
1875 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1876 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1877 dl, DAG, VA, Flags));
1881 if (!MemOpChains.empty())
1882 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1883 &MemOpChains[0], MemOpChains.size());
1885 // Build a sequence of copy-to-reg nodes chained together with token chain
1886 // and flag operands which copy the outgoing args into registers.
1888 // Tail call byval lowering might overwrite argument registers so in case of
1889 // tail call optimization the copies to registers are lowered later.
1891 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1892 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1893 RegsToPass[i].second, InFlag);
1894 InFlag = Chain.getValue(1);
1897 if (Subtarget->isPICStyleGOT()) {
1898 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1901 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1902 DAG.getNode(X86ISD::GlobalBaseReg,
1903 DebugLoc::getUnknownLoc(),
1906 InFlag = Chain.getValue(1);
1908 // If we are tail calling and generating PIC/GOT style code load the
1909 // address of the callee into ECX. The value in ecx is used as target of
1910 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1911 // for tail calls on PIC/GOT architectures. Normally we would just put the
1912 // address of GOT into ebx and then call target@PLT. But for tail calls
1913 // ebx would be restored (since ebx is callee saved) before jumping to the
1916 // Note: The actual moving to ECX is done further down.
1917 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1918 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1919 !G->getGlobal()->hasProtectedVisibility())
1920 Callee = LowerGlobalAddress(Callee, DAG);
1921 else if (isa<ExternalSymbolSDNode>(Callee))
1922 Callee = LowerExternalSymbol(Callee, DAG);
1926 if (Is64Bit && isVarArg) {
1927 // From AMD64 ABI document:
1928 // For calls that may call functions that use varargs or stdargs
1929 // (prototype-less calls or calls to functions containing ellipsis (...) in
1930 // the declaration) %al is used as hidden argument to specify the number
1931 // of SSE registers used. The contents of %al do not need to match exactly
1932 // the number of registers, but must be an ubound on the number of SSE
1933 // registers used and is in the range 0 - 8 inclusive.
1935 // FIXME: Verify this on Win64
1936 // Count the number of XMM registers allocated.
1937 static const unsigned XMMArgRegs[] = {
1938 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1939 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1941 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1942 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1943 && "SSE registers cannot be used when SSE is disabled");
1945 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1946 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1947 InFlag = Chain.getValue(1);
1951 // For tail calls lower the arguments to the 'real' stack slot.
1953 // Force all the incoming stack arguments to be loaded from the stack
1954 // before any new outgoing arguments are stored to the stack, because the
1955 // outgoing stack slots may alias the incoming argument stack slots, and
1956 // the alias isn't otherwise explicit. This is slightly more conservative
1957 // than necessary, because it means that each store effectively depends
1958 // on every argument instead of just those arguments it would clobber.
1959 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1961 SmallVector<SDValue, 8> MemOpChains2;
1964 // Do not flag preceeding copytoreg stuff together with the following stuff.
1966 if (GuaranteedTailCallOpt) {
1967 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1968 CCValAssign &VA = ArgLocs[i];
1971 assert(VA.isMemLoc());
1972 SDValue Arg = Outs[i].Val;
1973 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1974 // Create frame index.
1975 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1976 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1977 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1978 FIN = DAG.getFrameIndex(FI, getPointerTy());
1980 if (Flags.isByVal()) {
1981 // Copy relative to framepointer.
1982 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1983 if (StackPtr.getNode() == 0)
1984 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1986 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1988 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1992 // Store relative to framepointer.
1993 MemOpChains2.push_back(
1994 DAG.getStore(ArgChain, dl, Arg, FIN,
1995 PseudoSourceValue::getFixedStack(FI), 0,
2001 if (!MemOpChains2.empty())
2002 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2003 &MemOpChains2[0], MemOpChains2.size());
2005 // Copy arguments to their registers.
2006 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2007 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2008 RegsToPass[i].second, InFlag);
2009 InFlag = Chain.getValue(1);
2013 // Store the return address to the appropriate stack slot.
2014 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2018 bool WasGlobalOrExternal = false;
2019 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2020 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2021 // In the 64-bit large code model, we have to make all calls
2022 // through a register, since the call instruction's 32-bit
2023 // pc-relative offset may not be large enough to hold the whole
2025 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2026 WasGlobalOrExternal = true;
2027 // If the callee is a GlobalAddress node (quite common, every direct call
2028 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2031 // We should use extra load for direct calls to dllimported functions in
2033 GlobalValue *GV = G->getGlobal();
2034 if (!GV->hasDLLImportLinkage()) {
2035 unsigned char OpFlags = 0;
2037 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2038 // external symbols most go through the PLT in PIC mode. If the symbol
2039 // has hidden or protected visibility, or if it is static or local, then
2040 // we don't need to use the PLT - we can directly call it.
2041 if (Subtarget->isTargetELF() &&
2042 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2043 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2044 OpFlags = X86II::MO_PLT;
2045 } else if (Subtarget->isPICStyleStubAny() &&
2046 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2047 Subtarget->getDarwinVers() < 9) {
2048 // PC-relative references to external symbols should go through $stub,
2049 // unless we're building with the leopard linker or later, which
2050 // automatically synthesizes these stubs.
2051 OpFlags = X86II::MO_DARWIN_STUB;
2054 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2055 G->getOffset(), OpFlags);
2057 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2058 WasGlobalOrExternal = true;
2059 unsigned char OpFlags = 0;
2061 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2062 // symbols should go through the PLT.
2063 if (Subtarget->isTargetELF() &&
2064 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2065 OpFlags = X86II::MO_PLT;
2066 } else if (Subtarget->isPICStyleStubAny() &&
2067 Subtarget->getDarwinVers() < 9) {
2068 // PC-relative references to external symbols should go through $stub,
2069 // unless we're building with the leopard linker or later, which
2070 // automatically synthesizes these stubs.
2071 OpFlags = X86II::MO_DARWIN_STUB;
2074 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2078 if (isTailCall && !WasGlobalOrExternal) {
2079 // Force the address into a (call preserved) caller-saved register since
2080 // tailcall must happen after callee-saved registers are poped.
2081 // FIXME: Give it a special register class that contains caller-saved
2082 // register instead?
2083 unsigned TCReg = Is64Bit ? X86::R11 : X86::EAX;
2084 Chain = DAG.getCopyToReg(Chain, dl,
2085 DAG.getRegister(TCReg, getPointerTy()),
2087 Callee = DAG.getRegister(TCReg, getPointerTy());
2090 // Returns a chain & a flag for retval copy to use.
2091 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2092 SmallVector<SDValue, 8> Ops;
2094 if (!IsSibcall && isTailCall) {
2095 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2096 DAG.getIntPtrConstant(0, true), InFlag);
2097 InFlag = Chain.getValue(1);
2100 Ops.push_back(Chain);
2101 Ops.push_back(Callee);
2104 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2106 // Add argument registers to the end of the list so that they are known live
2108 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2109 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2110 RegsToPass[i].second.getValueType()));
2112 // Add an implicit use GOT pointer in EBX.
2113 if (!isTailCall && Subtarget->isPICStyleGOT())
2114 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2116 // Add an implicit use of AL for x86 vararg functions.
2117 if (Is64Bit && isVarArg)
2118 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2120 if (InFlag.getNode())
2121 Ops.push_back(InFlag);
2124 // If this is the first return lowered for this function, add the regs
2125 // to the liveout set for the function.
2126 if (MF.getRegInfo().liveout_empty()) {
2127 SmallVector<CCValAssign, 16> RVLocs;
2128 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2130 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2131 for (unsigned i = 0; i != RVLocs.size(); ++i)
2132 if (RVLocs[i].isRegLoc())
2133 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2136 assert(((Callee.getOpcode() == ISD::Register &&
2137 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2138 cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
2139 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2140 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2141 "Expecting a global address, external symbol, or scratch register");
2143 return DAG.getNode(X86ISD::TC_RETURN, dl,
2144 NodeTys, &Ops[0], Ops.size());
2147 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2148 InFlag = Chain.getValue(1);
2150 // Create the CALLSEQ_END node.
2151 unsigned NumBytesForCalleeToPush;
2152 if (IsCalleePop(isVarArg, CallConv))
2153 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2154 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2155 // If this is a call to a struct-return function, the callee
2156 // pops the hidden struct pointer, so we have to push it back.
2157 // This is common for Darwin/X86, Linux & Mingw32 targets.
2158 NumBytesForCalleeToPush = 4;
2160 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2162 // Returns a flag for retval copy to use.
2164 Chain = DAG.getCALLSEQ_END(Chain,
2165 DAG.getIntPtrConstant(NumBytes, true),
2166 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2169 InFlag = Chain.getValue(1);
2172 // Handle result values, copying them out of physregs into vregs that we
2174 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2175 Ins, dl, DAG, InVals);
2179 //===----------------------------------------------------------------------===//
2180 // Fast Calling Convention (tail call) implementation
2181 //===----------------------------------------------------------------------===//
2183 // Like std call, callee cleans arguments, convention except that ECX is
2184 // reserved for storing the tail called function address. Only 2 registers are
2185 // free for argument passing (inreg). Tail call optimization is performed
2187 // * tailcallopt is enabled
2188 // * caller/callee are fastcc
2189 // On X86_64 architecture with GOT-style position independent code only local
2190 // (within module) calls are supported at the moment.
2191 // To keep the stack aligned according to platform abi the function
2192 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2193 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2194 // If a tail called function callee has more arguments than the caller the
2195 // caller needs to make sure that there is room to move the RETADDR to. This is
2196 // achieved by reserving an area the size of the argument delta right after the
2197 // original REtADDR, but before the saved framepointer or the spilled registers
2198 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2210 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2211 /// for a 16 byte align requirement.
2212 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2213 SelectionDAG& DAG) {
2214 MachineFunction &MF = DAG.getMachineFunction();
2215 const TargetMachine &TM = MF.getTarget();
2216 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2217 unsigned StackAlignment = TFI.getStackAlignment();
2218 uint64_t AlignMask = StackAlignment - 1;
2219 int64_t Offset = StackSize;
2220 uint64_t SlotSize = TD->getPointerSize();
2221 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2222 // Number smaller than 12 so just add the difference.
2223 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2225 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2226 Offset = ((~AlignMask) & Offset) + StackAlignment +
2227 (StackAlignment-SlotSize);
2232 /// MatchingStackOffset - Return true if the given stack call argument is
2233 /// already available in the same position (relatively) of the caller's
2234 /// incoming argument stack.
2236 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2237 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2238 const X86InstrInfo *TII) {
2240 if (Arg.getOpcode() == ISD::CopyFromReg) {
2241 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2242 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2244 MachineInstr *Def = MRI->getVRegDef(VR);
2247 if (!Flags.isByVal()) {
2248 if (!TII->isLoadFromStackSlot(Def, FI))
2251 unsigned Opcode = Def->getOpcode();
2252 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2253 Def->getOperand(1).isFI()) {
2254 FI = Def->getOperand(1).getIndex();
2255 if (MFI->getObjectSize(FI) != Flags.getByValSize())
2261 LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg);
2264 SDValue Ptr = Ld->getBasePtr();
2265 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2268 FI = FINode->getIndex();
2271 if (!MFI->isFixedObjectIndex(FI))
2273 return Offset == MFI->getObjectOffset(FI);
2276 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2277 /// for tail call optimization. Targets which want to do tail call
2278 /// optimization should implement this function.
2280 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2281 CallingConv::ID CalleeCC,
2283 const SmallVectorImpl<ISD::OutputArg> &Outs,
2284 const SmallVectorImpl<ISD::InputArg> &Ins,
2285 SelectionDAG& DAG) const {
2286 if (CalleeCC != CallingConv::Fast &&
2287 CalleeCC != CallingConv::C)
2290 // If -tailcallopt is specified, make fastcc functions tail-callable.
2291 const Function *CallerF = DAG.getMachineFunction().getFunction();
2292 if (GuaranteedTailCallOpt) {
2293 if (CalleeCC == CallingConv::Fast &&
2294 CallerF->getCallingConv() == CalleeCC)
2299 // Look for obvious safe cases to perform tail call optimization that does not
2300 // requite ABI changes. This is what gcc calls sibcall.
2302 // Do not tail call optimize vararg calls for now.
2306 // If the callee takes no arguments then go on to check the results of the
2308 if (!Outs.empty()) {
2309 // Check if stack adjustment is needed. For now, do not do this if any
2310 // argument is passed on the stack.
2311 SmallVector<CCValAssign, 16> ArgLocs;
2312 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2313 ArgLocs, *DAG.getContext());
2314 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2315 if (CCInfo.getNextStackOffset()) {
2316 MachineFunction &MF = DAG.getMachineFunction();
2317 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2319 if (Subtarget->isTargetWin64())
2320 // Win64 ABI has additional complications.
2323 // Check if the arguments are already laid out in the right way as
2324 // the caller's fixed stack objects.
2325 MachineFrameInfo *MFI = MF.getFrameInfo();
2326 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2327 const X86InstrInfo *TII =
2328 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2329 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2330 CCValAssign &VA = ArgLocs[i];
2331 EVT RegVT = VA.getLocVT();
2332 SDValue Arg = Outs[i].Val;
2333 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2334 if (VA.getLocInfo() == CCValAssign::Indirect)
2336 if (!VA.isRegLoc()) {
2337 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2349 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2351 DenseMap<const Value *, unsigned> &vm,
2352 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2353 DenseMap<const AllocaInst *, int> &am
2355 , SmallSet<Instruction*, 8> &cil
2358 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2366 //===----------------------------------------------------------------------===//
2367 // Other Lowering Hooks
2368 //===----------------------------------------------------------------------===//
2371 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2372 MachineFunction &MF = DAG.getMachineFunction();
2373 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2374 int ReturnAddrIndex = FuncInfo->getRAIndex();
2376 if (ReturnAddrIndex == 0) {
2377 // Set up a frame object for the return address.
2378 uint64_t SlotSize = TD->getPointerSize();
2379 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2381 FuncInfo->setRAIndex(ReturnAddrIndex);
2384 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2388 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2389 bool hasSymbolicDisplacement) {
2390 // Offset should fit into 32 bit immediate field.
2391 if (!isInt32(Offset))
2394 // If we don't have a symbolic displacement - we don't have any extra
2396 if (!hasSymbolicDisplacement)
2399 // FIXME: Some tweaks might be needed for medium code model.
2400 if (M != CodeModel::Small && M != CodeModel::Kernel)
2403 // For small code model we assume that latest object is 16MB before end of 31
2404 // bits boundary. We may also accept pretty large negative constants knowing
2405 // that all objects are in the positive half of address space.
2406 if (M == CodeModel::Small && Offset < 16*1024*1024)
2409 // For kernel code model we know that all object resist in the negative half
2410 // of 32bits address space. We may not accept negative offsets, since they may
2411 // be just off and we may accept pretty large positive ones.
2412 if (M == CodeModel::Kernel && Offset > 0)
2418 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2419 /// specific condition code, returning the condition code and the LHS/RHS of the
2420 /// comparison to make.
2421 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2422 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2424 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2425 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2426 // X > -1 -> X == 0, jump !sign.
2427 RHS = DAG.getConstant(0, RHS.getValueType());
2428 return X86::COND_NS;
2429 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2430 // X < 0 -> X == 0, jump on sign.
2432 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2434 RHS = DAG.getConstant(0, RHS.getValueType());
2435 return X86::COND_LE;
2439 switch (SetCCOpcode) {
2440 default: llvm_unreachable("Invalid integer condition!");
2441 case ISD::SETEQ: return X86::COND_E;
2442 case ISD::SETGT: return X86::COND_G;
2443 case ISD::SETGE: return X86::COND_GE;
2444 case ISD::SETLT: return X86::COND_L;
2445 case ISD::SETLE: return X86::COND_LE;
2446 case ISD::SETNE: return X86::COND_NE;
2447 case ISD::SETULT: return X86::COND_B;
2448 case ISD::SETUGT: return X86::COND_A;
2449 case ISD::SETULE: return X86::COND_BE;
2450 case ISD::SETUGE: return X86::COND_AE;
2454 // First determine if it is required or is profitable to flip the operands.
2456 // If LHS is a foldable load, but RHS is not, flip the condition.
2457 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2458 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2459 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2460 std::swap(LHS, RHS);
2463 switch (SetCCOpcode) {
2469 std::swap(LHS, RHS);
2473 // On a floating point condition, the flags are set as follows:
2475 // 0 | 0 | 0 | X > Y
2476 // 0 | 0 | 1 | X < Y
2477 // 1 | 0 | 0 | X == Y
2478 // 1 | 1 | 1 | unordered
2479 switch (SetCCOpcode) {
2480 default: llvm_unreachable("Condcode should be pre-legalized away");
2482 case ISD::SETEQ: return X86::COND_E;
2483 case ISD::SETOLT: // flipped
2485 case ISD::SETGT: return X86::COND_A;
2486 case ISD::SETOLE: // flipped
2488 case ISD::SETGE: return X86::COND_AE;
2489 case ISD::SETUGT: // flipped
2491 case ISD::SETLT: return X86::COND_B;
2492 case ISD::SETUGE: // flipped
2494 case ISD::SETLE: return X86::COND_BE;
2496 case ISD::SETNE: return X86::COND_NE;
2497 case ISD::SETUO: return X86::COND_P;
2498 case ISD::SETO: return X86::COND_NP;
2500 case ISD::SETUNE: return X86::COND_INVALID;
2504 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2505 /// code. Current x86 isa includes the following FP cmov instructions:
2506 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2507 static bool hasFPCMov(unsigned X86CC) {
2523 /// isFPImmLegal - Returns true if the target can instruction select the
2524 /// specified FP immediate natively. If false, the legalizer will
2525 /// materialize the FP immediate as a load from a constant pool.
2526 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2527 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2528 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2534 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2535 /// the specified range (L, H].
2536 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2537 return (Val < 0) || (Val >= Low && Val < Hi);
2540 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2541 /// specified value.
2542 static bool isUndefOrEqual(int Val, int CmpVal) {
2543 if (Val < 0 || Val == CmpVal)
2548 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2549 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2550 /// the second operand.
2551 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2552 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2553 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2554 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2555 return (Mask[0] < 2 && Mask[1] < 2);
2559 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2560 SmallVector<int, 8> M;
2562 return ::isPSHUFDMask(M, N->getValueType(0));
2565 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2566 /// is suitable for input to PSHUFHW.
2567 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2568 if (VT != MVT::v8i16)
2571 // Lower quadword copied in order or undef.
2572 for (int i = 0; i != 4; ++i)
2573 if (Mask[i] >= 0 && Mask[i] != i)
2576 // Upper quadword shuffled.
2577 for (int i = 4; i != 8; ++i)
2578 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2584 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2585 SmallVector<int, 8> M;
2587 return ::isPSHUFHWMask(M, N->getValueType(0));
2590 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2591 /// is suitable for input to PSHUFLW.
2592 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2593 if (VT != MVT::v8i16)
2596 // Upper quadword copied in order.
2597 for (int i = 4; i != 8; ++i)
2598 if (Mask[i] >= 0 && Mask[i] != i)
2601 // Lower quadword shuffled.
2602 for (int i = 0; i != 4; ++i)
2609 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2610 SmallVector<int, 8> M;
2612 return ::isPSHUFLWMask(M, N->getValueType(0));
2615 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2616 /// is suitable for input to PALIGNR.
2617 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2619 int i, e = VT.getVectorNumElements();
2621 // Do not handle v2i64 / v2f64 shuffles with palignr.
2622 if (e < 4 || !hasSSSE3)
2625 for (i = 0; i != e; ++i)
2629 // All undef, not a palignr.
2633 // Determine if it's ok to perform a palignr with only the LHS, since we
2634 // don't have access to the actual shuffle elements to see if RHS is undef.
2635 bool Unary = Mask[i] < (int)e;
2636 bool NeedsUnary = false;
2638 int s = Mask[i] - i;
2640 // Check the rest of the elements to see if they are consecutive.
2641 for (++i; i != e; ++i) {
2646 Unary = Unary && (m < (int)e);
2647 NeedsUnary = NeedsUnary || (m < s);
2649 if (NeedsUnary && !Unary)
2651 if (Unary && m != ((s+i) & (e-1)))
2653 if (!Unary && m != (s+i))
2659 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2660 SmallVector<int, 8> M;
2662 return ::isPALIGNRMask(M, N->getValueType(0), true);
2665 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2666 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2667 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2668 int NumElems = VT.getVectorNumElements();
2669 if (NumElems != 2 && NumElems != 4)
2672 int Half = NumElems / 2;
2673 for (int i = 0; i < Half; ++i)
2674 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2676 for (int i = Half; i < NumElems; ++i)
2677 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2683 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2684 SmallVector<int, 8> M;
2686 return ::isSHUFPMask(M, N->getValueType(0));
2689 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2690 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2691 /// half elements to come from vector 1 (which would equal the dest.) and
2692 /// the upper half to come from vector 2.
2693 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2694 int NumElems = VT.getVectorNumElements();
2696 if (NumElems != 2 && NumElems != 4)
2699 int Half = NumElems / 2;
2700 for (int i = 0; i < Half; ++i)
2701 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2703 for (int i = Half; i < NumElems; ++i)
2704 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2709 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2710 SmallVector<int, 8> M;
2712 return isCommutedSHUFPMask(M, N->getValueType(0));
2715 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2716 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2717 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2718 if (N->getValueType(0).getVectorNumElements() != 4)
2721 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2722 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2723 isUndefOrEqual(N->getMaskElt(1), 7) &&
2724 isUndefOrEqual(N->getMaskElt(2), 2) &&
2725 isUndefOrEqual(N->getMaskElt(3), 3);
2728 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2729 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2731 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2732 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2737 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2738 isUndefOrEqual(N->getMaskElt(1), 3) &&
2739 isUndefOrEqual(N->getMaskElt(2), 2) &&
2740 isUndefOrEqual(N->getMaskElt(3), 3);
2743 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2744 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2745 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2746 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2748 if (NumElems != 2 && NumElems != 4)
2751 for (unsigned i = 0; i < NumElems/2; ++i)
2752 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2755 for (unsigned i = NumElems/2; i < NumElems; ++i)
2756 if (!isUndefOrEqual(N->getMaskElt(i), i))
2762 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2763 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2764 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2765 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2767 if (NumElems != 2 && NumElems != 4)
2770 for (unsigned i = 0; i < NumElems/2; ++i)
2771 if (!isUndefOrEqual(N->getMaskElt(i), i))
2774 for (unsigned i = 0; i < NumElems/2; ++i)
2775 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2781 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2782 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2783 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2784 bool V2IsSplat = false) {
2785 int NumElts = VT.getVectorNumElements();
2786 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2789 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2791 int BitI1 = Mask[i+1];
2792 if (!isUndefOrEqual(BitI, j))
2795 if (!isUndefOrEqual(BitI1, NumElts))
2798 if (!isUndefOrEqual(BitI1, j + NumElts))
2805 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2806 SmallVector<int, 8> M;
2808 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2811 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2812 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2813 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2814 bool V2IsSplat = false) {
2815 int NumElts = VT.getVectorNumElements();
2816 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2819 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2821 int BitI1 = Mask[i+1];
2822 if (!isUndefOrEqual(BitI, j + NumElts/2))
2825 if (isUndefOrEqual(BitI1, NumElts))
2828 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2835 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2836 SmallVector<int, 8> M;
2838 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2841 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2842 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2844 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2845 int NumElems = VT.getVectorNumElements();
2846 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2849 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2851 int BitI1 = Mask[i+1];
2852 if (!isUndefOrEqual(BitI, j))
2854 if (!isUndefOrEqual(BitI1, j))
2860 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2861 SmallVector<int, 8> M;
2863 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2866 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2867 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2869 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2870 int NumElems = VT.getVectorNumElements();
2871 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2874 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2876 int BitI1 = Mask[i+1];
2877 if (!isUndefOrEqual(BitI, j))
2879 if (!isUndefOrEqual(BitI1, j))
2885 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2886 SmallVector<int, 8> M;
2888 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2891 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2892 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2893 /// MOVSD, and MOVD, i.e. setting the lowest element.
2894 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2895 if (VT.getVectorElementType().getSizeInBits() < 32)
2898 int NumElts = VT.getVectorNumElements();
2900 if (!isUndefOrEqual(Mask[0], NumElts))
2903 for (int i = 1; i < NumElts; ++i)
2904 if (!isUndefOrEqual(Mask[i], i))
2910 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2911 SmallVector<int, 8> M;
2913 return ::isMOVLMask(M, N->getValueType(0));
2916 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2917 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2918 /// element of vector 2 and the other elements to come from vector 1 in order.
2919 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2920 bool V2IsSplat = false, bool V2IsUndef = false) {
2921 int NumOps = VT.getVectorNumElements();
2922 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2925 if (!isUndefOrEqual(Mask[0], 0))
2928 for (int i = 1; i < NumOps; ++i)
2929 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2930 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2931 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2937 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2938 bool V2IsUndef = false) {
2939 SmallVector<int, 8> M;
2941 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2944 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2945 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2946 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2947 if (N->getValueType(0).getVectorNumElements() != 4)
2950 // Expect 1, 1, 3, 3
2951 for (unsigned i = 0; i < 2; ++i) {
2952 int Elt = N->getMaskElt(i);
2953 if (Elt >= 0 && Elt != 1)
2958 for (unsigned i = 2; i < 4; ++i) {
2959 int Elt = N->getMaskElt(i);
2960 if (Elt >= 0 && Elt != 3)
2965 // Don't use movshdup if it can be done with a shufps.
2966 // FIXME: verify that matching u, u, 3, 3 is what we want.
