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 X8632_MachoTargetObjectFile();
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::SHL);
994 setTargetDAGCombine(ISD::SRA);
995 setTargetDAGCombine(ISD::SRL);
996 setTargetDAGCombine(ISD::OR);
997 setTargetDAGCombine(ISD::STORE);
998 setTargetDAGCombine(ISD::MEMBARRIER);
999 setTargetDAGCombine(ISD::ZERO_EXTEND);
1000 if (Subtarget->is64Bit())
1001 setTargetDAGCombine(ISD::MUL);
1003 computeRegisterProperties();
1005 // FIXME: These should be based on subtarget info. Plus, the values should
1006 // be smaller when we are in optimizing for size mode.
1007 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1008 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1009 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1010 setPrefLoopAlignment(16);
1011 benefitFromCodePlacementOpt = true;
1015 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1020 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1021 /// the desired ByVal argument alignment.
1022 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1025 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1026 if (VTy->getBitWidth() == 128)
1028 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1029 unsigned EltAlign = 0;
1030 getMaxByValAlign(ATy->getElementType(), EltAlign);
1031 if (EltAlign > MaxAlign)
1032 MaxAlign = EltAlign;
1033 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1034 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1035 unsigned EltAlign = 0;
1036 getMaxByValAlign(STy->getElementType(i), EltAlign);
1037 if (EltAlign > MaxAlign)
1038 MaxAlign = EltAlign;
1046 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1047 /// function arguments in the caller parameter area. For X86, aggregates
1048 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1049 /// are at 4-byte boundaries.
1050 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1051 if (Subtarget->is64Bit()) {
1052 // Max of 8 and alignment of type.
1053 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1060 if (Subtarget->hasSSE1())
1061 getMaxByValAlign(Ty, Align);
1065 /// getOptimalMemOpType - Returns the target specific optimal type for load
1066 /// and store operations as a result of memset, memcpy, and memmove
1067 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1070 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1071 bool isSrcConst, bool isSrcStr,
1072 SelectionDAG &DAG) const {
1073 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1074 // linux. This is because the stack realignment code can't handle certain
1075 // cases like PR2962. This should be removed when PR2962 is fixed.
1076 const Function *F = DAG.getMachineFunction().getFunction();
1077 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1078 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1079 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1081 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1084 if (Subtarget->is64Bit() && Size >= 8)
1089 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1090 /// current function. The returned value is a member of the
1091 /// MachineJumpTableInfo::JTEntryKind enum.
1092 unsigned X86TargetLowering::getJumpTableEncoding() const {
1093 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1095 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1096 Subtarget->isPICStyleGOT())
1097 return MachineJumpTableInfo::EK_Custom32;
1099 // Otherwise, use the normal jump table encoding heuristics.
1100 return TargetLowering::getJumpTableEncoding();
1103 /// getPICBaseSymbol - Return the X86-32 PIC base.
1105 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1106 MCContext &Ctx) const {
1107 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1108 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1109 Twine(MF->getFunctionNumber())+"$pb");
1114 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1115 const MachineBasicBlock *MBB,
1116 unsigned uid,MCContext &Ctx) const{
1117 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1118 Subtarget->isPICStyleGOT());
1119 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1121 return X86MCTargetExpr::Create(MBB->getSymbol(Ctx),
1122 X86MCTargetExpr::GOTOFF, Ctx);
1125 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1127 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1128 SelectionDAG &DAG) const {
1129 if (!Subtarget->is64Bit())
1130 // This doesn't have DebugLoc associated with it, but is not really the
1131 // same as a Register.
1132 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1137 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1138 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1140 const MCExpr *X86TargetLowering::
1141 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1142 MCContext &Ctx) const {
1143 // X86-64 uses RIP relative addressing based on the jump table label.
1144 if (Subtarget->isPICStyleRIPRel())
1145 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1147 // Otherwise, the reference is relative to the PIC base.
1148 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1151 /// getFunctionAlignment - Return the Log2 alignment of this function.
1152 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1153 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1156 //===----------------------------------------------------------------------===//
1157 // Return Value Calling Convention Implementation
1158 //===----------------------------------------------------------------------===//
1160 #include "X86GenCallingConv.inc"
1163 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1164 const SmallVectorImpl<EVT> &OutTys,
1165 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1166 SelectionDAG &DAG) {
1167 SmallVector<CCValAssign, 16> RVLocs;
1168 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1169 RVLocs, *DAG.getContext());
1170 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1174 X86TargetLowering::LowerReturn(SDValue Chain,
1175 CallingConv::ID CallConv, bool isVarArg,
1176 const SmallVectorImpl<ISD::OutputArg> &Outs,
1177 DebugLoc dl, SelectionDAG &DAG) {
1179 SmallVector<CCValAssign, 16> RVLocs;
1180 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1181 RVLocs, *DAG.getContext());
1182 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1184 // Add the regs to the liveout set for the function.
1185 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1186 for (unsigned i = 0; i != RVLocs.size(); ++i)
1187 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1188 MRI.addLiveOut(RVLocs[i].getLocReg());
1192 SmallVector<SDValue, 6> RetOps;
1193 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1194 // Operand #1 = Bytes To Pop
1195 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1197 // Copy the result values into the output registers.
1198 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1199 CCValAssign &VA = RVLocs[i];
1200 assert(VA.isRegLoc() && "Can only return in registers!");
1201 SDValue ValToCopy = Outs[i].Val;
1203 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1204 // the RET instruction and handled by the FP Stackifier.
1205 if (VA.getLocReg() == X86::ST0 ||
1206 VA.getLocReg() == X86::ST1) {
1207 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1208 // change the value to the FP stack register class.
1209 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1210 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1211 RetOps.push_back(ValToCopy);
1212 // Don't emit a copytoreg.
1216 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1217 // which is returned in RAX / RDX.
1218 if (Subtarget->is64Bit()) {
1219 EVT ValVT = ValToCopy.getValueType();
1220 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1221 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1222 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1223 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1227 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1228 Flag = Chain.getValue(1);
1231 // The x86-64 ABI for returning structs by value requires that we copy
1232 // the sret argument into %rax for the return. We saved the argument into
1233 // a virtual register in the entry block, so now we copy the value out
1235 if (Subtarget->is64Bit() &&
1236 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1237 MachineFunction &MF = DAG.getMachineFunction();
1238 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1239 unsigned Reg = FuncInfo->getSRetReturnReg();
1241 Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
1242 FuncInfo->setSRetReturnReg(Reg);
1244 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1246 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1247 Flag = Chain.getValue(1);
1249 // RAX now acts like a return value.
1250 MRI.addLiveOut(X86::RAX);
1253 RetOps[0] = Chain; // Update chain.
1255 // Add the flag if we have it.
1257 RetOps.push_back(Flag);
1259 return DAG.getNode(X86ISD::RET_FLAG, dl,
1260 MVT::Other, &RetOps[0], RetOps.size());
1263 /// LowerCallResult - Lower the result values of a call into the
1264 /// appropriate copies out of appropriate physical registers.
1267 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1268 CallingConv::ID CallConv, bool isVarArg,
1269 const SmallVectorImpl<ISD::InputArg> &Ins,
1270 DebugLoc dl, SelectionDAG &DAG,
1271 SmallVectorImpl<SDValue> &InVals) {
1273 // Assign locations to each value returned by this call.
1274 SmallVector<CCValAssign, 16> RVLocs;
1275 bool Is64Bit = Subtarget->is64Bit();
1276 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1277 RVLocs, *DAG.getContext());
1278 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1280 // Copy all of the result registers out of their specified physreg.
1281 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1282 CCValAssign &VA = RVLocs[i];
1283 EVT CopyVT = VA.getValVT();
1285 // If this is x86-64, and we disabled SSE, we can't return FP values
1286 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1287 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1288 llvm_report_error("SSE register return with SSE disabled");
1291 // If this is a call to a function that returns an fp value on the floating
1292 // point stack, but where we prefer to use the value in xmm registers, copy
1293 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1294 if ((VA.getLocReg() == X86::ST0 ||
1295 VA.getLocReg() == X86::ST1) &&
1296 isScalarFPTypeInSSEReg(VA.getValVT())) {
1301 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1302 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1303 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1304 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1305 MVT::v2i64, InFlag).getValue(1);
1306 Val = Chain.getValue(0);
1307 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1308 Val, DAG.getConstant(0, MVT::i64));
1310 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1311 MVT::i64, InFlag).getValue(1);
1312 Val = Chain.getValue(0);
1314 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1316 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1317 CopyVT, InFlag).getValue(1);
1318 Val = Chain.getValue(0);
1320 InFlag = Chain.getValue(2);
1322 if (CopyVT != VA.getValVT()) {
1323 // Round the F80 the right size, which also moves to the appropriate xmm
1325 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1326 // This truncation won't change the value.
1327 DAG.getIntPtrConstant(1));
1330 InVals.push_back(Val);
1337 //===----------------------------------------------------------------------===//
1338 // C & StdCall & Fast Calling Convention implementation
1339 //===----------------------------------------------------------------------===//
1340 // StdCall calling convention seems to be standard for many Windows' API
1341 // routines and around. It differs from C calling convention just a little:
1342 // callee should clean up the stack, not caller. Symbols should be also
1343 // decorated in some fancy way :) It doesn't support any vector arguments.
1344 // For info on fast calling convention see Fast Calling Convention (tail call)
1345 // implementation LowerX86_32FastCCCallTo.
1347 /// CallIsStructReturn - Determines whether a call uses struct return
1349 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1353 return Outs[0].Flags.isSRet();
1356 /// ArgsAreStructReturn - Determines whether a function uses struct
1357 /// return semantics.
1359 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1363 return Ins[0].Flags.isSRet();
1366 /// IsCalleePop - Determines whether the callee is required to pop its
1367 /// own arguments. Callee pop is necessary to support tail calls.
1368 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1372 switch (CallingConv) {
1375 case CallingConv::X86_StdCall:
1376 return !Subtarget->is64Bit();
1377 case CallingConv::X86_FastCall:
1378 return !Subtarget->is64Bit();
1379 case CallingConv::Fast:
1380 return GuaranteedTailCallOpt;
1384 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1385 /// given CallingConvention value.
1386 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1387 if (Subtarget->is64Bit()) {
1388 if (Subtarget->isTargetWin64())
1389 return CC_X86_Win64_C;
1394 if (CC == CallingConv::X86_FastCall)
1395 return CC_X86_32_FastCall;
1396 else if (CC == CallingConv::Fast)
1397 return CC_X86_32_FastCC;
1402 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1403 /// by "Src" to address "Dst" with size and alignment information specified by
1404 /// the specific parameter attribute. The copy will be passed as a byval
1405 /// function parameter.
1407 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1408 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1410 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1411 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1412 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1415 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1416 /// a tailcall target by changing its ABI.
1417 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1418 return GuaranteedTailCallOpt && CC == CallingConv::Fast;
1422 X86TargetLowering::LowerMemArgument(SDValue Chain,
1423 CallingConv::ID CallConv,
1424 const SmallVectorImpl<ISD::InputArg> &Ins,
1425 DebugLoc dl, SelectionDAG &DAG,
1426 const CCValAssign &VA,
1427 MachineFrameInfo *MFI,
1429 // Create the nodes corresponding to a load from this parameter slot.
1430 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1431 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1432 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1435 // If value is passed by pointer we have address passed instead of the value
1437 if (VA.getLocInfo() == CCValAssign::Indirect)
1438 ValVT = VA.getLocVT();
1440 ValVT = VA.getValVT();
1442 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1443 // changed with more analysis.
1444 // In case of tail call optimization mark all arguments mutable. Since they
1445 // could be overwritten by lowering of arguments in case of a tail call.
1446 if (Flags.isByVal()) {
1447 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1448 VA.getLocMemOffset(), isImmutable, false);
1449 return DAG.getFrameIndex(FI, getPointerTy());
1451 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1452 VA.getLocMemOffset(), isImmutable, false);
1453 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1454 return DAG.getLoad(ValVT, dl, Chain, FIN,
1455 PseudoSourceValue::getFixedStack(FI), 0,
1461 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1462 CallingConv::ID CallConv,
1464 const SmallVectorImpl<ISD::InputArg> &Ins,
1467 SmallVectorImpl<SDValue> &InVals) {
1469 MachineFunction &MF = DAG.getMachineFunction();
1470 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1472 const Function* Fn = MF.getFunction();
1473 if (Fn->hasExternalLinkage() &&
1474 Subtarget->isTargetCygMing() &&
1475 Fn->getName() == "main")
1476 FuncInfo->setForceFramePointer(true);
1478 MachineFrameInfo *MFI = MF.getFrameInfo();
1479 bool Is64Bit = Subtarget->is64Bit();
1480 bool IsWin64 = Subtarget->isTargetWin64();
1482 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1483 "Var args not supported with calling convention fastcc");
1485 // Assign locations to all of the incoming arguments.
1486 SmallVector<CCValAssign, 16> ArgLocs;
1487 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1488 ArgLocs, *DAG.getContext());
1489 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1491 unsigned LastVal = ~0U;
1493 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1494 CCValAssign &VA = ArgLocs[i];
1495 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1497 assert(VA.getValNo() != LastVal &&
1498 "Don't support value assigned to multiple locs yet");
1499 LastVal = VA.getValNo();
1501 if (VA.isRegLoc()) {
1502 EVT RegVT = VA.getLocVT();
1503 TargetRegisterClass *RC = NULL;
1504 if (RegVT == MVT::i32)
1505 RC = X86::GR32RegisterClass;
1506 else if (Is64Bit && RegVT == MVT::i64)
1507 RC = X86::GR64RegisterClass;
1508 else if (RegVT == MVT::f32)
1509 RC = X86::FR32RegisterClass;
1510 else if (RegVT == MVT::f64)
1511 RC = X86::FR64RegisterClass;
1512 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1513 RC = X86::VR128RegisterClass;
1514 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1515 RC = X86::VR64RegisterClass;
1517 llvm_unreachable("Unknown argument type!");
1519 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1520 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1522 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1523 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1525 if (VA.getLocInfo() == CCValAssign::SExt)
1526 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1527 DAG.getValueType(VA.getValVT()));
1528 else if (VA.getLocInfo() == CCValAssign::ZExt)
1529 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1530 DAG.getValueType(VA.getValVT()));
1531 else if (VA.getLocInfo() == CCValAssign::BCvt)
1532 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1534 if (VA.isExtInLoc()) {
1535 // Handle MMX values passed in XMM regs.
1536 if (RegVT.isVector()) {
1537 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1538 ArgValue, DAG.getConstant(0, MVT::i64));
1539 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1541 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1544 assert(VA.isMemLoc());
1545 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1548 // If value is passed via pointer - do a load.
1549 if (VA.getLocInfo() == CCValAssign::Indirect)
1550 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1553 InVals.push_back(ArgValue);
1556 // The x86-64 ABI for returning structs by value requires that we copy
1557 // the sret argument into %rax for the return. Save the argument into
1558 // a virtual register so that we can access it from the return points.
1559 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1560 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1561 unsigned Reg = FuncInfo->getSRetReturnReg();
1563 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1564 FuncInfo->setSRetReturnReg(Reg);
1566 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1567 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1570 unsigned StackSize = CCInfo.getNextStackOffset();
1571 // Align stack specially for tail calls.
1572 if (FuncIsMadeTailCallSafe(CallConv))
1573 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1575 // If the function takes variable number of arguments, make a frame index for
1576 // the start of the first vararg value... for expansion of llvm.va_start.
1578 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1579 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1582 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1584 // FIXME: We should really autogenerate these arrays
1585 static const unsigned GPR64ArgRegsWin64[] = {
1586 X86::RCX, X86::RDX, X86::R8, X86::R9
1588 static const unsigned XMMArgRegsWin64[] = {
1589 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1591 static const unsigned GPR64ArgRegs64Bit[] = {
1592 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1594 static const unsigned XMMArgRegs64Bit[] = {
1595 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1596 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1598 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1601 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1602 GPR64ArgRegs = GPR64ArgRegsWin64;
1603 XMMArgRegs = XMMArgRegsWin64;
1605 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1606 GPR64ArgRegs = GPR64ArgRegs64Bit;
1607 XMMArgRegs = XMMArgRegs64Bit;
1609 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1611 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1614 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1615 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1616 "SSE register cannot be used when SSE is disabled!");
1617 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1618 "SSE register cannot be used when SSE is disabled!");
1619 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1620 // Kernel mode asks for SSE to be disabled, so don't push them
1622 TotalNumXMMRegs = 0;
1624 // For X86-64, if there are vararg parameters that are passed via
1625 // registers, then we must store them to their spots on the stack so they
1626 // may be loaded by deferencing the result of va_next.
1627 VarArgsGPOffset = NumIntRegs * 8;
1628 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1629 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1630 TotalNumXMMRegs * 16, 16,
1633 // Store the integer parameter registers.
1634 SmallVector<SDValue, 8> MemOps;
1635 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1636 unsigned Offset = VarArgsGPOffset;
1637 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1638 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1639 DAG.getIntPtrConstant(Offset));
1640 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1641 X86::GR64RegisterClass);
1642 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1644 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1645 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1646 Offset, false, false, 0);
1647 MemOps.push_back(Store);
1651 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1652 // Now store the XMM (fp + vector) parameter registers.
1653 SmallVector<SDValue, 11> SaveXMMOps;
1654 SaveXMMOps.push_back(Chain);
1656 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1657 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1658 SaveXMMOps.push_back(ALVal);
1660 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1661 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1663 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1664 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1665 X86::VR128RegisterClass);
1666 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1667 SaveXMMOps.push_back(Val);
1669 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1671 &SaveXMMOps[0], SaveXMMOps.size()));
1674 if (!MemOps.empty())
1675 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1676 &MemOps[0], MemOps.size());
1680 // Some CCs need callee pop.
1681 if (IsCalleePop(isVarArg, CallConv)) {
1682 BytesToPopOnReturn = StackSize; // Callee pops everything.
1684 BytesToPopOnReturn = 0; // Callee pops nothing.
1685 // If this is an sret function, the return should pop the hidden pointer.
1686 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1687 BytesToPopOnReturn = 4;
1691 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1692 if (CallConv == CallingConv::X86_FastCall)
1693 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1696 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1702 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1703 SDValue StackPtr, SDValue Arg,
1704 DebugLoc dl, SelectionDAG &DAG,
1705 const CCValAssign &VA,
1706 ISD::ArgFlagsTy Flags) {
1707 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1708 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1709 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1710 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1711 if (Flags.isByVal()) {
1712 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1714 return DAG.getStore(Chain, dl, Arg, PtrOff,
1715 PseudoSourceValue::getStack(), LocMemOffset,
1719 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1720 /// optimization is performed and it is required.
1722 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1723 SDValue &OutRetAddr, SDValue Chain,
1724 bool IsTailCall, bool Is64Bit,
1725 int FPDiff, DebugLoc dl) {
1726 // Adjust the Return address stack slot.
1727 EVT VT = getPointerTy();
1728 OutRetAddr = getReturnAddressFrameIndex(DAG);
1730 // Load the "old" Return address.
1731 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1732 return SDValue(OutRetAddr.getNode(), 1);
1735 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1736 /// optimization is performed and it is required (FPDiff!=0).
1738 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1739 SDValue Chain, SDValue RetAddrFrIdx,
1740 bool Is64Bit, int FPDiff, DebugLoc dl) {
1741 // Store the return address to the appropriate stack slot.
1742 if (!FPDiff) return Chain;
1743 // Calculate the new stack slot for the return address.
1744 int SlotSize = Is64Bit ? 8 : 4;
1745 int NewReturnAddrFI =
1746 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, true,false);
1747 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1748 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1749 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1750 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1756 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1757 CallingConv::ID CallConv, bool isVarArg,
1759 const SmallVectorImpl<ISD::OutputArg> &Outs,
1760 const SmallVectorImpl<ISD::InputArg> &Ins,
1761 DebugLoc dl, SelectionDAG &DAG,
1762 SmallVectorImpl<SDValue> &InVals) {
1763 MachineFunction &MF = DAG.getMachineFunction();
1764 bool Is64Bit = Subtarget->is64Bit();
1765 bool IsStructRet = CallIsStructReturn(Outs);
1766 bool IsSibcall = false;
1769 // Check if it's really possible to do a tail call.
1770 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
1773 // Sibcalls are automatically detected tailcalls which do not require
1775 if (!GuaranteedTailCallOpt && isTailCall)
1782 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1783 "Var args not supported with calling convention fastcc");
1785 // Analyze operands of the call, assigning locations to each operand.
1786 SmallVector<CCValAssign, 16> ArgLocs;
1787 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1788 ArgLocs, *DAG.getContext());
1789 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1791 // Get a count of how many bytes are to be pushed on the stack.
1792 unsigned NumBytes = CCInfo.getNextStackOffset();
1794 // This is a sibcall. The memory operands are available in caller's
1795 // own caller's stack.
1797 else if (GuaranteedTailCallOpt && CallConv == CallingConv::Fast)
1798 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1801 if (isTailCall && !IsSibcall) {
1802 // Lower arguments at fp - stackoffset + fpdiff.
1803 unsigned NumBytesCallerPushed =
1804 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1805 FPDiff = NumBytesCallerPushed - NumBytes;
1807 // Set the delta of movement of the returnaddr stackslot.
1808 // But only set if delta is greater than previous delta.
1809 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1810 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1814 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1816 SDValue RetAddrFrIdx;
1817 // Load return adress for tail calls.
1818 if (isTailCall && FPDiff)
1819 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1820 Is64Bit, FPDiff, dl);
1822 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1823 SmallVector<SDValue, 8> MemOpChains;
1826 // Walk the register/memloc assignments, inserting copies/loads. In the case
1827 // of tail call optimization arguments are handle later.
1828 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1829 CCValAssign &VA = ArgLocs[i];
1830 EVT RegVT = VA.getLocVT();
1831 SDValue Arg = Outs[i].Val;
1832 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1833 bool isByVal = Flags.isByVal();
1835 // Promote the value if needed.
1836 switch (VA.getLocInfo()) {
1837 default: llvm_unreachable("Unknown loc info!");
1838 case CCValAssign::Full: break;
1839 case CCValAssign::SExt:
1840 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1842 case CCValAssign::ZExt:
1843 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1845 case CCValAssign::AExt:
1846 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1847 // Special case: passing MMX values in XMM registers.
1848 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1849 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1850 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1852 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1854 case CCValAssign::BCvt:
1855 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1857 case CCValAssign::Indirect: {
1858 // Store the argument.
1859 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1860 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1861 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1862 PseudoSourceValue::getFixedStack(FI), 0,
1869 if (VA.isRegLoc()) {
1870 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1871 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1872 assert(VA.isMemLoc());
1873 if (StackPtr.getNode() == 0)
1874 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1875 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1876 dl, DAG, VA, Flags));
1880 if (!MemOpChains.empty())
1881 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1882 &MemOpChains[0], MemOpChains.size());
1884 // Build a sequence of copy-to-reg nodes chained together with token chain
1885 // and flag operands which copy the outgoing args into registers.
1887 // Tail call byval lowering might overwrite argument registers so in case of
1888 // tail call optimization the copies to registers are lowered later.
1890 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1891 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1892 RegsToPass[i].second, InFlag);
1893 InFlag = Chain.getValue(1);
1896 if (Subtarget->isPICStyleGOT()) {
1897 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1900 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1901 DAG.getNode(X86ISD::GlobalBaseReg,
1902 DebugLoc::getUnknownLoc(),
1905 InFlag = Chain.getValue(1);
1907 // If we are tail calling and generating PIC/GOT style code load the
1908 // address of the callee into ECX. The value in ecx is used as target of
1909 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1910 // for tail calls on PIC/GOT architectures. Normally we would just put the
1911 // address of GOT into ebx and then call target@PLT. But for tail calls
1912 // ebx would be restored (since ebx is callee saved) before jumping to the
1915 // Note: The actual moving to ECX is done further down.
1916 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1917 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1918 !G->getGlobal()->hasProtectedVisibility())
1919 Callee = LowerGlobalAddress(Callee, DAG);
1920 else if (isa<ExternalSymbolSDNode>(Callee))
1921 Callee = LowerExternalSymbol(Callee, DAG);
1925 if (Is64Bit && isVarArg) {
1926 // From AMD64 ABI document:
1927 // For calls that may call functions that use varargs or stdargs
1928 // (prototype-less calls or calls to functions containing ellipsis (...) in
1929 // the declaration) %al is used as hidden argument to specify the number
1930 // of SSE registers used. The contents of %al do not need to match exactly
1931 // the number of registers, but must be an ubound on the number of SSE
1932 // registers used and is in the range 0 - 8 inclusive.
1934 // FIXME: Verify this on Win64
1935 // Count the number of XMM registers allocated.
1936 static const unsigned XMMArgRegs[] = {
1937 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1938 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1940 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1941 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1942 && "SSE registers cannot be used when SSE is disabled");
1944 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1945 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1946 InFlag = Chain.getValue(1);
1950 // For tail calls lower the arguments to the 'real' stack slot.
1952 // Force all the incoming stack arguments to be loaded from the stack
1953 // before any new outgoing arguments are stored to the stack, because the
1954 // outgoing stack slots may alias the incoming argument stack slots, and
1955 // the alias isn't otherwise explicit. This is slightly more conservative
1956 // than necessary, because it means that each store effectively depends
1957 // on every argument instead of just those arguments it would clobber.
1958 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1960 SmallVector<SDValue, 8> MemOpChains2;
1963 // Do not flag preceeding copytoreg stuff together with the following stuff.
1965 if (GuaranteedTailCallOpt) {
1966 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1967 CCValAssign &VA = ArgLocs[i];
1970 assert(VA.isMemLoc());
1971 SDValue Arg = Outs[i].Val;
1972 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1973 // Create frame index.
1974 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1975 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1976 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1977 FIN = DAG.getFrameIndex(FI, getPointerTy());
1979 if (Flags.isByVal()) {
1980 // Copy relative to framepointer.
1981 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1982 if (StackPtr.getNode() == 0)
1983 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1985 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1987 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1991 // Store relative to framepointer.
1992 MemOpChains2.push_back(
1993 DAG.getStore(ArgChain, dl, Arg, FIN,
1994 PseudoSourceValue::getFixedStack(FI), 0,
2000 if (!MemOpChains2.empty())
2001 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2002 &MemOpChains2[0], MemOpChains2.size());
2004 // Copy arguments to their registers.
2005 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2006 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2007 RegsToPass[i].second, InFlag);
2008 InFlag = Chain.getValue(1);
2012 // Store the return address to the appropriate stack slot.
2013 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2017 bool WasGlobalOrExternal = false;
2018 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2019 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2020 // In the 64-bit large code model, we have to make all calls
2021 // through a register, since the call instruction's 32-bit
2022 // pc-relative offset may not be large enough to hold the whole
2024 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2025 WasGlobalOrExternal = true;
2026 // If the callee is a GlobalAddress node (quite common, every direct call
2027 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2030 // We should use extra load for direct calls to dllimported functions in
2032 GlobalValue *GV = G->getGlobal();
2033 if (!GV->hasDLLImportLinkage()) {
2034 unsigned char OpFlags = 0;
2036 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2037 // external symbols most go through the PLT in PIC mode. If the symbol
2038 // has hidden or protected visibility, or if it is static or local, then
2039 // we don't need to use the PLT - we can directly call it.
2040 if (Subtarget->isTargetELF() &&
2041 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2042 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2043 OpFlags = X86II::MO_PLT;
2044 } else if (Subtarget->isPICStyleStubAny() &&
2045 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2046 Subtarget->getDarwinVers() < 9) {
2047 // PC-relative references to external symbols should go through $stub,
2048 // unless we're building with the leopard linker or later, which
2049 // automatically synthesizes these stubs.
2050 OpFlags = X86II::MO_DARWIN_STUB;
2053 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2054 G->getOffset(), OpFlags);
2056 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2057 WasGlobalOrExternal = true;
2058 unsigned char OpFlags = 0;
2060 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2061 // symbols should go through the PLT.
2062 if (Subtarget->isTargetELF() &&
2063 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2064 OpFlags = X86II::MO_PLT;
2065 } else if (Subtarget->isPICStyleStubAny() &&
2066 Subtarget->getDarwinVers() < 9) {
2067 // PC-relative references to external symbols should go through $stub,
2068 // unless we're building with the leopard linker or later, which
2069 // automatically synthesizes these stubs.
2070 OpFlags = X86II::MO_DARWIN_STUB;
2073 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2077 if (isTailCall && !WasGlobalOrExternal) {
2078 // Force the address into a (call preserved) caller-saved register since
2079 // tailcall must happen after callee-saved registers are poped.
2080 // FIXME: Give it a special register class that contains caller-saved
2081 // register instead?
2082 unsigned TCReg = Is64Bit ? X86::R11 : X86::EAX;
2083 Chain = DAG.getCopyToReg(Chain, dl,
2084 DAG.getRegister(TCReg, getPointerTy()),
2086 Callee = DAG.getRegister(TCReg, getPointerTy());
2089 // Returns a chain & a flag for retval copy to use.
2090 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2091 SmallVector<SDValue, 8> Ops;
2093 if (!IsSibcall && isTailCall) {
2094 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2095 DAG.getIntPtrConstant(0, true), InFlag);
2096 InFlag = Chain.getValue(1);
2099 Ops.push_back(Chain);
2100 Ops.push_back(Callee);
2103 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2105 // Add argument registers to the end of the list so that they are known live
2107 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2108 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2109 RegsToPass[i].second.getValueType()));
2111 // Add an implicit use GOT pointer in EBX.
2112 if (!isTailCall && Subtarget->isPICStyleGOT())
2113 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2115 // Add an implicit use of AL for x86 vararg functions.
2116 if (Is64Bit && isVarArg)
2117 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2119 if (InFlag.getNode())
2120 Ops.push_back(InFlag);
2123 // If this is the first return lowered for this function, add the regs
2124 // to the liveout set for the function.
2125 if (MF.getRegInfo().liveout_empty()) {
2126 SmallVector<CCValAssign, 16> RVLocs;
2127 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2129 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2130 for (unsigned i = 0; i != RVLocs.size(); ++i)
2131 if (RVLocs[i].isRegLoc())
2132 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2135 assert(((Callee.getOpcode() == ISD::Register &&
2136 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2137 cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
2138 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2139 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2140 "Expecting a global address, external symbol, or scratch register");
2142 return DAG.getNode(X86ISD::TC_RETURN, dl,
2143 NodeTys, &Ops[0], Ops.size());
2146 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2147 InFlag = Chain.getValue(1);
2149 // Create the CALLSEQ_END node.
2150 unsigned NumBytesForCalleeToPush;
2151 if (IsCalleePop(isVarArg, CallConv))
2152 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2153 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2154 // If this is a call to a struct-return function, the callee
2155 // pops the hidden struct pointer, so we have to push it back.
2156 // This is common for Darwin/X86, Linux & Mingw32 targets.
2157 NumBytesForCalleeToPush = 4;
2159 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2161 // Returns a flag for retval copy to use.
2163 Chain = DAG.getCALLSEQ_END(Chain,
2164 DAG.getIntPtrConstant(NumBytes, true),
2165 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2168 InFlag = Chain.getValue(1);
2171 // Handle result values, copying them out of physregs into vregs that we
2173 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2174 Ins, dl, DAG, InVals);
2178 //===----------------------------------------------------------------------===//
2179 // Fast Calling Convention (tail call) implementation
2180 //===----------------------------------------------------------------------===//
2182 // Like std call, callee cleans arguments, convention except that ECX is
2183 // reserved for storing the tail called function address. Only 2 registers are
2184 // free for argument passing (inreg). Tail call optimization is performed
2186 // * tailcallopt is enabled
2187 // * caller/callee are fastcc
2188 // On X86_64 architecture with GOT-style position independent code only local
2189 // (within module) calls are supported at the moment.
2190 // To keep the stack aligned according to platform abi the function
2191 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2192 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2193 // If a tail called function callee has more arguments than the caller the
2194 // caller needs to make sure that there is room to move the RETADDR to. This is
2195 // achieved by reserving an area the size of the argument delta right after the
2196 // original REtADDR, but before the saved framepointer or the spilled registers
2197 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2209 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2210 /// for a 16 byte align requirement.
2211 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2212 SelectionDAG& DAG) {
2213 MachineFunction &MF = DAG.getMachineFunction();
2214 const TargetMachine &TM = MF.getTarget();
2215 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2216 unsigned StackAlignment = TFI.getStackAlignment();
2217 uint64_t AlignMask = StackAlignment - 1;
2218 int64_t Offset = StackSize;
2219 uint64_t SlotSize = TD->getPointerSize();
2220 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2221 // Number smaller than 12 so just add the difference.
2222 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2224 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2225 Offset = ((~AlignMask) & Offset) + StackAlignment +
2226 (StackAlignment-SlotSize);
2231 /// MatchingStackOffset - Return true if the given stack call argument is
2232 /// already available in the same position (relatively) of the caller's
2233 /// incoming argument stack.
2235 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2236 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2237 const X86InstrInfo *TII) {
2239 if (Arg.getOpcode() == ISD::CopyFromReg) {
2240 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2241 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2243 MachineInstr *Def = MRI->getVRegDef(VR);
2246 if (!Flags.isByVal()) {
2247 if (!TII->isLoadFromStackSlot(Def, FI))
2250 unsigned Opcode = Def->getOpcode();
2251 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2252 Def->getOperand(1).isFI()) {
2253 FI = Def->getOperand(1).getIndex();
2254 if (MFI->getObjectSize(FI) != Flags.getByValSize())
2260 LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg);
2263 SDValue Ptr = Ld->getBasePtr();
2264 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2267 FI = FINode->getIndex();
2270 if (!MFI->isFixedObjectIndex(FI))
2272 return Offset == MFI->getObjectOffset(FI);
2275 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2276 /// for tail call optimization. Targets which want to do tail call
2277 /// optimization should implement this function.
2279 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2280 CallingConv::ID CalleeCC,
2282 const SmallVectorImpl<ISD::OutputArg> &Outs,
2283 const SmallVectorImpl<ISD::InputArg> &Ins,
2284 SelectionDAG& DAG) const {
2285 if (CalleeCC != CallingConv::Fast &&
2286 CalleeCC != CallingConv::C)
2289 // If -tailcallopt is specified, make fastcc functions tail-callable.