2970 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2971 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2972 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2973 if (N->getValueType(0).getVectorNumElements() != 4)
2976 // Expect 0, 0, 2, 2
2977 for (unsigned i = 0; i < 2; ++i)
2978 if (N->getMaskElt(i) > 0)
2982 for (unsigned i = 2; i < 4; ++i) {
2983 int Elt = N->getMaskElt(i);
2984 if (Elt >= 0 && Elt != 2)
2989 // Don't use movsldup if it can be done with a shufps.
2993 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2994 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2995 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2996 int e = N->getValueType(0).getVectorNumElements() / 2;
2998 for (int i = 0; i < e; ++i)
2999 if (!isUndefOrEqual(N->getMaskElt(i), i))
3001 for (int i = 0; i < e; ++i)
3002 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3007 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3008 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3009 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3010 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3011 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3013 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3015 for (int i = 0; i < NumOperands; ++i) {
3016 int Val = SVOp->getMaskElt(NumOperands-i-1);
3017 if (Val < 0) Val = 0;
3018 if (Val >= NumOperands) Val -= NumOperands;
3020 if (i != NumOperands - 1)
3026 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3027 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3028 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3029 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3031 // 8 nodes, but we only care about the last 4.
3032 for (unsigned i = 7; i >= 4; --i) {
3033 int Val = SVOp->getMaskElt(i);
3042 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3043 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3044 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3045 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3047 // 8 nodes, but we only care about the first 4.
3048 for (int i = 3; i >= 0; --i) {
3049 int Val = SVOp->getMaskElt(i);
3058 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3059 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3060 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3061 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3062 EVT VVT = N->getValueType(0);
3063 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3067 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3068 Val = SVOp->getMaskElt(i);
3072 return (Val - i) * EltSize;
3075 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3077 bool X86::isZeroNode(SDValue Elt) {
3078 return ((isa<ConstantSDNode>(Elt) &&
3079 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3080 (isa<ConstantFPSDNode>(Elt) &&
3081 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3084 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3085 /// their permute mask.
3086 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3087 SelectionDAG &DAG) {
3088 EVT VT = SVOp->getValueType(0);
3089 unsigned NumElems = VT.getVectorNumElements();
3090 SmallVector<int, 8> MaskVec;
3092 for (unsigned i = 0; i != NumElems; ++i) {
3093 int idx = SVOp->getMaskElt(i);
3095 MaskVec.push_back(idx);
3096 else if (idx < (int)NumElems)
3097 MaskVec.push_back(idx + NumElems);
3099 MaskVec.push_back(idx - NumElems);
3101 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3102 SVOp->getOperand(0), &MaskVec[0]);
3105 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3106 /// the two vector operands have swapped position.
3107 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3108 unsigned NumElems = VT.getVectorNumElements();
3109 for (unsigned i = 0; i != NumElems; ++i) {
3113 else if (idx < (int)NumElems)
3114 Mask[i] = idx + NumElems;
3116 Mask[i] = idx - NumElems;
3120 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3121 /// match movhlps. The lower half elements should come from upper half of
3122 /// V1 (and in order), and the upper half elements should come from the upper
3123 /// half of V2 (and in order).
3124 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3125 if (Op->getValueType(0).getVectorNumElements() != 4)
3127 for (unsigned i = 0, e = 2; i != e; ++i)
3128 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3130 for (unsigned i = 2; i != 4; ++i)
3131 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3136 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3137 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3139 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3140 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3142 N = N->getOperand(0).getNode();
3143 if (!ISD::isNON_EXTLoad(N))
3146 *LD = cast<LoadSDNode>(N);
3150 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3151 /// match movlp{s|d}. The lower half elements should come from lower half of
3152 /// V1 (and in order), and the upper half elements should come from the upper
3153 /// half of V2 (and in order). And since V1 will become the source of the
3154 /// MOVLP, it must be either a vector load or a scalar load to vector.
3155 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3156 ShuffleVectorSDNode *Op) {
3157 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3159 // Is V2 is a vector load, don't do this transformation. We will try to use
3160 // load folding shufps op.
3161 if (ISD::isNON_EXTLoad(V2))
3164 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3166 if (NumElems != 2 && NumElems != 4)
3168 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3169 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3171 for (unsigned i = NumElems/2; i != NumElems; ++i)
3172 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3177 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3179 static bool isSplatVector(SDNode *N) {
3180 if (N->getOpcode() != ISD::BUILD_VECTOR)
3183 SDValue SplatValue = N->getOperand(0);
3184 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3185 if (N->getOperand(i) != SplatValue)
3190 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3191 /// to an zero vector.
3192 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3193 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3194 SDValue V1 = N->getOperand(0);
3195 SDValue V2 = N->getOperand(1);
3196 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3197 for (unsigned i = 0; i != NumElems; ++i) {
3198 int Idx = N->getMaskElt(i);
3199 if (Idx >= (int)NumElems) {
3200 unsigned Opc = V2.getOpcode();
3201 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3203 if (Opc != ISD::BUILD_VECTOR ||
3204 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3206 } else if (Idx >= 0) {
3207 unsigned Opc = V1.getOpcode();
3208 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3210 if (Opc != ISD::BUILD_VECTOR ||
3211 !X86::isZeroNode(V1.getOperand(Idx)))
3218 /// getZeroVector - Returns a vector of specified type with all zero elements.
3220 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3222 assert(VT.isVector() && "Expected a vector type");
3224 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3225 // type. This ensures they get CSE'd.
3227 if (VT.getSizeInBits() == 64) { // MMX
3228 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3229 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3230 } else if (HasSSE2) { // SSE2
3231 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3232 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3234 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3235 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3237 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3240 /// getOnesVector - Returns a vector of specified type with all bits set.
3242 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3243 assert(VT.isVector() && "Expected a vector type");
3245 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3246 // type. This ensures they get CSE'd.
3247 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3249 if (VT.getSizeInBits() == 64) // MMX
3250 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3252 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3253 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3257 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3258 /// that point to V2 points to its first element.
3259 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3260 EVT VT = SVOp->getValueType(0);
3261 unsigned NumElems = VT.getVectorNumElements();
3263 bool Changed = false;
3264 SmallVector<int, 8> MaskVec;
3265 SVOp->getMask(MaskVec);
3267 for (unsigned i = 0; i != NumElems; ++i) {
3268 if (MaskVec[i] > (int)NumElems) {
3269 MaskVec[i] = NumElems;
3274 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3275 SVOp->getOperand(1), &MaskVec[0]);
3276 return SDValue(SVOp, 0);
3279 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3280 /// operation of specified width.
3281 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3283 unsigned NumElems = VT.getVectorNumElements();
3284 SmallVector<int, 8> Mask;
3285 Mask.push_back(NumElems);
3286 for (unsigned i = 1; i != NumElems; ++i)
3288 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3291 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3292 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3294 unsigned NumElems = VT.getVectorNumElements();
3295 SmallVector<int, 8> Mask;
3296 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3298 Mask.push_back(i + NumElems);
3300 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3303 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3304 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3306 unsigned NumElems = VT.getVectorNumElements();
3307 unsigned Half = NumElems/2;
3308 SmallVector<int, 8> Mask;
3309 for (unsigned i = 0; i != Half; ++i) {
3310 Mask.push_back(i + Half);
3311 Mask.push_back(i + NumElems + Half);
3313 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3316 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3317 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3319 if (SV->getValueType(0).getVectorNumElements() <= 4)
3320 return SDValue(SV, 0);
3322 EVT PVT = MVT::v4f32;
3323 EVT VT = SV->getValueType(0);
3324 DebugLoc dl = SV->getDebugLoc();
3325 SDValue V1 = SV->getOperand(0);
3326 int NumElems = VT.getVectorNumElements();
3327 int EltNo = SV->getSplatIndex();
3329 // unpack elements to the correct location
3330 while (NumElems > 4) {
3331 if (EltNo < NumElems/2) {
3332 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3334 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3335 EltNo -= NumElems/2;
3340 // Perform the splat.
3341 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3342 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3343 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3344 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3347 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3348 /// vector of zero or undef vector. This produces a shuffle where the low
3349 /// element of V2 is swizzled into the zero/undef vector, landing at element
3350 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3351 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3352 bool isZero, bool HasSSE2,
3353 SelectionDAG &DAG) {
3354 EVT VT = V2.getValueType();
3356 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3357 unsigned NumElems = VT.getVectorNumElements();
3358 SmallVector<int, 16> MaskVec;
3359 for (unsigned i = 0; i != NumElems; ++i)
3360 // If this is the insertion idx, put the low elt of V2 here.
3361 MaskVec.push_back(i == Idx ? NumElems : i);
3362 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3365 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3366 /// a shuffle that is zero.
3368 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3369 bool Low, SelectionDAG &DAG) {
3370 unsigned NumZeros = 0;
3371 for (int i = 0; i < NumElems; ++i) {
3372 unsigned Index = Low ? i : NumElems-i-1;
3373 int Idx = SVOp->getMaskElt(Index);
3378 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3379 if (Elt.getNode() && X86::isZeroNode(Elt))
3387 /// isVectorShift - Returns true if the shuffle can be implemented as a
3388 /// logical left or right shift of a vector.
3389 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3390 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3391 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3392 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3395 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3398 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3402 bool SeenV1 = false;
3403 bool SeenV2 = false;
3404 for (int i = NumZeros; i < NumElems; ++i) {
3405 int Val = isLeft ? (i - NumZeros) : i;
3406 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3418 if (SeenV1 && SeenV2)
3421 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3427 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3429 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3430 unsigned NumNonZero, unsigned NumZero,
3431 SelectionDAG &DAG, TargetLowering &TLI) {
3435 DebugLoc dl = Op.getDebugLoc();
3438 for (unsigned i = 0; i < 16; ++i) {
3439 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3440 if (ThisIsNonZero && First) {
3442 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3444 V = DAG.getUNDEF(MVT::v8i16);
3449 SDValue ThisElt(0, 0), LastElt(0, 0);
3450 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3451 if (LastIsNonZero) {
3452 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3453 MVT::i16, Op.getOperand(i-1));
3455 if (ThisIsNonZero) {
3456 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3457 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3458 ThisElt, DAG.getConstant(8, MVT::i8));
3460 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3464 if (ThisElt.getNode())
3465 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3466 DAG.getIntPtrConstant(i/2));
3470 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3473 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3475 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3476 unsigned NumNonZero, unsigned NumZero,
3477 SelectionDAG &DAG, TargetLowering &TLI) {
3481 DebugLoc dl = Op.getDebugLoc();
3484 for (unsigned i = 0; i < 8; ++i) {
3485 bool isNonZero = (NonZeros & (1 << i)) != 0;
3489 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3491 V = DAG.getUNDEF(MVT::v8i16);
3494 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3495 MVT::v8i16, V, Op.getOperand(i),
3496 DAG.getIntPtrConstant(i));
3503 /// getVShift - Return a vector logical shift node.
3505 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3506 unsigned NumBits, SelectionDAG &DAG,
3507 const TargetLowering &TLI, DebugLoc dl) {
3508 bool isMMX = VT.getSizeInBits() == 64;
3509 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3510 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3511 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3512 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3513 DAG.getNode(Opc, dl, ShVT, SrcOp,
3514 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3518 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3519 SelectionDAG &DAG) {
3521 // Check if the scalar load can be widened into a vector load. And if
3522 // the address is "base + cst" see if the cst can be "absorbed" into
3523 // the shuffle mask.
3524 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3525 SDValue Ptr = LD->getBasePtr();
3526 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3528 EVT PVT = LD->getValueType(0);
3529 if (PVT != MVT::i32 && PVT != MVT::f32)
3534 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3535 FI = FINode->getIndex();
3537 } else if (Ptr.getOpcode() == ISD::ADD &&
3538 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3539 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3540 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3541 Offset = Ptr.getConstantOperandVal(1);
3542 Ptr = Ptr.getOperand(0);
3547 SDValue Chain = LD->getChain();
3548 // Make sure the stack object alignment is at least 16.
3549 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3550 if (DAG.InferPtrAlignment(Ptr) < 16) {
3551 if (MFI->isFixedObjectIndex(FI)) {
3552 // Can't change the alignment. FIXME: It's possible to compute
3553 // the exact stack offset and reference FI + adjust offset instead.
3554 // If someone *really* cares about this. That's the way to implement it.
3557 MFI->setObjectAlignment(FI, 16);
3561 // (Offset % 16) must be multiple of 4. Then address is then
3562 // Ptr + (Offset & ~15).
3565 if ((Offset % 16) & 3)
3567 int64_t StartOffset = Offset & ~15;
3569 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3570 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3572 int EltNo = (Offset - StartOffset) >> 2;
3573 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3574 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3575 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3577 // Canonicalize it to a v4i32 shuffle.
3578 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3579 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3580 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3581 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3588 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3589 DebugLoc dl = Op.getDebugLoc();
3590 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3591 if (ISD::isBuildVectorAllZeros(Op.getNode())
3592 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3593 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3594 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3595 // eliminated on x86-32 hosts.
3596 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3599 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3600 return getOnesVector(Op.getValueType(), DAG, dl);
3601 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3604 EVT VT = Op.getValueType();
3605 EVT ExtVT = VT.getVectorElementType();
3606 unsigned EVTBits = ExtVT.getSizeInBits();
3608 unsigned NumElems = Op.getNumOperands();
3609 unsigned NumZero = 0;
3610 unsigned NumNonZero = 0;
3611 unsigned NonZeros = 0;
3612 bool IsAllConstants = true;
3613 SmallSet<SDValue, 8> Values;
3614 for (unsigned i = 0; i < NumElems; ++i) {
3615 SDValue Elt = Op.getOperand(i);
3616 if (Elt.getOpcode() == ISD::UNDEF)
3619 if (Elt.getOpcode() != ISD::Constant &&
3620 Elt.getOpcode() != ISD::ConstantFP)
3621 IsAllConstants = false;
3622 if (X86::isZeroNode(Elt))
3625 NonZeros |= (1 << i);
3630 if (NumNonZero == 0) {
3631 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3632 return DAG.getUNDEF(VT);
3635 // Special case for single non-zero, non-undef, element.
3636 if (NumNonZero == 1) {
3637 unsigned Idx = CountTrailingZeros_32(NonZeros);
3638 SDValue Item = Op.getOperand(Idx);
3640 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3641 // the value are obviously zero, truncate the value to i32 and do the
3642 // insertion that way. Only do this if the value is non-constant or if the
3643 // value is a constant being inserted into element 0. It is cheaper to do
3644 // a constant pool load than it is to do a movd + shuffle.
3645 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3646 (!IsAllConstants || Idx == 0)) {
3647 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3648 // Handle MMX and SSE both.
3649 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3650 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3652 // Truncate the value (which may itself be a constant) to i32, and
3653 // convert it to a vector with movd (S2V+shuffle to zero extend).
3654 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3655 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3656 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3657 Subtarget->hasSSE2(), DAG);
3659 // Now we have our 32-bit value zero extended in the low element of
3660 // a vector. If Idx != 0, swizzle it into place.
3662 SmallVector<int, 4> Mask;
3663 Mask.push_back(Idx);
3664 for (unsigned i = 1; i != VecElts; ++i)
3666 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3667 DAG.getUNDEF(Item.getValueType()),
3670 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3674 // If we have a constant or non-constant insertion into the low element of
3675 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3676 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3677 // depending on what the source datatype is.
3680 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3681 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3682 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3683 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3684 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3685 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3687 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3688 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3689 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3690 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3691 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3692 Subtarget->hasSSE2(), DAG);
3693 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3697 // Is it a vector logical left shift?
3698 if (NumElems == 2 && Idx == 1 &&
3699 X86::isZeroNode(Op.getOperand(0)) &&
3700 !X86::isZeroNode(Op.getOperand(1))) {
3701 unsigned NumBits = VT.getSizeInBits();
3702 return getVShift(true, VT,
3703 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3704 VT, Op.getOperand(1)),
3705 NumBits/2, DAG, *this, dl);
3708 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3711 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3712 // is a non-constant being inserted into an element other than the low one,
3713 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3714 // movd/movss) to move this into the low element, then shuffle it into
3716 if (EVTBits == 32) {
3717 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3719 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3720 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3721 Subtarget->hasSSE2(), DAG);
3722 SmallVector<int, 8> MaskVec;
3723 for (unsigned i = 0; i < NumElems; i++)
3724 MaskVec.push_back(i == Idx ? 0 : 1);
3725 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3729 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3730 if (Values.size() == 1) {
3731 if (EVTBits == 32) {
3732 // Instead of a shuffle like this:
3733 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3734 // Check if it's possible to issue this instead.
3735 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3736 unsigned Idx = CountTrailingZeros_32(NonZeros);
3737 SDValue Item = Op.getOperand(Idx);
3738 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3739 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3744 // A vector full of immediates; various special cases are already
3745 // handled, so this is best done with a single constant-pool load.
3749 // Let legalizer expand 2-wide build_vectors.
3750 if (EVTBits == 64) {
3751 if (NumNonZero == 1) {
3752 // One half is zero or undef.
3753 unsigned Idx = CountTrailingZeros_32(NonZeros);
3754 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3755 Op.getOperand(Idx));
3756 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3757 Subtarget->hasSSE2(), DAG);
3762 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3763 if (EVTBits == 8 && NumElems == 16) {
3764 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3766 if (V.getNode()) return V;
3769 if (EVTBits == 16 && NumElems == 8) {
3770 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3772 if (V.getNode()) return V;
3775 // If element VT is == 32 bits, turn it into a number of shuffles.
3776 SmallVector<SDValue, 8> V;
3778 if (NumElems == 4 && NumZero > 0) {
3779 for (unsigned i = 0; i < 4; ++i) {
3780 bool isZero = !(NonZeros & (1 << i));
3782 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3784 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3787 for (unsigned i = 0; i < 2; ++i) {
3788 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3791 V[i] = V[i*2]; // Must be a zero vector.
3794 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3797 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3800 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3805 SmallVector<int, 8> MaskVec;
3806 bool Reverse = (NonZeros & 0x3) == 2;
3807 for (unsigned i = 0; i < 2; ++i)
3808 MaskVec.push_back(Reverse ? 1-i : i);
3809 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3810 for (unsigned i = 0; i < 2; ++i)
3811 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3812 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3815 if (Values.size() > 2) {
3816 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3817 // values to be inserted is equal to the number of elements, in which case
3818 // use the unpack code below in the hopes of matching the consecutive elts
3819 // load merge pattern for shuffles.
3820 // FIXME: We could probably just check that here directly.
3821 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3822 getSubtarget()->hasSSE41()) {
3823 V[0] = DAG.getUNDEF(VT);
3824 for (unsigned i = 0; i < NumElems; ++i)
3825 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3826 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3827 Op.getOperand(i), DAG.getIntPtrConstant(i));
3830 // Expand into a number of unpckl*.
3832 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3833 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3834 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3835 for (unsigned i = 0; i < NumElems; ++i)
3836 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3838 while (NumElems != 0) {
3839 for (unsigned i = 0; i < NumElems; ++i)
3840 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3850 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3851 // We support concatenate two MMX registers and place them in a MMX
3852 // register. This is better than doing a stack convert.
3853 DebugLoc dl = Op.getDebugLoc();
3854 EVT ResVT = Op.getValueType();
3855 assert(Op.getNumOperands() == 2);
3856 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3857 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3859 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3860 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3861 InVec = Op.getOperand(1);
3862 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3863 unsigned NumElts = ResVT.getVectorNumElements();
3864 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3865 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3866 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3868 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3869 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3870 Mask[0] = 0; Mask[1] = 2;
3871 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3873 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3876 // v8i16 shuffles - Prefer shuffles in the following order:
3877 // 1. [all] pshuflw, pshufhw, optional move
3878 // 2. [ssse3] 1 x pshufb
3879 // 3. [ssse3] 2 x pshufb + 1 x por
3880 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3882 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3883 SelectionDAG &DAG, X86TargetLowering &TLI) {
3884 SDValue V1 = SVOp->getOperand(0);
3885 SDValue V2 = SVOp->getOperand(1);
3886 DebugLoc dl = SVOp->getDebugLoc();
3887 SmallVector<int, 8> MaskVals;
3889 // Determine if more than 1 of the words in each of the low and high quadwords
3890 // of the result come from the same quadword of one of the two inputs. Undef
3891 // mask values count as coming from any quadword, for better codegen.
3892 SmallVector<unsigned, 4> LoQuad(4);
3893 SmallVector<unsigned, 4> HiQuad(4);
3894 BitVector InputQuads(4);
3895 for (unsigned i = 0; i < 8; ++i) {
3896 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3897 int EltIdx = SVOp->getMaskElt(i);
3898 MaskVals.push_back(EltIdx);
3907 InputQuads.set(EltIdx / 4);
3910 int BestLoQuad = -1;
3911 unsigned MaxQuad = 1;
3912 for (unsigned i = 0; i < 4; ++i) {
3913 if (LoQuad[i] > MaxQuad) {
3915 MaxQuad = LoQuad[i];
3919 int BestHiQuad = -1;
3921 for (unsigned i = 0; i < 4; ++i) {
3922 if (HiQuad[i] > MaxQuad) {
3924 MaxQuad = HiQuad[i];
3928 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3929 // of the two input vectors, shuffle them into one input vector so only a
3930 // single pshufb instruction is necessary. If There are more than 2 input
3931 // quads, disable the next transformation since it does not help SSSE3.
3932 bool V1Used = InputQuads[0] || InputQuads[1];
3933 bool V2Used = InputQuads[2] || InputQuads[3];
3934 if (TLI.getSubtarget()->hasSSSE3()) {
3935 if (InputQuads.count() == 2 && V1Used && V2Used) {
3936 BestLoQuad = InputQuads.find_first();
3937 BestHiQuad = InputQuads.find_next(BestLoQuad);
3939 if (InputQuads.count() > 2) {
3945 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3946 // the shuffle mask. If a quad is scored as -1, that means that it contains
3947 // words from all 4 input quadwords.
3949 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3950 SmallVector<int, 8> MaskV;
3951 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3952 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3953 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3954 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3955 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3956 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3958 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3959 // source words for the shuffle, to aid later transformations.
3960 bool AllWordsInNewV = true;
3961 bool InOrder[2] = { true, true };
3962 for (unsigned i = 0; i != 8; ++i) {
3963 int idx = MaskVals[i];
3965 InOrder[i/4] = false;
3966 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3968 AllWordsInNewV = false;
3972 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3973 if (AllWordsInNewV) {
3974 for (int i = 0; i != 8; ++i) {
3975 int idx = MaskVals[i];
3978 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3979 if ((idx != i) && idx < 4)
3981 if ((idx != i) && idx > 3)
3990 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3991 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3992 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3993 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3994 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3998 // If we have SSSE3, and all words of the result are from 1 input vector,
3999 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
4000 // is present, fall back to case 4.
4001 if (TLI.getSubtarget()->hasSSSE3()) {
4002 SmallVector<SDValue,16> pshufbMask;
4004 // If we have elements from both input vectors, set the high bit of the
4005 // shuffle mask element to zero out elements that come from V2 in the V1
4006 // mask, and elements that come from V1 in the V2 mask, so that the two
4007 // results can be OR'd together.
4008 bool TwoInputs = V1Used && V2Used;
4009 for (unsigned i = 0; i != 8; ++i) {
4010 int EltIdx = MaskVals[i] * 2;
4011 if (TwoInputs && (EltIdx >= 16)) {
4012 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4013 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4016 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4017 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4019 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4020 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4021 DAG.getNode(ISD::BUILD_VECTOR, dl,
4022 MVT::v16i8, &pshufbMask[0], 16));
4024 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4026 // Calculate the shuffle mask for the second input, shuffle it, and
4027 // OR it with the first shuffled input.
4029 for (unsigned i = 0; i != 8; ++i) {
4030 int EltIdx = MaskVals[i] * 2;
4032 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4033 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4036 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4037 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4039 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4040 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4041 DAG.getNode(ISD::BUILD_VECTOR, dl,
4042 MVT::v16i8, &pshufbMask[0], 16));
4043 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4044 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4047 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4048 // and update MaskVals with new element order.
4049 BitVector InOrder(8);
4050 if (BestLoQuad >= 0) {
4051 SmallVector<int, 8> MaskV;
4052 for (int i = 0; i != 4; ++i) {
4053 int idx = MaskVals[i];
4055 MaskV.push_back(-1);
4057 } else if ((idx / 4) == BestLoQuad) {
4058 MaskV.push_back(idx & 3);
4061 MaskV.push_back(-1);
4064 for (unsigned i = 4; i != 8; ++i)
4066 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4070 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4071 // and update MaskVals with the new element order.
4072 if (BestHiQuad >= 0) {
4073 SmallVector<int, 8> MaskV;
4074 for (unsigned i = 0; i != 4; ++i)
4076 for (unsigned i = 4; i != 8; ++i) {
4077 int idx = MaskVals[i];
4079 MaskV.push_back(-1);
4081 } else if ((idx / 4) == BestHiQuad) {
4082 MaskV.push_back((idx & 3) + 4);
4085 MaskV.push_back(-1);
4088 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4092 // In case BestHi & BestLo were both -1, which means each quadword has a word
4093 // from each of the four input quadwords, calculate the InOrder bitvector now
4094 // before falling through to the insert/extract cleanup.
4095 if (BestLoQuad == -1 && BestHiQuad == -1) {
4097 for (int i = 0; i != 8; ++i)
4098 if (MaskVals[i] < 0 || MaskVals[i] == i)
4102 // The other elements are put in the right place using pextrw and pinsrw.
4103 for (unsigned i = 0; i != 8; ++i) {
4106 int EltIdx = MaskVals[i];
4109 SDValue ExtOp = (EltIdx < 8)
4110 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4111 DAG.getIntPtrConstant(EltIdx))
4112 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4113 DAG.getIntPtrConstant(EltIdx - 8));
4114 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4115 DAG.getIntPtrConstant(i));
4120 // v16i8 shuffles - Prefer shuffles in the following order:
4121 // 1. [ssse3] 1 x pshufb
4122 // 2. [ssse3] 2 x pshufb + 1 x por
4123 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4125 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4126 SelectionDAG &DAG, X86TargetLowering &TLI) {
4127 SDValue V1 = SVOp->getOperand(0);
4128 SDValue V2 = SVOp->getOperand(1);
4129 DebugLoc dl = SVOp->getDebugLoc();
4130 SmallVector<int, 16> MaskVals;
4131 SVOp->getMask(MaskVals);
4133 // If we have SSSE3, case 1 is generated when all result bytes come from
4134 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4135 // present, fall back to case 3.
4136 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4139 for (unsigned i = 0; i < 16; ++i) {
4140 int EltIdx = MaskVals[i];
4149 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4150 if (TLI.getSubtarget()->hasSSSE3()) {
4151 SmallVector<SDValue,16> pshufbMask;
4153 // If all result elements are from one input vector, then only translate
4154 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4156 // Otherwise, we have elements from both input vectors, and must zero out
4157 // elements that come from V2 in the first mask, and V1 in the second mask
4158 // so that we can OR them together.
4159 bool TwoInputs = !(V1Only || V2Only);
4160 for (unsigned i = 0; i != 16; ++i) {
4161 int EltIdx = MaskVals[i];
4162 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4163 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4166 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4168 // If all the elements are from V2, assign it to V1 and return after
4169 // building the first pshufb.
4172 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4173 DAG.getNode(ISD::BUILD_VECTOR, dl,
4174 MVT::v16i8, &pshufbMask[0], 16));
4178 // Calculate the shuffle mask for the second input, shuffle it, and
4179 // OR it with the first shuffled input.
4181 for (unsigned i = 0; i != 16; ++i) {
4182 int EltIdx = MaskVals[i];
4184 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4187 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4189 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4190 DAG.getNode(ISD::BUILD_VECTOR, dl,
4191 MVT::v16i8, &pshufbMask[0], 16));
4192 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4195 // No SSSE3 - Calculate in place words and then fix all out of place words
4196 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4197 // the 16 different words that comprise the two doublequadword input vectors.
4198 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4199 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4200 SDValue NewV = V2Only ? V2 : V1;
4201 for (int i = 0; i != 8; ++i) {
4202 int Elt0 = MaskVals[i*2];
4203 int Elt1 = MaskVals[i*2+1];
4205 // This word of the result is all undef, skip it.
4206 if (Elt0 < 0 && Elt1 < 0)
4209 // This word of the result is already in the correct place, skip it.
4210 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4212 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4215 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4216 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4219 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4220 // using a single extract together, load it and store it.
4221 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4222 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4223 DAG.getIntPtrConstant(Elt1 / 2));
4224 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4225 DAG.getIntPtrConstant(i));
4229 // If Elt1 is defined, extract it from the appropriate source. If the
4230 // source byte is not also odd, shift the extracted word left 8 bits
4231 // otherwise clear the bottom 8 bits if we need to do an or.
4233 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4234 DAG.getIntPtrConstant(Elt1 / 2));
4235 if ((Elt1 & 1) == 0)
4236 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4237 DAG.getConstant(8, TLI.getShiftAmountTy()));
4239 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4240 DAG.getConstant(0xFF00, MVT::i16));
4242 // If Elt0 is defined, extract it from the appropriate source. If the
4243 // source byte is not also even, shift the extracted word right 8 bits. If
4244 // Elt1 was also defined, OR the extracted values together before
4245 // inserting them in the result.
4247 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4248 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4249 if ((Elt0 & 1) != 0)
4250 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4251 DAG.getConstant(8, TLI.getShiftAmountTy()));
4253 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4254 DAG.getConstant(0x00FF, MVT::i16));
4255 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4258 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4259 DAG.getIntPtrConstant(i));
4261 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4264 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4265 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4266 /// done when every pair / quad of shuffle mask elements point to elements in
4267 /// the right sequence. e.g.
4268 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4270 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4272 TargetLowering &TLI, DebugLoc dl) {
4273 EVT VT = SVOp->getValueType(0);
4274 SDValue V1 = SVOp->getOperand(0);
4275 SDValue V2 = SVOp->getOperand(1);
4276 unsigned NumElems = VT.getVectorNumElements();
4277 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4278 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4279 EVT MaskEltVT = MaskVT.getVectorElementType();
4281 switch (VT.getSimpleVT().SimpleTy) {
4282 default: assert(false && "Unexpected!");
4283 case MVT::v4f32: NewVT = MVT::v2f64; break;
4284 case MVT::v4i32: NewVT = MVT::v2i64; break;
4285 case MVT::v8i16: NewVT = MVT::v4i32; break;
4286 case MVT::v16i8: NewVT = MVT::v4i32; break;
4289 if (NewWidth == 2) {
4295 int Scale = NumElems / NewWidth;
4296 SmallVector<int, 8> MaskVec;
4297 for (unsigned i = 0; i < NumElems; i += Scale) {
4299 for (int j = 0; j < Scale; ++j) {
4300 int EltIdx = SVOp->getMaskElt(i+j);
4304 StartIdx = EltIdx - (EltIdx % Scale);
4305 if (EltIdx != StartIdx + j)
4309 MaskVec.push_back(-1);
4311 MaskVec.push_back(StartIdx / Scale);
4314 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4315 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4316 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4319 /// getVZextMovL - Return a zero-extending vector move low node.
4321 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4322 SDValue SrcOp, SelectionDAG &DAG,
4323 const X86Subtarget *Subtarget, DebugLoc dl) {
4324 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4325 LoadSDNode *LD = NULL;
4326 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4327 LD = dyn_cast<LoadSDNode>(SrcOp);
4329 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4331 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4332 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4333 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4334 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4335 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4337 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4338 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4339 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4340 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4348 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4349 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4350 DAG.getNode(ISD::BIT_CONVERT, dl,
4354 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4357 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4358 SDValue V1 = SVOp->getOperand(0);
4359 SDValue V2 = SVOp->getOperand(1);
4360 DebugLoc dl = SVOp->getDebugLoc();
4361 EVT VT = SVOp->getValueType(0);
4363 SmallVector<std::pair<int, int>, 8> Locs;
4365 SmallVector<int, 8> Mask1(4U, -1);
4366 SmallVector<int, 8> PermMask;
4367 SVOp->getMask(PermMask);
4371 for (unsigned i = 0; i != 4; ++i) {
4372 int Idx = PermMask[i];
4374 Locs[i] = std::make_pair(-1, -1);
4376 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4378 Locs[i] = std::make_pair(0, NumLo);
4382 Locs[i] = std::make_pair(1, NumHi);
4384 Mask1[2+NumHi] = Idx;
4390 if (NumLo <= 2 && NumHi <= 2) {
4391 // If no more than two elements come from either vector. This can be
4392 // implemented with two shuffles. First shuffle gather the elements.
4393 // The second shuffle, which takes the first shuffle as both of its
4394 // vector operands, put the elements into the right order.
4395 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4397 SmallVector<int, 8> Mask2(4U, -1);
4399 for (unsigned i = 0; i != 4; ++i) {
4400 if (Locs[i].first == -1)
4403 unsigned Idx = (i < 2) ? 0 : 4;
4404 Idx += Locs[i].first * 2 + Locs[i].second;
4409 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4410 } else if (NumLo == 3 || NumHi == 3) {
4411 // Otherwise, we must have three elements from one vector, call it X, and
4412 // one element from the other, call it Y. First, use a shufps to build an
4413 // intermediate vector with the one element from Y and the element from X
4414 // that will be in the same half in the final destination (the indexes don't
4415 // matter). Then, use a shufps to build the final vector, taking the half
4416 // containing the element from Y from the intermediate, and the other half
4419 // Normalize it so the 3 elements come from V1.
4420 CommuteVectorShuffleMask(PermMask, VT);
4424 // Find the element from V2.
4426 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4427 int Val = PermMask[HiIndex];
4434 Mask1[0] = PermMask[HiIndex];
4436 Mask1[2] = PermMask[HiIndex^1];
4438 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4441 Mask1[0] = PermMask[0];
4442 Mask1[1] = PermMask[1];
4443 Mask1[2] = HiIndex & 1 ? 6 : 4;
4444 Mask1[3] = HiIndex & 1 ? 4 : 6;
4445 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4447 Mask1[0] = HiIndex & 1 ? 2 : 0;
4448 Mask1[1] = HiIndex & 1 ? 0 : 2;
4449 Mask1[2] = PermMask[2];
4450 Mask1[3] = PermMask[3];
4455 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4459 // Break it into (shuffle shuffle_hi, shuffle_lo).
4461 SmallVector<int,8> LoMask(4U, -1);
4462 SmallVector<int,8> HiMask(4U, -1);
4464 SmallVector<int,8> *MaskPtr = &LoMask;
4465 unsigned MaskIdx = 0;
4468 for (unsigned i = 0; i != 4; ++i) {
4475 int Idx = PermMask[i];
4477 Locs[i] = std::make_pair(-1, -1);
4478 } else if (Idx < 4) {
4479 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4480 (*MaskPtr)[LoIdx] = Idx;
4483 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4484 (*MaskPtr)[HiIdx] = Idx;
4489 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4490 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4491 SmallVector<int, 8> MaskOps;
4492 for (unsigned i = 0; i != 4; ++i) {
4493 if (Locs[i].first == -1) {
4494 MaskOps.push_back(-1);
4496 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4497 MaskOps.push_back(Idx);
4500 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4504 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4505 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4506 SDValue V1 = Op.getOperand(0);
4507 SDValue V2 = Op.getOperand(1);
4508 EVT VT = Op.getValueType();
4509 DebugLoc dl = Op.getDebugLoc();
4510 unsigned NumElems = VT.getVectorNumElements();
4511 bool isMMX = VT.getSizeInBits() == 64;
4512 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4513 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4514 bool V1IsSplat = false;
4515 bool V2IsSplat = false;
4517 if (isZeroShuffle(SVOp))
4518 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4520 // Promote splats to v4f32.
4521 if (SVOp->isSplat()) {
4522 if (isMMX || NumElems < 4)
4524 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4527 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4529 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4530 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4531 if (NewOp.getNode())
4532 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4533 LowerVECTOR_SHUFFLE(NewOp, DAG));
4534 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4535 // FIXME: Figure out a cleaner way to do this.
4536 // Try to make use of movq to zero out the top part.
4537 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4538 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4539 if (NewOp.getNode()) {
4540 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4541 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4542 DAG, Subtarget, dl);
4544 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4545 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4546 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4547 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4548 DAG, Subtarget, dl);
4552 if (X86::isPSHUFDMask(SVOp))
4555 // Check if this can be converted into a logical shift.
4556 bool isLeft = false;
4559 bool isShift = getSubtarget()->hasSSE2() &&
4560 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4561 if (isShift && ShVal.hasOneUse()) {
4562 // If the shifted value has multiple uses, it may be cheaper to use
4563 // v_set0 + movlhps or movhlps, etc.
4564 EVT EltVT = VT.getVectorElementType();
4565 ShAmt *= EltVT.getSizeInBits();
4566 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4569 if (X86::isMOVLMask(SVOp)) {
4572 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4573 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4578 // FIXME: fold these into legal mask.
4579 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4580 X86::isMOVSLDUPMask(SVOp) ||
4581 X86::isMOVHLPSMask(SVOp) ||
4582 X86::isMOVLHPSMask(SVOp) ||
4583 X86::isMOVLPMask(SVOp)))
4586 if (ShouldXformToMOVHLPS(SVOp) ||
4587 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4588 return CommuteVectorShuffle(SVOp, DAG);
4591 // No better options. Use a vshl / vsrl.
4592 EVT EltVT = VT.getVectorElementType();
4593 ShAmt *= EltVT.getSizeInBits();
4594 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4597 bool Commuted = false;
4598 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4599 // 1,1,1,1 -> v8i16 though.
4600 V1IsSplat = isSplatVector(V1.getNode());
4601 V2IsSplat = isSplatVector(V2.getNode());
4603 // Canonicalize the splat or undef, if present, to be on the RHS.
4604 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4605 Op = CommuteVectorShuffle(SVOp, DAG);
4606 SVOp = cast<ShuffleVectorSDNode>(Op);
4607 V1 = SVOp->getOperand(0);
4608 V2 = SVOp->getOperand(1);
4609 std::swap(V1IsSplat, V2IsSplat);
4610 std::swap(V1IsUndef, V2IsUndef);
4614 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4615 // Shuffling low element of v1 into undef, just return v1.
4618 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4619 // the instruction selector will not match, so get a canonical MOVL with
4620 // swapped operands to undo the commute.
4621 return getMOVL(DAG, dl, VT, V2, V1);
4624 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4625 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4626 X86::isUNPCKLMask(SVOp) ||
4627 X86::isUNPCKHMask(SVOp))
4631 // Normalize mask so all entries that point to V2 points to its first
4632 // element then try to match unpck{h|l} again. If match, return a
4633 // new vector_shuffle with the corrected mask.
4634 SDValue NewMask = NormalizeMask(SVOp, DAG);
4635 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4636 if (NSVOp != SVOp) {
4637 if (X86::isUNPCKLMask(NSVOp, true)) {
4639 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4646 // Commute is back and try unpck* again.
4647 // FIXME: this seems wrong.
4648 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4649 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4650 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4651 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4652 X86::isUNPCKLMask(NewSVOp) ||
4653 X86::isUNPCKHMask(NewSVOp))
4657 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4659 // Normalize the node to match x86 shuffle ops if needed
4660 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4661 return CommuteVectorShuffle(SVOp, DAG);
4663 // Check for legal shuffle and return?
4664 SmallVector<int, 16> PermMask;
4665 SVOp->getMask(PermMask);
4666 if (isShuffleMaskLegal(PermMask, VT))
4669 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4670 if (VT == MVT::v8i16) {
4671 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4672 if (NewOp.getNode())
4676 if (VT == MVT::v16i8) {
4677 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4678 if (NewOp.getNode())
4682 // Handle all 4 wide cases with a number of shuffles except for MMX.
4683 if (NumElems == 4 && !isMMX)
4684 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4690 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4691 SelectionDAG &DAG) {
4692 EVT VT = Op.getValueType();
4693 DebugLoc dl = Op.getDebugLoc();
4694 if (VT.getSizeInBits() == 8) {
4695 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4696 Op.getOperand(0), Op.getOperand(1));
4697 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4698 DAG.getValueType(VT));
4699 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4700 } else if (VT.getSizeInBits() == 16) {
4701 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4702 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4704 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4705 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4706 DAG.getNode(ISD::BIT_CONVERT, dl,
4710 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4711 Op.getOperand(0), Op.getOperand(1));
4712 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4713 DAG.getValueType(VT));
4714 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4715 } else if (VT == MVT::f32) {
4716 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4717 // the result back to FR32 register. It's only worth matching if the
4718 // result has a single use which is a store or a bitcast to i32. And in
4719 // the case of a store, it's not worth it if the index is a constant 0,
4720 // because a MOVSSmr can be used instead, which is smaller and faster.
4721 if (!Op.hasOneUse())
4723 SDNode *User = *Op.getNode()->use_begin();
4724 if ((User->getOpcode() != ISD::STORE ||
4725 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4726 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4727 (User->getOpcode() != ISD::BIT_CONVERT ||
4728 User->getValueType(0) != MVT::i32))
4730 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4731 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4734 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4735 } else if (VT == MVT::i32) {
4736 // ExtractPS works with constant index.
4737 if (isa<ConstantSDNode>(Op.getOperand(1)))
4745 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4746 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4749 if (Subtarget->hasSSE41()) {
4750 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4755 EVT VT = Op.getValueType();
4756 DebugLoc dl = Op.getDebugLoc();
4757 // TODO: handle v16i8.
4758 if (VT.getSizeInBits() == 16) {
4759 SDValue Vec = Op.getOperand(0);
4760 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4762 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4763 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4764 DAG.getNode(ISD::BIT_CONVERT, dl,
4767 // Transform it so it match pextrw which produces a 32-bit result.
4768 EVT EltVT = MVT::i32;
4769 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4770 Op.getOperand(0), Op.getOperand(1));
4771 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4772 DAG.getValueType(VT));
4773 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4774 } else if (VT.getSizeInBits() == 32) {
4775 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4779 // SHUFPS the element to the lowest double word, then movss.
4780 int Mask[4] = { Idx, -1, -1, -1 };
4781 EVT VVT = Op.getOperand(0).getValueType();
4782 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4783 DAG.getUNDEF(VVT), Mask);
4784 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4785 DAG.getIntPtrConstant(0));
4786 } else if (VT.getSizeInBits() == 64) {
4787 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4788 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4789 // to match extract_elt for f64.
4790 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4794 // UNPCKHPD the element to the lowest double word, then movsd.
4795 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4796 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4797 int Mask[2] = { 1, -1 };
4798 EVT VVT = Op.getOperand(0).getValueType();
4799 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4800 DAG.getUNDEF(VVT), Mask);
4801 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4802 DAG.getIntPtrConstant(0));
4809 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4810 EVT VT = Op.getValueType();
4811 EVT EltVT = VT.getVectorElementType();
4812 DebugLoc dl = Op.getDebugLoc();
4814 SDValue N0 = Op.getOperand(0);
4815 SDValue N1 = Op.getOperand(1);
4816 SDValue N2 = Op.getOperand(2);
4818 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4819 isa<ConstantSDNode>(N2)) {
4821 if (VT == MVT::v8i16)
4822 Opc = X86ISD::PINSRW;
4823 else if (VT == MVT::v4i16)
4824 Opc = X86ISD::MMX_PINSRW;
4825 else if (VT == MVT::v16i8)
4826 Opc = X86ISD::PINSRB;
4828 Opc = X86ISD::PINSRB;
4830 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4832 if (N1.getValueType() != MVT::i32)
4833 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4834 if (N2.getValueType() != MVT::i32)
4835 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4836 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4837 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4838 // Bits [7:6] of the constant are the source select. This will always be
4839 // zero here. The DAG Combiner may combine an extract_elt index into these
4840 // bits. For example (insert (extract, 3), 2) could be matched by putting
4841 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4842 // Bits [5:4] of the constant are the destination select. This is the
4843 // value of the incoming immediate.
4844 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4845 // combine either bitwise AND or insert of float 0.0 to set these bits.
4846 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4847 // Create this as a scalar to vector..
4848 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4849 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4850 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4851 // PINSR* works with constant index.
4858 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4859 EVT VT = Op.getValueType();
4860 EVT EltVT = VT.getVectorElementType();
4862 if (Subtarget->hasSSE41())
4863 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4865 if (EltVT == MVT::i8)
4868 DebugLoc dl = Op.getDebugLoc();
4869 SDValue N0 = Op.getOperand(0);
4870 SDValue N1 = Op.getOperand(1);
4871 SDValue N2 = Op.getOperand(2);
4873 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4874 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4875 // as its second argument.
4876 if (N1.getValueType() != MVT::i32)
4877 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4878 if (N2.getValueType() != MVT::i32)
4879 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4880 return DAG.getNode(VT == MVT::v8i16 ? X86ISD::PINSRW : X86ISD::MMX_PINSRW,
4881 dl, VT, N0, N1, N2);
4887 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4888 DebugLoc dl = Op.getDebugLoc();
4889 if (Op.getValueType() == MVT::v2f32)
4890 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4891 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4892 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4893 Op.getOperand(0))));
4895 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4896 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4898 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4899 EVT VT = MVT::v2i32;
4900 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4907 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4908 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4911 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4912 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4913 // one of the above mentioned nodes. It has to be wrapped because otherwise
4914 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4915 // be used to form addressing mode. These wrapped nodes will be selected
4918 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4919 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4921 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4923 unsigned char OpFlag = 0;
4924 unsigned WrapperKind = X86ISD::Wrapper;
4925 CodeModel::Model M = getTargetMachine().getCodeModel();
4927 if (Subtarget->isPICStyleRIPRel() &&
4928 (M == CodeModel::Small || M == CodeModel::Kernel))
4929 WrapperKind = X86ISD::WrapperRIP;
4930 else if (Subtarget->isPICStyleGOT())
4931 OpFlag = X86II::MO_GOTOFF;
4932 else if (Subtarget->isPICStyleStubPIC())
4933 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4935 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4937 CP->getOffset(), OpFlag);
4938 DebugLoc DL = CP->getDebugLoc();
4939 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4940 // With PIC, the address is actually $g + Offset.
4942 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4943 DAG.getNode(X86ISD::GlobalBaseReg,
4944 DebugLoc::getUnknownLoc(), getPointerTy()),
4951 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4952 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4954 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4956 unsigned char OpFlag = 0;
4957 unsigned WrapperKind = X86ISD::Wrapper;
4958 CodeModel::Model M = getTargetMachine().getCodeModel();
4960 if (Subtarget->isPICStyleRIPRel() &&
4961 (M == CodeModel::Small || M == CodeModel::Kernel))
4962 WrapperKind = X86ISD::WrapperRIP;
4963 else if (Subtarget->isPICStyleGOT())
4964 OpFlag = X86II::MO_GOTOFF;
4965 else if (Subtarget->isPICStyleStubPIC())
4966 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4968 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4970 DebugLoc DL = JT->getDebugLoc();
4971 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4973 // With PIC, the address is actually $g + Offset.
4975 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4976 DAG.getNode(X86ISD::GlobalBaseReg,
4977 DebugLoc::getUnknownLoc(), getPointerTy()),
4985 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4986 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4988 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4990 unsigned char OpFlag = 0;
4991 unsigned WrapperKind = X86ISD::Wrapper;
4992 CodeModel::Model M = getTargetMachine().getCodeModel();
4994 if (Subtarget->isPICStyleRIPRel() &&
4995 (M == CodeModel::Small || M == CodeModel::Kernel))
4996 WrapperKind = X86ISD::WrapperRIP;
4997 else if (Subtarget->isPICStyleGOT())
4998 OpFlag = X86II::MO_GOTOFF;
4999 else if (Subtarget->isPICStyleStubPIC())
5000 OpFlag = X86II::MO_PIC_BASE_OFFSET;
5002 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
5004 DebugLoc DL = Op.getDebugLoc();
5005 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
5008 // With PIC, the address is actually $g + Offset.
5009 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5010 !Subtarget->is64Bit()) {
5011 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5012 DAG.getNode(X86ISD::GlobalBaseReg,
5013 DebugLoc::getUnknownLoc(),
5022 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
5023 // Create the TargetBlockAddressAddress node.
5024 unsigned char OpFlags =
5025 Subtarget->ClassifyBlockAddressReference();
5026 CodeModel::Model M = getTargetMachine().getCodeModel();
5027 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5028 DebugLoc dl = Op.getDebugLoc();
5029 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5030 /*isTarget=*/true, OpFlags);
5032 if (Subtarget->isPICStyleRIPRel() &&
5033 (M == CodeModel::Small || M == CodeModel::Kernel))
5034 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5036 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5038 // With PIC, the address is actually $g + Offset.
5039 if (isGlobalRelativeToPICBase(OpFlags)) {
5040 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5041 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5049 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5051 SelectionDAG &DAG) const {
5052 // Create the TargetGlobalAddress node, folding in the constant
5053 // offset if it is legal.
5054 unsigned char OpFlags =
5055 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5056 CodeModel::Model M = getTargetMachine().getCodeModel();
5058 if (OpFlags == X86II::MO_NO_FLAG &&
5059 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5060 // A direct static reference to a global.
5061 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5064 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5067 if (Subtarget->isPICStyleRIPRel() &&
5068 (M == CodeModel::Small || M == CodeModel::Kernel))
5069 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5071 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5073 // With PIC, the address is actually $g + Offset.
5074 if (isGlobalRelativeToPICBase(OpFlags)) {
5075 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5076 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5080 // For globals that require a load from a stub to get the address, emit the
5082 if (isGlobalStubReference(OpFlags))
5083 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5084 PseudoSourceValue::getGOT(), 0, false, false, 0);
5086 // If there was a non-zero offset that we didn't fold, create an explicit
5089 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5090 DAG.getConstant(Offset, getPointerTy()));
5096 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5097 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5098 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5099 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5103 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5104 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5105 unsigned char OperandFlags) {
5106 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5107 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5108 DebugLoc dl = GA->getDebugLoc();
5109 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5110 GA->getValueType(0),
5114 SDValue Ops[] = { Chain, TGA, *InFlag };
5115 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5117 SDValue Ops[] = { Chain, TGA };
5118 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5121 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5122 MFI->setHasCalls(true);
5124 SDValue Flag = Chain.getValue(1);
5125 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5128 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5130 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5133 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5134 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5135 DAG.getNode(X86ISD::GlobalBaseReg,
5136 DebugLoc::getUnknownLoc(),
5138 InFlag = Chain.getValue(1);
5140 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5143 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5145 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5147 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5148 X86::RAX, X86II::MO_TLSGD);
5151 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5152 // "local exec" model.
5153 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5154 const EVT PtrVT, TLSModel::Model model,
5156 DebugLoc dl = GA->getDebugLoc();
5157 // Get the Thread Pointer
5158 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5159 DebugLoc::getUnknownLoc(), PtrVT,
5160 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5163 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5164 NULL, 0, false, false, 0);
5166 unsigned char OperandFlags = 0;
5167 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5169 unsigned WrapperKind = X86ISD::Wrapper;
5170 if (model == TLSModel::LocalExec) {
5171 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5172 } else if (is64Bit) {
5173 assert(model == TLSModel::InitialExec);
5174 OperandFlags = X86II::MO_GOTTPOFF;
5175 WrapperKind = X86ISD::WrapperRIP;
5177 assert(model == TLSModel::InitialExec);
5178 OperandFlags = X86II::MO_INDNTPOFF;
5181 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5183 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5184 GA->getOffset(), OperandFlags);
5185 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5187 if (model == TLSModel::InitialExec)
5188 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5189 PseudoSourceValue::getGOT(), 0, false, false, 0);
5191 // The address of the thread local variable is the add of the thread
5192 // pointer with the offset of the variable.