2290 const Function *CallerF = DAG.getMachineFunction().getFunction();
2291 if (GuaranteedTailCallOpt) {
2292 if (CalleeCC == CallingConv::Fast &&
2293 CallerF->getCallingConv() == CalleeCC)
2298 // Look for obvious safe cases to perform tail call optimization that does not
2299 // requite ABI changes. This is what gcc calls sibcall.
2301 // Do not tail call optimize vararg calls for now.
2305 // If the callee takes no arguments then go on to check the results of the
2307 if (!Outs.empty()) {
2308 // Check if stack adjustment is needed. For now, do not do this if any
2309 // argument is passed on the stack.
2310 SmallVector<CCValAssign, 16> ArgLocs;
2311 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2312 ArgLocs, *DAG.getContext());
2313 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2314 if (CCInfo.getNextStackOffset()) {
2315 MachineFunction &MF = DAG.getMachineFunction();
2316 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2318 if (Subtarget->isTargetWin64())
2319 // Win64 ABI has additional complications.
2322 // Check if the arguments are already laid out in the right way as
2323 // the caller's fixed stack objects.
2324 MachineFrameInfo *MFI = MF.getFrameInfo();
2325 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2326 const X86InstrInfo *TII =
2327 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2328 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2329 CCValAssign &VA = ArgLocs[i];
2330 EVT RegVT = VA.getLocVT();
2331 SDValue Arg = Outs[i].Val;
2332 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2333 if (VA.getLocInfo() == CCValAssign::Indirect)
2335 if (!VA.isRegLoc()) {
2336 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2348 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2350 DenseMap<const Value *, unsigned> &vm,
2351 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2352 DenseMap<const AllocaInst *, int> &am
2354 , SmallSet<Instruction*, 8> &cil
2357 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2365 //===----------------------------------------------------------------------===//
2366 // Other Lowering Hooks
2367 //===----------------------------------------------------------------------===//
2370 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2371 MachineFunction &MF = DAG.getMachineFunction();
2372 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2373 int ReturnAddrIndex = FuncInfo->getRAIndex();
2375 if (ReturnAddrIndex == 0) {
2376 // Set up a frame object for the return address.
2377 uint64_t SlotSize = TD->getPointerSize();
2378 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2380 FuncInfo->setRAIndex(ReturnAddrIndex);
2383 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2387 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2388 bool hasSymbolicDisplacement) {
2389 // Offset should fit into 32 bit immediate field.
2390 if (!isInt32(Offset))
2393 // If we don't have a symbolic displacement - we don't have any extra
2395 if (!hasSymbolicDisplacement)
2398 // FIXME: Some tweaks might be needed for medium code model.
2399 if (M != CodeModel::Small && M != CodeModel::Kernel)
2402 // For small code model we assume that latest object is 16MB before end of 31
2403 // bits boundary. We may also accept pretty large negative constants knowing
2404 // that all objects are in the positive half of address space.
2405 if (M == CodeModel::Small && Offset < 16*1024*1024)
2408 // For kernel code model we know that all object resist in the negative half
2409 // of 32bits address space. We may not accept negative offsets, since they may
2410 // be just off and we may accept pretty large positive ones.
2411 if (M == CodeModel::Kernel && Offset > 0)
2417 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2418 /// specific condition code, returning the condition code and the LHS/RHS of the
2419 /// comparison to make.
2420 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2421 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2423 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2424 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2425 // X > -1 -> X == 0, jump !sign.
2426 RHS = DAG.getConstant(0, RHS.getValueType());
2427 return X86::COND_NS;
2428 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2429 // X < 0 -> X == 0, jump on sign.
2431 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2433 RHS = DAG.getConstant(0, RHS.getValueType());
2434 return X86::COND_LE;
2438 switch (SetCCOpcode) {
2439 default: llvm_unreachable("Invalid integer condition!");
2440 case ISD::SETEQ: return X86::COND_E;
2441 case ISD::SETGT: return X86::COND_G;
2442 case ISD::SETGE: return X86::COND_GE;
2443 case ISD::SETLT: return X86::COND_L;
2444 case ISD::SETLE: return X86::COND_LE;
2445 case ISD::SETNE: return X86::COND_NE;
2446 case ISD::SETULT: return X86::COND_B;
2447 case ISD::SETUGT: return X86::COND_A;
2448 case ISD::SETULE: return X86::COND_BE;
2449 case ISD::SETUGE: return X86::COND_AE;
2453 // First determine if it is required or is profitable to flip the operands.
2455 // If LHS is a foldable load, but RHS is not, flip the condition.
2456 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2457 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2458 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2459 std::swap(LHS, RHS);
2462 switch (SetCCOpcode) {
2468 std::swap(LHS, RHS);
2472 // On a floating point condition, the flags are set as follows:
2474 // 0 | 0 | 0 | X > Y
2475 // 0 | 0 | 1 | X < Y
2476 // 1 | 0 | 0 | X == Y
2477 // 1 | 1 | 1 | unordered
2478 switch (SetCCOpcode) {
2479 default: llvm_unreachable("Condcode should be pre-legalized away");
2481 case ISD::SETEQ: return X86::COND_E;
2482 case ISD::SETOLT: // flipped
2484 case ISD::SETGT: return X86::COND_A;
2485 case ISD::SETOLE: // flipped
2487 case ISD::SETGE: return X86::COND_AE;
2488 case ISD::SETUGT: // flipped
2490 case ISD::SETLT: return X86::COND_B;
2491 case ISD::SETUGE: // flipped
2493 case ISD::SETLE: return X86::COND_BE;
2495 case ISD::SETNE: return X86::COND_NE;
2496 case ISD::SETUO: return X86::COND_P;
2497 case ISD::SETO: return X86::COND_NP;
2499 case ISD::SETUNE: return X86::COND_INVALID;
2503 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2504 /// code. Current x86 isa includes the following FP cmov instructions:
2505 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2506 static bool hasFPCMov(unsigned X86CC) {
2522 /// isFPImmLegal - Returns true if the target can instruction select the
2523 /// specified FP immediate natively. If false, the legalizer will
2524 /// materialize the FP immediate as a load from a constant pool.
2525 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2526 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2527 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2533 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2534 /// the specified range (L, H].
2535 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2536 return (Val < 0) || (Val >= Low && Val < Hi);
2539 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2540 /// specified value.
2541 static bool isUndefOrEqual(int Val, int CmpVal) {
2542 if (Val < 0 || Val == CmpVal)
2547 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2548 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2549 /// the second operand.
2550 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2551 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2552 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2553 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2554 return (Mask[0] < 2 && Mask[1] < 2);
2558 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2559 SmallVector<int, 8> M;
2561 return ::isPSHUFDMask(M, N->getValueType(0));
2564 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2565 /// is suitable for input to PSHUFHW.
2566 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2567 if (VT != MVT::v8i16)
2570 // Lower quadword copied in order or undef.
2571 for (int i = 0; i != 4; ++i)
2572 if (Mask[i] >= 0 && Mask[i] != i)
2575 // Upper quadword shuffled.
2576 for (int i = 4; i != 8; ++i)
2577 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2583 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2584 SmallVector<int, 8> M;
2586 return ::isPSHUFHWMask(M, N->getValueType(0));
2589 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2590 /// is suitable for input to PSHUFLW.
2591 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2592 if (VT != MVT::v8i16)
2595 // Upper quadword copied in order.
2596 for (int i = 4; i != 8; ++i)
2597 if (Mask[i] >= 0 && Mask[i] != i)
2600 // Lower quadword shuffled.
2601 for (int i = 0; i != 4; ++i)
2608 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2609 SmallVector<int, 8> M;
2611 return ::isPSHUFLWMask(M, N->getValueType(0));
2614 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2615 /// is suitable for input to PALIGNR.
2616 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2618 int i, e = VT.getVectorNumElements();
2620 // Do not handle v2i64 / v2f64 shuffles with palignr.
2621 if (e < 4 || !hasSSSE3)
2624 for (i = 0; i != e; ++i)
2628 // All undef, not a palignr.
2632 // Determine if it's ok to perform a palignr with only the LHS, since we
2633 // don't have access to the actual shuffle elements to see if RHS is undef.
2634 bool Unary = Mask[i] < (int)e;
2635 bool NeedsUnary = false;
2637 int s = Mask[i] - i;
2639 // Check the rest of the elements to see if they are consecutive.
2640 for (++i; i != e; ++i) {
2645 Unary = Unary && (m < (int)e);
2646 NeedsUnary = NeedsUnary || (m < s);
2648 if (NeedsUnary && !Unary)
2650 if (Unary && m != ((s+i) & (e-1)))
2652 if (!Unary && m != (s+i))
2658 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2659 SmallVector<int, 8> M;
2661 return ::isPALIGNRMask(M, N->getValueType(0), true);
2664 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2665 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2666 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2667 int NumElems = VT.getVectorNumElements();
2668 if (NumElems != 2 && NumElems != 4)
2671 int Half = NumElems / 2;
2672 for (int i = 0; i < Half; ++i)
2673 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2675 for (int i = Half; i < NumElems; ++i)
2676 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2682 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2683 SmallVector<int, 8> M;
2685 return ::isSHUFPMask(M, N->getValueType(0));
2688 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2689 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2690 /// half elements to come from vector 1 (which would equal the dest.) and
2691 /// the upper half to come from vector 2.
2692 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2693 int NumElems = VT.getVectorNumElements();
2695 if (NumElems != 2 && NumElems != 4)
2698 int Half = NumElems / 2;
2699 for (int i = 0; i < Half; ++i)
2700 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2702 for (int i = Half; i < NumElems; ++i)
2703 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2708 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2709 SmallVector<int, 8> M;
2711 return isCommutedSHUFPMask(M, N->getValueType(0));
2714 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2715 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2716 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2717 if (N->getValueType(0).getVectorNumElements() != 4)
2720 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2721 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2722 isUndefOrEqual(N->getMaskElt(1), 7) &&
2723 isUndefOrEqual(N->getMaskElt(2), 2) &&
2724 isUndefOrEqual(N->getMaskElt(3), 3);
2727 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2728 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2730 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2731 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2736 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2737 isUndefOrEqual(N->getMaskElt(1), 3) &&
2738 isUndefOrEqual(N->getMaskElt(2), 2) &&
2739 isUndefOrEqual(N->getMaskElt(3), 3);
2742 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2743 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2744 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2745 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2747 if (NumElems != 2 && NumElems != 4)
2750 for (unsigned i = 0; i < NumElems/2; ++i)
2751 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2754 for (unsigned i = NumElems/2; i < NumElems; ++i)
2755 if (!isUndefOrEqual(N->getMaskElt(i), i))
2761 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2762 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2763 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2764 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2766 if (NumElems != 2 && NumElems != 4)
2769 for (unsigned i = 0; i < NumElems/2; ++i)
2770 if (!isUndefOrEqual(N->getMaskElt(i), i))
2773 for (unsigned i = 0; i < NumElems/2; ++i)
2774 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2780 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2781 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2782 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2783 bool V2IsSplat = false) {
2784 int NumElts = VT.getVectorNumElements();
2785 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2788 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2790 int BitI1 = Mask[i+1];
2791 if (!isUndefOrEqual(BitI, j))
2794 if (!isUndefOrEqual(BitI1, NumElts))
2797 if (!isUndefOrEqual(BitI1, j + NumElts))
2804 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2805 SmallVector<int, 8> M;
2807 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2810 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2811 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2812 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2813 bool V2IsSplat = false) {
2814 int NumElts = VT.getVectorNumElements();
2815 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2818 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2820 int BitI1 = Mask[i+1];
2821 if (!isUndefOrEqual(BitI, j + NumElts/2))
2824 if (isUndefOrEqual(BitI1, NumElts))
2827 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2834 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2835 SmallVector<int, 8> M;
2837 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2840 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2841 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2843 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2844 int NumElems = VT.getVectorNumElements();
2845 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2848 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2850 int BitI1 = Mask[i+1];
2851 if (!isUndefOrEqual(BitI, j))
2853 if (!isUndefOrEqual(BitI1, j))
2859 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2860 SmallVector<int, 8> M;
2862 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2865 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2866 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2868 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2869 int NumElems = VT.getVectorNumElements();
2870 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2873 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2875 int BitI1 = Mask[i+1];
2876 if (!isUndefOrEqual(BitI, j))
2878 if (!isUndefOrEqual(BitI1, j))
2884 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2885 SmallVector<int, 8> M;
2887 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2890 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2891 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2892 /// MOVSD, and MOVD, i.e. setting the lowest element.
2893 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2894 if (VT.getVectorElementType().getSizeInBits() < 32)
2897 int NumElts = VT.getVectorNumElements();
2899 if (!isUndefOrEqual(Mask[0], NumElts))
2902 for (int i = 1; i < NumElts; ++i)
2903 if (!isUndefOrEqual(Mask[i], i))
2909 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2910 SmallVector<int, 8> M;
2912 return ::isMOVLMask(M, N->getValueType(0));
2915 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2916 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2917 /// element of vector 2 and the other elements to come from vector 1 in order.
2918 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2919 bool V2IsSplat = false, bool V2IsUndef = false) {
2920 int NumOps = VT.getVectorNumElements();
2921 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2924 if (!isUndefOrEqual(Mask[0], 0))
2927 for (int i = 1; i < NumOps; ++i)
2928 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2929 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2930 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2936 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2937 bool V2IsUndef = false) {
2938 SmallVector<int, 8> M;
2940 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2943 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2944 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2945 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2946 if (N->getValueType(0).getVectorNumElements() != 4)
2949 // Expect 1, 1, 3, 3
2950 for (unsigned i = 0; i < 2; ++i) {
2951 int Elt = N->getMaskElt(i);
2952 if (Elt >= 0 && Elt != 1)
2957 for (unsigned i = 2; i < 4; ++i) {
2958 int Elt = N->getMaskElt(i);
2959 if (Elt >= 0 && Elt != 3)
2964 // Don't use movshdup if it can be done with a shufps.
2965 // FIXME: verify that matching u, u, 3, 3 is what we want.
2969 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2970 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2971 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2972 if (N->getValueType(0).getVectorNumElements() != 4)
2975 // Expect 0, 0, 2, 2
2976 for (unsigned i = 0; i < 2; ++i)
2977 if (N->getMaskElt(i) > 0)
2981 for (unsigned i = 2; i < 4; ++i) {
2982 int Elt = N->getMaskElt(i);
2983 if (Elt >= 0 && Elt != 2)
2988 // Don't use movsldup if it can be done with a shufps.
2992 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2993 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2994 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2995 int e = N->getValueType(0).getVectorNumElements() / 2;
2997 for (int i = 0; i < e; ++i)
2998 if (!isUndefOrEqual(N->getMaskElt(i), i))
3000 for (int i = 0; i < e; ++i)
3001 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3006 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3007 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3008 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3009 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3010 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3012 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3014 for (int i = 0; i < NumOperands; ++i) {
3015 int Val = SVOp->getMaskElt(NumOperands-i-1);
3016 if (Val < 0) Val = 0;
3017 if (Val >= NumOperands) Val -= NumOperands;
3019 if (i != NumOperands - 1)
3025 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3026 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3027 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3028 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3030 // 8 nodes, but we only care about the last 4.
3031 for (unsigned i = 7; i >= 4; --i) {
3032 int Val = SVOp->getMaskElt(i);
3041 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3042 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3043 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3044 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3046 // 8 nodes, but we only care about the first 4.
3047 for (int i = 3; i >= 0; --i) {
3048 int Val = SVOp->getMaskElt(i);
3057 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3058 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3059 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3060 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3061 EVT VVT = N->getValueType(0);
3062 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3066 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3067 Val = SVOp->getMaskElt(i);
3071 return (Val - i) * EltSize;
3074 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3076 bool X86::isZeroNode(SDValue Elt) {
3077 return ((isa<ConstantSDNode>(Elt) &&
3078 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3079 (isa<ConstantFPSDNode>(Elt) &&
3080 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3083 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3084 /// their permute mask.
3085 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3086 SelectionDAG &DAG) {
3087 EVT VT = SVOp->getValueType(0);
3088 unsigned NumElems = VT.getVectorNumElements();
3089 SmallVector<int, 8> MaskVec;
3091 for (unsigned i = 0; i != NumElems; ++i) {
3092 int idx = SVOp->getMaskElt(i);
3094 MaskVec.push_back(idx);
3095 else if (idx < (int)NumElems)
3096 MaskVec.push_back(idx + NumElems);
3098 MaskVec.push_back(idx - NumElems);
3100 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3101 SVOp->getOperand(0), &MaskVec[0]);
3104 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3105 /// the two vector operands have swapped position.
3106 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3107 unsigned NumElems = VT.getVectorNumElements();
3108 for (unsigned i = 0; i != NumElems; ++i) {
3112 else if (idx < (int)NumElems)
3113 Mask[i] = idx + NumElems;
3115 Mask[i] = idx - NumElems;
3119 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3120 /// match movhlps. The lower half elements should come from upper half of
3121 /// V1 (and in order), and the upper half elements should come from the upper
3122 /// half of V2 (and in order).
3123 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3124 if (Op->getValueType(0).getVectorNumElements() != 4)
3126 for (unsigned i = 0, e = 2; i != e; ++i)
3127 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3129 for (unsigned i = 2; i != 4; ++i)
3130 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3135 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3136 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3138 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3139 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3141 N = N->getOperand(0).getNode();
3142 if (!ISD::isNON_EXTLoad(N))
3145 *LD = cast<LoadSDNode>(N);
3149 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3150 /// match movlp{s|d}. The lower half elements should come from lower half of
3151 /// V1 (and in order), and the upper half elements should come from the upper
3152 /// half of V2 (and in order). And since V1 will become the source of the
3153 /// MOVLP, it must be either a vector load or a scalar load to vector.