5193 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5197 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5198 // TODO: implement the "local dynamic" model
5199 // TODO: implement the "initial exec"model for pic executables
5200 assert(Subtarget->isTargetELF() &&
5201 "TLS not implemented for non-ELF targets");
5202 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5203 const GlobalValue *GV = GA->getGlobal();
5205 // If GV is an alias then use the aliasee for determining
5206 // thread-localness.
5207 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5208 GV = GA->resolveAliasedGlobal(false);
5210 TLSModel::Model model = getTLSModel(GV,
5211 getTargetMachine().getRelocationModel());
5214 case TLSModel::GeneralDynamic:
5215 case TLSModel::LocalDynamic: // not implemented
5216 if (Subtarget->is64Bit())
5217 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5218 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5220 case TLSModel::InitialExec:
5221 case TLSModel::LocalExec:
5222 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5223 Subtarget->is64Bit());
5226 llvm_unreachable("Unreachable");
5231 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5232 /// take a 2 x i32 value to shift plus a shift amount.
5233 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5234 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5235 EVT VT = Op.getValueType();
5236 unsigned VTBits = VT.getSizeInBits();
5237 DebugLoc dl = Op.getDebugLoc();
5238 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5239 SDValue ShOpLo = Op.getOperand(0);
5240 SDValue ShOpHi = Op.getOperand(1);
5241 SDValue ShAmt = Op.getOperand(2);
5242 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5243 DAG.getConstant(VTBits - 1, MVT::i8))
5244 : DAG.getConstant(0, VT);
5247 if (Op.getOpcode() == ISD::SHL_PARTS) {
5248 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5249 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5251 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5252 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5255 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5256 DAG.getConstant(VTBits, MVT::i8));
5257 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
5258 AndNode, DAG.getConstant(0, MVT::i8));
5261 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5262 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5263 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5265 if (Op.getOpcode() == ISD::SHL_PARTS) {
5266 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5267 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5269 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5270 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5273 SDValue Ops[2] = { Lo, Hi };
5274 return DAG.getMergeValues(Ops, 2, dl);
5277 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5278 EVT SrcVT = Op.getOperand(0).getValueType();
5280 if (SrcVT.isVector()) {
5281 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5287 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5288 "Unknown SINT_TO_FP to lower!");
5290 // These are really Legal; return the operand so the caller accepts it as
5292 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5294 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5295 Subtarget->is64Bit()) {
5299 DebugLoc dl = Op.getDebugLoc();
5300 unsigned Size = SrcVT.getSizeInBits()/8;
5301 MachineFunction &MF = DAG.getMachineFunction();
5302 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5303 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5304 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5306 PseudoSourceValue::getFixedStack(SSFI), 0,
5308 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5311 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5313 SelectionDAG &DAG) {
5315 DebugLoc dl = Op.getDebugLoc();
5317 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5319 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5321 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5322 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5323 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5324 Tys, Ops, array_lengthof(Ops));
5327 Chain = Result.getValue(1);
5328 SDValue InFlag = Result.getValue(2);
5330 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5331 // shouldn't be necessary except that RFP cannot be live across
5332 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5333 MachineFunction &MF = DAG.getMachineFunction();
5334 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5335 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5336 Tys = DAG.getVTList(MVT::Other);
5338 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5340 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5341 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5342 PseudoSourceValue::getFixedStack(SSFI), 0,
5349 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5350 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5351 // This algorithm is not obvious. Here it is in C code, more or less:
5353 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5354 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5355 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5357 // Copy ints to xmm registers.
5358 __m128i xh = _mm_cvtsi32_si128( hi );
5359 __m128i xl = _mm_cvtsi32_si128( lo );
5361 // Combine into low half of a single xmm register.
5362 __m128i x = _mm_unpacklo_epi32( xh, xl );
5366 // Merge in appropriate exponents to give the integer bits the right
5368 x = _mm_unpacklo_epi32( x, exp );
5370 // Subtract away the biases to deal with the IEEE-754 double precision
5372 d = _mm_sub_pd( (__m128d) x, bias );
5374 // All conversions up to here are exact. The correctly rounded result is
5375 // calculated using the current rounding mode using the following
5377 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5378 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5379 // store doesn't really need to be here (except
5380 // maybe to zero the other double)
5385 DebugLoc dl = Op.getDebugLoc();
5386 LLVMContext *Context = DAG.getContext();
5388 // Build some magic constants.
5389 std::vector<Constant*> CV0;
5390 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5391 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5392 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5393 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5394 Constant *C0 = ConstantVector::get(CV0);
5395 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5397 std::vector<Constant*> CV1;
5399 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5401 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5402 Constant *C1 = ConstantVector::get(CV1);
5403 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5405 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5406 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5408 DAG.getIntPtrConstant(1)));
5409 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5410 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5412 DAG.getIntPtrConstant(0)));
5413 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5414 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5415 PseudoSourceValue::getConstantPool(), 0,
5417 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5418 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5419 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5420 PseudoSourceValue::getConstantPool(), 0,
5422 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5424 // Add the halves; easiest way is to swap them into another reg first.
5425 int ShufMask[2] = { 1, -1 };
5426 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5427 DAG.getUNDEF(MVT::v2f64), ShufMask);
5428 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5429 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5430 DAG.getIntPtrConstant(0));
5433 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5434 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5435 DebugLoc dl = Op.getDebugLoc();
5436 // FP constant to bias correct the final result.
5437 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5440 // Load the 32-bit value into an XMM register.
5441 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5442 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5444 DAG.getIntPtrConstant(0)));
5446 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5447 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5448 DAG.getIntPtrConstant(0));
5450 // Or the load with the bias.
5451 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5452 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5453 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5455 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5456 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5457 MVT::v2f64, Bias)));
5458 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5459 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5460 DAG.getIntPtrConstant(0));
5462 // Subtract the bias.
5463 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5465 // Handle final rounding.
5466 EVT DestVT = Op.getValueType();
5468 if (DestVT.bitsLT(MVT::f64)) {
5469 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5470 DAG.getIntPtrConstant(0));
5471 } else if (DestVT.bitsGT(MVT::f64)) {
5472 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5475 // Handle final rounding.
5479 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5480 SDValue N0 = Op.getOperand(0);
5481 DebugLoc dl = Op.getDebugLoc();
5483 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5484 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5485 // the optimization here.
5486 if (DAG.SignBitIsZero(N0))
5487 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5489 EVT SrcVT = N0.getValueType();
5490 if (SrcVT == MVT::i64) {
5491 // We only handle SSE2 f64 target here; caller can expand the rest.
5492 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5495 return LowerUINT_TO_FP_i64(Op, DAG);
5496 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5497 return LowerUINT_TO_FP_i32(Op, DAG);
5500 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5502 // Make a 64-bit buffer, and use it to build an FILD.
5503 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5504 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5505 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5506 getPointerTy(), StackSlot, WordOff);
5507 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5508 StackSlot, NULL, 0, false, false, 0);
5509 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5510 OffsetSlot, NULL, 0, false, false, 0);
5511 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5514 std::pair<SDValue,SDValue> X86TargetLowering::
5515 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5516 DebugLoc dl = Op.getDebugLoc();
5518 EVT DstTy = Op.getValueType();
5521 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5525 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5526 DstTy.getSimpleVT() >= MVT::i16 &&
5527 "Unknown FP_TO_SINT to lower!");
5529 // These are really Legal.
5530 if (DstTy == MVT::i32 &&
5531 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5532 return std::make_pair(SDValue(), SDValue());
5533 if (Subtarget->is64Bit() &&
5534 DstTy == MVT::i64 &&
5535 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5536 return std::make_pair(SDValue(), SDValue());
5538 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5540 MachineFunction &MF = DAG.getMachineFunction();
5541 unsigned MemSize = DstTy.getSizeInBits()/8;
5542 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5543 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5546 switch (DstTy.getSimpleVT().SimpleTy) {
5547 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5548 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5549 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5550 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5553 SDValue Chain = DAG.getEntryNode();
5554 SDValue Value = Op.getOperand(0);
5555 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5556 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5557 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5558 PseudoSourceValue::getFixedStack(SSFI), 0,
5560 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5562 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5564 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5565 Chain = Value.getValue(1);
5566 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5567 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5570 // Build the FP_TO_INT*_IN_MEM
5571 SDValue Ops[] = { Chain, Value, StackSlot };
5572 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5574 return std::make_pair(FIST, StackSlot);
5577 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5578 if (Op.getValueType().isVector()) {
5579 if (Op.getValueType() == MVT::v2i32 &&
5580 Op.getOperand(0).getValueType() == MVT::v2f64) {
5586 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5587 SDValue FIST = Vals.first, StackSlot = Vals.second;
5588 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5589 if (FIST.getNode() == 0) return Op;
5592 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5593 FIST, StackSlot, NULL, 0, false, false, 0);
5596 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5597 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5598 SDValue FIST = Vals.first, StackSlot = Vals.second;
5599 assert(FIST.getNode() && "Unexpected failure");
5602 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5603 FIST, StackSlot, NULL, 0, false, false, 0);
5606 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5607 LLVMContext *Context = DAG.getContext();
5608 DebugLoc dl = Op.getDebugLoc();
5609 EVT VT = Op.getValueType();
5612 EltVT = VT.getVectorElementType();
5613 std::vector<Constant*> CV;
5614 if (EltVT == MVT::f64) {
5615 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5619 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5625 Constant *C = ConstantVector::get(CV);
5626 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5627 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5628 PseudoSourceValue::getConstantPool(), 0,
5630 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5633 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5634 LLVMContext *Context = DAG.getContext();
5635 DebugLoc dl = Op.getDebugLoc();
5636 EVT VT = Op.getValueType();
5639 EltVT = VT.getVectorElementType();
5640 std::vector<Constant*> CV;
5641 if (EltVT == MVT::f64) {
5642 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5646 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5652 Constant *C = ConstantVector::get(CV);
5653 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5654 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5655 PseudoSourceValue::getConstantPool(), 0,
5657 if (VT.isVector()) {
5658 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5659 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5660 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5662 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5664 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5668 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5669 LLVMContext *Context = DAG.getContext();
5670 SDValue Op0 = Op.getOperand(0);
5671 SDValue Op1 = Op.getOperand(1);
5672 DebugLoc dl = Op.getDebugLoc();
5673 EVT VT = Op.getValueType();
5674 EVT SrcVT = Op1.getValueType();
5676 // If second operand is smaller, extend it first.
5677 if (SrcVT.bitsLT(VT)) {
5678 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5681 // And if it is bigger, shrink it first.
5682 if (SrcVT.bitsGT(VT)) {
5683 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5687 // At this point the operands and the result should have the same
5688 // type, and that won't be f80 since that is not custom lowered.
5690 // First get the sign bit of second operand.
5691 std::vector<Constant*> CV;
5692 if (SrcVT == MVT::f64) {
5693 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5694 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5696 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5697 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5698 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5699 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5701 Constant *C = ConstantVector::get(CV);
5702 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5703 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5704 PseudoSourceValue::getConstantPool(), 0,
5706 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5708 // Shift sign bit right or left if the two operands have different types.
5709 if (SrcVT.bitsGT(VT)) {
5710 // Op0 is MVT::f32, Op1 is MVT::f64.
5711 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5712 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5713 DAG.getConstant(32, MVT::i32));
5714 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5715 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5716 DAG.getIntPtrConstant(0));
5719 // Clear first operand sign bit.
5721 if (VT == MVT::f64) {
5722 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5723 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5725 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5726 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5727 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5728 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5730 C = ConstantVector::get(CV);
5731 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5732 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5733 PseudoSourceValue::getConstantPool(), 0,
5735 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5737 // Or the value with the sign bit.
5738 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5741 /// Emit nodes that will be selected as "test Op0,Op0", or something
5743 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5744 SelectionDAG &DAG) {
5745 DebugLoc dl = Op.getDebugLoc();
5747 // CF and OF aren't always set the way we want. Determine which
5748 // of these we need.
5749 bool NeedCF = false;
5750 bool NeedOF = false;
5752 case X86::COND_A: case X86::COND_AE:
5753 case X86::COND_B: case X86::COND_BE:
5756 case X86::COND_G: case X86::COND_GE:
5757 case X86::COND_L: case X86::COND_LE:
5758 case X86::COND_O: case X86::COND_NO:
5764 // See if we can use the EFLAGS value from the operand instead of
5765 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5766 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5767 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5768 unsigned Opcode = 0;
5769 unsigned NumOperands = 0;
5770 switch (Op.getNode()->getOpcode()) {
5772 // Due to an isel shortcoming, be conservative if this add is likely to
5773 // be selected as part of a load-modify-store instruction. When the root
5774 // node in a match is a store, isel doesn't know how to remap non-chain
5775 // non-flag uses of other nodes in the match, such as the ADD in this
5776 // case. This leads to the ADD being left around and reselected, with
5777 // the result being two adds in the output.
5778 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5779 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5780 if (UI->getOpcode() == ISD::STORE)
5782 if (ConstantSDNode *C =
5783 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5784 // An add of one will be selected as an INC.
5785 if (C->getAPIntValue() == 1) {
5786 Opcode = X86ISD::INC;
5790 // An add of negative one (subtract of one) will be selected as a DEC.
5791 if (C->getAPIntValue().isAllOnesValue()) {
5792 Opcode = X86ISD::DEC;
5797 // Otherwise use a regular EFLAGS-setting add.
5798 Opcode = X86ISD::ADD;
5802 // If the primary and result isn't used, don't bother using X86ISD::AND,
5803 // because a TEST instruction will be better.
5804 bool NonFlagUse = false;
5805 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5806 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5808 unsigned UOpNo = UI.getOperandNo();
5809 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5810 // Look pass truncate.
5811 UOpNo = User->use_begin().getOperandNo();
5812 User = *User->use_begin();
5814 if (User->getOpcode() != ISD::BRCOND &&
5815 User->getOpcode() != ISD::SETCC &&
5816 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5828 // Due to the ISEL shortcoming noted above, be conservative if this op is
5829 // likely to be selected as part of a load-modify-store instruction.
5830 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5831 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5832 if (UI->getOpcode() == ISD::STORE)
5834 // Otherwise use a regular EFLAGS-setting instruction.
5835 switch (Op.getNode()->getOpcode()) {
5836 case ISD::SUB: Opcode = X86ISD::SUB; break;
5837 case ISD::OR: Opcode = X86ISD::OR; break;
5838 case ISD::XOR: Opcode = X86ISD::XOR; break;
5839 case ISD::AND: Opcode = X86ISD::AND; break;
5840 default: llvm_unreachable("unexpected operator!");
5851 return SDValue(Op.getNode(), 1);
5857 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5858 SmallVector<SDValue, 4> Ops;
5859 for (unsigned i = 0; i != NumOperands; ++i)
5860 Ops.push_back(Op.getOperand(i));
5861 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5862 DAG.ReplaceAllUsesWith(Op, New);
5863 return SDValue(New.getNode(), 1);
5867 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5868 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5869 DAG.getConstant(0, Op.getValueType()));
5872 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5874 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5875 SelectionDAG &DAG) {
5876 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5877 if (C->getAPIntValue() == 0)
5878 return EmitTest(Op0, X86CC, DAG);
5880 DebugLoc dl = Op0.getDebugLoc();
5881 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5884 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5885 /// if it's possible.
5886 static SDValue LowerToBT(SDValue And, ISD::CondCode CC,
5887 DebugLoc dl, SelectionDAG &DAG) {
5888 SDValue Op0 = And.getOperand(0);
5889 SDValue Op1 = And.getOperand(1);
5890 if (Op0.getOpcode() == ISD::TRUNCATE)
5891 Op0 = Op0.getOperand(0);
5892 if (Op1.getOpcode() == ISD::TRUNCATE)
5893 Op1 = Op1.getOperand(0);
5896 if (Op1.getOpcode() == ISD::SHL) {
5897 if (ConstantSDNode *And10C = dyn_cast<ConstantSDNode>(Op1.getOperand(0)))
5898 if (And10C->getZExtValue() == 1) {
5900 RHS = Op1.getOperand(1);
5902 } else if (Op0.getOpcode() == ISD::SHL) {
5903 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
5904 if (And00C->getZExtValue() == 1) {
5906 RHS = Op0.getOperand(1);
5908 } else if (Op1.getOpcode() == ISD::Constant) {
5909 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
5910 SDValue AndLHS = Op0;
5911 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5912 LHS = AndLHS.getOperand(0);
5913 RHS = AndLHS.getOperand(1);
5917 if (LHS.getNode()) {
5918 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5919 // instruction. Since the shift amount is in-range-or-undefined, we know
5920 // that doing a bittest on the i16 value is ok. We extend to i32 because
5921 // the encoding for the i16 version is larger than the i32 version.
5922 if (LHS.getValueType() == MVT::i8)
5923 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5925 // If the operand types disagree, extend the shift amount to match. Since
5926 // BT ignores high bits (like shifts) we can use anyextend.
5927 if (LHS.getValueType() != RHS.getValueType())
5928 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5930 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5931 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5932 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5933 DAG.getConstant(Cond, MVT::i8), BT);
5939 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5940 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5941 SDValue Op0 = Op.getOperand(0);
5942 SDValue Op1 = Op.getOperand(1);
5943 DebugLoc dl = Op.getDebugLoc();
5944 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5946 // Optimize to BT if possible.
5947 // Lower (X & (1 << N)) == 0 to BT(X, N).
5948 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5949 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5950 if (Op0.getOpcode() == ISD::AND &&
5952 Op1.getOpcode() == ISD::Constant &&
5953 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5954 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5955 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5956 if (NewSetCC.getNode())
5960 // Look for "(setcc) == / != 1" to avoid unncessary setcc.
5961 if (Op0.getOpcode() == X86ISD::SETCC &&
5962 Op1.getOpcode() == ISD::Constant &&
5963 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
5964 cast<ConstantSDNode>(Op1)->isNullValue()) &&
5965 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5966 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
5967 bool Invert = (CC == ISD::SETNE) ^
5968 cast<ConstantSDNode>(Op1)->isNullValue();
5970 CCode = X86::GetOppositeBranchCondition(CCode);
5971 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5972 DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
5975 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5976 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5977 if (X86CC == X86::COND_INVALID)
5980 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5982 // Use sbb x, x to materialize carry bit into a GPR.
5983 if (X86CC == X86::COND_B)
5984 return DAG.getNode(ISD::AND, dl, MVT::i8,
5985 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5986 DAG.getConstant(X86CC, MVT::i8), Cond),
5987 DAG.getConstant(1, MVT::i8));
5989 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5990 DAG.getConstant(X86CC, MVT::i8), Cond);
5993 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5995 SDValue Op0 = Op.getOperand(0);
5996 SDValue Op1 = Op.getOperand(1);
5997 SDValue CC = Op.getOperand(2);
5998 EVT VT = Op.getValueType();
5999 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
6000 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
6001 DebugLoc dl = Op.getDebugLoc();
6005 EVT VT0 = Op0.getValueType();
6006 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
6007 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
6010 switch (SetCCOpcode) {
6013 case ISD::SETEQ: SSECC = 0; break;
6015 case ISD::SETGT: Swap = true; // Fallthrough
6017 case ISD::SETOLT: SSECC = 1; break;
6019 case ISD::SETGE: Swap = true; // Fallthrough
6021 case ISD::SETOLE: SSECC = 2; break;
6022 case ISD::SETUO: SSECC = 3; break;
6024 case ISD::SETNE: SSECC = 4; break;
6025 case ISD::SETULE: Swap = true;
6026 case ISD::SETUGE: SSECC = 5; break;
6027 case ISD::SETULT: Swap = true;
6028 case ISD::SETUGT: SSECC = 6; break;
6029 case ISD::SETO: SSECC = 7; break;
6032 std::swap(Op0, Op1);
6034 // In the two special cases we can't handle, emit two comparisons.
6036 if (SetCCOpcode == ISD::SETUEQ) {
6038 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6039 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6040 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6042 else if (SetCCOpcode == ISD::SETONE) {
6044 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6045 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6046 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6048 llvm_unreachable("Illegal FP comparison");
6050 // Handle all other FP comparisons here.
6051 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6054 // We are handling one of the integer comparisons here. Since SSE only has
6055 // GT and EQ comparisons for integer, swapping operands and multiple
6056 // operations may be required for some comparisons.
6057 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6058 bool Swap = false, Invert = false, FlipSigns = false;
6060 switch (VT.getSimpleVT().SimpleTy) {
6063 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6065 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6067 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6068 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6071 switch (SetCCOpcode) {
6073 case ISD::SETNE: Invert = true;
6074 case ISD::SETEQ: Opc = EQOpc; break;
6075 case ISD::SETLT: Swap = true;
6076 case ISD::SETGT: Opc = GTOpc; break;
6077 case ISD::SETGE: Swap = true;
6078 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6079 case ISD::SETULT: Swap = true;
6080 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6081 case ISD::SETUGE: Swap = true;
6082 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6085 std::swap(Op0, Op1);
6087 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6088 // bits of the inputs before performing those operations.
6090 EVT EltVT = VT.getVectorElementType();
6091 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6093 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6094 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6096 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6097 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6100 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6102 // If the logical-not of the result is required, perform that now.
6104 Result = DAG.getNOT(dl, Result, VT);
6109 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6110 static bool isX86LogicalCmp(SDValue Op) {
6111 unsigned Opc = Op.getNode()->getOpcode();
6112 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6114 if (Op.getResNo() == 1 &&
6115 (Opc == X86ISD::ADD ||
6116 Opc == X86ISD::SUB ||
6117 Opc == X86ISD::SMUL ||
6118 Opc == X86ISD::UMUL ||
6119 Opc == X86ISD::INC ||
6120 Opc == X86ISD::DEC ||
6121 Opc == X86ISD::OR ||
6122 Opc == X86ISD::XOR ||
6123 Opc == X86ISD::AND))
6129 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6130 bool addTest = true;
6131 SDValue Cond = Op.getOperand(0);
6132 DebugLoc dl = Op.getDebugLoc();
6135 if (Cond.getOpcode() == ISD::SETCC) {
6136 SDValue NewCond = LowerSETCC(Cond, DAG);
6137 if (NewCond.getNode())
6141 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6142 SDValue Op1 = Op.getOperand(1);
6143 SDValue Op2 = Op.getOperand(2);
6144 if (Cond.getOpcode() == X86ISD::SETCC &&
6145 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6146 SDValue Cmp = Cond.getOperand(1);
6147 if (Cmp.getOpcode() == X86ISD::CMP) {
6148 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6149 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6150 ConstantSDNode *RHSC =
6151 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6152 if (N1C && N1C->isAllOnesValue() &&
6153 N2C && N2C->isNullValue() &&
6154 RHSC && RHSC->isNullValue()) {
6155 SDValue CmpOp0 = Cmp.getOperand(0);
6156 Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
6157 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6158 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6159 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6164 // Look pass (and (setcc_carry (cmp ...)), 1).
6165 if (Cond.getOpcode() == ISD::AND &&
6166 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6167 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6168 if (C && C->getAPIntValue() == 1)
6169 Cond = Cond.getOperand(0);
6172 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6173 // setting operand in place of the X86ISD::SETCC.
6174 if (Cond.getOpcode() == X86ISD::SETCC ||
6175 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6176 CC = Cond.getOperand(0);
6178 SDValue Cmp = Cond.getOperand(1);
6179 unsigned Opc = Cmp.getOpcode();
6180 EVT VT = Op.getValueType();
6182 bool IllegalFPCMov = false;
6183 if (VT.isFloatingPoint() && !VT.isVector() &&
6184 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6185 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6187 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6188 Opc == X86ISD::BT) { // FIXME
6195 // Look pass the truncate.
6196 if (Cond.getOpcode() == ISD::TRUNCATE)
6197 Cond = Cond.getOperand(0);
6199 // We know the result of AND is compared against zero. Try to match
6201 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6202 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6203 if (NewSetCC.getNode()) {
6204 CC = NewSetCC.getOperand(0);
6205 Cond = NewSetCC.getOperand(1);
6212 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6213 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6216 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6217 // condition is true.
6218 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6219 SDValue Ops[] = { Op2, Op1, CC, Cond };
6220 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6223 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6224 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6225 // from the AND / OR.
6226 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6227 Opc = Op.getOpcode();
6228 if (Opc != ISD::OR && Opc != ISD::AND)
6230 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6231 Op.getOperand(0).hasOneUse() &&
6232 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6233 Op.getOperand(1).hasOneUse());
6236 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6237 // 1 and that the SETCC node has a single use.
6238 static bool isXor1OfSetCC(SDValue Op) {
6239 if (Op.getOpcode() != ISD::XOR)
6241 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6242 if (N1C && N1C->getAPIntValue() == 1) {
6243 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6244 Op.getOperand(0).hasOneUse();
6249 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6250 bool addTest = true;
6251 SDValue Chain = Op.getOperand(0);
6252 SDValue Cond = Op.getOperand(1);
6253 SDValue Dest = Op.getOperand(2);
6254 DebugLoc dl = Op.getDebugLoc();
6257 if (Cond.getOpcode() == ISD::SETCC) {
6258 SDValue NewCond = LowerSETCC(Cond, DAG);
6259 if (NewCond.getNode())
6263 // FIXME: LowerXALUO doesn't handle these!!
6264 else if (Cond.getOpcode() == X86ISD::ADD ||
6265 Cond.getOpcode() == X86ISD::SUB ||
6266 Cond.getOpcode() == X86ISD::SMUL ||
6267 Cond.getOpcode() == X86ISD::UMUL)
6268 Cond = LowerXALUO(Cond, DAG);
6271 // Look pass (and (setcc_carry (cmp ...)), 1).
6272 if (Cond.getOpcode() == ISD::AND &&
6273 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6274 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6275 if (C && C->getAPIntValue() == 1)
6276 Cond = Cond.getOperand(0);
6279 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6280 // setting operand in place of the X86ISD::SETCC.
6281 if (Cond.getOpcode() == X86ISD::SETCC ||
6282 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6283 CC = Cond.getOperand(0);
6285 SDValue Cmp = Cond.getOperand(1);
6286 unsigned Opc = Cmp.getOpcode();
6287 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6288 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6292 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6296 // These can only come from an arithmetic instruction with overflow,
6297 // e.g. SADDO, UADDO.
6298 Cond = Cond.getNode()->getOperand(1);
6305 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6306 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6307 if (CondOpc == ISD::OR) {
6308 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6309 // two branches instead of an explicit OR instruction with a
6311 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6312 isX86LogicalCmp(Cmp)) {
6313 CC = Cond.getOperand(0).getOperand(0);
6314 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6315 Chain, Dest, CC, Cmp);
6316 CC = Cond.getOperand(1).getOperand(0);
6320 } else { // ISD::AND
6321 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6322 // two branches instead of an explicit AND instruction with a
6323 // separate test. However, we only do this if this block doesn't
6324 // have a fall-through edge, because this requires an explicit
6325 // jmp when the condition is false.
6326 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6327 isX86LogicalCmp(Cmp) &&
6328 Op.getNode()->hasOneUse()) {
6329 X86::CondCode CCode =
6330 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6331 CCode = X86::GetOppositeBranchCondition(CCode);
6332 CC = DAG.getConstant(CCode, MVT::i8);
6333 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6334 // Look for an unconditional branch following this conditional branch.
6335 // We need this because we need to reverse the successors in order
6336 // to implement FCMP_OEQ.
6337 if (User.getOpcode() == ISD::BR) {
6338 SDValue FalseBB = User.getOperand(1);
6340 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6341 assert(NewBR == User);
6344 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6345 Chain, Dest, CC, Cmp);
6346 X86::CondCode CCode =
6347 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6348 CCode = X86::GetOppositeBranchCondition(CCode);
6349 CC = DAG.getConstant(CCode, MVT::i8);
6355 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6356 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6357 // It should be transformed during dag combiner except when the condition
6358 // is set by a arithmetics with overflow node.
6359 X86::CondCode CCode =
6360 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6361 CCode = X86::GetOppositeBranchCondition(CCode);
6362 CC = DAG.getConstant(CCode, MVT::i8);
6363 Cond = Cond.getOperand(0).getOperand(1);
6369 // Look pass the truncate.
6370 if (Cond.getOpcode() == ISD::TRUNCATE)
6371 Cond = Cond.getOperand(0);
6373 // We know the result of AND is compared against zero. Try to match
6375 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6376 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6377 if (NewSetCC.getNode()) {
6378 CC = NewSetCC.getOperand(0);
6379 Cond = NewSetCC.getOperand(1);
6386 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6387 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6389 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6390 Chain, Dest, CC, Cond);
6394 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6395 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6396 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6397 // that the guard pages used by the OS virtual memory manager are allocated in
6398 // correct sequence.
6400 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6401 SelectionDAG &DAG) {
6402 assert(Subtarget->isTargetCygMing() &&
6403 "This should be used only on Cygwin/Mingw targets");
6404 DebugLoc dl = Op.getDebugLoc();
6407 SDValue Chain = Op.getOperand(0);
6408 SDValue Size = Op.getOperand(1);
6409 // FIXME: Ensure alignment here
6413 EVT IntPtr = getPointerTy();
6414 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6416 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6418 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6419 Flag = Chain.getValue(1);
6421 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6422 SDValue Ops[] = { Chain,
6423 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6424 DAG.getRegister(X86::EAX, IntPtr),
6425 DAG.getRegister(X86StackPtr, SPTy),
6427 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6428 Flag = Chain.getValue(1);
6430 Chain = DAG.getCALLSEQ_END(Chain,
6431 DAG.getIntPtrConstant(0, true),
6432 DAG.getIntPtrConstant(0, true),
6435 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6437 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6438 return DAG.getMergeValues(Ops1, 2, dl);
6442 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6444 SDValue Dst, SDValue Src,
6445 SDValue Size, unsigned Align,
6447 uint64_t DstSVOff) {
6448 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6450 // If not DWORD aligned or size is more than the threshold, call the library.
6451 // The libc version is likely to be faster for these cases. It can use the
6452 // address value and run time information about the CPU.
6453 if ((Align & 3) != 0 ||
6455 ConstantSize->getZExtValue() >
6456 getSubtarget()->getMaxInlineSizeThreshold()) {
6457 SDValue InFlag(0, 0);
6459 // Check to see if there is a specialized entry-point for memory zeroing.
6460 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6462 if (const char *bzeroEntry = V &&
6463 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6464 EVT IntPtr = getPointerTy();
6465 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6466 TargetLowering::ArgListTy Args;
6467 TargetLowering::ArgListEntry Entry;
6469 Entry.Ty = IntPtrTy;
6470 Args.push_back(Entry);
6472 Args.push_back(Entry);
6473 std::pair<SDValue,SDValue> CallResult =
6474 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6475 false, false, false, false,
6476 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6477 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6478 DAG.GetOrdering(Chain.getNode()));
6479 return CallResult.second;
6482 // Otherwise have the target-independent code call memset.
6486 uint64_t SizeVal = ConstantSize->getZExtValue();
6487 SDValue InFlag(0, 0);
6490 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6491 unsigned BytesLeft = 0;
6492 bool TwoRepStos = false;
6495 uint64_t Val = ValC->getZExtValue() & 255;
6497 // If the value is a constant, then we can potentially use larger sets.
6498 switch (Align & 3) {
6499 case 2: // WORD aligned
6502 Val = (Val << 8) | Val;
6504 case 0: // DWORD aligned
6507 Val = (Val << 8) | Val;
6508 Val = (Val << 16) | Val;
6509 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6512 Val = (Val << 32) | Val;
6515 default: // Byte aligned
6518 Count = DAG.getIntPtrConstant(SizeVal);
6522 if (AVT.bitsGT(MVT::i8)) {
6523 unsigned UBytes = AVT.getSizeInBits() / 8;
6524 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6525 BytesLeft = SizeVal % UBytes;
6528 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6530 InFlag = Chain.getValue(1);
6533 Count = DAG.getIntPtrConstant(SizeVal);
6534 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6535 InFlag = Chain.getValue(1);
6538 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6541 InFlag = Chain.getValue(1);
6542 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6545 InFlag = Chain.getValue(1);
6547 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6548 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6549 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6552 InFlag = Chain.getValue(1);
6554 EVT CVT = Count.getValueType();
6555 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6556 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6557 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6560 InFlag = Chain.getValue(1);
6561 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6562 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6563 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6564 } else if (BytesLeft) {
6565 // Handle the last 1 - 7 bytes.
6566 unsigned Offset = SizeVal - BytesLeft;
6567 EVT AddrVT = Dst.getValueType();
6568 EVT SizeVT = Size.getValueType();
6570 Chain = DAG.getMemset(Chain, dl,
6571 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6572 DAG.getConstant(Offset, AddrVT)),
6574 DAG.getConstant(BytesLeft, SizeVT),
6575 Align, DstSV, DstSVOff + Offset);
6578 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6583 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6584 SDValue Chain, SDValue Dst, SDValue Src,
6585 SDValue Size, unsigned Align,
6587 const Value *DstSV, uint64_t DstSVOff,
6588 const Value *SrcSV, uint64_t SrcSVOff) {
6589 // This requires the copy size to be a constant, preferrably
6590 // within a subtarget-specific limit.
6591 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6594 uint64_t SizeVal = ConstantSize->getZExtValue();
6595 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6598 /// If not DWORD aligned, call the library.
6599 if ((Align & 3) != 0)
6604 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6607 unsigned UBytes = AVT.getSizeInBits() / 8;
6608 unsigned CountVal = SizeVal / UBytes;
6609 SDValue Count = DAG.getIntPtrConstant(CountVal);
6610 unsigned BytesLeft = SizeVal % UBytes;
6612 SDValue InFlag(0, 0);
6613 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6616 InFlag = Chain.getValue(1);
6617 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6620 InFlag = Chain.getValue(1);
6621 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6624 InFlag = Chain.getValue(1);
6626 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6627 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6628 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6629 array_lengthof(Ops));
6631 SmallVector<SDValue, 4> Results;
6632 Results.push_back(RepMovs);
6634 // Handle the last 1 - 7 bytes.
6635 unsigned Offset = SizeVal - BytesLeft;
6636 EVT DstVT = Dst.getValueType();
6637 EVT SrcVT = Src.getValueType();
6638 EVT SizeVT = Size.getValueType();
6639 Results.push_back(DAG.getMemcpy(Chain, dl,
6640 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6641 DAG.getConstant(Offset, DstVT)),
6642 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6643 DAG.getConstant(Offset, SrcVT)),
6644 DAG.getConstant(BytesLeft, SizeVT),
6645 Align, AlwaysInline,
6646 DstSV, DstSVOff + Offset,
6647 SrcSV, SrcSVOff + Offset));
6650 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6651 &Results[0], Results.size());
6654 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6655 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6656 DebugLoc dl = Op.getDebugLoc();
6658 if (!Subtarget->is64Bit()) {
6659 // vastart just stores the address of the VarArgsFrameIndex slot into the
6660 // memory location argument.
6661 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6662 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6667 // gp_offset (0 - 6 * 8)
6668 // fp_offset (48 - 48 + 8 * 16)
6669 // overflow_arg_area (point to parameters coming in memory).
6671 SmallVector<SDValue, 8> MemOps;
6672 SDValue FIN = Op.getOperand(1);
6674 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6675 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6676 FIN, SV, 0, false, false, 0);
6677 MemOps.push_back(Store);
6680 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6681 FIN, DAG.getIntPtrConstant(4));
6682 Store = DAG.getStore(Op.getOperand(0), dl,
6683 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6684 FIN, SV, 0, false, false, 0);
6685 MemOps.push_back(Store);
6687 // Store ptr to overflow_arg_area
6688 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6689 FIN, DAG.getIntPtrConstant(4));
6690 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6691 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6693 MemOps.push_back(Store);
6695 // Store ptr to reg_save_area.
6696 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6697 FIN, DAG.getIntPtrConstant(8));
6698 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6699 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6701 MemOps.push_back(Store);
6702 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6703 &MemOps[0], MemOps.size());
6706 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6707 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6708 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6709 SDValue Chain = Op.getOperand(0);
6710 SDValue SrcPtr = Op.getOperand(1);
6711 SDValue SrcSV = Op.getOperand(2);
6713 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6717 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6718 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6719 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6720 SDValue Chain = Op.getOperand(0);
6721 SDValue DstPtr = Op.getOperand(1);
6722 SDValue SrcPtr = Op.getOperand(2);
6723 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6724 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6725 DebugLoc dl = Op.getDebugLoc();
6727 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6728 DAG.getIntPtrConstant(24), 8, false,
6729 DstSV, 0, SrcSV, 0);
6733 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6734 DebugLoc dl = Op.getDebugLoc();
6735 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6737 default: return SDValue(); // Don't custom lower most intrinsics.
6738 // Comparison intrinsics.
6739 case Intrinsic::x86_sse_comieq_ss:
6740 case Intrinsic::x86_sse_comilt_ss:
6741 case Intrinsic::x86_sse_comile_ss:
6742 case Intrinsic::x86_sse_comigt_ss:
6743 case Intrinsic::x86_sse_comige_ss:
6744 case Intrinsic::x86_sse_comineq_ss:
6745 case Intrinsic::x86_sse_ucomieq_ss:
6746 case Intrinsic::x86_sse_ucomilt_ss:
6747 case Intrinsic::x86_sse_ucomile_ss:
6748 case Intrinsic::x86_sse_ucomigt_ss:
6749 case Intrinsic::x86_sse_ucomige_ss:
6750 case Intrinsic::x86_sse_ucomineq_ss:
6751 case Intrinsic::x86_sse2_comieq_sd:
6752 case Intrinsic::x86_sse2_comilt_sd:
6753 case Intrinsic::x86_sse2_comile_sd:
6754 case Intrinsic::x86_sse2_comigt_sd:
6755 case Intrinsic::x86_sse2_comige_sd:
6756 case Intrinsic::x86_sse2_comineq_sd:
6757 case Intrinsic::x86_sse2_ucomieq_sd:
6758 case Intrinsic::x86_sse2_ucomilt_sd:
6759 case Intrinsic::x86_sse2_ucomile_sd:
6760 case Intrinsic::x86_sse2_ucomigt_sd:
6761 case Intrinsic::x86_sse2_ucomige_sd:
6762 case Intrinsic::x86_sse2_ucomineq_sd: {
6764 ISD::CondCode CC = ISD::SETCC_INVALID;
6767 case Intrinsic::x86_sse_comieq_ss:
6768 case Intrinsic::x86_sse2_comieq_sd:
6772 case Intrinsic::x86_sse_comilt_ss:
6773 case Intrinsic::x86_sse2_comilt_sd:
6777 case Intrinsic::x86_sse_comile_ss:
6778 case Intrinsic::x86_sse2_comile_sd:
6782 case Intrinsic::x86_sse_comigt_ss:
6783 case Intrinsic::x86_sse2_comigt_sd:
6787 case Intrinsic::x86_sse_comige_ss:
6788 case Intrinsic::x86_sse2_comige_sd:
6792 case Intrinsic::x86_sse_comineq_ss:
6793 case Intrinsic::x86_sse2_comineq_sd:
6797 case Intrinsic::x86_sse_ucomieq_ss:
6798 case Intrinsic::x86_sse2_ucomieq_sd:
6799 Opc = X86ISD::UCOMI;
6802 case Intrinsic::x86_sse_ucomilt_ss:
6803 case Intrinsic::x86_sse2_ucomilt_sd:
6804 Opc = X86ISD::UCOMI;
6807 case Intrinsic::x86_sse_ucomile_ss:
6808 case Intrinsic::x86_sse2_ucomile_sd:
6809 Opc = X86ISD::UCOMI;
6812 case Intrinsic::x86_sse_ucomigt_ss:
6813 case Intrinsic::x86_sse2_ucomigt_sd:
6814 Opc = X86ISD::UCOMI;
6817 case Intrinsic::x86_sse_ucomige_ss:
6818 case Intrinsic::x86_sse2_ucomige_sd:
6819 Opc = X86ISD::UCOMI;
6822 case Intrinsic::x86_sse_ucomineq_ss:
6823 case Intrinsic::x86_sse2_ucomineq_sd:
6824 Opc = X86ISD::UCOMI;
6829 SDValue LHS = Op.getOperand(1);
6830 SDValue RHS = Op.getOperand(2);
6831 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6832 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6833 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6834 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6835 DAG.getConstant(X86CC, MVT::i8), Cond);
6836 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6838 // ptest intrinsics. The intrinsic these come from are designed to return
6839 // an integer value, not just an instruction so lower it to the ptest
6840 // pattern and a setcc for the result.
6841 case Intrinsic::x86_sse41_ptestz:
6842 case Intrinsic::x86_sse41_ptestc:
6843 case Intrinsic::x86_sse41_ptestnzc:{
6846 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6847 case Intrinsic::x86_sse41_ptestz:
6849 X86CC = X86::COND_E;
6851 case Intrinsic::x86_sse41_ptestc:
6853 X86CC = X86::COND_B;
6855 case Intrinsic::x86_sse41_ptestnzc:
6857 X86CC = X86::COND_A;
6861 SDValue LHS = Op.getOperand(1);
6862 SDValue RHS = Op.getOperand(2);
6863 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6864 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6865 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6866 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6869 // Fix vector shift instructions where the last operand is a non-immediate
6871 case Intrinsic::x86_sse2_pslli_w:
6872 case Intrinsic::x86_sse2_pslli_d:
6873 case Intrinsic::x86_sse2_pslli_q:
6874 case Intrinsic::x86_sse2_psrli_w:
6875 case Intrinsic::x86_sse2_psrli_d:
6876 case Intrinsic::x86_sse2_psrli_q:
6877 case Intrinsic::x86_sse2_psrai_w:
6878 case Intrinsic::x86_sse2_psrai_d:
6879 case Intrinsic::x86_mmx_pslli_w:
6880 case Intrinsic::x86_mmx_pslli_d:
6881 case Intrinsic::x86_mmx_pslli_q:
6882 case Intrinsic::x86_mmx_psrli_w:
6883 case Intrinsic::x86_mmx_psrli_d:
6884 case Intrinsic::x86_mmx_psrli_q:
6885 case Intrinsic::x86_mmx_psrai_w:
6886 case Intrinsic::x86_mmx_psrai_d: {
6887 SDValue ShAmt = Op.getOperand(2);
6888 if (isa<ConstantSDNode>(ShAmt))
6891 unsigned NewIntNo = 0;
6892 EVT ShAmtVT = MVT::v4i32;
6894 case Intrinsic::x86_sse2_pslli_w:
6895 NewIntNo = Intrinsic::x86_sse2_psll_w;
6897 case Intrinsic::x86_sse2_pslli_d:
6898 NewIntNo = Intrinsic::x86_sse2_psll_d;
6900 case Intrinsic::x86_sse2_pslli_q:
6901 NewIntNo = Intrinsic::x86_sse2_psll_q;
6903 case Intrinsic::x86_sse2_psrli_w:
6904 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6906 case Intrinsic::x86_sse2_psrli_d:
6907 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6909 case Intrinsic::x86_sse2_psrli_q:
6910 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6912 case Intrinsic::x86_sse2_psrai_w:
6913 NewIntNo = Intrinsic::x86_sse2_psra_w;
6915 case Intrinsic::x86_sse2_psrai_d:
6916 NewIntNo = Intrinsic::x86_sse2_psra_d;
6919 ShAmtVT = MVT::v2i32;
6921 case Intrinsic::x86_mmx_pslli_w:
6922 NewIntNo = Intrinsic::x86_mmx_psll_w;
6924 case Intrinsic::x86_mmx_pslli_d:
6925 NewIntNo = Intrinsic::x86_mmx_psll_d;
6927 case Intrinsic::x86_mmx_pslli_q:
6928 NewIntNo = Intrinsic::x86_mmx_psll_q;
6930 case Intrinsic::x86_mmx_psrli_w:
6931 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6933 case Intrinsic::x86_mmx_psrli_d:
6934 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6936 case Intrinsic::x86_mmx_psrli_q:
6937 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6939 case Intrinsic::x86_mmx_psrai_w:
6940 NewIntNo = Intrinsic::x86_mmx_psra_w;
6942 case Intrinsic::x86_mmx_psrai_d:
6943 NewIntNo = Intrinsic::x86_mmx_psra_d;
6945 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6951 // The vector shift intrinsics with scalars uses 32b shift amounts but
6952 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6956 ShOps[1] = DAG.getConstant(0, MVT::i32);
6957 if (ShAmtVT == MVT::v4i32) {
6958 ShOps[2] = DAG.getUNDEF(MVT::i32);
6959 ShOps[3] = DAG.getUNDEF(MVT::i32);
6960 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6962 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6965 EVT VT = Op.getValueType();
6966 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6967 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6968 DAG.getConstant(NewIntNo, MVT::i32),
6969 Op.getOperand(1), ShAmt);
6974 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6975 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6976 DebugLoc dl = Op.getDebugLoc();
6979 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6981 DAG.getConstant(TD->getPointerSize(),
6982 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6983 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6984 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6986 NULL, 0, false, false, 0);
6989 // Just load the return address.
6990 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6991 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6992 RetAddrFI, NULL, 0, false, false, 0);
6995 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6996 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6997 MFI->setFrameAddressIsTaken(true);
6998 EVT VT = Op.getValueType();
6999 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
7000 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
7001 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
7002 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
7004 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
7009 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
7010 SelectionDAG &DAG) {
7011 return DAG.getIntPtrConstant(2*TD->getPointerSize());
7014 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
7016 MachineFunction &MF = DAG.getMachineFunction();
7017 SDValue Chain = Op.getOperand(0);
7018 SDValue Offset = Op.getOperand(1);
7019 SDValue Handler = Op.getOperand(2);
7020 DebugLoc dl = Op.getDebugLoc();
7022 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
7024 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
7026 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
7027 DAG.getIntPtrConstant(-TD->getPointerSize()));
7028 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
7029 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7030 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7031 MF.getRegInfo().addLiveOut(StoreAddrReg);
7033 return DAG.getNode(X86ISD::EH_RETURN, dl,
7035 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7038 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7039 SelectionDAG &DAG) {
7040 SDValue Root = Op.getOperand(0);
7041 SDValue Trmp = Op.getOperand(1); // trampoline
7042 SDValue FPtr = Op.getOperand(2); // nested function
7043 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7044 DebugLoc dl = Op.getDebugLoc();
7046 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7048 if (Subtarget->is64Bit()) {
7049 SDValue OutChains[6];
7051 // Large code-model.
7052 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7053 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7055 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7056 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7058 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7060 // Load the pointer to the nested function into R11.
7061 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7062 SDValue Addr = Trmp;
7063 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7064 Addr, TrmpAddr, 0, false, false, 0);
7066 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7067 DAG.getConstant(2, MVT::i64));
7068 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7071 // Load the 'nest' parameter value into R10.
7072 // R10 is specified in X86CallingConv.td
7073 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7074 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7075 DAG.getConstant(10, MVT::i64));
7076 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7077 Addr, TrmpAddr, 10, false, false, 0);
7079 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7080 DAG.getConstant(12, MVT::i64));
7081 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7084 // Jump to the nested function.
7085 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7086 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7087 DAG.getConstant(20, MVT::i64));
7088 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7089 Addr, TrmpAddr, 20, false, false, 0);
7091 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7092 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7093 DAG.getConstant(22, MVT::i64));
7094 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7095 TrmpAddr, 22, false, false, 0);
7098 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7099 return DAG.getMergeValues(Ops, 2, dl);
7101 const Function *Func =
7102 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7103 CallingConv::ID CC = Func->getCallingConv();
7108 llvm_unreachable("Unsupported calling convention");
7109 case CallingConv::C:
7110 case CallingConv::X86_StdCall: {
7111 // Pass 'nest' parameter in ECX.
7112 // Must be kept in sync with X86CallingConv.td
7115 // Check that ECX wasn't needed by an 'inreg' parameter.
7116 const FunctionType *FTy = Func->getFunctionType();
7117 const AttrListPtr &Attrs = Func->getAttributes();
7119 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7120 unsigned InRegCount = 0;
7123 for (FunctionType::param_iterator I = FTy->param_begin(),
7124 E = FTy->param_end(); I != E; ++I, ++Idx)
7125 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7126 // FIXME: should only count parameters that are lowered to integers.
7127 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7129 if (InRegCount > 2) {
7130 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7135 case CallingConv::X86_FastCall:
7136 case CallingConv::Fast:
7137 // Pass 'nest' parameter in EAX.
7138 // Must be kept in sync with X86CallingConv.td
7143 SDValue OutChains[4];
7146 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7147 DAG.getConstant(10, MVT::i32));
7148 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7150 // This is storing the opcode for MOV32ri.
7151 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7152 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7153 OutChains[0] = DAG.getStore(Root, dl,
7154 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7155 Trmp, TrmpAddr, 0, false, false, 0);
7157 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7158 DAG.getConstant(1, MVT::i32));
7159 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7162 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7163 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7164 DAG.getConstant(5, MVT::i32));
7165 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7166 TrmpAddr, 5, false, false, 1);
7168 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7169 DAG.getConstant(6, MVT::i32));
7170 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7174 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7175 return DAG.getMergeValues(Ops, 2, dl);
7179 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7181 The rounding mode is in bits 11:10 of FPSR, and has the following
7188 FLT_ROUNDS, on the other hand, expects the following:
7195 To perform the conversion, we do:
7196 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7199 MachineFunction &MF = DAG.getMachineFunction();
7200 const TargetMachine &TM = MF.getTarget();
7201 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7202 unsigned StackAlignment = TFI.getStackAlignment();
7203 EVT VT = Op.getValueType();
7204 DebugLoc dl = Op.getDebugLoc();
7206 // Save FP Control Word to stack slot
7207 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7208 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7210 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7211 DAG.getEntryNode(), StackSlot);
7213 // Load FP Control Word from stack slot
7214 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7217 // Transform as necessary
7219 DAG.getNode(ISD::SRL, dl, MVT::i16,
7220 DAG.getNode(ISD::AND, dl, MVT::i16,
7221 CWD, DAG.getConstant(0x800, MVT::i16)),
7222 DAG.getConstant(11, MVT::i8));
7224 DAG.getNode(ISD::SRL, dl, MVT::i16,
7225 DAG.getNode(ISD::AND, dl, MVT::i16,
7226 CWD, DAG.getConstant(0x400, MVT::i16)),
7227 DAG.getConstant(9, MVT::i8));
7230 DAG.getNode(ISD::AND, dl, MVT::i16,
7231 DAG.getNode(ISD::ADD, dl, MVT::i16,
7232 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7233 DAG.getConstant(1, MVT::i16)),
7234 DAG.getConstant(3, MVT::i16));
7237 return DAG.getNode((VT.getSizeInBits() < 16 ?
7238 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7241 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7242 EVT VT = Op.getValueType();
7244 unsigned NumBits = VT.getSizeInBits();
7245 DebugLoc dl = Op.getDebugLoc();
7247 Op = Op.getOperand(0);
7248 if (VT == MVT::i8) {
7249 // Zero extend to i32 since there is not an i8 bsr.
7251 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7254 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7255 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7256 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7258 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7261 DAG.getConstant(NumBits+NumBits-1, OpVT),
7262 DAG.getConstant(X86::COND_E, MVT::i8),
7265 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7267 // Finally xor with NumBits-1.