3154 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3155 ShuffleVectorSDNode *Op) {
3156 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3158 // Is V2 is a vector load, don't do this transformation. We will try to use
3159 // load folding shufps op.
3160 if (ISD::isNON_EXTLoad(V2))
3163 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3165 if (NumElems != 2 && NumElems != 4)
3167 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3168 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3170 for (unsigned i = NumElems/2; i != NumElems; ++i)
3171 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3176 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3178 static bool isSplatVector(SDNode *N) {
3179 if (N->getOpcode() != ISD::BUILD_VECTOR)
3182 SDValue SplatValue = N->getOperand(0);
3183 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3184 if (N->getOperand(i) != SplatValue)
3189 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3190 /// to an zero vector.
3191 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3192 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3193 SDValue V1 = N->getOperand(0);
3194 SDValue V2 = N->getOperand(1);
3195 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3196 for (unsigned i = 0; i != NumElems; ++i) {
3197 int Idx = N->getMaskElt(i);
3198 if (Idx >= (int)NumElems) {
3199 unsigned Opc = V2.getOpcode();
3200 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3202 if (Opc != ISD::BUILD_VECTOR ||
3203 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3205 } else if (Idx >= 0) {
3206 unsigned Opc = V1.getOpcode();
3207 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3209 if (Opc != ISD::BUILD_VECTOR ||
3210 !X86::isZeroNode(V1.getOperand(Idx)))
3217 /// getZeroVector - Returns a vector of specified type with all zero elements.
3219 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3221 assert(VT.isVector() && "Expected a vector type");
3223 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3224 // type. This ensures they get CSE'd.
3226 if (VT.getSizeInBits() == 64) { // MMX
3227 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3228 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3229 } else if (HasSSE2) { // SSE2
3230 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3231 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3233 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3234 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3236 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3239 /// getOnesVector - Returns a vector of specified type with all bits set.
3241 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3242 assert(VT.isVector() && "Expected a vector type");
3244 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3245 // type. This ensures they get CSE'd.
3246 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3248 if (VT.getSizeInBits() == 64) // MMX
3249 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3251 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3252 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3256 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3257 /// that point to V2 points to its first element.
3258 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3259 EVT VT = SVOp->getValueType(0);
3260 unsigned NumElems = VT.getVectorNumElements();
3262 bool Changed = false;
3263 SmallVector<int, 8> MaskVec;
3264 SVOp->getMask(MaskVec);
3266 for (unsigned i = 0; i != NumElems; ++i) {
3267 if (MaskVec[i] > (int)NumElems) {
3268 MaskVec[i] = NumElems;
3273 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3274 SVOp->getOperand(1), &MaskVec[0]);
3275 return SDValue(SVOp, 0);
3278 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3279 /// operation of specified width.
3280 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3282 unsigned NumElems = VT.getVectorNumElements();
3283 SmallVector<int, 8> Mask;
3284 Mask.push_back(NumElems);
3285 for (unsigned i = 1; i != NumElems; ++i)
3287 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3290 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3291 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3293 unsigned NumElems = VT.getVectorNumElements();
3294 SmallVector<int, 8> Mask;
3295 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3297 Mask.push_back(i + NumElems);
3299 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3302 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3303 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3305 unsigned NumElems = VT.getVectorNumElements();
3306 unsigned Half = NumElems/2;
3307 SmallVector<int, 8> Mask;
3308 for (unsigned i = 0; i != Half; ++i) {
3309 Mask.push_back(i + Half);
3310 Mask.push_back(i + NumElems + Half);
3312 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3315 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3316 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3318 if (SV->getValueType(0).getVectorNumElements() <= 4)
3319 return SDValue(SV, 0);
3321 EVT PVT = MVT::v4f32;
3322 EVT VT = SV->getValueType(0);
3323 DebugLoc dl = SV->getDebugLoc();
3324 SDValue V1 = SV->getOperand(0);
3325 int NumElems = VT.getVectorNumElements();
3326 int EltNo = SV->getSplatIndex();
3328 // unpack elements to the correct location
3329 while (NumElems > 4) {
3330 if (EltNo < NumElems/2) {
3331 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3333 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3334 EltNo -= NumElems/2;
3339 // Perform the splat.
3340 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3341 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3342 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3343 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3346 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3347 /// vector of zero or undef vector. This produces a shuffle where the low
3348 /// element of V2 is swizzled into the zero/undef vector, landing at element
3349 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3350 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3351 bool isZero, bool HasSSE2,
3352 SelectionDAG &DAG) {
3353 EVT VT = V2.getValueType();
3355 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3356 unsigned NumElems = VT.getVectorNumElements();
3357 SmallVector<int, 16> MaskVec;
3358 for (unsigned i = 0; i != NumElems; ++i)
3359 // If this is the insertion idx, put the low elt of V2 here.
3360 MaskVec.push_back(i == Idx ? NumElems : i);
3361 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3364 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3365 /// a shuffle that is zero.
3367 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3368 bool Low, SelectionDAG &DAG) {
3369 unsigned NumZeros = 0;
3370 for (int i = 0; i < NumElems; ++i) {
3371 unsigned Index = Low ? i : NumElems-i-1;
3372 int Idx = SVOp->getMaskElt(Index);
3377 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3378 if (Elt.getNode() && X86::isZeroNode(Elt))
3386 /// isVectorShift - Returns true if the shuffle can be implemented as a
3387 /// logical left or right shift of a vector.
3388 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3389 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3390 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3391 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3394 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3397 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3401 bool SeenV1 = false;
3402 bool SeenV2 = false;
3403 for (int i = NumZeros; i < NumElems; ++i) {
3404 int Val = isLeft ? (i - NumZeros) : i;
3405 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3417 if (SeenV1 && SeenV2)
3420 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3426 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3428 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3429 unsigned NumNonZero, unsigned NumZero,
3430 SelectionDAG &DAG, TargetLowering &TLI) {
3434 DebugLoc dl = Op.getDebugLoc();
3437 for (unsigned i = 0; i < 16; ++i) {
3438 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3439 if (ThisIsNonZero && First) {
3441 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3443 V = DAG.getUNDEF(MVT::v8i16);
3448 SDValue ThisElt(0, 0), LastElt(0, 0);
3449 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3450 if (LastIsNonZero) {
3451 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3452 MVT::i16, Op.getOperand(i-1));
3454 if (ThisIsNonZero) {
3455 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3456 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3457 ThisElt, DAG.getConstant(8, MVT::i8));
3459 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3463 if (ThisElt.getNode())
3464 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3465 DAG.getIntPtrConstant(i/2));
3469 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3472 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3474 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3475 unsigned NumNonZero, unsigned NumZero,
3476 SelectionDAG &DAG, TargetLowering &TLI) {
3480 DebugLoc dl = Op.getDebugLoc();
3483 for (unsigned i = 0; i < 8; ++i) {
3484 bool isNonZero = (NonZeros & (1 << i)) != 0;
3488 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3490 V = DAG.getUNDEF(MVT::v8i16);
3493 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3494 MVT::v8i16, V, Op.getOperand(i),
3495 DAG.getIntPtrConstant(i));
3502 /// getVShift - Return a vector logical shift node.
3504 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3505 unsigned NumBits, SelectionDAG &DAG,
3506 const TargetLowering &TLI, DebugLoc dl) {
3507 bool isMMX = VT.getSizeInBits() == 64;
3508 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3509 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3510 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3511 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3512 DAG.getNode(Opc, dl, ShVT, SrcOp,
3513 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3517 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3518 SelectionDAG &DAG) {
3520 // Check if the scalar load can be widened into a vector load. And if
3521 // the address is "base + cst" see if the cst can be "absorbed" into
3522 // the shuffle mask.
3523 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3524 SDValue Ptr = LD->getBasePtr();
3525 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3527 EVT PVT = LD->getValueType(0);
3528 if (PVT != MVT::i32 && PVT != MVT::f32)
3533 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3534 FI = FINode->getIndex();
3536 } else if (Ptr.getOpcode() == ISD::ADD &&
3537 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3538 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3539 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3540 Offset = Ptr.getConstantOperandVal(1);
3541 Ptr = Ptr.getOperand(0);
3546 SDValue Chain = LD->getChain();
3547 // Make sure the stack object alignment is at least 16.
3548 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3549 if (DAG.InferPtrAlignment(Ptr) < 16) {
3550 if (MFI->isFixedObjectIndex(FI)) {
3551 // Can't change the alignment. FIXME: It's possible to compute
3552 // the exact stack offset and reference FI + adjust offset instead.
3553 // If someone *really* cares about this. That's the way to implement it.
3556 MFI->setObjectAlignment(FI, 16);
3560 // (Offset % 16) must be multiple of 4. Then address is then
3561 // Ptr + (Offset & ~15).
3564 if ((Offset % 16) & 3)
3566 int64_t StartOffset = Offset & ~15;
3568 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3569 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3571 int EltNo = (Offset - StartOffset) >> 2;
3572 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3573 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3574 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3576 // Canonicalize it to a v4i32 shuffle.
3577 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3578 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3579 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3580 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3587 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3588 DebugLoc dl = Op.getDebugLoc();
3589 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3590 if (ISD::isBuildVectorAllZeros(Op.getNode())
3591 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3592 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3593 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3594 // eliminated on x86-32 hosts.
3595 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3598 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3599 return getOnesVector(Op.getValueType(), DAG, dl);
3600 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3603 EVT VT = Op.getValueType();
3604 EVT ExtVT = VT.getVectorElementType();
3605 unsigned EVTBits = ExtVT.getSizeInBits();
3607 unsigned NumElems = Op.getNumOperands();
3608 unsigned NumZero = 0;
3609 unsigned NumNonZero = 0;
3610 unsigned NonZeros = 0;
3611 bool IsAllConstants = true;
3612 SmallSet<SDValue, 8> Values;
3613 for (unsigned i = 0; i < NumElems; ++i) {
3614 SDValue Elt = Op.getOperand(i);
3615 if (Elt.getOpcode() == ISD::UNDEF)
3618 if (Elt.getOpcode() != ISD::Constant &&
3619 Elt.getOpcode() != ISD::ConstantFP)
3620 IsAllConstants = false;
3621 if (X86::isZeroNode(Elt))
3624 NonZeros |= (1 << i);
3629 if (NumNonZero == 0) {
3630 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3631 return DAG.getUNDEF(VT);
3634 // Special case for single non-zero, non-undef, element.
3635 if (NumNonZero == 1) {
3636 unsigned Idx = CountTrailingZeros_32(NonZeros);
3637 SDValue Item = Op.getOperand(Idx);
3639 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3640 // the value are obviously zero, truncate the value to i32 and do the
3641 // insertion that way. Only do this if the value is non-constant or if the
3642 // value is a constant being inserted into element 0. It is cheaper to do
3643 // a constant pool load than it is to do a movd + shuffle.
3644 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3645 (!IsAllConstants || Idx == 0)) {
3646 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3647 // Handle MMX and SSE both.
3648 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3649 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3651 // Truncate the value (which may itself be a constant) to i32, and
3652 // convert it to a vector with movd (S2V+shuffle to zero extend).
3653 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3654 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3655 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3656 Subtarget->hasSSE2(), DAG);
3658 // Now we have our 32-bit value zero extended in the low element of
3659 // a vector. If Idx != 0, swizzle it into place.
3661 SmallVector<int, 4> Mask;
3662 Mask.push_back(Idx);
3663 for (unsigned i = 1; i != VecElts; ++i)
3665 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3666 DAG.getUNDEF(Item.getValueType()),
3669 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3673 // If we have a constant or non-constant insertion into the low element of
3674 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3675 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3676 // depending on what the source datatype is.
3679 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3680 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3681 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3682 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3683 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3684 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3686 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3687 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3688 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3689 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3690 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3691 Subtarget->hasSSE2(), DAG);
3692 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3696 // Is it a vector logical left shift?
3697 if (NumElems == 2 && Idx == 1 &&
3698 X86::isZeroNode(Op.getOperand(0)) &&
3699 !X86::isZeroNode(Op.getOperand(1))) {
3700 unsigned NumBits = VT.getSizeInBits();
3701 return getVShift(true, VT,
3702 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3703 VT, Op.getOperand(1)),
3704 NumBits/2, DAG, *this, dl);
3707 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3710 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3711 // is a non-constant being inserted into an element other than the low one,
3712 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3713 // movd/movss) to move this into the low element, then shuffle it into
3715 if (EVTBits == 32) {
3716 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3718 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3719 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3720 Subtarget->hasSSE2(), DAG);
3721 SmallVector<int, 8> MaskVec;
3722 for (unsigned i = 0; i < NumElems; i++)
3723 MaskVec.push_back(i == Idx ? 0 : 1);
3724 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3728 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3729 if (Values.size() == 1) {
3730 if (EVTBits == 32) {
3731 // Instead of a shuffle like this:
3732 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3733 // Check if it's possible to issue this instead.
3734 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3735 unsigned Idx = CountTrailingZeros_32(NonZeros);
3736 SDValue Item = Op.getOperand(Idx);
3737 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3738 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3743 // A vector full of immediates; various special cases are already
3744 // handled, so this is best done with a single constant-pool load.
3748 // Let legalizer expand 2-wide build_vectors.
3749 if (EVTBits == 64) {
3750 if (NumNonZero == 1) {
3751 // One half is zero or undef.
3752 unsigned Idx = CountTrailingZeros_32(NonZeros);
3753 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3754 Op.getOperand(Idx));
3755 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3756 Subtarget->hasSSE2(), DAG);
3761 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3762 if (EVTBits == 8 && NumElems == 16) {
3763 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3765 if (V.getNode()) return V;
3768 if (EVTBits == 16 && NumElems == 8) {
3769 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3771 if (V.getNode()) return V;
3774 // If element VT is == 32 bits, turn it into a number of shuffles.
3775 SmallVector<SDValue, 8> V;
3777 if (NumElems == 4 && NumZero > 0) {
3778 for (unsigned i = 0; i < 4; ++i) {
3779 bool isZero = !(NonZeros & (1 << i));
3781 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3783 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3786 for (unsigned i = 0; i < 2; ++i) {
3787 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3790 V[i] = V[i*2]; // Must be a zero vector.
3793 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3796 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3799 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3804 SmallVector<int, 8> MaskVec;
3805 bool Reverse = (NonZeros & 0x3) == 2;
3806 for (unsigned i = 0; i < 2; ++i)
3807 MaskVec.push_back(Reverse ? 1-i : i);
3808 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3809 for (unsigned i = 0; i < 2; ++i)
3810 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3811 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3814 if (Values.size() > 2) {
3815 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3816 // values to be inserted is equal to the number of elements, in which case
3817 // use the unpack code below in the hopes of matching the consecutive elts
3818 // load merge pattern for shuffles.
3819 // FIXME: We could probably just check that here directly.
3820 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3821 getSubtarget()->hasSSE41()) {
3822 V[0] = DAG.getUNDEF(VT);
3823 for (unsigned i = 0; i < NumElems; ++i)
3824 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3825 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3826 Op.getOperand(i), DAG.getIntPtrConstant(i));
3829 // Expand into a number of unpckl*.
3831 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3832 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3833 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3834 for (unsigned i = 0; i < NumElems; ++i)
3835 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3837 while (NumElems != 0) {
3838 for (unsigned i = 0; i < NumElems; ++i)
3839 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3849 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3850 // We support concatenate two MMX registers and place them in a MMX
3851 // register. This is better than doing a stack convert.
3852 DebugLoc dl = Op.getDebugLoc();
3853 EVT ResVT = Op.getValueType();
3854 assert(Op.getNumOperands() == 2);
3855 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3856 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3858 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3859 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3860 InVec = Op.getOperand(1);
3861 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3862 unsigned NumElts = ResVT.getVectorNumElements();
3863 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3864 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3865 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3867 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3868 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3869 Mask[0] = 0; Mask[1] = 2;
3870 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3872 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3875 // v8i16 shuffles - Prefer shuffles in the following order:
3876 // 1. [all] pshuflw, pshufhw, optional move
3877 // 2. [ssse3] 1 x pshufb
3878 // 3. [ssse3] 2 x pshufb + 1 x por
3879 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3881 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3882 SelectionDAG &DAG, X86TargetLowering &TLI) {
3883 SDValue V1 = SVOp->getOperand(0);
3884 SDValue V2 = SVOp->getOperand(1);
3885 DebugLoc dl = SVOp->getDebugLoc();
3886 SmallVector<int, 8> MaskVals;
3888 // Determine if more than 1 of the words in each of the low and high quadwords
3889 // of the result come from the same quadword of one of the two inputs. Undef
3890 // mask values count as coming from any quadword, for better codegen.