7268 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7271 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7275 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7276 EVT VT = Op.getValueType();
7278 unsigned NumBits = VT.getSizeInBits();
7279 DebugLoc dl = Op.getDebugLoc();
7281 Op = Op.getOperand(0);
7282 if (VT == MVT::i8) {
7284 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7287 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7288 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7289 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7291 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7294 DAG.getConstant(NumBits, OpVT),
7295 DAG.getConstant(X86::COND_E, MVT::i8),
7298 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7301 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7305 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7306 EVT VT = Op.getValueType();
7307 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7308 DebugLoc dl = Op.getDebugLoc();
7310 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7311 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7312 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7313 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7314 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7316 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7317 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7318 // return AloBlo + AloBhi + AhiBlo;
7320 SDValue A = Op.getOperand(0);
7321 SDValue B = Op.getOperand(1);
7323 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7324 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7325 A, DAG.getConstant(32, MVT::i32));
7326 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7327 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7328 B, DAG.getConstant(32, MVT::i32));
7329 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7330 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7332 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7333 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7335 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7336 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7338 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7339 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7340 AloBhi, DAG.getConstant(32, MVT::i32));
7341 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7342 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7343 AhiBlo, DAG.getConstant(32, MVT::i32));
7344 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7345 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7350 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7351 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7352 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7353 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7354 // has only one use.
7355 SDNode *N = Op.getNode();
7356 SDValue LHS = N->getOperand(0);
7357 SDValue RHS = N->getOperand(1);
7358 unsigned BaseOp = 0;
7360 DebugLoc dl = Op.getDebugLoc();
7362 switch (Op.getOpcode()) {
7363 default: llvm_unreachable("Unknown ovf instruction!");
7365 // A subtract of one will be selected as a INC. Note that INC doesn't
7366 // set CF, so we can't do this for UADDO.
7367 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7368 if (C->getAPIntValue() == 1) {
7369 BaseOp = X86ISD::INC;
7373 BaseOp = X86ISD::ADD;
7377 BaseOp = X86ISD::ADD;
7381 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7382 // set CF, so we can't do this for USUBO.
7383 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7384 if (C->getAPIntValue() == 1) {
7385 BaseOp = X86ISD::DEC;
7389 BaseOp = X86ISD::SUB;
7393 BaseOp = X86ISD::SUB;
7397 BaseOp = X86ISD::SMUL;
7401 BaseOp = X86ISD::UMUL;
7406 // Also sets EFLAGS.
7407 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7408 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7411 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7412 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7414 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7418 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7419 EVT T = Op.getValueType();
7420 DebugLoc dl = Op.getDebugLoc();
7423 switch(T.getSimpleVT().SimpleTy) {
7425 assert(false && "Invalid value type!");
7426 case MVT::i8: Reg = X86::AL; size = 1; break;
7427 case MVT::i16: Reg = X86::AX; size = 2; break;
7428 case MVT::i32: Reg = X86::EAX; size = 4; break;
7430 assert(Subtarget->is64Bit() && "Node not type legal!");
7431 Reg = X86::RAX; size = 8;
7434 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7435 Op.getOperand(2), SDValue());
7436 SDValue Ops[] = { cpIn.getValue(0),
7439 DAG.getTargetConstant(size, MVT::i8),
7441 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7442 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7444 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7448 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7449 SelectionDAG &DAG) {
7450 assert(Subtarget->is64Bit() && "Result not type legalized?");
7451 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7452 SDValue TheChain = Op.getOperand(0);
7453 DebugLoc dl = Op.getDebugLoc();
7454 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7455 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7456 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7458 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7459 DAG.getConstant(32, MVT::i8));
7461 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7464 return DAG.getMergeValues(Ops, 2, dl);
7467 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7468 SDNode *Node = Op.getNode();
7469 DebugLoc dl = Node->getDebugLoc();
7470 EVT T = Node->getValueType(0);
7471 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7472 DAG.getConstant(0, T), Node->getOperand(2));
7473 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7474 cast<AtomicSDNode>(Node)->getMemoryVT(),
7475 Node->getOperand(0),
7476 Node->getOperand(1), negOp,
7477 cast<AtomicSDNode>(Node)->getSrcValue(),
7478 cast<AtomicSDNode>(Node)->getAlignment());
7481 /// LowerOperation - Provide custom lowering hooks for some operations.
7483 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7484 switch (Op.getOpcode()) {
7485 default: llvm_unreachable("Should not custom lower this!");
7486 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7487 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7488 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7489 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7490 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7491 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7492 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7493 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7494 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7495 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7496 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7497 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7498 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7499 case ISD::SHL_PARTS:
7500 case ISD::SRA_PARTS:
7501 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7502 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7503 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7504 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7505 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7506 case ISD::FABS: return LowerFABS(Op, DAG);
7507 case ISD::FNEG: return LowerFNEG(Op, DAG);
7508 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7509 case ISD::SETCC: return LowerSETCC(Op, DAG);
7510 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7511 case ISD::SELECT: return LowerSELECT(Op, DAG);
7512 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7513 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7514 case ISD::VASTART: return LowerVASTART(Op, DAG);
7515 case ISD::VAARG: return LowerVAARG(Op, DAG);
7516 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7517 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7518 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7519 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7520 case ISD::FRAME_TO_ARGS_OFFSET:
7521 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7522 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7523 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7524 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7525 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7526 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7527 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7528 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7534 case ISD::UMULO: return LowerXALUO(Op, DAG);
7535 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7539 void X86TargetLowering::
7540 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7541 SelectionDAG &DAG, unsigned NewOp) {
7542 EVT T = Node->getValueType(0);
7543 DebugLoc dl = Node->getDebugLoc();
7544 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7546 SDValue Chain = Node->getOperand(0);
7547 SDValue In1 = Node->getOperand(1);
7548 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7549 Node->getOperand(2), DAG.getIntPtrConstant(0));
7550 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7551 Node->getOperand(2), DAG.getIntPtrConstant(1));
7552 SDValue Ops[] = { Chain, In1, In2L, In2H };
7553 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7555 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7556 cast<MemSDNode>(Node)->getMemOperand());
7557 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7558 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7559 Results.push_back(Result.getValue(2));
7562 /// ReplaceNodeResults - Replace a node with an illegal result type
7563 /// with a new node built out of custom code.
7564 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7565 SmallVectorImpl<SDValue>&Results,
7566 SelectionDAG &DAG) {
7567 DebugLoc dl = N->getDebugLoc();
7568 switch (N->getOpcode()) {
7570 assert(false && "Do not know how to custom type legalize this operation!");
7572 case ISD::FP_TO_SINT: {
7573 std::pair<SDValue,SDValue> Vals =
7574 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7575 SDValue FIST = Vals.first, StackSlot = Vals.second;
7576 if (FIST.getNode() != 0) {
7577 EVT VT = N->getValueType(0);
7578 // Return a load from the stack slot.
7579 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7584 case ISD::READCYCLECOUNTER: {
7585 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7586 SDValue TheChain = N->getOperand(0);
7587 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7588 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7590 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7592 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7593 SDValue Ops[] = { eax, edx };
7594 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7595 Results.push_back(edx.getValue(1));
7598 case ISD::ATOMIC_CMP_SWAP: {
7599 EVT T = N->getValueType(0);
7600 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7601 SDValue cpInL, cpInH;
7602 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7603 DAG.getConstant(0, MVT::i32));
7604 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7605 DAG.getConstant(1, MVT::i32));
7606 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7607 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7609 SDValue swapInL, swapInH;
7610 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7611 DAG.getConstant(0, MVT::i32));
7612 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7613 DAG.getConstant(1, MVT::i32));
7614 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7616 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7617 swapInL.getValue(1));
7618 SDValue Ops[] = { swapInH.getValue(0),
7620 swapInH.getValue(1) };
7621 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7622 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7623 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7624 MVT::i32, Result.getValue(1));
7625 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7626 MVT::i32, cpOutL.getValue(2));
7627 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7628 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7629 Results.push_back(cpOutH.getValue(1));
7632 case ISD::ATOMIC_LOAD_ADD:
7633 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7635 case ISD::ATOMIC_LOAD_AND:
7636 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7638 case ISD::ATOMIC_LOAD_NAND:
7639 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7641 case ISD::ATOMIC_LOAD_OR:
7642 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7644 case ISD::ATOMIC_LOAD_SUB:
7645 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7647 case ISD::ATOMIC_LOAD_XOR:
7648 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7650 case ISD::ATOMIC_SWAP:
7651 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7656 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7658 default: return NULL;
7659 case X86ISD::BSF: return "X86ISD::BSF";
7660 case X86ISD::BSR: return "X86ISD::BSR";
7661 case X86ISD::SHLD: return "X86ISD::SHLD";
7662 case X86ISD::SHRD: return "X86ISD::SHRD";
7663 case X86ISD::FAND: return "X86ISD::FAND";
7664 case X86ISD::FOR: return "X86ISD::FOR";
7665 case X86ISD::FXOR: return "X86ISD::FXOR";
7666 case X86ISD::FSRL: return "X86ISD::FSRL";
7667 case X86ISD::FILD: return "X86ISD::FILD";
7668 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7669 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7670 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7671 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7672 case X86ISD::FLD: return "X86ISD::FLD";
7673 case X86ISD::FST: return "X86ISD::FST";
7674 case X86ISD::CALL: return "X86ISD::CALL";
7675 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7676 case X86ISD::BT: return "X86ISD::BT";
7677 case X86ISD::CMP: return "X86ISD::CMP";
7678 case X86ISD::COMI: return "X86ISD::COMI";
7679 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7680 case X86ISD::SETCC: return "X86ISD::SETCC";
7681 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7682 case X86ISD::CMOV: return "X86ISD::CMOV";
7683 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7684 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7685 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7686 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7687 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7688 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7689 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7690 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7691 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7692 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7693 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7694 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7695 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
7696 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7697 case X86ISD::FMAX: return "X86ISD::FMAX";
7698 case X86ISD::FMIN: return "X86ISD::FMIN";
7699 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7700 case X86ISD::FRCP: return "X86ISD::FRCP";
7701 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7702 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7703 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7704 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7705 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7706 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7707 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7708 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7709 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7710 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7711 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7712 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7713 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7714 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7715 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7716 case X86ISD::VSHL: return "X86ISD::VSHL";
7717 case X86ISD::VSRL: return "X86ISD::VSRL";
7718 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7719 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7720 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7721 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7722 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7723 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7724 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7725 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7726 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7727 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7728 case X86ISD::ADD: return "X86ISD::ADD";
7729 case X86ISD::SUB: return "X86ISD::SUB";
7730 case X86ISD::SMUL: return "X86ISD::SMUL";
7731 case X86ISD::UMUL: return "X86ISD::UMUL";
7732 case X86ISD::INC: return "X86ISD::INC";
7733 case X86ISD::DEC: return "X86ISD::DEC";
7734 case X86ISD::OR: return "X86ISD::OR";
7735 case X86ISD::XOR: return "X86ISD::XOR";
7736 case X86ISD::AND: return "X86ISD::AND";
7737 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7738 case X86ISD::PTEST: return "X86ISD::PTEST";
7739 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7743 // isLegalAddressingMode - Return true if the addressing mode represented
7744 // by AM is legal for this target, for a load/store of the specified type.
7745 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7746 const Type *Ty) const {
7747 // X86 supports extremely general addressing modes.
7748 CodeModel::Model M = getTargetMachine().getCodeModel();
7750 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7751 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7756 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7758 // If a reference to this global requires an extra load, we can't fold it.
7759 if (isGlobalStubReference(GVFlags))
7762 // If BaseGV requires a register for the PIC base, we cannot also have a
7763 // BaseReg specified.
7764 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7767 // If lower 4G is not available, then we must use rip-relative addressing.
7768 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7778 // These scales always work.
7783 // These scales are formed with basereg+scalereg. Only accept if there is
7788 default: // Other stuff never works.
7796 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7797 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7799 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7800 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7801 if (NumBits1 <= NumBits2)
7803 return Subtarget->is64Bit() || NumBits1 < 64;
7806 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7807 if (!VT1.isInteger() || !VT2.isInteger())
7809 unsigned NumBits1 = VT1.getSizeInBits();
7810 unsigned NumBits2 = VT2.getSizeInBits();
7811 if (NumBits1 <= NumBits2)
7813 return Subtarget->is64Bit() || NumBits1 < 64;
7816 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7817 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7818 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7821 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7822 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7823 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7826 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7827 // i16 instructions are longer (0x66 prefix) and potentially slower.
7828 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7831 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7832 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7833 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7834 /// are assumed to be legal.
7836 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7838 // Only do shuffles on 128-bit vector types for now.
7839 if (VT.getSizeInBits() == 64)
7842 // FIXME: pshufb, blends, shifts.
7843 return (VT.getVectorNumElements() == 2 ||
7844 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7845 isMOVLMask(M, VT) ||
7846 isSHUFPMask(M, VT) ||
7847 isPSHUFDMask(M, VT) ||
7848 isPSHUFHWMask(M, VT) ||
7849 isPSHUFLWMask(M, VT) ||
7850 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7851 isUNPCKLMask(M, VT) ||
7852 isUNPCKHMask(M, VT) ||
7853 isUNPCKL_v_undef_Mask(M, VT) ||
7854 isUNPCKH_v_undef_Mask(M, VT));
7858 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7860 unsigned NumElts = VT.getVectorNumElements();
7861 // FIXME: This collection of masks seems suspect.
7864 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7865 return (isMOVLMask(Mask, VT) ||
7866 isCommutedMOVLMask(Mask, VT, true) ||
7867 isSHUFPMask(Mask, VT) ||
7868 isCommutedSHUFPMask(Mask, VT));
7873 //===----------------------------------------------------------------------===//
7874 // X86 Scheduler Hooks
7875 //===----------------------------------------------------------------------===//
7877 // private utility function
7879 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7880 MachineBasicBlock *MBB,
7888 TargetRegisterClass *RC,
7889 bool invSrc) const {
7890 // For the atomic bitwise operator, we generate
7893 // ld t1 = [bitinstr.addr]
7894 // op t2 = t1, [bitinstr.val]
7896 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7898 // fallthrough -->nextMBB
7899 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7900 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7901 MachineFunction::iterator MBBIter = MBB;
7904 /// First build the CFG
7905 MachineFunction *F = MBB->getParent();
7906 MachineBasicBlock *thisMBB = MBB;
7907 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7908 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7909 F->insert(MBBIter, newMBB);
7910 F->insert(MBBIter, nextMBB);
7912 // Move all successors to thisMBB to nextMBB
7913 nextMBB->transferSuccessors(thisMBB);
7915 // Update thisMBB to fall through to newMBB
7916 thisMBB->addSuccessor(newMBB);
7918 // newMBB jumps to itself and fall through to nextMBB
7919 newMBB->addSuccessor(nextMBB);
7920 newMBB->addSuccessor(newMBB);
7922 // Insert instructions into newMBB based on incoming instruction
7923 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7924 "unexpected number of operands");
7925 DebugLoc dl = bInstr->getDebugLoc();
7926 MachineOperand& destOper = bInstr->getOperand(0);
7927 MachineOperand* argOpers[2 + X86AddrNumOperands];
7928 int numArgs = bInstr->getNumOperands() - 1;
7929 for (int i=0; i < numArgs; ++i)
7930 argOpers[i] = &bInstr->getOperand(i+1);
7932 // x86 address has 4 operands: base, index, scale, and displacement
7933 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7934 int valArgIndx = lastAddrIndx + 1;
7936 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7937 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7938 for (int i=0; i <= lastAddrIndx; ++i)
7939 (*MIB).addOperand(*argOpers[i]);
7941 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7943 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7948 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7949 assert((argOpers[valArgIndx]->isReg() ||
7950 argOpers[valArgIndx]->isImm()) &&
7952 if (argOpers[valArgIndx]->isReg())
7953 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7955 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7957 (*MIB).addOperand(*argOpers[valArgIndx]);
7959 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7962 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7963 for (int i=0; i <= lastAddrIndx; ++i)
7964 (*MIB).addOperand(*argOpers[i]);
7966 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7967 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7968 bInstr->memoperands_end());
7970 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7974 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
7976 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7980 // private utility function: 64 bit atomics on 32 bit host.
7982 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7983 MachineBasicBlock *MBB,
7988 bool invSrc) const {
7989 // For the atomic bitwise operator, we generate
7990 // thisMBB (instructions are in pairs, except cmpxchg8b)
7991 // ld t1,t2 = [bitinstr.addr]
7993 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7994 // op t5, t6 <- out1, out2, [bitinstr.val]
7995 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7996 // mov ECX, EBX <- t5, t6
7997 // mov EAX, EDX <- t1, t2
7998 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7999 // mov t3, t4 <- EAX, EDX
8001 // result in out1, out2
8002 // fallthrough -->nextMBB
8004 const TargetRegisterClass *RC = X86::GR32RegisterClass;
8005 const unsigned LoadOpc = X86::MOV32rm;
8006 const unsigned copyOpc = X86::MOV32rr;
8007 const unsigned NotOpc = X86::NOT32r;
8008 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8009 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8010 MachineFunction::iterator MBBIter = MBB;
8013 /// First build the CFG
8014 MachineFunction *F = MBB->getParent();
8015 MachineBasicBlock *thisMBB = MBB;
8016 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8017 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8018 F->insert(MBBIter, newMBB);
8019 F->insert(MBBIter, nextMBB);
8021 // Move all successors to thisMBB to nextMBB
8022 nextMBB->transferSuccessors(thisMBB);
8024 // Update thisMBB to fall through to newMBB
8025 thisMBB->addSuccessor(newMBB);
8027 // newMBB jumps to itself and fall through to nextMBB
8028 newMBB->addSuccessor(nextMBB);
8029 newMBB->addSuccessor(newMBB);
8031 DebugLoc dl = bInstr->getDebugLoc();
8032 // Insert instructions into newMBB based on incoming instruction
8033 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8034 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8035 "unexpected number of operands");
8036 MachineOperand& dest1Oper = bInstr->getOperand(0);
8037 MachineOperand& dest2Oper = bInstr->getOperand(1);
8038 MachineOperand* argOpers[2 + X86AddrNumOperands];
8039 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8040 argOpers[i] = &bInstr->getOperand(i+2);
8042 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8043 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8045 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8046 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8047 for (int i=0; i <= lastAddrIndx; ++i)
8048 (*MIB).addOperand(*argOpers[i]);
8049 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8050 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8051 // add 4 to displacement.
8052 for (int i=0; i <= lastAddrIndx-2; ++i)
8053 (*MIB).addOperand(*argOpers[i]);
8054 MachineOperand newOp3 = *(argOpers[3]);
8056 newOp3.setImm(newOp3.getImm()+4);
8058 newOp3.setOffset(newOp3.getOffset()+4);
8059 (*MIB).addOperand(newOp3);
8060 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8062 // t3/4 are defined later, at the bottom of the loop
8063 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8064 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8065 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8066 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8067 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8068 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8070 // The subsequent operations should be using the destination registers of
8071 //the PHI instructions.
8073 t1 = F->getRegInfo().createVirtualRegister(RC);
8074 t2 = F->getRegInfo().createVirtualRegister(RC);
8075 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8076 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8078 t1 = dest1Oper.getReg();
8079 t2 = dest2Oper.getReg();
8082 int valArgIndx = lastAddrIndx + 1;
8083 assert((argOpers[valArgIndx]->isReg() ||
8084 argOpers[valArgIndx]->isImm()) &&
8086 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8087 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8088 if (argOpers[valArgIndx]->isReg())
8089 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8091 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8092 if (regOpcL != X86::MOV32rr)
8094 (*MIB).addOperand(*argOpers[valArgIndx]);
8095 assert(argOpers[valArgIndx + 1]->isReg() ==
8096 argOpers[valArgIndx]->isReg());
8097 assert(argOpers[valArgIndx + 1]->isImm() ==
8098 argOpers[valArgIndx]->isImm());
8099 if (argOpers[valArgIndx + 1]->isReg())
8100 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8102 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8103 if (regOpcH != X86::MOV32rr)
8105 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8107 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8109 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8112 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8114 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8117 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8118 for (int i=0; i <= lastAddrIndx; ++i)
8119 (*MIB).addOperand(*argOpers[i]);
8121 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8122 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8123 bInstr->memoperands_end());
8125 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8126 MIB.addReg(X86::EAX);
8127 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8128 MIB.addReg(X86::EDX);
8131 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8133 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8137 // private utility function
8139 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8140 MachineBasicBlock *MBB,
8141 unsigned cmovOpc) const {
8142 // For the atomic min/max operator, we generate
8145 // ld t1 = [min/max.addr]
8146 // mov t2 = [min/max.val]
8148 // cmov[cond] t2 = t1
8150 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8152 // fallthrough -->nextMBB
8154 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8155 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8156 MachineFunction::iterator MBBIter = MBB;
8159 /// First build the CFG
8160 MachineFunction *F = MBB->getParent();
8161 MachineBasicBlock *thisMBB = MBB;
8162 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8163 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8164 F->insert(MBBIter, newMBB);
8165 F->insert(MBBIter, nextMBB);
8167 // Move all successors of thisMBB to nextMBB
8168 nextMBB->transferSuccessors(thisMBB);
8170 // Update thisMBB to fall through to newMBB
8171 thisMBB->addSuccessor(newMBB);
8173 // newMBB jumps to newMBB and fall through to nextMBB
8174 newMBB->addSuccessor(nextMBB);
8175 newMBB->addSuccessor(newMBB);
8177 DebugLoc dl = mInstr->getDebugLoc();
8178 // Insert instructions into newMBB based on incoming instruction
8179 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8180 "unexpected number of operands");
8181 MachineOperand& destOper = mInstr->getOperand(0);
8182 MachineOperand* argOpers[2 + X86AddrNumOperands];
8183 int numArgs = mInstr->getNumOperands() - 1;
8184 for (int i=0; i < numArgs; ++i)
8185 argOpers[i] = &mInstr->getOperand(i+1);
8187 // x86 address has 4 operands: base, index, scale, and displacement
8188 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8189 int valArgIndx = lastAddrIndx + 1;
8191 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8192 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8193 for (int i=0; i <= lastAddrIndx; ++i)
8194 (*MIB).addOperand(*argOpers[i]);
8196 // We only support register and immediate values
8197 assert((argOpers[valArgIndx]->isReg() ||
8198 argOpers[valArgIndx]->isImm()) &&
8201 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8202 if (argOpers[valArgIndx]->isReg())
8203 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8205 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8206 (*MIB).addOperand(*argOpers[valArgIndx]);
8208 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8211 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8216 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8217 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8221 // Cmp and exchange if none has modified the memory location
8222 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8223 for (int i=0; i <= lastAddrIndx; ++i)
8224 (*MIB).addOperand(*argOpers[i]);
8226 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8227 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8228 mInstr->memoperands_end());
8230 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8231 MIB.addReg(X86::EAX);
8234 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8236 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8240 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8241 // all of this code can be replaced with that in the .td file.
8243 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8244 unsigned numArgs, bool memArg) const {
8246 MachineFunction *F = BB->getParent();
8247 DebugLoc dl = MI->getDebugLoc();
8248 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8252 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8254 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8256 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8258 for (unsigned i = 0; i < numArgs; ++i) {
8259 MachineOperand &Op = MI->getOperand(i+1);
8261 if (!(Op.isReg() && Op.isImplicit()))
8265 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8268 F->DeleteMachineInstr(MI);
8274 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8276 MachineBasicBlock *MBB) const {
8277 // Emit code to save XMM registers to the stack. The ABI says that the
8278 // number of registers to save is given in %al, so it's theoretically
8279 // possible to do an indirect jump trick to avoid saving all of them,
8280 // however this code takes a simpler approach and just executes all
8281 // of the stores if %al is non-zero. It's less code, and it's probably
8282 // easier on the hardware branch predictor, and stores aren't all that
8283 // expensive anyway.
8285 // Create the new basic blocks. One block contains all the XMM stores,
8286 // and one block is the final destination regardless of whether any
8287 // stores were performed.
8288 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8289 MachineFunction *F = MBB->getParent();
8290 MachineFunction::iterator MBBIter = MBB;
8292 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8293 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8294 F->insert(MBBIter, XMMSaveMBB);
8295 F->insert(MBBIter, EndMBB);
8298 // Move any original successors of MBB to the end block.
8299 EndMBB->transferSuccessors(MBB);
8300 // The original block will now fall through to the XMM save block.
8301 MBB->addSuccessor(XMMSaveMBB);
8302 // The XMMSaveMBB will fall through to the end block.
8303 XMMSaveMBB->addSuccessor(EndMBB);
8305 // Now add the instructions.
8306 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8307 DebugLoc DL = MI->getDebugLoc();
8309 unsigned CountReg = MI->getOperand(0).getReg();
8310 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8311 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8313 if (!Subtarget->isTargetWin64()) {
8314 // If %al is 0, branch around the XMM save block.
8315 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8316 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8317 MBB->addSuccessor(EndMBB);
8320 // In the XMM save block, save all the XMM argument registers.
8321 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8322 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8323 MachineMemOperand *MMO =
8324 F->getMachineMemOperand(
8325 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8326 MachineMemOperand::MOStore, Offset,
8327 /*Size=*/16, /*Align=*/16);
8328 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8329 .addFrameIndex(RegSaveFrameIndex)
8330 .addImm(/*Scale=*/1)
8331 .addReg(/*IndexReg=*/0)
8332 .addImm(/*Disp=*/Offset)
8333 .addReg(/*Segment=*/0)
8334 .addReg(MI->getOperand(i).getReg())
8335 .addMemOperand(MMO);
8338 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8344 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8345 MachineBasicBlock *BB,
8346 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8347 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8348 DebugLoc DL = MI->getDebugLoc();
8350 // To "insert" a SELECT_CC instruction, we actually have to insert the
8351 // diamond control-flow pattern. The incoming instruction knows the
8352 // destination vreg to set, the condition code register to branch on, the
8353 // true/false values to select between, and a branch opcode to use.
8354 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8355 MachineFunction::iterator It = BB;
8361 // cmpTY ccX, r1, r2
8363 // fallthrough --> copy0MBB
8364 MachineBasicBlock *thisMBB = BB;
8365 MachineFunction *F = BB->getParent();
8366 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8367 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8369 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8370 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8371 F->insert(It, copy0MBB);
8372 F->insert(It, sinkMBB);
8373 // Update machine-CFG edges by first adding all successors of the current
8374 // block to the new block which will contain the Phi node for the select.
8375 // Also inform sdisel of the edge changes.
8376 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8377 E = BB->succ_end(); I != E; ++I) {
8378 EM->insert(std::make_pair(*I, sinkMBB));
8379 sinkMBB->addSuccessor(*I);
8381 // Next, remove all successors of the current block, and add the true
8382 // and fallthrough blocks as its successors.
8383 while (!BB->succ_empty())
8384 BB->removeSuccessor(BB->succ_begin());
8385 // Add the true and fallthrough blocks as its successors.
8386 BB->addSuccessor(copy0MBB);
8387 BB->addSuccessor(sinkMBB);
8390 // %FalseValue = ...
8391 // # fallthrough to sinkMBB
8394 // Update machine-CFG edges
8395 BB->addSuccessor(sinkMBB);
8398 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8401 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8402 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8403 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8405 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8411 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8412 MachineBasicBlock *BB,
8413 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8414 switch (MI->getOpcode()) {
8415 default: assert(false && "Unexpected instr type to insert");
8417 case X86::CMOV_V1I64:
8418 case X86::CMOV_FR32:
8419 case X86::CMOV_FR64:
8420 case X86::CMOV_V4F32:
8421 case X86::CMOV_V2F64:
8422 case X86::CMOV_V2I64:
8423 return EmitLoweredSelect(MI, BB, EM);
8425 case X86::FP32_TO_INT16_IN_MEM:
8426 case X86::FP32_TO_INT32_IN_MEM:
8427 case X86::FP32_TO_INT64_IN_MEM:
8428 case X86::FP64_TO_INT16_IN_MEM:
8429 case X86::FP64_TO_INT32_IN_MEM:
8430 case X86::FP64_TO_INT64_IN_MEM:
8431 case X86::FP80_TO_INT16_IN_MEM:
8432 case X86::FP80_TO_INT32_IN_MEM:
8433 case X86::FP80_TO_INT64_IN_MEM: {
8434 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8435 DebugLoc DL = MI->getDebugLoc();
8437 // Change the floating point control register to use "round towards zero"
8438 // mode when truncating to an integer value.
8439 MachineFunction *F = BB->getParent();
8440 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8441 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8443 // Load the old value of the high byte of the control word...
8445 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8446 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8449 // Set the high part to be round to zero...
8450 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8453 // Reload the modified control word now...
8454 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8456 // Restore the memory image of control word to original value
8457 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8460 // Get the X86 opcode to use.
8462 switch (MI->getOpcode()) {
8463 default: llvm_unreachable("illegal opcode!");
8464 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8465 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8466 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8467 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8468 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8469 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8470 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8471 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8472 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8476 MachineOperand &Op = MI->getOperand(0);
8478 AM.BaseType = X86AddressMode::RegBase;
8479 AM.Base.Reg = Op.getReg();
8481 AM.BaseType = X86AddressMode::FrameIndexBase;
8482 AM.Base.FrameIndex = Op.getIndex();
8484 Op = MI->getOperand(1);
8486 AM.Scale = Op.getImm();
8487 Op = MI->getOperand(2);
8489 AM.IndexReg = Op.getImm();
8490 Op = MI->getOperand(3);
8491 if (Op.isGlobal()) {
8492 AM.GV = Op.getGlobal();
8494 AM.Disp = Op.getImm();
8496 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8497 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8499 // Reload the original control word now.
8500 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8502 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8505 // String/text processing lowering.
8506 case X86::PCMPISTRM128REG:
8507 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8508 case X86::PCMPISTRM128MEM:
8509 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8510 case X86::PCMPESTRM128REG:
8511 return EmitPCMP(MI, BB, 5, false /* in mem */);
8512 case X86::PCMPESTRM128MEM:
8513 return EmitPCMP(MI, BB, 5, true /* in mem */);
8516 case X86::ATOMAND32:
8517 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8518 X86::AND32ri, X86::MOV32rm,
8519 X86::LCMPXCHG32, X86::MOV32rr,
8520 X86::NOT32r, X86::EAX,
8521 X86::GR32RegisterClass);
8523 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8524 X86::OR32ri, X86::MOV32rm,
8525 X86::LCMPXCHG32, X86::MOV32rr,
8526 X86::NOT32r, X86::EAX,
8527 X86::GR32RegisterClass);
8528 case X86::ATOMXOR32:
8529 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8530 X86::XOR32ri, X86::MOV32rm,
8531 X86::LCMPXCHG32, X86::MOV32rr,
8532 X86::NOT32r, X86::EAX,
8533 X86::GR32RegisterClass);
8534 case X86::ATOMNAND32:
8535 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8536 X86::AND32ri, X86::MOV32rm,
8537 X86::LCMPXCHG32, X86::MOV32rr,
8538 X86::NOT32r, X86::EAX,
8539 X86::GR32RegisterClass, true);
8540 case X86::ATOMMIN32:
8541 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8542 case X86::ATOMMAX32:
8543 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8544 case X86::ATOMUMIN32:
8545 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8546 case X86::ATOMUMAX32:
8547 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8549 case X86::ATOMAND16:
8550 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8551 X86::AND16ri, X86::MOV16rm,
8552 X86::LCMPXCHG16, X86::MOV16rr,
8553 X86::NOT16r, X86::AX,
8554 X86::GR16RegisterClass);
8556 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8557 X86::OR16ri, X86::MOV16rm,
8558 X86::LCMPXCHG16, X86::MOV16rr,
8559 X86::NOT16r, X86::AX,
8560 X86::GR16RegisterClass);
8561 case X86::ATOMXOR16:
8562 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8563 X86::XOR16ri, X86::MOV16rm,
8564 X86::LCMPXCHG16, X86::MOV16rr,
8565 X86::NOT16r, X86::AX,
8566 X86::GR16RegisterClass);
8567 case X86::ATOMNAND16:
8568 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8569 X86::AND16ri, X86::MOV16rm,
8570 X86::LCMPXCHG16, X86::MOV16rr,
8571 X86::NOT16r, X86::AX,
8572 X86::GR16RegisterClass, true);
8573 case X86::ATOMMIN16:
8574 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8575 case X86::ATOMMAX16:
8576 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8577 case X86::ATOMUMIN16:
8578 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8579 case X86::ATOMUMAX16:
8580 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8583 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8584 X86::AND8ri, X86::MOV8rm,
8585 X86::LCMPXCHG8, X86::MOV8rr,
8586 X86::NOT8r, X86::AL,
8587 X86::GR8RegisterClass);
8589 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8590 X86::OR8ri, X86::MOV8rm,
8591 X86::LCMPXCHG8, X86::MOV8rr,
8592 X86::NOT8r, X86::AL,
8593 X86::GR8RegisterClass);
8595 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8596 X86::XOR8ri, X86::MOV8rm,
8597 X86::LCMPXCHG8, X86::MOV8rr,
8598 X86::NOT8r, X86::AL,
8599 X86::GR8RegisterClass);
8600 case X86::ATOMNAND8:
8601 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8602 X86::AND8ri, X86::MOV8rm,
8603 X86::LCMPXCHG8, X86::MOV8rr,
8604 X86::NOT8r, X86::AL,
8605 X86::GR8RegisterClass, true);
8606 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8607 // This group is for 64-bit host.
8608 case X86::ATOMAND64:
8609 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8610 X86::AND64ri32, X86::MOV64rm,
8611 X86::LCMPXCHG64, X86::MOV64rr,
8612 X86::NOT64r, X86::RAX,
8613 X86::GR64RegisterClass);
8615 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8616 X86::OR64ri32, X86::MOV64rm,
8617 X86::LCMPXCHG64, X86::MOV64rr,
8618 X86::NOT64r, X86::RAX,
8619 X86::GR64RegisterClass);
8620 case X86::ATOMXOR64:
8621 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8622 X86::XOR64ri32, X86::MOV64rm,
8623 X86::LCMPXCHG64, X86::MOV64rr,
8624 X86::NOT64r, X86::RAX,
8625 X86::GR64RegisterClass);
8626 case X86::ATOMNAND64:
8627 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8628 X86::AND64ri32, X86::MOV64rm,
8629 X86::LCMPXCHG64, X86::MOV64rr,
8630 X86::NOT64r, X86::RAX,
8631 X86::GR64RegisterClass, true);
8632 case X86::ATOMMIN64:
8633 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8634 case X86::ATOMMAX64:
8635 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8636 case X86::ATOMUMIN64:
8637 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8638 case X86::ATOMUMAX64:
8639 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8641 // This group does 64-bit operations on a 32-bit host.
8642 case X86::ATOMAND6432:
8643 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8644 X86::AND32rr, X86::AND32rr,
8645 X86::AND32ri, X86::AND32ri,
8647 case X86::ATOMOR6432:
8648 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8649 X86::OR32rr, X86::OR32rr,
8650 X86::OR32ri, X86::OR32ri,
8652 case X86::ATOMXOR6432:
8653 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8654 X86::XOR32rr, X86::XOR32rr,
8655 X86::XOR32ri, X86::XOR32ri,
8657 case X86::ATOMNAND6432:
8658 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8659 X86::AND32rr, X86::AND32rr,
8660 X86::AND32ri, X86::AND32ri,
8662 case X86::ATOMADD6432:
8663 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8664 X86::ADD32rr, X86::ADC32rr,
8665 X86::ADD32ri, X86::ADC32ri,
8667 case X86::ATOMSUB6432:
8668 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8669 X86::SUB32rr, X86::SBB32rr,
8670 X86::SUB32ri, X86::SBB32ri,
8672 case X86::ATOMSWAP6432:
8673 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8674 X86::MOV32rr, X86::MOV32rr,
8675 X86::MOV32ri, X86::MOV32ri,
8677 case X86::VASTART_SAVE_XMM_REGS:
8678 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8682 //===----------------------------------------------------------------------===//
8683 // X86 Optimization Hooks
8684 //===----------------------------------------------------------------------===//
8686 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8690 const SelectionDAG &DAG,
8691 unsigned Depth) const {
8692 unsigned Opc = Op.getOpcode();
8693 assert((Opc >= ISD::BUILTIN_OP_END ||
8694 Opc == ISD::INTRINSIC_WO_CHAIN ||
8695 Opc == ISD::INTRINSIC_W_CHAIN ||
8696 Opc == ISD::INTRINSIC_VOID) &&
8697 "Should use MaskedValueIsZero if you don't know whether Op"
8698 " is a target node!");
8700 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8712 // These nodes' second result is a boolean.
8713 if (Op.getResNo() == 0)
8717 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8718 Mask.getBitWidth() - 1);
8723 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8724 /// node is a GlobalAddress + offset.
8725 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8726 GlobalValue* &GA, int64_t &Offset) const{
8727 if (N->getOpcode() == X86ISD::Wrapper) {
8728 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8729 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8730 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8734 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8737 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8738 EVT EltVT, LoadSDNode *&LDBase,
8739 unsigned &LastLoadedElt,
8740 SelectionDAG &DAG, MachineFrameInfo *MFI,
8741 const TargetLowering &TLI) {
8743 LastLoadedElt = -1U;
8744 for (unsigned i = 0; i < NumElems; ++i) {
8745 if (N->getMaskElt(i) < 0) {
8751 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8752 if (!Elt.getNode() ||
8753 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8756 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8758 LDBase = cast<LoadSDNode>(Elt.getNode());
8762 if (Elt.getOpcode() == ISD::UNDEF)
8765 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8766 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8773 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8774 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8775 /// if the load addresses are consecutive, non-overlapping, and in the right
8776 /// order. In the case of v2i64, it will see if it can rewrite the
8777 /// shuffle to be an appropriate build vector so it can take advantage of
8778 // performBuildVectorCombine.
8779 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8780 const TargetLowering &TLI) {
8781 DebugLoc dl = N->getDebugLoc();
8782 EVT VT = N->getValueType(0);
8783 EVT EltVT = VT.getVectorElementType();
8784 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8785 unsigned NumElems = VT.getVectorNumElements();
8787 if (VT.getSizeInBits() != 128)
8790 // Try to combine a vector_shuffle into a 128-bit load.
8791 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8792 LoadSDNode *LD = NULL;
8793 unsigned LastLoadedElt;
8794 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8798 if (LastLoadedElt == NumElems - 1) {
8799 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8800 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8801 LD->getSrcValue(), LD->getSrcValueOffset(),
8802 LD->isVolatile(), LD->isNonTemporal(), 0);
8803 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8804 LD->getSrcValue(), LD->getSrcValueOffset(),
8805 LD->isVolatile(), LD->isNonTemporal(),
8806 LD->getAlignment());
8807 } else if (NumElems == 4 && LastLoadedElt == 1) {
8808 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8809 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8810 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8811 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8816 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8817 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8818 const X86Subtarget *Subtarget) {
8819 DebugLoc DL = N->getDebugLoc();
8820 SDValue Cond = N->getOperand(0);
8821 // Get the LHS/RHS of the select.
8822 SDValue LHS = N->getOperand(1);
8823 SDValue RHS = N->getOperand(2);
8825 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8826 // instructions match the semantics of the common C idiom x<y?x:y but not
8827 // x<=y?x:y, because of how they handle negative zero (which can be
8828 // ignored in unsafe-math mode).
8829 if (Subtarget->hasSSE2() &&
8830 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8831 Cond.getOpcode() == ISD::SETCC) {
8832 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8834 unsigned Opcode = 0;
8835 // Check for x CC y ? x : y.
8836 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8840 if (!UnsafeFPMath) break;
8841 Opcode = X86ISD::FMIN;
8844 if (!UnsafeFPMath) break;
8845 Opcode = X86ISD::FMIN;
8851 Opcode = X86ISD::FMIN;
8855 if (!UnsafeFPMath) break;
8856 Opcode = X86ISD::FMAX;
8859 if (!UnsafeFPMath) break;
8860 Opcode = X86ISD::FMAX;
8866 Opcode = X86ISD::FMAX;
8869 // Check for x CC y ? y : x -- a min/max with reversed arms.
8870 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8874 if (!UnsafeFPMath) break;
8875 Opcode = X86ISD::FMIN;
8878 if (!UnsafeFPMath) break;
8879 Opcode = X86ISD::FMIN;
8885 Opcode = X86ISD::FMIN;
8889 if (!UnsafeFPMath) break;
8890 Opcode = X86ISD::FMAX;
8893 if (!UnsafeFPMath) break;
8894 Opcode = X86ISD::FMAX;
8900 Opcode = X86ISD::FMAX;
8906 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8909 // If this is a select between two integer constants, try to do some
8911 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8912 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8913 // Don't do this for crazy integer types.
8914 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8915 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8916 // so that TrueC (the true value) is larger than FalseC.
8917 bool NeedsCondInvert = false;
8919 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8920 // Efficiently invertible.
8921 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8922 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8923 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8924 NeedsCondInvert = true;
8925 std::swap(TrueC, FalseC);
8928 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8929 if (FalseC->getAPIntValue() == 0 &&
8930 TrueC->getAPIntValue().isPowerOf2()) {
8931 if (NeedsCondInvert) // Invert the condition if needed.
8932 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8933 DAG.getConstant(1, Cond.getValueType()));
8935 // Zero extend the condition if needed.
8936 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8938 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8939 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8940 DAG.getConstant(ShAmt, MVT::i8));
8943 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8944 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8945 if (NeedsCondInvert) // Invert the condition if needed.
8946 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8947 DAG.getConstant(1, Cond.getValueType()));
8949 // Zero extend the condition if needed.
8950 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8951 FalseC->getValueType(0), Cond);
8952 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8953 SDValue(FalseC, 0));
8956 // Optimize cases that will turn into an LEA instruction. This requires
8957 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8958 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8959 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8960 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8962 bool isFastMultiplier = false;
8964 switch ((unsigned char)Diff) {
8966 case 1: // result = add base, cond
8967 case 2: // result = lea base( , cond*2)
8968 case 3: // result = lea base(cond, cond*2)
8969 case 4: // result = lea base( , cond*4)
8970 case 5: // result = lea base(cond, cond*4)
8971 case 8: // result = lea base( , cond*8)
8972 case 9: // result = lea base(cond, cond*8)
8973 isFastMultiplier = true;
8978 if (isFastMultiplier) {
8979 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8980 if (NeedsCondInvert) // Invert the condition if needed.
8981 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8982 DAG.getConstant(1, Cond.getValueType()));
8984 // Zero extend the condition if needed.
8985 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8987 // Scale the condition by the difference.
8989 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8990 DAG.getConstant(Diff, Cond.getValueType()));
8992 // Add the base if non-zero.
8993 if (FalseC->getAPIntValue() != 0)
8994 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8995 SDValue(FalseC, 0));
9005 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9006 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9007 TargetLowering::DAGCombinerInfo &DCI) {
9008 DebugLoc DL = N->getDebugLoc();
9010 // If the flag operand isn't dead, don't touch this CMOV.
9011 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9014 // If this is a select between two integer constants, try to do some
9015 // optimizations. Note that the operands are ordered the opposite of SELECT
9017 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9018 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9019 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9020 // larger than FalseC (the false value).
9021 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9023 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9024 CC = X86::GetOppositeBranchCondition(CC);
9025 std::swap(TrueC, FalseC);
9028 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9029 // This is efficient for any integer data type (including i8/i16) and
9031 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9032 SDValue Cond = N->getOperand(3);
9033 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9034 DAG.getConstant(CC, MVT::i8), Cond);
9036 // Zero extend the condition if needed.
9037 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9039 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9040 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9041 DAG.getConstant(ShAmt, MVT::i8));
9042 if (N->getNumValues() == 2) // Dead flag value?
9043 return DCI.CombineTo(N, Cond, SDValue());
9047 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9048 // for any integer data type, including i8/i16.
9049 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9050 SDValue Cond = N->getOperand(3);
9051 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9052 DAG.getConstant(CC, MVT::i8), Cond);
9054 // Zero extend the condition if needed.
9055 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9056 FalseC->getValueType(0), Cond);
9057 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9058 SDValue(FalseC, 0));
9060 if (N->getNumValues() == 2) // Dead flag value?
9061 return DCI.CombineTo(N, Cond, SDValue());
9065 // Optimize cases that will turn into an LEA instruction. This requires
9066 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9067 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9068 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9069 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9071 bool isFastMultiplier = false;
9073 switch ((unsigned char)Diff) {
9075 case 1: // result = add base, cond
9076 case 2: // result = lea base( , cond*2)
9077 case 3: // result = lea base(cond, cond*2)
9078 case 4: // result = lea base( , cond*4)
9079 case 5: // result = lea base(cond, cond*4)
9080 case 8: // result = lea base( , cond*8)
9081 case 9: // result = lea base(cond, cond*8)
9082 isFastMultiplier = true;
9087 if (isFastMultiplier) {
9088 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9089 SDValue Cond = N->getOperand(3);
9090 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9091 DAG.getConstant(CC, MVT::i8), Cond);
9092 // Zero extend the condition if needed.
9093 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9095 // Scale the condition by the difference.
9097 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9098 DAG.getConstant(Diff, Cond.getValueType()));
9100 // Add the base if non-zero.
9101 if (FalseC->getAPIntValue() != 0)
9102 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9103 SDValue(FalseC, 0));
9104 if (N->getNumValues() == 2) // Dead flag value?
9105 return DCI.CombineTo(N, Cond, SDValue());
9114 /// PerformANDCombine - Look for SSE and instructions of this form:
9115 /// (and x, (build_vector signbit,signbit,signbit,signbit)). If there
9116 /// exists a use of a build_vector that's the bitwise complement of the mask,
9117 /// then transform the node to
9118 /// (and (xor x, (build_vector -1,-1,-1,-1)), (build_vector ~sb,~sb,~sb,~sb)).
9119 static SDValue PerformANDCombine(SDNode *N, SelectionDAG &DAG,
9120 TargetLowering::DAGCombinerInfo &DCI) {
9121 EVT VT = N->getValueType(0);
9122 if (!VT.isVector() || !VT.isInteger())
9125 SDValue N0 = N->getOperand(0);
9126 SDValue N1 = N->getOperand(1);
9127 if (N0.getOpcode() == ISD::XOR || !N1.hasOneUse())
9130 if (N1.getOpcode() == ISD::BUILD_VECTOR) {
9131 unsigned NumElts = VT.getVectorNumElements();
9132 EVT EltVT = VT.getVectorElementType();
9133 SmallVector<SDValue, 8> Mask;
9134 Mask.reserve(NumElts);
9135 for (unsigned i = 0; i != NumElts; ++i) {
9136 SDValue Arg = N1.getOperand(i);
9137 if (Arg.getOpcode() == ISD::UNDEF) {
9138 Mask.push_back(Arg);
9141 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Arg);
9144 if (!C->getAPIntValue().isSignBit() &&
9145 !C->getAPIntValue().isMaxSignedValue())
9147 Mask.push_back(DAG.getConstant(~C->getAPIntValue(), EltVT));
9149 N1 = DAG.getNode(ISD::BUILD_VECTOR, N1.getDebugLoc(), VT,
9151 if (!N1.use_empty()) {
9152 unsigned Bits = EltVT.getSizeInBits();
9154 for (unsigned i = 0; i != NumElts; ++i)
9155 Mask.push_back(DAG.getConstant(APInt::getAllOnesValue(Bits), EltVT));
9156 SDValue NewMask = DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
9157 VT, &Mask[0], NumElts);
9158 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9159 DAG.getNode(ISD::XOR, N->getDebugLoc(), VT,
9167 /// PerformMulCombine - Optimize a single multiply with constant into two
9168 /// in order to implement it with two cheaper instructions, e.g.
9169 /// LEA + SHL, LEA + LEA.
9170 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9171 TargetLowering::DAGCombinerInfo &DCI) {
9172 if (DAG.getMachineFunction().
9173 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
9176 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9179 EVT VT = N->getValueType(0);
9183 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9186 uint64_t MulAmt = C->getZExtValue();
9187 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9190 uint64_t MulAmt1 = 0;
9191 uint64_t MulAmt2 = 0;
9192 if ((MulAmt % 9) == 0) {
9194 MulAmt2 = MulAmt / 9;
9195 } else if ((MulAmt % 5) == 0) {
9197 MulAmt2 = MulAmt / 5;
9198 } else if ((MulAmt % 3) == 0) {
9200 MulAmt2 = MulAmt / 3;
9203 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9204 DebugLoc DL = N->getDebugLoc();
9206 if (isPowerOf2_64(MulAmt2) &&
9207 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9208 // If second multiplifer is pow2, issue it first. We want the multiply by
9209 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9211 std::swap(MulAmt1, MulAmt2);
9214 if (isPowerOf2_64(MulAmt1))
9215 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9216 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9218 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9219 DAG.getConstant(MulAmt1, VT));
9221 if (isPowerOf2_64(MulAmt2))
9222 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9223 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9225 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9226 DAG.getConstant(MulAmt2, VT));
9228 // Do not add new nodes to DAG combiner worklist.
9229 DCI.CombineTo(N, NewMul, false);
9234 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9235 SDValue N0 = N->getOperand(0);
9236 SDValue N1 = N->getOperand(1);
9237 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9238 EVT VT = N0.getValueType();
9240 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9241 // since the result of setcc_c is all zero's or all ones.
9242 if (N1C && N0.getOpcode() == ISD::AND &&
9243 N0.getOperand(1).getOpcode() == ISD::Constant) {
9244 SDValue N00 = N0.getOperand(0);
9245 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9246 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9247 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9248 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9249 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9250 APInt ShAmt = N1C->getAPIntValue();
9251 Mask = Mask.shl(ShAmt);
9253 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9254 N00, DAG.getConstant(Mask, VT));
9261 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9263 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9264 const X86Subtarget *Subtarget) {
9265 EVT VT = N->getValueType(0);
9266 if (!VT.isVector() && VT.isInteger() &&
9267 N->getOpcode() == ISD::SHL)
9268 return PerformSHLCombine(N, DAG);
9270 // On X86 with SSE2 support, we can transform this to a vector shift if
9271 // all elements are shifted by the same amount. We can't do this in legalize
9272 // because the a constant vector is typically transformed to a constant pool
9273 // so we have no knowledge of the shift amount.
9274 if (!Subtarget->hasSSE2())
9277 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9280 SDValue ShAmtOp = N->getOperand(1);
9281 EVT EltVT = VT.getVectorElementType();
9282 DebugLoc DL = N->getDebugLoc();
9283 SDValue BaseShAmt = SDValue();
9284 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9285 unsigned NumElts = VT.getVectorNumElements();
9287 for (; i != NumElts; ++i) {
9288 SDValue Arg = ShAmtOp.getOperand(i);
9289 if (Arg.getOpcode() == ISD::UNDEF) continue;
9293 for (; i != NumElts; ++i) {
9294 SDValue Arg = ShAmtOp.getOperand(i);
9295 if (Arg.getOpcode() == ISD::UNDEF) continue;
9296 if (Arg != BaseShAmt) {
9300 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9301 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9302 SDValue InVec = ShAmtOp.getOperand(0);
9303 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9304 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9306 for (; i != NumElts; ++i) {
9307 SDValue Arg = InVec.getOperand(i);
9308 if (Arg.getOpcode() == ISD::UNDEF) continue;
9312 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9313 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9314 unsigned SplatIdx= cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9315 if (C->getZExtValue() == SplatIdx)
9316 BaseShAmt = InVec.getOperand(1);
9319 if (BaseShAmt.getNode() == 0)
9320 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9321 DAG.getIntPtrConstant(0));
9325 // The shift amount is an i32.