3891 SmallVector<unsigned, 4> LoQuad(4);
3892 SmallVector<unsigned, 4> HiQuad(4);
3893 BitVector InputQuads(4);
3894 for (unsigned i = 0; i < 8; ++i) {
3895 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3896 int EltIdx = SVOp->getMaskElt(i);
3897 MaskVals.push_back(EltIdx);
3906 InputQuads.set(EltIdx / 4);
3909 int BestLoQuad = -1;
3910 unsigned MaxQuad = 1;
3911 for (unsigned i = 0; i < 4; ++i) {
3912 if (LoQuad[i] > MaxQuad) {
3914 MaxQuad = LoQuad[i];
3918 int BestHiQuad = -1;
3920 for (unsigned i = 0; i < 4; ++i) {
3921 if (HiQuad[i] > MaxQuad) {
3923 MaxQuad = HiQuad[i];
3927 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3928 // of the two input vectors, shuffle them into one input vector so only a
3929 // single pshufb instruction is necessary. If There are more than 2 input
3930 // quads, disable the next transformation since it does not help SSSE3.
3931 bool V1Used = InputQuads[0] || InputQuads[1];
3932 bool V2Used = InputQuads[2] || InputQuads[3];
3933 if (TLI.getSubtarget()->hasSSSE3()) {
3934 if (InputQuads.count() == 2 && V1Used && V2Used) {
3935 BestLoQuad = InputQuads.find_first();
3936 BestHiQuad = InputQuads.find_next(BestLoQuad);
3938 if (InputQuads.count() > 2) {
3944 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3945 // the shuffle mask. If a quad is scored as -1, that means that it contains
3946 // words from all 4 input quadwords.
3948 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3949 SmallVector<int, 8> MaskV;
3950 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3951 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3952 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3953 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3954 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3955 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3957 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3958 // source words for the shuffle, to aid later transformations.
3959 bool AllWordsInNewV = true;
3960 bool InOrder[2] = { true, true };
3961 for (unsigned i = 0; i != 8; ++i) {
3962 int idx = MaskVals[i];
3964 InOrder[i/4] = false;
3965 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3967 AllWordsInNewV = false;
3971 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3972 if (AllWordsInNewV) {
3973 for (int i = 0; i != 8; ++i) {
3974 int idx = MaskVals[i];
3977 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3978 if ((idx != i) && idx < 4)
3980 if ((idx != i) && idx > 3)
3989 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3990 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3991 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3992 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3993 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3997 // If we have SSSE3, and all words of the result are from 1 input vector,
3998 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3999 // is present, fall back to case 4.
4000 if (TLI.getSubtarget()->hasSSSE3()) {
4001 SmallVector<SDValue,16> pshufbMask;
4003 // If we have elements from both input vectors, set the high bit of the
4004 // shuffle mask element to zero out elements that come from V2 in the V1
4005 // mask, and elements that come from V1 in the V2 mask, so that the two
4006 // results can be OR'd together.
4007 bool TwoInputs = V1Used && V2Used;
4008 for (unsigned i = 0; i != 8; ++i) {
4009 int EltIdx = MaskVals[i] * 2;
4010 if (TwoInputs && (EltIdx >= 16)) {
4011 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4012 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4015 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4016 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4018 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4019 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4020 DAG.getNode(ISD::BUILD_VECTOR, dl,
4021 MVT::v16i8, &pshufbMask[0], 16));
4023 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4025 // Calculate the shuffle mask for the second input, shuffle it, and
4026 // OR it with the first shuffled input.
4028 for (unsigned i = 0; i != 8; ++i) {
4029 int EltIdx = MaskVals[i] * 2;
4031 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4032 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4035 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4036 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4038 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4039 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4040 DAG.getNode(ISD::BUILD_VECTOR, dl,
4041 MVT::v16i8, &pshufbMask[0], 16));
4042 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4043 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4046 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4047 // and update MaskVals with new element order.
4048 BitVector InOrder(8);
4049 if (BestLoQuad >= 0) {
4050 SmallVector<int, 8> MaskV;
4051 for (int i = 0; i != 4; ++i) {
4052 int idx = MaskVals[i];
4054 MaskV.push_back(-1);
4056 } else if ((idx / 4) == BestLoQuad) {
4057 MaskV.push_back(idx & 3);
4060 MaskV.push_back(-1);
4063 for (unsigned i = 4; i != 8; ++i)
4065 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4069 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4070 // and update MaskVals with the new element order.
4071 if (BestHiQuad >= 0) {
4072 SmallVector<int, 8> MaskV;
4073 for (unsigned i = 0; i != 4; ++i)
4075 for (unsigned i = 4; i != 8; ++i) {
4076 int idx = MaskVals[i];
4078 MaskV.push_back(-1);
4080 } else if ((idx / 4) == BestHiQuad) {
4081 MaskV.push_back((idx & 3) + 4);
4084 MaskV.push_back(-1);
4087 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4091 // In case BestHi & BestLo were both -1, which means each quadword has a word
4092 // from each of the four input quadwords, calculate the InOrder bitvector now
4093 // before falling through to the insert/extract cleanup.
4094 if (BestLoQuad == -1 && BestHiQuad == -1) {
4096 for (int i = 0; i != 8; ++i)
4097 if (MaskVals[i] < 0 || MaskVals[i] == i)
4101 // The other elements are put in the right place using pextrw and pinsrw.
4102 for (unsigned i = 0; i != 8; ++i) {
4105 int EltIdx = MaskVals[i];
4108 SDValue ExtOp = (EltIdx < 8)
4109 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4110 DAG.getIntPtrConstant(EltIdx))
4111 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4112 DAG.getIntPtrConstant(EltIdx - 8));
4113 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4114 DAG.getIntPtrConstant(i));
4119 // v16i8 shuffles - Prefer shuffles in the following order:
4120 // 1. [ssse3] 1 x pshufb
4121 // 2. [ssse3] 2 x pshufb + 1 x por
4122 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4124 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4125 SelectionDAG &DAG, X86TargetLowering &TLI) {
4126 SDValue V1 = SVOp->getOperand(0);
4127 SDValue V2 = SVOp->getOperand(1);
4128 DebugLoc dl = SVOp->getDebugLoc();
4129 SmallVector<int, 16> MaskVals;
4130 SVOp->getMask(MaskVals);
4132 // If we have SSSE3, case 1 is generated when all result bytes come from
4133 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4134 // present, fall back to case 3.
4135 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4138 for (unsigned i = 0; i < 16; ++i) {
4139 int EltIdx = MaskVals[i];
4148 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4149 if (TLI.getSubtarget()->hasSSSE3()) {
4150 SmallVector<SDValue,16> pshufbMask;
4152 // If all result elements are from one input vector, then only translate
4153 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4155 // Otherwise, we have elements from both input vectors, and must zero out
4156 // elements that come from V2 in the first mask, and V1 in the second mask
4157 // so that we can OR them together.
4158 bool TwoInputs = !(V1Only || V2Only);
4159 for (unsigned i = 0; i != 16; ++i) {
4160 int EltIdx = MaskVals[i];
4161 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4162 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4165 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4167 // If all the elements are from V2, assign it to V1 and return after
4168 // building the first pshufb.
4171 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4172 DAG.getNode(ISD::BUILD_VECTOR, dl,
4173 MVT::v16i8, &pshufbMask[0], 16));
4177 // Calculate the shuffle mask for the second input, shuffle it, and
4178 // OR it with the first shuffled input.
4180 for (unsigned i = 0; i != 16; ++i) {
4181 int EltIdx = MaskVals[i];
4183 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4186 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4188 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4189 DAG.getNode(ISD::BUILD_VECTOR, dl,
4190 MVT::v16i8, &pshufbMask[0], 16));
4191 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4194 // No SSSE3 - Calculate in place words and then fix all out of place words
4195 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4196 // the 16 different words that comprise the two doublequadword input vectors.
4197 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4198 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4199 SDValue NewV = V2Only ? V2 : V1;
4200 for (int i = 0; i != 8; ++i) {
4201 int Elt0 = MaskVals[i*2];
4202 int Elt1 = MaskVals[i*2+1];
4204 // This word of the result is all undef, skip it.
4205 if (Elt0 < 0 && Elt1 < 0)
4208 // This word of the result is already in the correct place, skip it.
4209 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4211 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4214 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4215 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4218 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4219 // using a single extract together, load it and store it.
4220 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4221 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4222 DAG.getIntPtrConstant(Elt1 / 2));
4223 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4224 DAG.getIntPtrConstant(i));
4228 // If Elt1 is defined, extract it from the appropriate source. If the
4229 // source byte is not also odd, shift the extracted word left 8 bits
4230 // otherwise clear the bottom 8 bits if we need to do an or.
4232 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4233 DAG.getIntPtrConstant(Elt1 / 2));
4234 if ((Elt1 & 1) == 0)
4235 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4236 DAG.getConstant(8, TLI.getShiftAmountTy()));
4238 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4239 DAG.getConstant(0xFF00, MVT::i16));
4241 // If Elt0 is defined, extract it from the appropriate source. If the
4242 // source byte is not also even, shift the extracted word right 8 bits. If
4243 // Elt1 was also defined, OR the extracted values together before
4244 // inserting them in the result.
4246 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4247 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4248 if ((Elt0 & 1) != 0)
4249 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4250 DAG.getConstant(8, TLI.getShiftAmountTy()));
4252 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4253 DAG.getConstant(0x00FF, MVT::i16));
4254 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4257 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4258 DAG.getIntPtrConstant(i));
4260 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4263 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4264 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4265 /// done when every pair / quad of shuffle mask elements point to elements in
4266 /// the right sequence. e.g.
4267 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4269 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4271 TargetLowering &TLI, DebugLoc dl) {
4272 EVT VT = SVOp->getValueType(0);
4273 SDValue V1 = SVOp->getOperand(0);
4274 SDValue V2 = SVOp->getOperand(1);
4275 unsigned NumElems = VT.getVectorNumElements();
4276 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4277 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4278 EVT MaskEltVT = MaskVT.getVectorElementType();
4280 switch (VT.getSimpleVT().SimpleTy) {
4281 default: assert(false && "Unexpected!");
4282 case MVT::v4f32: NewVT = MVT::v2f64; break;
4283 case MVT::v4i32: NewVT = MVT::v2i64; break;
4284 case MVT::v8i16: NewVT = MVT::v4i32; break;
4285 case MVT::v16i8: NewVT = MVT::v4i32; break;
4288 if (NewWidth == 2) {
4294 int Scale = NumElems / NewWidth;
4295 SmallVector<int, 8> MaskVec;
4296 for (unsigned i = 0; i < NumElems; i += Scale) {
4298 for (int j = 0; j < Scale; ++j) {
4299 int EltIdx = SVOp->getMaskElt(i+j);
4303 StartIdx = EltIdx - (EltIdx % Scale);
4304 if (EltIdx != StartIdx + j)
4308 MaskVec.push_back(-1);
4310 MaskVec.push_back(StartIdx / Scale);
4313 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4314 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4315 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4318 /// getVZextMovL - Return a zero-extending vector move low node.
4320 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4321 SDValue SrcOp, SelectionDAG &DAG,
4322 const X86Subtarget *Subtarget, DebugLoc dl) {
4323 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4324 LoadSDNode *LD = NULL;
4325 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4326 LD = dyn_cast<LoadSDNode>(SrcOp);
4328 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4330 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4331 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4332 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4333 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4334 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4336 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4337 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4338 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4339 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4347 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4348 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4349 DAG.getNode(ISD::BIT_CONVERT, dl,
4353 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4356 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4357 SDValue V1 = SVOp->getOperand(0);
4358 SDValue V2 = SVOp->getOperand(1);
4359 DebugLoc dl = SVOp->getDebugLoc();
4360 EVT VT = SVOp->getValueType(0);
4362 SmallVector<std::pair<int, int>, 8> Locs;
4364 SmallVector<int, 8> Mask1(4U, -1);
4365 SmallVector<int, 8> PermMask;
4366 SVOp->getMask(PermMask);
4370 for (unsigned i = 0; i != 4; ++i) {
4371 int Idx = PermMask[i];
4373 Locs[i] = std::make_pair(-1, -1);
4375 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4377 Locs[i] = std::make_pair(0, NumLo);
4381 Locs[i] = std::make_pair(1, NumHi);
4383 Mask1[2+NumHi] = Idx;
4389 if (NumLo <= 2 && NumHi <= 2) {
4390 // If no more than two elements come from either vector. This can be
4391 // implemented with two shuffles. First shuffle gather the elements.
4392 // The second shuffle, which takes the first shuffle as both of its
4393 // vector operands, put the elements into the right order.
4394 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4396 SmallVector<int, 8> Mask2(4U, -1);
4398 for (unsigned i = 0; i != 4; ++i) {
4399 if (Locs[i].first == -1)
4402 unsigned Idx = (i < 2) ? 0 : 4;
4403 Idx += Locs[i].first * 2 + Locs[i].second;
4408 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4409 } else if (NumLo == 3 || NumHi == 3) {
4410 // Otherwise, we must have three elements from one vector, call it X, and
4411 // one element from the other, call it Y. First, use a shufps to build an
4412 // intermediate vector with the one element from Y and the element from X
4413 // that will be in the same half in the final destination (the indexes don't
4414 // matter). Then, use a shufps to build the final vector, taking the half
4415 // containing the element from Y from the intermediate, and the other half
4418 // Normalize it so the 3 elements come from V1.
4419 CommuteVectorShuffleMask(PermMask, VT);
4423 // Find the element from V2.
4425 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4426 int Val = PermMask[HiIndex];
4433 Mask1[0] = PermMask[HiIndex];
4435 Mask1[2] = PermMask[HiIndex^1];
4437 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4440 Mask1[0] = PermMask[0];
4441 Mask1[1] = PermMask[1];
4442 Mask1[2] = HiIndex & 1 ? 6 : 4;
4443 Mask1[3] = HiIndex & 1 ? 4 : 6;
4444 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4446 Mask1[0] = HiIndex & 1 ? 2 : 0;
4447 Mask1[1] = HiIndex & 1 ? 0 : 2;
4448 Mask1[2] = PermMask[2];
4449 Mask1[3] = PermMask[3];
4454 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4458 // Break it into (shuffle shuffle_hi, shuffle_lo).
4460 SmallVector<int,8> LoMask(4U, -1);
4461 SmallVector<int,8> HiMask(4U, -1);
4463 SmallVector<int,8> *MaskPtr = &LoMask;
4464 unsigned MaskIdx = 0;
4467 for (unsigned i = 0; i != 4; ++i) {
4474 int Idx = PermMask[i];
4476 Locs[i] = std::make_pair(-1, -1);
4477 } else if (Idx < 4) {
4478 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4479 (*MaskPtr)[LoIdx] = Idx;
4482 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4483 (*MaskPtr)[HiIdx] = Idx;
4488 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4489 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4490 SmallVector<int, 8> MaskOps;
4491 for (unsigned i = 0; i != 4; ++i) {
4492 if (Locs[i].first == -1) {
4493 MaskOps.push_back(-1);
4495 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4496 MaskOps.push_back(Idx);
4499 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4503 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4504 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4505 SDValue V1 = Op.getOperand(0);
4506 SDValue V2 = Op.getOperand(1);
4507 EVT VT = Op.getValueType();
4508 DebugLoc dl = Op.getDebugLoc();
4509 unsigned NumElems = VT.getVectorNumElements();
4510 bool isMMX = VT.getSizeInBits() == 64;
4511 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4512 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4513 bool V1IsSplat = false;
4514 bool V2IsSplat = false;
4516 if (isZeroShuffle(SVOp))
4517 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4519 // Promote splats to v4f32.
4520 if (SVOp->isSplat()) {
4521 if (isMMX || NumElems < 4)
4523 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4526 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4528 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4529 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4530 if (NewOp.getNode())
4531 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4532 LowerVECTOR_SHUFFLE(NewOp, DAG));
4533 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4534 // FIXME: Figure out a cleaner way to do this.
4535 // Try to make use of movq to zero out the top part.
4536 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4537 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4538 if (NewOp.getNode()) {
4539 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4540 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4541 DAG, Subtarget, dl);
4543 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4544 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4545 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4546 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4547 DAG, Subtarget, dl);
4551 if (X86::isPSHUFDMask(SVOp))
4554 // Check if this can be converted into a logical shift.
4555 bool isLeft = false;
4558 bool isShift = getSubtarget()->hasSSE2() &&
4559 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4560 if (isShift && ShVal.hasOneUse()) {
4561 // If the shifted value has multiple uses, it may be cheaper to use
4562 // v_set0 + movlhps or movhlps, etc.
4563 EVT EltVT = VT.getVectorElementType();
4564 ShAmt *= EltVT.getSizeInBits();
4565 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4568 if (X86::isMOVLMask(SVOp)) {
4571 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4572 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4577 // FIXME: fold these into legal mask.
4578 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4579 X86::isMOVSLDUPMask(SVOp) ||
4580 X86::isMOVHLPSMask(SVOp) ||
4581 X86::isMOVLHPSMask(SVOp) ||
4582 X86::isMOVLPMask(SVOp)))
4585 if (ShouldXformToMOVHLPS(SVOp) ||
4586 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4587 return CommuteVectorShuffle(SVOp, DAG);
4590 // No better options. Use a vshl / vsrl.
4591 EVT EltVT = VT.getVectorElementType();
4592 ShAmt *= EltVT.getSizeInBits();
4593 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4596 bool Commuted = false;
4597 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4598 // 1,1,1,1 -> v8i16 though.
4599 V1IsSplat = isSplatVector(V1.getNode());
4600 V2IsSplat = isSplatVector(V2.getNode());
4602 // Canonicalize the splat or undef, if present, to be on the RHS.
4603 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4604 Op = CommuteVectorShuffle(SVOp, DAG);
4605 SVOp = cast<ShuffleVectorSDNode>(Op);
4606 V1 = SVOp->getOperand(0);
4607 V2 = SVOp->getOperand(1);
4608 std::swap(V1IsSplat, V2IsSplat);
4609 std::swap(V1IsUndef, V2IsUndef);
4613 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4614 // Shuffling low element of v1 into undef, just return v1.
4617 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4618 // the instruction selector will not match, so get a canonical MOVL with
4619 // swapped operands to undo the commute.
4620 return getMOVL(DAG, dl, VT, V2, V1);
4623 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4624 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4625 X86::isUNPCKLMask(SVOp) ||
4626 X86::isUNPCKHMask(SVOp))
4630 // Normalize mask so all entries that point to V2 points to its first
4631 // element then try to match unpck{h|l} again. If match, return a
4632 // new vector_shuffle with the corrected mask.
4633 SDValue NewMask = NormalizeMask(SVOp, DAG);
4634 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4635 if (NSVOp != SVOp) {
4636 if (X86::isUNPCKLMask(NSVOp, true)) {
4638 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4645 // Commute is back and try unpck* again.
4646 // FIXME: this seems wrong.
4647 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4648 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4649 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4650 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4651 X86::isUNPCKLMask(NewSVOp) ||
4652 X86::isUNPCKHMask(NewSVOp))
4656 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4658 // Normalize the node to match x86 shuffle ops if needed
4659 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4660 return CommuteVectorShuffle(SVOp, DAG);
4662 // Check for legal shuffle and return?
4663 SmallVector<int, 16> PermMask;
4664 SVOp->getMask(PermMask);
4665 if (isShuffleMaskLegal(PermMask, VT))
4668 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4669 if (VT == MVT::v8i16) {
4670 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4671 if (NewOp.getNode())
4675 if (VT == MVT::v16i8) {
4676 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4677 if (NewOp.getNode())
4681 // Handle all 4 wide cases with a number of shuffles except for MMX.
4682 if (NumElems == 4 && !isMMX)
4683 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4689 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4690 SelectionDAG &DAG) {
4691 EVT VT = Op.getValueType();
4692 DebugLoc dl = Op.getDebugLoc();
4693 if (VT.getSizeInBits() == 8) {
4694 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4695 Op.getOperand(0), Op.getOperand(1));
4696 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4697 DAG.getValueType(VT));
4698 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4699 } else if (VT.getSizeInBits() == 16) {
4700 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4701 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4703 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4704 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4705 DAG.getNode(ISD::BIT_CONVERT, dl,
4709 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4710 Op.getOperand(0), Op.getOperand(1));
4711 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4712 DAG.getValueType(VT));
4713 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4714 } else if (VT == MVT::f32) {
4715 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4716 // the result back to FR32 register. It's only worth matching if the
4717 // result has a single use which is a store or a bitcast to i32. And in
4718 // the case of a store, it's not worth it if the index is a constant 0,
4719 // because a MOVSSmr can be used instead, which is smaller and faster.
4720 if (!Op.hasOneUse())
4722 SDNode *User = *Op.getNode()->use_begin();
4723 if ((User->getOpcode() != ISD::STORE ||
4724 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4725 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4726 (User->getOpcode() != ISD::BIT_CONVERT ||
4727 User->getValueType(0) != MVT::i32))
4729 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4730 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4733 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4734 } else if (VT == MVT::i32) {
4735 // ExtractPS works with constant index.
4736 if (isa<ConstantSDNode>(Op.getOperand(1)))
4744 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4745 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4748 if (Subtarget->hasSSE41()) {
4749 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4754 EVT VT = Op.getValueType();
4755 DebugLoc dl = Op.getDebugLoc();
4756 // TODO: handle v16i8.
4757 if (VT.getSizeInBits() == 16) {
4758 SDValue Vec = Op.getOperand(0);
4759 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4761 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4762 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4763 DAG.getNode(ISD::BIT_CONVERT, dl,
4766 // Transform it so it match pextrw which produces a 32-bit result.
4767 EVT EltVT = MVT::i32;
4768 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4769 Op.getOperand(0), Op.getOperand(1));
4770 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4771 DAG.getValueType(VT));
4772 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4773 } else if (VT.getSizeInBits() == 32) {
4774 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4778 // SHUFPS the element to the lowest double word, then movss.
4779 int Mask[4] = { Idx, -1, -1, -1 };
4780 EVT VVT = Op.getOperand(0).getValueType();
4781 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4782 DAG.getUNDEF(VVT), Mask);
4783 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4784 DAG.getIntPtrConstant(0));
4785 } else if (VT.getSizeInBits() == 64) {
4786 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4787 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4788 // to match extract_elt for f64.
4789 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4793 // UNPCKHPD the element to the lowest double word, then movsd.
4794 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4795 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4796 int Mask[2] = { 1, -1 };
4797 EVT VVT = Op.getOperand(0).getValueType();
4798 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4799 DAG.getUNDEF(VVT), Mask);
4800 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4801 DAG.getIntPtrConstant(0));
4808 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4809 EVT VT = Op.getValueType();
4810 EVT EltVT = VT.getVectorElementType();
4811 DebugLoc dl = Op.getDebugLoc();
4813 SDValue N0 = Op.getOperand(0);
4814 SDValue N1 = Op.getOperand(1);
4815 SDValue N2 = Op.getOperand(2);
4817 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4818 isa<ConstantSDNode>(N2)) {
4819 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4821 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4823 if (N1.getValueType() != MVT::i32)
4824 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4825 if (N2.getValueType() != MVT::i32)
4826 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4827 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4828 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4829 // Bits [7:6] of the constant are the source select. This will always be
4830 // zero here. The DAG Combiner may combine an extract_elt index into these
4831 // bits. For example (insert (extract, 3), 2) could be matched by putting
4832 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4833 // Bits [5:4] of the constant are the destination select. This is the
4834 // value of the incoming immediate.
4835 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4836 // combine either bitwise AND or insert of float 0.0 to set these bits.
4837 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4838 // Create this as a scalar to vector..
4839 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4840 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4841 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4842 // PINSR* works with constant index.
4849 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4850 EVT VT = Op.getValueType();
4851 EVT EltVT = VT.getVectorElementType();
4853 if (Subtarget->hasSSE41())
4854 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4856 if (EltVT == MVT::i8)
4859 DebugLoc dl = Op.getDebugLoc();
4860 SDValue N0 = Op.getOperand(0);
4861 SDValue N1 = Op.getOperand(1);
4862 SDValue N2 = Op.getOperand(2);
4864 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4865 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4866 // as its second argument.
4867 if (N1.getValueType() != MVT::i32)
4868 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4869 if (N2.getValueType() != MVT::i32)
4870 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4871 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4877 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4878 DebugLoc dl = Op.getDebugLoc();
4879 if (Op.getValueType() == MVT::v2f32)
4880 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4881 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4882 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4883 Op.getOperand(0))));
4885 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4886 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4888 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4889 EVT VT = MVT::v2i32;
4890 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4897 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4898 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4901 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4902 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4903 // one of the above mentioned nodes. It has to be wrapped because otherwise
4904 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4905 // be used to form addressing mode. These wrapped nodes will be selected
4908 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4909 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4911 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4913 unsigned char OpFlag = 0;
4914 unsigned WrapperKind = X86ISD::Wrapper;
4915 CodeModel::Model M = getTargetMachine().getCodeModel();
4917 if (Subtarget->isPICStyleRIPRel() &&
4918 (M == CodeModel::Small || M == CodeModel::Kernel))
4919 WrapperKind = X86ISD::WrapperRIP;
4920 else if (Subtarget->isPICStyleGOT())
4921 OpFlag = X86II::MO_GOTOFF;
4922 else if (Subtarget->isPICStyleStubPIC())
4923 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4925 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4927 CP->getOffset(), OpFlag);
4928 DebugLoc DL = CP->getDebugLoc();
4929 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4930 // With PIC, the address is actually $g + Offset.
4932 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4933 DAG.getNode(X86ISD::GlobalBaseReg,
4934 DebugLoc::getUnknownLoc(), getPointerTy()),
4941 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4942 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4944 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4946 unsigned char OpFlag = 0;
4947 unsigned WrapperKind = X86ISD::Wrapper;
4948 CodeModel::Model M = getTargetMachine().getCodeModel();
4950 if (Subtarget->isPICStyleRIPRel() &&
4951 (M == CodeModel::Small || M == CodeModel::Kernel))
4952 WrapperKind = X86ISD::WrapperRIP;
4953 else if (Subtarget->isPICStyleGOT())
4954 OpFlag = X86II::MO_GOTOFF;
4955 else if (Subtarget->isPICStyleStubPIC())
4956 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4958 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4960 DebugLoc DL = JT->getDebugLoc();
4961 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4963 // With PIC, the address is actually $g + Offset.
4965 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4966 DAG.getNode(X86ISD::GlobalBaseReg,
4967 DebugLoc::getUnknownLoc(), getPointerTy()),
4975 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4976 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4978 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4980 unsigned char OpFlag = 0;
4981 unsigned WrapperKind = X86ISD::Wrapper;
4982 CodeModel::Model M = getTargetMachine().getCodeModel();
4984 if (Subtarget->isPICStyleRIPRel() &&
4985 (M == CodeModel::Small || M == CodeModel::Kernel))
4986 WrapperKind = X86ISD::WrapperRIP;
4987 else if (Subtarget->isPICStyleGOT())
4988 OpFlag = X86II::MO_GOTOFF;
4989 else if (Subtarget->isPICStyleStubPIC())
4990 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4992 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4994 DebugLoc DL = Op.getDebugLoc();
4995 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4998 // With PIC, the address is actually $g + Offset.
4999 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
5000 !Subtarget->is64Bit()) {
5001 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5002 DAG.getNode(X86ISD::GlobalBaseReg,
5003 DebugLoc::getUnknownLoc(),
5012 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
5013 // Create the TargetBlockAddressAddress node.
5014 unsigned char OpFlags =
5015 Subtarget->ClassifyBlockAddressReference();
5016 CodeModel::Model M = getTargetMachine().getCodeModel();
5017 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5018 DebugLoc dl = Op.getDebugLoc();
5019 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5020 /*isTarget=*/true, OpFlags);
5022 if (Subtarget->isPICStyleRIPRel() &&
5023 (M == CodeModel::Small || M == CodeModel::Kernel))
5024 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5026 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5028 // With PIC, the address is actually $g + Offset.
5029 if (isGlobalRelativeToPICBase(OpFlags)) {
5030 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5031 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5039 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5041 SelectionDAG &DAG) const {
5042 // Create the TargetGlobalAddress node, folding in the constant
5043 // offset if it is legal.
5044 unsigned char OpFlags =
5045 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5046 CodeModel::Model M = getTargetMachine().getCodeModel();
5048 if (OpFlags == X86II::MO_NO_FLAG &&
5049 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5050 // A direct static reference to a global.
5051 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5054 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5057 if (Subtarget->isPICStyleRIPRel() &&
5058 (M == CodeModel::Small || M == CodeModel::Kernel))
5059 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5061 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5063 // With PIC, the address is actually $g + Offset.
5064 if (isGlobalRelativeToPICBase(OpFlags)) {
5065 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5066 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5070 // For globals that require a load from a stub to get the address, emit the
5072 if (isGlobalStubReference(OpFlags))
5073 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5074 PseudoSourceValue::getGOT(), 0, false, false, 0);
5076 // If there was a non-zero offset that we didn't fold, create an explicit
5079 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5080 DAG.getConstant(Offset, getPointerTy()));
5086 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5087 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5088 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5089 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5093 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5094 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5095 unsigned char OperandFlags) {
5096 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5097 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5098 DebugLoc dl = GA->getDebugLoc();
5099 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5100 GA->getValueType(0),
5104 SDValue Ops[] = { Chain, TGA, *InFlag };
5105 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5107 SDValue Ops[] = { Chain, TGA };
5108 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5111 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5112 MFI->setHasCalls(true);
5114 SDValue Flag = Chain.getValue(1);
5115 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5118 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5120 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5123 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5124 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5125 DAG.getNode(X86ISD::GlobalBaseReg,
5126 DebugLoc::getUnknownLoc(),
5128 InFlag = Chain.getValue(1);
5130 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5133 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5135 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5137 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5138 X86::RAX, X86II::MO_TLSGD);
5141 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5142 // "local exec" model.
5143 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5144 const EVT PtrVT, TLSModel::Model model,
5146 DebugLoc dl = GA->getDebugLoc();
5147 // Get the Thread Pointer
5148 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5149 DebugLoc::getUnknownLoc(), PtrVT,
5150 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5153 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5154 NULL, 0, false, false, 0);
5156 unsigned char OperandFlags = 0;
5157 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5159 unsigned WrapperKind = X86ISD::Wrapper;
5160 if (model == TLSModel::LocalExec) {
5161 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5162 } else if (is64Bit) {
5163 assert(model == TLSModel::InitialExec);
5164 OperandFlags = X86II::MO_GOTTPOFF;
5165 WrapperKind = X86ISD::WrapperRIP;
5167 assert(model == TLSModel::InitialExec);
5168 OperandFlags = X86II::MO_INDNTPOFF;
5171 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5173 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5174 GA->getOffset(), OperandFlags);
5175 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5177 if (model == TLSModel::InitialExec)
5178 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5179 PseudoSourceValue::getGOT(), 0, false, false, 0);
5181 // The address of the thread local variable is the add of the thread
5182 // pointer with the offset of the variable.
5183 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5187 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5188 // TODO: implement the "local dynamic" model
5189 // TODO: implement the "initial exec"model for pic executables
5190 assert(Subtarget->isTargetELF() &&
5191 "TLS not implemented for non-ELF targets");
5192 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5193 const GlobalValue *GV = GA->getGlobal();
5195 // If GV is an alias then use the aliasee for determining
5196 // thread-localness.
5197 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5198 GV = GA->resolveAliasedGlobal(false);
5200 TLSModel::Model model = getTLSModel(GV,
5201 getTargetMachine().getRelocationModel());
5204 case TLSModel::GeneralDynamic:
5205 case TLSModel::LocalDynamic: // not implemented
5206 if (Subtarget->is64Bit())
5207 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5208 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5210 case TLSModel::InitialExec:
5211 case TLSModel::LocalExec:
5212 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5213 Subtarget->is64Bit());
5216 llvm_unreachable("Unreachable");
5221 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5222 /// take a 2 x i32 value to shift plus a shift amount.
5223 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5224 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5225 EVT VT = Op.getValueType();
5226 unsigned VTBits = VT.getSizeInBits();
5227 DebugLoc dl = Op.getDebugLoc();
5228 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5229 SDValue ShOpLo = Op.getOperand(0);
5230 SDValue ShOpHi = Op.getOperand(1);
5231 SDValue ShAmt = Op.getOperand(2);
5232 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5233 DAG.getConstant(VTBits - 1, MVT::i8))
5234 : DAG.getConstant(0, VT);
5237 if (Op.getOpcode() == ISD::SHL_PARTS) {
5238 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5239 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5241 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5242 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5245 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5246 DAG.getConstant(VTBits, MVT::i8));
5247 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
5248 AndNode, DAG.getConstant(0, MVT::i8));
5251 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5252 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5253 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5255 if (Op.getOpcode() == ISD::SHL_PARTS) {
5256 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5257 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5259 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5260 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5263 SDValue Ops[2] = { Lo, Hi };
5264 return DAG.getMergeValues(Ops, 2, dl);
5267 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5268 EVT SrcVT = Op.getOperand(0).getValueType();
5270 if (SrcVT.isVector()) {
5271 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5277 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5278 "Unknown SINT_TO_FP to lower!");
5280 // These are really Legal; return the operand so the caller accepts it as
5282 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5284 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5285 Subtarget->is64Bit()) {
5289 DebugLoc dl = Op.getDebugLoc();
5290 unsigned Size = SrcVT.getSizeInBits()/8;
5291 MachineFunction &MF = DAG.getMachineFunction();
5292 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5293 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5294 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5296 PseudoSourceValue::getFixedStack(SSFI), 0,
5298 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5301 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5303 SelectionDAG &DAG) {
5305 DebugLoc dl = Op.getDebugLoc();
5307 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5309 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5311 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5312 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5313 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5314 Tys, Ops, array_lengthof(Ops));
5317 Chain = Result.getValue(1);
5318 SDValue InFlag = Result.getValue(2);
5320 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5321 // shouldn't be necessary except that RFP cannot be live across
5322 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5323 MachineFunction &MF = DAG.getMachineFunction();
5324 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5325 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5326 Tys = DAG.getVTList(MVT::Other);
5328 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5330 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5331 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5332 PseudoSourceValue::getFixedStack(SSFI), 0,
5339 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5340 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5341 // This algorithm is not obvious. Here it is in C code, more or less:
5343 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5344 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5345 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5347 // Copy ints to xmm registers.
5348 __m128i xh = _mm_cvtsi32_si128( hi );
5349 __m128i xl = _mm_cvtsi32_si128( lo );
5351 // Combine into low half of a single xmm register.