9326 if (EltVT.bitsGT(MVT::i32))
9327 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9328 else if (EltVT.bitsLT(MVT::i32))
9329 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9331 // The shift amount is identical so we can do a vector shift.
9332 SDValue ValOp = N->getOperand(0);
9333 switch (N->getOpcode()) {
9335 llvm_unreachable("Unknown shift opcode!");
9338 if (VT == MVT::v2i64)
9339 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9340 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9342 if (VT == MVT::v4i32)
9343 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9344 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9346 if (VT == MVT::v8i16)
9347 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9348 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9352 if (VT == MVT::v4i32)
9353 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9354 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9356 if (VT == MVT::v8i16)
9357 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9358 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9362 if (VT == MVT::v2i64)
9363 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9364 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9366 if (VT == MVT::v4i32)
9367 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9368 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9370 if (VT == MVT::v8i16)
9371 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9372 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9379 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9380 const X86Subtarget *Subtarget) {
9381 EVT VT = N->getValueType(0);
9382 if (VT != MVT::i64 || !Subtarget->is64Bit())
9385 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9386 SDValue N0 = N->getOperand(0);
9387 SDValue N1 = N->getOperand(1);
9388 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9390 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9393 SDValue ShAmt0 = N0.getOperand(1);
9394 if (ShAmt0.getValueType() != MVT::i8)
9396 SDValue ShAmt1 = N1.getOperand(1);
9397 if (ShAmt1.getValueType() != MVT::i8)
9399 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9400 ShAmt0 = ShAmt0.getOperand(0);
9401 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9402 ShAmt1 = ShAmt1.getOperand(0);
9404 DebugLoc DL = N->getDebugLoc();
9405 unsigned Opc = X86ISD::SHLD;
9406 SDValue Op0 = N0.getOperand(0);
9407 SDValue Op1 = N1.getOperand(0);
9408 if (ShAmt0.getOpcode() == ISD::SUB) {
9410 std::swap(Op0, Op1);
9411 std::swap(ShAmt0, ShAmt1);
9414 if (ShAmt1.getOpcode() == ISD::SUB) {
9415 SDValue Sum = ShAmt1.getOperand(0);
9416 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9417 if (SumC->getSExtValue() == 64 &&
9418 ShAmt1.getOperand(1) == ShAmt0)
9419 return DAG.getNode(Opc, DL, VT,
9421 DAG.getNode(ISD::TRUNCATE, DL,
9424 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9425 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9427 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9428 return DAG.getNode(Opc, DL, VT,
9429 N0.getOperand(0), N1.getOperand(0),
9430 DAG.getNode(ISD::TRUNCATE, DL,
9437 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9438 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9439 const X86Subtarget *Subtarget) {
9440 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9441 // the FP state in cases where an emms may be missing.
9442 // A preferable solution to the general problem is to figure out the right
9443 // places to insert EMMS. This qualifies as a quick hack.
9445 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9446 StoreSDNode *St = cast<StoreSDNode>(N);
9447 EVT VT = St->getValue().getValueType();
9448 if (VT.getSizeInBits() != 64)
9451 const Function *F = DAG.getMachineFunction().getFunction();
9452 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9453 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9454 && Subtarget->hasSSE2();
9455 if ((VT.isVector() ||
9456 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9457 isa<LoadSDNode>(St->getValue()) &&
9458 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9459 St->getChain().hasOneUse() && !St->isVolatile()) {
9460 SDNode* LdVal = St->getValue().getNode();
9462 int TokenFactorIndex = -1;
9463 SmallVector<SDValue, 8> Ops;
9464 SDNode* ChainVal = St->getChain().getNode();
9465 // Must be a store of a load. We currently handle two cases: the load
9466 // is a direct child, and it's under an intervening TokenFactor. It is
9467 // possible to dig deeper under nested TokenFactors.
9468 if (ChainVal == LdVal)
9469 Ld = cast<LoadSDNode>(St->getChain());
9470 else if (St->getValue().hasOneUse() &&
9471 ChainVal->getOpcode() == ISD::TokenFactor) {
9472 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9473 if (ChainVal->getOperand(i).getNode() == LdVal) {
9474 TokenFactorIndex = i;
9475 Ld = cast<LoadSDNode>(St->getValue());
9477 Ops.push_back(ChainVal->getOperand(i));
9481 if (!Ld || !ISD::isNormalLoad(Ld))
9484 // If this is not the MMX case, i.e. we are just turning i64 load/store
9485 // into f64 load/store, avoid the transformation if there are multiple
9486 // uses of the loaded value.
9487 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9490 DebugLoc LdDL = Ld->getDebugLoc();
9491 DebugLoc StDL = N->getDebugLoc();
9492 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9493 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9495 if (Subtarget->is64Bit() || F64IsLegal) {
9496 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9497 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9498 Ld->getBasePtr(), Ld->getSrcValue(),
9499 Ld->getSrcValueOffset(), Ld->isVolatile(),
9500 Ld->isNonTemporal(), Ld->getAlignment());
9501 SDValue NewChain = NewLd.getValue(1);
9502 if (TokenFactorIndex != -1) {
9503 Ops.push_back(NewChain);
9504 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9507 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9508 St->getSrcValue(), St->getSrcValueOffset(),
9509 St->isVolatile(), St->isNonTemporal(),
9510 St->getAlignment());
9513 // Otherwise, lower to two pairs of 32-bit loads / stores.
9514 SDValue LoAddr = Ld->getBasePtr();
9515 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9516 DAG.getConstant(4, MVT::i32));
9518 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9519 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9520 Ld->isVolatile(), Ld->isNonTemporal(),
9521 Ld->getAlignment());
9522 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9523 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9524 Ld->isVolatile(), Ld->isNonTemporal(),
9525 MinAlign(Ld->getAlignment(), 4));
9527 SDValue NewChain = LoLd.getValue(1);
9528 if (TokenFactorIndex != -1) {
9529 Ops.push_back(LoLd);
9530 Ops.push_back(HiLd);
9531 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9535 LoAddr = St->getBasePtr();
9536 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9537 DAG.getConstant(4, MVT::i32));
9539 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9540 St->getSrcValue(), St->getSrcValueOffset(),
9541 St->isVolatile(), St->isNonTemporal(),
9542 St->getAlignment());
9543 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9545 St->getSrcValueOffset() + 4,
9547 St->isNonTemporal(),
9548 MinAlign(St->getAlignment(), 4));
9549 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9554 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9555 /// X86ISD::FXOR nodes.
9556 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9557 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9558 // F[X]OR(0.0, x) -> x
9559 // F[X]OR(x, 0.0) -> x
9560 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9561 if (C->getValueAPF().isPosZero())
9562 return N->getOperand(1);
9563 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9564 if (C->getValueAPF().isPosZero())
9565 return N->getOperand(0);
9569 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9570 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9571 // FAND(0.0, x) -> 0.0
9572 // FAND(x, 0.0) -> 0.0
9573 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9574 if (C->getValueAPF().isPosZero())
9575 return N->getOperand(0);
9576 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9577 if (C->getValueAPF().isPosZero())
9578 return N->getOperand(1);
9582 static SDValue PerformBTCombine(SDNode *N,
9584 TargetLowering::DAGCombinerInfo &DCI) {
9585 // BT ignores high bits in the bit index operand.
9586 SDValue Op1 = N->getOperand(1);
9587 if (Op1.hasOneUse()) {
9588 unsigned BitWidth = Op1.getValueSizeInBits();
9589 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9590 APInt KnownZero, KnownOne;
9591 TargetLowering::TargetLoweringOpt TLO(DAG);
9592 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9593 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9594 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9595 DCI.CommitTargetLoweringOpt(TLO);
9600 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9601 SDValue Op = N->getOperand(0);
9602 if (Op.getOpcode() == ISD::BIT_CONVERT)
9603 Op = Op.getOperand(0);
9604 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9605 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9606 VT.getVectorElementType().getSizeInBits() ==
9607 OpVT.getVectorElementType().getSizeInBits()) {
9608 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9613 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9614 // Locked instructions, in turn, have implicit fence semantics (all memory
9615 // operations are flushed before issuing the locked instruction, and the
9616 // are not buffered), so we can fold away the common pattern of
9617 // fence-atomic-fence.
9618 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9619 SDValue atomic = N->getOperand(0);
9620 switch (atomic.getOpcode()) {
9621 case ISD::ATOMIC_CMP_SWAP:
9622 case ISD::ATOMIC_SWAP:
9623 case ISD::ATOMIC_LOAD_ADD:
9624 case ISD::ATOMIC_LOAD_SUB:
9625 case ISD::ATOMIC_LOAD_AND:
9626 case ISD::ATOMIC_LOAD_OR:
9627 case ISD::ATOMIC_LOAD_XOR:
9628 case ISD::ATOMIC_LOAD_NAND:
9629 case ISD::ATOMIC_LOAD_MIN:
9630 case ISD::ATOMIC_LOAD_MAX:
9631 case ISD::ATOMIC_LOAD_UMIN:
9632 case ISD::ATOMIC_LOAD_UMAX:
9638 SDValue fence = atomic.getOperand(0);
9639 if (fence.getOpcode() != ISD::MEMBARRIER)
9642 switch (atomic.getOpcode()) {
9643 case ISD::ATOMIC_CMP_SWAP:
9644 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9645 atomic.getOperand(1), atomic.getOperand(2),
9646 atomic.getOperand(3));
9647 case ISD::ATOMIC_SWAP:
9648 case ISD::ATOMIC_LOAD_ADD:
9649 case ISD::ATOMIC_LOAD_SUB:
9650 case ISD::ATOMIC_LOAD_AND:
9651 case ISD::ATOMIC_LOAD_OR:
9652 case ISD::ATOMIC_LOAD_XOR:
9653 case ISD::ATOMIC_LOAD_NAND:
9654 case ISD::ATOMIC_LOAD_MIN:
9655 case ISD::ATOMIC_LOAD_MAX:
9656 case ISD::ATOMIC_LOAD_UMIN:
9657 case ISD::ATOMIC_LOAD_UMAX:
9658 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9659 atomic.getOperand(1), atomic.getOperand(2));
9665 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9666 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9667 // (and (i32 x86isd::setcc_carry), 1)
9668 // This eliminates the zext. This transformation is necessary because
9669 // ISD::SETCC is always legalized to i8.
9670 DebugLoc dl = N->getDebugLoc();
9671 SDValue N0 = N->getOperand(0);
9672 EVT VT = N->getValueType(0);
9673 if (N0.getOpcode() == ISD::AND &&
9675 N0.getOperand(0).hasOneUse()) {
9676 SDValue N00 = N0.getOperand(0);
9677 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9679 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9680 if (!C || C->getZExtValue() != 1)
9682 return DAG.getNode(ISD::AND, dl, VT,
9683 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9684 N00.getOperand(0), N00.getOperand(1)),
9685 DAG.getConstant(1, VT));
9691 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9692 DAGCombinerInfo &DCI) const {
9693 SelectionDAG &DAG = DCI.DAG;
9694 switch (N->getOpcode()) {
9696 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9697 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9698 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9699 case ISD::AND: return PerformANDCombine(N, DAG, DCI);
9700 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9703 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9704 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9705 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9707 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9708 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9709 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9710 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9711 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9712 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9718 //===----------------------------------------------------------------------===//
9719 // X86 Inline Assembly Support
9720 //===----------------------------------------------------------------------===//
9722 static bool LowerToBSwap(CallInst *CI) {
9723 // FIXME: this should verify that we are targetting a 486 or better. If not,
9724 // we will turn this bswap into something that will be lowered to logical ops
9725 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9726 // so don't worry about this.
9728 // Verify this is a simple bswap.
9729 if (CI->getNumOperands() != 2 ||
9730 CI->getType() != CI->getOperand(1)->getType() ||
9731 !CI->getType()->isIntegerTy())
9734 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9735 if (!Ty || Ty->getBitWidth() % 16 != 0)
9738 // Okay, we can do this xform, do so now.
9739 const Type *Tys[] = { Ty };
9740 Module *M = CI->getParent()->getParent()->getParent();
9741 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9743 Value *Op = CI->getOperand(1);
9744 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9746 CI->replaceAllUsesWith(Op);
9747 CI->eraseFromParent();
9751 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9752 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9753 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9755 std::string AsmStr = IA->getAsmString();
9757 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9758 SmallVector<StringRef, 4> AsmPieces;
9759 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9761 switch (AsmPieces.size()) {
9762 default: return false;
9764 AsmStr = AsmPieces[0];
9766 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9769 if (AsmPieces.size() == 2 &&
9770 (AsmPieces[0] == "bswap" ||
9771 AsmPieces[0] == "bswapq" ||
9772 AsmPieces[0] == "bswapl") &&
9773 (AsmPieces[1] == "$0" ||
9774 AsmPieces[1] == "${0:q}")) {
9775 // No need to check constraints, nothing other than the equivalent of
9776 // "=r,0" would be valid here.
9777 return LowerToBSwap(CI);
9779 // rorw $$8, ${0:w} --> llvm.bswap.i16
9780 if (CI->getType()->isIntegerTy(16) &&
9781 AsmPieces.size() == 3 &&
9782 AsmPieces[0] == "rorw" &&
9783 AsmPieces[1] == "$$8," &&
9784 AsmPieces[2] == "${0:w}" &&
9785 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9786 return LowerToBSwap(CI);
9790 if (CI->getType()->isIntegerTy(64) &&
9791 Constraints.size() >= 2 &&
9792 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9793 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9794 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9795 SmallVector<StringRef, 4> Words;
9796 SplitString(AsmPieces[0], Words, " \t");
9797 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9799 SplitString(AsmPieces[1], Words, " \t");
9800 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9802 SplitString(AsmPieces[2], Words, " \t,");
9803 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9804 Words[2] == "%edx") {
9805 return LowerToBSwap(CI);
9817 /// getConstraintType - Given a constraint letter, return the type of
9818 /// constraint it is for this target.
9819 X86TargetLowering::ConstraintType
9820 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9821 if (Constraint.size() == 1) {
9822 switch (Constraint[0]) {
9834 return C_RegisterClass;
9842 return TargetLowering::getConstraintType(Constraint);
9845 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9846 /// with another that has more specific requirements based on the type of the
9847 /// corresponding operand.
9848 const char *X86TargetLowering::
9849 LowerXConstraint(EVT ConstraintVT) const {
9850 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9851 // 'f' like normal targets.
9852 if (ConstraintVT.isFloatingPoint()) {
9853 if (Subtarget->hasSSE2())
9855 if (Subtarget->hasSSE1())
9859 return TargetLowering::LowerXConstraint(ConstraintVT);
9862 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9863 /// vector. If it is invalid, don't add anything to Ops.
9864 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9867 std::vector<SDValue>&Ops,
9868 SelectionDAG &DAG) const {
9869 SDValue Result(0, 0);
9871 switch (Constraint) {
9874 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9875 if (C->getZExtValue() <= 31) {
9876 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9882 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9883 if (C->getZExtValue() <= 63) {
9884 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9890 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9891 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9892 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9898 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9899 if (C->getZExtValue() <= 255) {
9900 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9906 // 32-bit signed value
9907 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9908 const ConstantInt *CI = C->getConstantIntValue();
9909 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9910 C->getSExtValue())) {
9911 // Widen to 64 bits here to get it sign extended.
9912 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9915 // FIXME gcc accepts some relocatable values here too, but only in certain
9916 // memory models; it's complicated.
9921 // 32-bit unsigned value
9922 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9923 const ConstantInt *CI = C->getConstantIntValue();
9924 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9925 C->getZExtValue())) {
9926 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9930 // FIXME gcc accepts some relocatable values here too, but only in certain
9931 // memory models; it's complicated.
9935 // Literal immediates are always ok.
9936 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9937 // Widen to 64 bits here to get it sign extended.
9938 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9942 // If we are in non-pic codegen mode, we allow the address of a global (with
9943 // an optional displacement) to be used with 'i'.
9944 GlobalAddressSDNode *GA = 0;
9947 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9949 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9950 Offset += GA->getOffset();
9952 } else if (Op.getOpcode() == ISD::ADD) {
9953 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9954 Offset += C->getZExtValue();
9955 Op = Op.getOperand(0);
9958 } else if (Op.getOpcode() == ISD::SUB) {
9959 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9960 Offset += -C->getZExtValue();
9961 Op = Op.getOperand(0);
9966 // Otherwise, this isn't something we can handle, reject it.
9970 GlobalValue *GV = GA->getGlobal();
9971 // If we require an extra load to get this address, as in PIC mode, we
9973 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9974 getTargetMachine())))
9978 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9980 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9986 if (Result.getNode()) {
9987 Ops.push_back(Result);
9990 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9994 std::vector<unsigned> X86TargetLowering::
9995 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9997 if (Constraint.size() == 1) {
9998 // FIXME: not handling fp-stack yet!
9999 switch (Constraint[0]) { // GCC X86 Constraint Letters
10000 default: break; // Unknown constraint letter
10001 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
10002 if (Subtarget->is64Bit()) {
10003 if (VT == MVT::i32)
10004 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
10005 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
10006 X86::R10D,X86::R11D,X86::R12D,
10007 X86::R13D,X86::R14D,X86::R15D,
10008 X86::EBP, X86::ESP, 0);
10009 else if (VT == MVT::i16)
10010 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10011 X86::SI, X86::DI, X86::R8W,X86::R9W,
10012 X86::R10W,X86::R11W,X86::R12W,
10013 X86::R13W,X86::R14W,X86::R15W,
10014 X86::BP, X86::SP, 0);
10015 else if (VT == MVT::i8)
10016 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10017 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10018 X86::R10B,X86::R11B,X86::R12B,
10019 X86::R13B,X86::R14B,X86::R15B,
10020 X86::BPL, X86::SPL, 0);
10022 else if (VT == MVT::i64)
10023 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10024 X86::RSI, X86::RDI, X86::R8, X86::R9,
10025 X86::R10, X86::R11, X86::R12,
10026 X86::R13, X86::R14, X86::R15,
10027 X86::RBP, X86::RSP, 0);
10031 // 32-bit fallthrough
10032 case 'Q': // Q_REGS
10033 if (VT == MVT::i32)
10034 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10035 else if (VT == MVT::i16)
10036 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10037 else if (VT == MVT::i8)
10038 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10039 else if (VT == MVT::i64)
10040 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10045 return std::vector<unsigned>();
10048 std::pair<unsigned, const TargetRegisterClass*>
10049 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10051 // First, see if this is a constraint that directly corresponds to an LLVM
10053 if (Constraint.size() == 1) {
10054 // GCC Constraint Letters
10055 switch (Constraint[0]) {
10057 case 'r': // GENERAL_REGS
10058 case 'l': // INDEX_REGS
10060 return std::make_pair(0U, X86::GR8RegisterClass);
10061 if (VT == MVT::i16)
10062 return std::make_pair(0U, X86::GR16RegisterClass);
10063 if (VT == MVT::i32 || !Subtarget->is64Bit())
10064 return std::make_pair(0U, X86::GR32RegisterClass);
10065 return std::make_pair(0U, X86::GR64RegisterClass);
10066 case 'R': // LEGACY_REGS
10068 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10069 if (VT == MVT::i16)
10070 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10071 if (VT == MVT::i32 || !Subtarget->is64Bit())
10072 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10073 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10074 case 'f': // FP Stack registers.
10075 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10076 // value to the correct fpstack register class.
10077 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10078 return std::make_pair(0U, X86::RFP32RegisterClass);
10079 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10080 return std::make_pair(0U, X86::RFP64RegisterClass);
10081 return std::make_pair(0U, X86::RFP80RegisterClass);
10082 case 'y': // MMX_REGS if MMX allowed.
10083 if (!Subtarget->hasMMX()) break;
10084 return std::make_pair(0U, X86::VR64RegisterClass);
10085 case 'Y': // SSE_REGS if SSE2 allowed
10086 if (!Subtarget->hasSSE2()) break;
10088 case 'x': // SSE_REGS if SSE1 allowed
10089 if (!Subtarget->hasSSE1()) break;
10091 switch (VT.getSimpleVT().SimpleTy) {
10093 // Scalar SSE types.
10096 return std::make_pair(0U, X86::FR32RegisterClass);
10099 return std::make_pair(0U, X86::FR64RegisterClass);
10107 return std::make_pair(0U, X86::VR128RegisterClass);
10113 // Use the default implementation in TargetLowering to convert the register
10114 // constraint into a member of a register class.
10115 std::pair<unsigned, const TargetRegisterClass*> Res;
10116 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10118 // Not found as a standard register?
10119 if (Res.second == 0) {
10120 // Map st(0) -> st(7) -> ST0
10121 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10122 tolower(Constraint[1]) == 's' &&
10123 tolower(Constraint[2]) == 't' &&
10124 Constraint[3] == '(' &&
10125 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10126 Constraint[5] == ')' &&
10127 Constraint[6] == '}') {
10129 Res.first = X86::ST0+Constraint[4]-'0';
10130 Res.second = X86::RFP80RegisterClass;
10134 // GCC allows "st(0)" to be called just plain "st".
10135 if (StringRef("{st}").equals_lower(Constraint)) {
10136 Res.first = X86::ST0;
10137 Res.second = X86::RFP80RegisterClass;
10142 if (StringRef("{flags}").equals_lower(Constraint)) {
10143 Res.first = X86::EFLAGS;
10144 Res.second = X86::CCRRegisterClass;
10148 // 'A' means EAX + EDX.
10149 if (Constraint == "A") {
10150 Res.first = X86::EAX;
10151 Res.second = X86::GR32_ADRegisterClass;
10157 // Otherwise, check to see if this is a register class of the wrong value
10158 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10159 // turn into {ax},{dx}.
10160 if (Res.second->hasType(VT))
10161 return Res; // Correct type already, nothing to do.
10163 // All of the single-register GCC register classes map their values onto
10164 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10165 // really want an 8-bit or 32-bit register, map to the appropriate register
10166 // class and return the appropriate register.
10167 if (Res.second == X86::GR16RegisterClass) {
10168 if (VT == MVT::i8) {
10169 unsigned DestReg = 0;
10170 switch (Res.first) {
10172 case X86::AX: DestReg = X86::AL; break;
10173 case X86::DX: DestReg = X86::DL; break;
10174 case X86::CX: DestReg = X86::CL; break;
10175 case X86::BX: DestReg = X86::BL; break;
10178 Res.first = DestReg;
10179 Res.second = X86::GR8RegisterClass;
10181 } else if (VT == MVT::i32) {
10182 unsigned DestReg = 0;
10183 switch (Res.first) {
10185 case X86::AX: DestReg = X86::EAX; break;
10186 case X86::DX: DestReg = X86::EDX; break;
10187 case X86::CX: DestReg = X86::ECX; break;
10188 case X86::BX: DestReg = X86::EBX; break;
10189 case X86::SI: DestReg = X86::ESI; break;
10190 case X86::DI: DestReg = X86::EDI; break;
10191 case X86::BP: DestReg = X86::EBP; break;
10192 case X86::SP: DestReg = X86::ESP; break;
10195 Res.first = DestReg;
10196 Res.second = X86::GR32RegisterClass;
10198 } else if (VT == MVT::i64) {
10199 unsigned DestReg = 0;
10200 switch (Res.first) {
10202 case X86::AX: DestReg = X86::RAX; break;
10203 case X86::DX: DestReg = X86::RDX; break;
10204 case X86::CX: DestReg = X86::RCX; break;
10205 case X86::BX: DestReg = X86::RBX; break;
10206 case X86::SI: DestReg = X86::RSI; break;
10207 case X86::DI: DestReg = X86::RDI; break;
10208 case X86::BP: DestReg = X86::RBP; break;
10209 case X86::SP: DestReg = X86::RSP; break;
10212 Res.first = DestReg;
10213 Res.second = X86::GR64RegisterClass;
10216 } else if (Res.second == X86::FR32RegisterClass ||
10217 Res.second == X86::FR64RegisterClass ||
10218 Res.second == X86::VR128RegisterClass) {
10219 // Handle references to XMM physical registers that got mapped into the
10220 // wrong class. This can happen with constraints like {xmm0} where the
10221 // target independent register mapper will just pick the first match it can
10222 // find, ignoring the required type.
10223 if (VT == MVT::f32)
10224 Res.second = X86::FR32RegisterClass;
10225 else if (VT == MVT::f64)
10226 Res.second = X86::FR64RegisterClass;
10227 else if (X86::VR128RegisterClass->hasType(VT))
10228 Res.second = X86::VR128RegisterClass;
10234 //===----------------------------------------------------------------------===//
10235 // X86 Widen vector type
10236 //===----------------------------------------------------------------------===//
10238 /// getWidenVectorType: given a vector type, returns the type to widen
10239 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10240 /// If there is no vector type that we want to widen to, returns MVT::Other
10241 /// When and where to widen is target dependent based on the cost of
10242 /// scalarizing vs using the wider vector type.
10244 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10245 assert(VT.isVector());
10246 if (isTypeLegal(VT))
10249 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10250 // type based on element type. This would speed up our search (though
10251 // it may not be worth it since the size of the list is relatively
10253 EVT EltVT = VT.getVectorElementType();
10254 unsigned NElts = VT.getVectorNumElements();
10256 // On X86, it make sense to widen any vector wider than 1
10260 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10261 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10262 EVT SVT = (MVT::SimpleValueType)nVT;
10264 if (isTypeLegal(SVT) &&
10265 SVT.getVectorElementType() == EltVT &&
10266 SVT.getVectorNumElements() > NElts)