5352 __m128i x = _mm_unpacklo_epi32( xh, xl );
5356 // Merge in appropriate exponents to give the integer bits the right
5358 x = _mm_unpacklo_epi32( x, exp );
5360 // Subtract away the biases to deal with the IEEE-754 double precision
5362 d = _mm_sub_pd( (__m128d) x, bias );
5364 // All conversions up to here are exact. The correctly rounded result is
5365 // calculated using the current rounding mode using the following
5367 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5368 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5369 // store doesn't really need to be here (except
5370 // maybe to zero the other double)
5375 DebugLoc dl = Op.getDebugLoc();
5376 LLVMContext *Context = DAG.getContext();
5378 // Build some magic constants.
5379 std::vector<Constant*> CV0;
5380 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5381 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5382 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5383 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5384 Constant *C0 = ConstantVector::get(CV0);
5385 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5387 std::vector<Constant*> CV1;
5389 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5391 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5392 Constant *C1 = ConstantVector::get(CV1);
5393 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5395 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5396 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5398 DAG.getIntPtrConstant(1)));
5399 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5400 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5402 DAG.getIntPtrConstant(0)));
5403 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5404 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5405 PseudoSourceValue::getConstantPool(), 0,
5407 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5408 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5409 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5410 PseudoSourceValue::getConstantPool(), 0,
5412 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5414 // Add the halves; easiest way is to swap them into another reg first.
5415 int ShufMask[2] = { 1, -1 };
5416 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5417 DAG.getUNDEF(MVT::v2f64), ShufMask);
5418 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5419 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5420 DAG.getIntPtrConstant(0));
5423 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5424 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5425 DebugLoc dl = Op.getDebugLoc();
5426 // FP constant to bias correct the final result.
5427 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5430 // Load the 32-bit value into an XMM register.
5431 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5432 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5434 DAG.getIntPtrConstant(0)));
5436 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5437 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5438 DAG.getIntPtrConstant(0));
5440 // Or the load with the bias.
5441 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5442 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5443 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5445 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5446 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5447 MVT::v2f64, Bias)));
5448 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5449 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5450 DAG.getIntPtrConstant(0));
5452 // Subtract the bias.
5453 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5455 // Handle final rounding.
5456 EVT DestVT = Op.getValueType();
5458 if (DestVT.bitsLT(MVT::f64)) {
5459 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5460 DAG.getIntPtrConstant(0));
5461 } else if (DestVT.bitsGT(MVT::f64)) {
5462 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5465 // Handle final rounding.
5469 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5470 SDValue N0 = Op.getOperand(0);
5471 DebugLoc dl = Op.getDebugLoc();
5473 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5474 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5475 // the optimization here.
5476 if (DAG.SignBitIsZero(N0))
5477 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5479 EVT SrcVT = N0.getValueType();
5480 if (SrcVT == MVT::i64) {
5481 // We only handle SSE2 f64 target here; caller can expand the rest.
5482 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5485 return LowerUINT_TO_FP_i64(Op, DAG);
5486 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5487 return LowerUINT_TO_FP_i32(Op, DAG);
5490 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5492 // Make a 64-bit buffer, and use it to build an FILD.
5493 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5494 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5495 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5496 getPointerTy(), StackSlot, WordOff);
5497 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5498 StackSlot, NULL, 0, false, false, 0);
5499 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5500 OffsetSlot, NULL, 0, false, false, 0);
5501 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5504 std::pair<SDValue,SDValue> X86TargetLowering::
5505 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5506 DebugLoc dl = Op.getDebugLoc();
5508 EVT DstTy = Op.getValueType();
5511 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5515 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5516 DstTy.getSimpleVT() >= MVT::i16 &&
5517 "Unknown FP_TO_SINT to lower!");
5519 // These are really Legal.
5520 if (DstTy == MVT::i32 &&
5521 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5522 return std::make_pair(SDValue(), SDValue());
5523 if (Subtarget->is64Bit() &&
5524 DstTy == MVT::i64 &&
5525 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5526 return std::make_pair(SDValue(), SDValue());
5528 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5530 MachineFunction &MF = DAG.getMachineFunction();
5531 unsigned MemSize = DstTy.getSizeInBits()/8;
5532 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5533 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5536 switch (DstTy.getSimpleVT().SimpleTy) {
5537 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5538 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5539 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5540 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5543 SDValue Chain = DAG.getEntryNode();
5544 SDValue Value = Op.getOperand(0);
5545 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5546 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5547 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5548 PseudoSourceValue::getFixedStack(SSFI), 0,
5550 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5552 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5554 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5555 Chain = Value.getValue(1);
5556 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5557 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5560 // Build the FP_TO_INT*_IN_MEM
5561 SDValue Ops[] = { Chain, Value, StackSlot };
5562 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5564 return std::make_pair(FIST, StackSlot);
5567 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5568 if (Op.getValueType().isVector()) {
5569 if (Op.getValueType() == MVT::v2i32 &&
5570 Op.getOperand(0).getValueType() == MVT::v2f64) {
5576 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5577 SDValue FIST = Vals.first, StackSlot = Vals.second;
5578 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5579 if (FIST.getNode() == 0) return Op;
5582 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5583 FIST, StackSlot, NULL, 0, false, false, 0);
5586 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5587 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5588 SDValue FIST = Vals.first, StackSlot = Vals.second;
5589 assert(FIST.getNode() && "Unexpected failure");
5592 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5593 FIST, StackSlot, NULL, 0, false, false, 0);
5596 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5597 LLVMContext *Context = DAG.getContext();
5598 DebugLoc dl = Op.getDebugLoc();
5599 EVT VT = Op.getValueType();
5602 EltVT = VT.getVectorElementType();
5603 std::vector<Constant*> CV;
5604 if (EltVT == MVT::f64) {
5605 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5609 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5615 Constant *C = ConstantVector::get(CV);
5616 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5617 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5618 PseudoSourceValue::getConstantPool(), 0,
5620 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5623 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5624 LLVMContext *Context = DAG.getContext();
5625 DebugLoc dl = Op.getDebugLoc();
5626 EVT VT = Op.getValueType();
5629 EltVT = VT.getVectorElementType();
5630 std::vector<Constant*> CV;
5631 if (EltVT == MVT::f64) {
5632 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5636 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5642 Constant *C = ConstantVector::get(CV);
5643 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5644 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5645 PseudoSourceValue::getConstantPool(), 0,
5647 if (VT.isVector()) {
5648 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5649 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5650 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5652 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5654 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5658 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5659 LLVMContext *Context = DAG.getContext();
5660 SDValue Op0 = Op.getOperand(0);
5661 SDValue Op1 = Op.getOperand(1);
5662 DebugLoc dl = Op.getDebugLoc();
5663 EVT VT = Op.getValueType();
5664 EVT SrcVT = Op1.getValueType();
5666 // If second operand is smaller, extend it first.
5667 if (SrcVT.bitsLT(VT)) {
5668 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5671 // And if it is bigger, shrink it first.
5672 if (SrcVT.bitsGT(VT)) {
5673 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5677 // At this point the operands and the result should have the same
5678 // type, and that won't be f80 since that is not custom lowered.
5680 // First get the sign bit of second operand.
5681 std::vector<Constant*> CV;
5682 if (SrcVT == MVT::f64) {
5683 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5684 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5686 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5687 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5688 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5689 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5691 Constant *C = ConstantVector::get(CV);
5692 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5693 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5694 PseudoSourceValue::getConstantPool(), 0,
5696 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5698 // Shift sign bit right or left if the two operands have different types.
5699 if (SrcVT.bitsGT(VT)) {
5700 // Op0 is MVT::f32, Op1 is MVT::f64.
5701 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5702 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5703 DAG.getConstant(32, MVT::i32));
5704 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5705 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5706 DAG.getIntPtrConstant(0));
5709 // Clear first operand sign bit.
5711 if (VT == MVT::f64) {
5712 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5713 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5715 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5716 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5717 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5718 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5720 C = ConstantVector::get(CV);
5721 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5722 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5723 PseudoSourceValue::getConstantPool(), 0,
5725 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5727 // Or the value with the sign bit.
5728 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5731 /// Emit nodes that will be selected as "test Op0,Op0", or something
5733 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5734 SelectionDAG &DAG) {
5735 DebugLoc dl = Op.getDebugLoc();
5737 // CF and OF aren't always set the way we want. Determine which
5738 // of these we need.
5739 bool NeedCF = false;
5740 bool NeedOF = false;
5742 case X86::COND_A: case X86::COND_AE:
5743 case X86::COND_B: case X86::COND_BE:
5746 case X86::COND_G: case X86::COND_GE:
5747 case X86::COND_L: case X86::COND_LE:
5748 case X86::COND_O: case X86::COND_NO:
5754 // See if we can use the EFLAGS value from the operand instead of
5755 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5756 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5757 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5758 unsigned Opcode = 0;
5759 unsigned NumOperands = 0;
5760 switch (Op.getNode()->getOpcode()) {
5762 // Due to an isel shortcoming, be conservative if this add is likely to
5763 // be selected as part of a load-modify-store instruction. When the root
5764 // node in a match is a store, isel doesn't know how to remap non-chain
5765 // non-flag uses of other nodes in the match, such as the ADD in this
5766 // case. This leads to the ADD being left around and reselected, with
5767 // the result being two adds in the output.
5768 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5769 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5770 if (UI->getOpcode() == ISD::STORE)
5772 if (ConstantSDNode *C =
5773 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5774 // An add of one will be selected as an INC.
5775 if (C->getAPIntValue() == 1) {
5776 Opcode = X86ISD::INC;
5780 // An add of negative one (subtract of one) will be selected as a DEC.
5781 if (C->getAPIntValue().isAllOnesValue()) {
5782 Opcode = X86ISD::DEC;
5787 // Otherwise use a regular EFLAGS-setting add.
5788 Opcode = X86ISD::ADD;
5792 // If the primary and result isn't used, don't bother using X86ISD::AND,
5793 // because a TEST instruction will be better.
5794 bool NonFlagUse = false;
5795 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5796 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5798 unsigned UOpNo = UI.getOperandNo();
5799 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5800 // Look pass truncate.
5801 UOpNo = User->use_begin().getOperandNo();
5802 User = *User->use_begin();
5804 if (User->getOpcode() != ISD::BRCOND &&
5805 User->getOpcode() != ISD::SETCC &&
5806 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5818 // Due to the ISEL shortcoming noted above, be conservative if this op is
5819 // likely to be selected as part of a load-modify-store instruction.
5820 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5821 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5822 if (UI->getOpcode() == ISD::STORE)
5824 // Otherwise use a regular EFLAGS-setting instruction.
5825 switch (Op.getNode()->getOpcode()) {
5826 case ISD::SUB: Opcode = X86ISD::SUB; break;
5827 case ISD::OR: Opcode = X86ISD::OR; break;
5828 case ISD::XOR: Opcode = X86ISD::XOR; break;
5829 case ISD::AND: Opcode = X86ISD::AND; break;
5830 default: llvm_unreachable("unexpected operator!");
5841 return SDValue(Op.getNode(), 1);
5847 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5848 SmallVector<SDValue, 4> Ops;
5849 for (unsigned i = 0; i != NumOperands; ++i)
5850 Ops.push_back(Op.getOperand(i));
5851 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5852 DAG.ReplaceAllUsesWith(Op, New);
5853 return SDValue(New.getNode(), 1);
5857 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5858 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5859 DAG.getConstant(0, Op.getValueType()));
5862 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5864 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5865 SelectionDAG &DAG) {
5866 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5867 if (C->getAPIntValue() == 0)
5868 return EmitTest(Op0, X86CC, DAG);
5870 DebugLoc dl = Op0.getDebugLoc();
5871 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5874 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5875 /// if it's possible.
5876 static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
5877 DebugLoc dl, SelectionDAG &DAG) {
5879 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5880 if (ConstantSDNode *Op010C =
5881 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5882 if (Op010C->getZExtValue() == 1) {
5883 LHS = Op0.getOperand(0);
5884 RHS = Op0.getOperand(1).getOperand(1);
5886 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5887 if (ConstantSDNode *Op000C =
5888 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5889 if (Op000C->getZExtValue() == 1) {
5890 LHS = Op0.getOperand(1);
5891 RHS = Op0.getOperand(0).getOperand(1);
5893 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5894 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5895 SDValue AndLHS = Op0.getOperand(0);
5896 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5897 LHS = AndLHS.getOperand(0);
5898 RHS = AndLHS.getOperand(1);
5902 if (LHS.getNode()) {
5903 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5904 // instruction. Since the shift amount is in-range-or-undefined, we know
5905 // that doing a bittest on the i16 value is ok. We extend to i32 because
5906 // the encoding for the i16 version is larger than the i32 version.
5907 if (LHS.getValueType() == MVT::i8)
5908 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5910 // If the operand types disagree, extend the shift amount to match. Since
5911 // BT ignores high bits (like shifts) we can use anyextend.
5912 if (LHS.getValueType() != RHS.getValueType())
5913 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5915 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5916 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5917 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5918 DAG.getConstant(Cond, MVT::i8), BT);
5924 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5925 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5926 SDValue Op0 = Op.getOperand(0);
5927 SDValue Op1 = Op.getOperand(1);
5928 DebugLoc dl = Op.getDebugLoc();
5929 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5931 // Optimize to BT if possible.
5932 // Lower (X & (1 << N)) == 0 to BT(X, N).
5933 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5934 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5935 if (Op0.getOpcode() == ISD::AND &&
5937 Op1.getOpcode() == ISD::Constant &&
5938 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5939 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5940 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5941 if (NewSetCC.getNode())
5945 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5946 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5947 if (X86CC == X86::COND_INVALID)
5950 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5952 // Use sbb x, x to materialize carry bit into a GPR.
5953 if (X86CC == X86::COND_B)
5954 return DAG.getNode(ISD::AND, dl, MVT::i8,
5955 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5956 DAG.getConstant(X86CC, MVT::i8), Cond),
5957 DAG.getConstant(1, MVT::i8));
5959 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5960 DAG.getConstant(X86CC, MVT::i8), Cond);
5963 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5965 SDValue Op0 = Op.getOperand(0);
5966 SDValue Op1 = Op.getOperand(1);
5967 SDValue CC = Op.getOperand(2);
5968 EVT VT = Op.getValueType();
5969 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5970 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5971 DebugLoc dl = Op.getDebugLoc();
5975 EVT VT0 = Op0.getValueType();
5976 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5977 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5980 switch (SetCCOpcode) {
5983 case ISD::SETEQ: SSECC = 0; break;
5985 case ISD::SETGT: Swap = true; // Fallthrough
5987 case ISD::SETOLT: SSECC = 1; break;
5989 case ISD::SETGE: Swap = true; // Fallthrough
5991 case ISD::SETOLE: SSECC = 2; break;
5992 case ISD::SETUO: SSECC = 3; break;
5994 case ISD::SETNE: SSECC = 4; break;
5995 case ISD::SETULE: Swap = true;
5996 case ISD::SETUGE: SSECC = 5; break;
5997 case ISD::SETULT: Swap = true;
5998 case ISD::SETUGT: SSECC = 6; break;
5999 case ISD::SETO: SSECC = 7; break;
6002 std::swap(Op0, Op1);
6004 // In the two special cases we can't handle, emit two comparisons.
6006 if (SetCCOpcode == ISD::SETUEQ) {
6008 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6009 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6010 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6012 else if (SetCCOpcode == ISD::SETONE) {
6014 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6015 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6016 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6018 llvm_unreachable("Illegal FP comparison");
6020 // Handle all other FP comparisons here.
6021 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6024 // We are handling one of the integer comparisons here. Since SSE only has
6025 // GT and EQ comparisons for integer, swapping operands and multiple
6026 // operations may be required for some comparisons.
6027 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6028 bool Swap = false, Invert = false, FlipSigns = false;
6030 switch (VT.getSimpleVT().SimpleTy) {
6033 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6035 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6037 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6038 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6041 switch (SetCCOpcode) {
6043 case ISD::SETNE: Invert = true;
6044 case ISD::SETEQ: Opc = EQOpc; break;
6045 case ISD::SETLT: Swap = true;
6046 case ISD::SETGT: Opc = GTOpc; break;
6047 case ISD::SETGE: Swap = true;
6048 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6049 case ISD::SETULT: Swap = true;
6050 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6051 case ISD::SETUGE: Swap = true;
6052 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6055 std::swap(Op0, Op1);
6057 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6058 // bits of the inputs before performing those operations.
6060 EVT EltVT = VT.getVectorElementType();
6061 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6063 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6064 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6066 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6067 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6070 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6072 // If the logical-not of the result is required, perform that now.
6074 Result = DAG.getNOT(dl, Result, VT);
6079 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6080 static bool isX86LogicalCmp(SDValue Op) {
6081 unsigned Opc = Op.getNode()->getOpcode();
6082 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6084 if (Op.getResNo() == 1 &&
6085 (Opc == X86ISD::ADD ||
6086 Opc == X86ISD::SUB ||
6087 Opc == X86ISD::SMUL ||
6088 Opc == X86ISD::UMUL ||
6089 Opc == X86ISD::INC ||
6090 Opc == X86ISD::DEC ||
6091 Opc == X86ISD::OR ||
6092 Opc == X86ISD::XOR ||
6093 Opc == X86ISD::AND))
6099 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6100 bool addTest = true;
6101 SDValue Cond = Op.getOperand(0);
6102 DebugLoc dl = Op.getDebugLoc();
6105 if (Cond.getOpcode() == ISD::SETCC) {
6106 SDValue NewCond = LowerSETCC(Cond, DAG);
6107 if (NewCond.getNode())
6111 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6112 SDValue Op1 = Op.getOperand(1);
6113 SDValue Op2 = Op.getOperand(2);
6114 if (Cond.getOpcode() == X86ISD::SETCC &&
6115 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6116 SDValue Cmp = Cond.getOperand(1);
6117 if (Cmp.getOpcode() == X86ISD::CMP) {
6118 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6119 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6120 ConstantSDNode *RHSC =
6121 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6122 if (N1C && N1C->isAllOnesValue() &&
6123 N2C && N2C->isNullValue() &&
6124 RHSC && RHSC->isNullValue()) {
6125 SDValue CmpOp0 = Cmp.getOperand(0);
6126 Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
6127 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6128 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6129 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6134 // Look pass (and (setcc_carry (cmp ...)), 1).
6135 if (Cond.getOpcode() == ISD::AND &&
6136 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6137 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6138 if (C && C->getAPIntValue() == 1)
6139 Cond = Cond.getOperand(0);
6142 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6143 // setting operand in place of the X86ISD::SETCC.
6144 if (Cond.getOpcode() == X86ISD::SETCC ||
6145 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6146 CC = Cond.getOperand(0);
6148 SDValue Cmp = Cond.getOperand(1);
6149 unsigned Opc = Cmp.getOpcode();
6150 EVT VT = Op.getValueType();
6152 bool IllegalFPCMov = false;
6153 if (VT.isFloatingPoint() && !VT.isVector() &&
6154 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6155 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6157 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6158 Opc == X86ISD::BT) { // FIXME
6165 // Look pass the truncate.
6166 if (Cond.getOpcode() == ISD::TRUNCATE)
6167 Cond = Cond.getOperand(0);
6169 // We know the result of AND is compared against zero. Try to match
6171 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6172 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6173 if (NewSetCC.getNode()) {
6174 CC = NewSetCC.getOperand(0);
6175 Cond = NewSetCC.getOperand(1);
6182 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6183 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6186 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6187 // condition is true.
6188 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6189 SDValue Ops[] = { Op2, Op1, CC, Cond };
6190 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6193 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6194 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6195 // from the AND / OR.
6196 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6197 Opc = Op.getOpcode();
6198 if (Opc != ISD::OR && Opc != ISD::AND)
6200 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6201 Op.getOperand(0).hasOneUse() &&
6202 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6203 Op.getOperand(1).hasOneUse());
6206 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6207 // 1 and that the SETCC node has a single use.
6208 static bool isXor1OfSetCC(SDValue Op) {
6209 if (Op.getOpcode() != ISD::XOR)
6211 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6212 if (N1C && N1C->getAPIntValue() == 1) {
6213 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6214 Op.getOperand(0).hasOneUse();
6219 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6220 bool addTest = true;
6221 SDValue Chain = Op.getOperand(0);
6222 SDValue Cond = Op.getOperand(1);
6223 SDValue Dest = Op.getOperand(2);
6224 DebugLoc dl = Op.getDebugLoc();
6227 if (Cond.getOpcode() == ISD::SETCC) {
6228 SDValue NewCond = LowerSETCC(Cond, DAG);
6229 if (NewCond.getNode())
6233 // FIXME: LowerXALUO doesn't handle these!!
6234 else if (Cond.getOpcode() == X86ISD::ADD ||
6235 Cond.getOpcode() == X86ISD::SUB ||
6236 Cond.getOpcode() == X86ISD::SMUL ||
6237 Cond.getOpcode() == X86ISD::UMUL)
6238 Cond = LowerXALUO(Cond, DAG);
6241 // Look pass (and (setcc_carry (cmp ...)), 1).
6242 if (Cond.getOpcode() == ISD::AND &&
6243 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6244 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6245 if (C && C->getAPIntValue() == 1)
6246 Cond = Cond.getOperand(0);
6249 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6250 // setting operand in place of the X86ISD::SETCC.
6251 if (Cond.getOpcode() == X86ISD::SETCC ||
6252 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6253 CC = Cond.getOperand(0);
6255 SDValue Cmp = Cond.getOperand(1);
6256 unsigned Opc = Cmp.getOpcode();
6257 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6258 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6262 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6266 // These can only come from an arithmetic instruction with overflow,
6267 // e.g. SADDO, UADDO.
6268 Cond = Cond.getNode()->getOperand(1);
6275 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6276 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6277 if (CondOpc == ISD::OR) {
6278 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6279 // two branches instead of an explicit OR instruction with a
6281 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6282 isX86LogicalCmp(Cmp)) {
6283 CC = Cond.getOperand(0).getOperand(0);
6284 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6285 Chain, Dest, CC, Cmp);
6286 CC = Cond.getOperand(1).getOperand(0);
6290 } else { // ISD::AND
6291 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6292 // two branches instead of an explicit AND instruction with a
6293 // separate test. However, we only do this if this block doesn't
6294 // have a fall-through edge, because this requires an explicit
6295 // jmp when the condition is false.
6296 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6297 isX86LogicalCmp(Cmp) &&
6298 Op.getNode()->hasOneUse()) {
6299 X86::CondCode CCode =
6300 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6301 CCode = X86::GetOppositeBranchCondition(CCode);
6302 CC = DAG.getConstant(CCode, MVT::i8);
6303 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6304 // Look for an unconditional branch following this conditional branch.
6305 // We need this because we need to reverse the successors in order
6306 // to implement FCMP_OEQ.
6307 if (User.getOpcode() == ISD::BR) {
6308 SDValue FalseBB = User.getOperand(1);
6310 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6311 assert(NewBR == User);
6314 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6315 Chain, Dest, CC, Cmp);
6316 X86::CondCode CCode =
6317 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6318 CCode = X86::GetOppositeBranchCondition(CCode);
6319 CC = DAG.getConstant(CCode, MVT::i8);
6325 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6326 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6327 // It should be transformed during dag combiner except when the condition
6328 // is set by a arithmetics with overflow node.
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 Cond = Cond.getOperand(0).getOperand(1);
6339 // Look pass the truncate.
6340 if (Cond.getOpcode() == ISD::TRUNCATE)
6341 Cond = Cond.getOperand(0);
6343 // We know the result of AND is compared against zero. Try to match
6345 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6346 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6347 if (NewSetCC.getNode()) {
6348 CC = NewSetCC.getOperand(0);
6349 Cond = NewSetCC.getOperand(1);
6356 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6357 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6359 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6360 Chain, Dest, CC, Cond);
6364 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6365 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6366 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6367 // that the guard pages used by the OS virtual memory manager are allocated in
6368 // correct sequence.
6370 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6371 SelectionDAG &DAG) {
6372 assert(Subtarget->isTargetCygMing() &&
6373 "This should be used only on Cygwin/Mingw targets");
6374 DebugLoc dl = Op.getDebugLoc();
6377 SDValue Chain = Op.getOperand(0);
6378 SDValue Size = Op.getOperand(1);
6379 // FIXME: Ensure alignment here
6383 EVT IntPtr = getPointerTy();
6384 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6386 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6388 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6389 Flag = Chain.getValue(1);
6391 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6392 SDValue Ops[] = { Chain,
6393 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6394 DAG.getRegister(X86::EAX, IntPtr),
6395 DAG.getRegister(X86StackPtr, SPTy),
6397 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6398 Flag = Chain.getValue(1);
6400 Chain = DAG.getCALLSEQ_END(Chain,
6401 DAG.getIntPtrConstant(0, true),
6402 DAG.getIntPtrConstant(0, true),
6405 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6407 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6408 return DAG.getMergeValues(Ops1, 2, dl);
6412 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6414 SDValue Dst, SDValue Src,
6415 SDValue Size, unsigned Align,
6417 uint64_t DstSVOff) {
6418 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6420 // If not DWORD aligned or size is more than the threshold, call the library.
6421 // The libc version is likely to be faster for these cases. It can use the
6422 // address value and run time information about the CPU.
6423 if ((Align & 3) != 0 ||
6425 ConstantSize->getZExtValue() >
6426 getSubtarget()->getMaxInlineSizeThreshold()) {
6427 SDValue InFlag(0, 0);
6429 // Check to see if there is a specialized entry-point for memory zeroing.
6430 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6432 if (const char *bzeroEntry = V &&
6433 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6434 EVT IntPtr = getPointerTy();
6435 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6436 TargetLowering::ArgListTy Args;
6437 TargetLowering::ArgListEntry Entry;
6439 Entry.Ty = IntPtrTy;
6440 Args.push_back(Entry);
6442 Args.push_back(Entry);
6443 std::pair<SDValue,SDValue> CallResult =
6444 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6445 false, false, false, false,
6446 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6447 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6448 DAG.GetOrdering(Chain.getNode()));
6449 return CallResult.second;
6452 // Otherwise have the target-independent code call memset.
6456 uint64_t SizeVal = ConstantSize->getZExtValue();
6457 SDValue InFlag(0, 0);
6460 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6461 unsigned BytesLeft = 0;
6462 bool TwoRepStos = false;
6465 uint64_t Val = ValC->getZExtValue() & 255;
6467 // If the value is a constant, then we can potentially use larger sets.
6468 switch (Align & 3) {
6469 case 2: // WORD aligned
6472 Val = (Val << 8) | Val;
6474 case 0: // DWORD aligned
6477 Val = (Val << 8) | Val;
6478 Val = (Val << 16) | Val;
6479 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6482 Val = (Val << 32) | Val;
6485 default: // Byte aligned
6488 Count = DAG.getIntPtrConstant(SizeVal);
6492 if (AVT.bitsGT(MVT::i8)) {
6493 unsigned UBytes = AVT.getSizeInBits() / 8;
6494 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6495 BytesLeft = SizeVal % UBytes;
6498 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6500 InFlag = Chain.getValue(1);
6503 Count = DAG.getIntPtrConstant(SizeVal);
6504 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6505 InFlag = Chain.getValue(1);
6508 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6511 InFlag = Chain.getValue(1);
6512 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6515 InFlag = Chain.getValue(1);
6517 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6518 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6519 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6522 InFlag = Chain.getValue(1);
6524 EVT CVT = Count.getValueType();
6525 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6526 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6527 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6530 InFlag = Chain.getValue(1);
6531 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6532 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6533 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6534 } else if (BytesLeft) {
6535 // Handle the last 1 - 7 bytes.
6536 unsigned Offset = SizeVal - BytesLeft;
6537 EVT AddrVT = Dst.getValueType();
6538 EVT SizeVT = Size.getValueType();
6540 Chain = DAG.getMemset(Chain, dl,
6541 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6542 DAG.getConstant(Offset, AddrVT)),
6544 DAG.getConstant(BytesLeft, SizeVT),
6545 Align, DstSV, DstSVOff + Offset);
6548 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6553 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6554 SDValue Chain, SDValue Dst, SDValue Src,
6555 SDValue Size, unsigned Align,
6557 const Value *DstSV, uint64_t DstSVOff,
6558 const Value *SrcSV, uint64_t SrcSVOff) {
6559 // This requires the copy size to be a constant, preferrably
6560 // within a subtarget-specific limit.
6561 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6564 uint64_t SizeVal = ConstantSize->getZExtValue();
6565 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6568 /// If not DWORD aligned, call the library.
6569 if ((Align & 3) != 0)
6574 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6577 unsigned UBytes = AVT.getSizeInBits() / 8;
6578 unsigned CountVal = SizeVal / UBytes;
6579 SDValue Count = DAG.getIntPtrConstant(CountVal);
6580 unsigned BytesLeft = SizeVal % UBytes;
6582 SDValue InFlag(0, 0);
6583 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6586 InFlag = Chain.getValue(1);
6587 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6590 InFlag = Chain.getValue(1);
6591 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6594 InFlag = Chain.getValue(1);
6596 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6597 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6598 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6599 array_lengthof(Ops));
6601 SmallVector<SDValue, 4> Results;
6602 Results.push_back(RepMovs);
6604 // Handle the last 1 - 7 bytes.
6605 unsigned Offset = SizeVal - BytesLeft;
6606 EVT DstVT = Dst.getValueType();
6607 EVT SrcVT = Src.getValueType();
6608 EVT SizeVT = Size.getValueType();
6609 Results.push_back(DAG.getMemcpy(Chain, dl,
6610 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6611 DAG.getConstant(Offset, DstVT)),
6612 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6613 DAG.getConstant(Offset, SrcVT)),
6614 DAG.getConstant(BytesLeft, SizeVT),
6615 Align, AlwaysInline,
6616 DstSV, DstSVOff + Offset,
6617 SrcSV, SrcSVOff + Offset));
6620 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6621 &Results[0], Results.size());
6624 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6625 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6626 DebugLoc dl = Op.getDebugLoc();
6628 if (!Subtarget->is64Bit()) {
6629 // vastart just stores the address of the VarArgsFrameIndex slot into the
6630 // memory location argument.
6631 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6632 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6637 // gp_offset (0 - 6 * 8)
6638 // fp_offset (48 - 48 + 8 * 16)
6639 // overflow_arg_area (point to parameters coming in memory).
6641 SmallVector<SDValue, 8> MemOps;
6642 SDValue FIN = Op.getOperand(1);
6644 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6645 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6646 FIN, SV, 0, false, false, 0);
6647 MemOps.push_back(Store);
6650 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6651 FIN, DAG.getIntPtrConstant(4));
6652 Store = DAG.getStore(Op.getOperand(0), dl,
6653 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6654 FIN, SV, 0, false, false, 0);
6655 MemOps.push_back(Store);
6657 // Store ptr to overflow_arg_area
6658 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6659 FIN, DAG.getIntPtrConstant(4));
6660 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6661 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6663 MemOps.push_back(Store);
6665 // Store ptr to reg_save_area.
6666 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6667 FIN, DAG.getIntPtrConstant(8));
6668 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6669 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6671 MemOps.push_back(Store);
6672 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6673 &MemOps[0], MemOps.size());
6676 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6677 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6678 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6679 SDValue Chain = Op.getOperand(0);
6680 SDValue SrcPtr = Op.getOperand(1);
6681 SDValue SrcSV = Op.getOperand(2);
6683 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6687 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6688 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6689 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6690 SDValue Chain = Op.getOperand(0);
6691 SDValue DstPtr = Op.getOperand(1);
6692 SDValue SrcPtr = Op.getOperand(2);
6693 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6694 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6695 DebugLoc dl = Op.getDebugLoc();
6697 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6698 DAG.getIntPtrConstant(24), 8, false,
6699 DstSV, 0, SrcSV, 0);
6703 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6704 DebugLoc dl = Op.getDebugLoc();
6705 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6707 default: return SDValue(); // Don't custom lower most intrinsics.
6708 // Comparison intrinsics.
6709 case Intrinsic::x86_sse_comieq_ss:
6710 case Intrinsic::x86_sse_comilt_ss:
6711 case Intrinsic::x86_sse_comile_ss:
6712 case Intrinsic::x86_sse_comigt_ss:
6713 case Intrinsic::x86_sse_comige_ss:
6714 case Intrinsic::x86_sse_comineq_ss:
6715 case Intrinsic::x86_sse_ucomieq_ss:
6716 case Intrinsic::x86_sse_ucomilt_ss:
6717 case Intrinsic::x86_sse_ucomile_ss:
6718 case Intrinsic::x86_sse_ucomigt_ss:
6719 case Intrinsic::x86_sse_ucomige_ss:
6720 case Intrinsic::x86_sse_ucomineq_ss:
6721 case Intrinsic::x86_sse2_comieq_sd:
6722 case Intrinsic::x86_sse2_comilt_sd:
6723 case Intrinsic::x86_sse2_comile_sd:
6724 case Intrinsic::x86_sse2_comigt_sd:
6725 case Intrinsic::x86_sse2_comige_sd:
6726 case Intrinsic::x86_sse2_comineq_sd:
6727 case Intrinsic::x86_sse2_ucomieq_sd:
6728 case Intrinsic::x86_sse2_ucomilt_sd:
6729 case Intrinsic::x86_sse2_ucomile_sd:
6730 case Intrinsic::x86_sse2_ucomigt_sd:
6731 case Intrinsic::x86_sse2_ucomige_sd:
6732 case Intrinsic::x86_sse2_ucomineq_sd: {
6734 ISD::CondCode CC = ISD::SETCC_INVALID;
6737 case Intrinsic::x86_sse_comieq_ss:
6738 case Intrinsic::x86_sse2_comieq_sd:
6742 case Intrinsic::x86_sse_comilt_ss:
6743 case Intrinsic::x86_sse2_comilt_sd:
6747 case Intrinsic::x86_sse_comile_ss:
6748 case Intrinsic::x86_sse2_comile_sd:
6752 case Intrinsic::x86_sse_comigt_ss:
6753 case Intrinsic::x86_sse2_comigt_sd:
6757 case Intrinsic::x86_sse_comige_ss:
6758 case Intrinsic::x86_sse2_comige_sd:
6762 case Intrinsic::x86_sse_comineq_ss:
6763 case Intrinsic::x86_sse2_comineq_sd:
6767 case Intrinsic::x86_sse_ucomieq_ss:
6768 case Intrinsic::x86_sse2_ucomieq_sd:
6769 Opc = X86ISD::UCOMI;
6772 case Intrinsic::x86_sse_ucomilt_ss:
6773 case Intrinsic::x86_sse2_ucomilt_sd:
6774 Opc = X86ISD::UCOMI;
6777 case Intrinsic::x86_sse_ucomile_ss:
6778 case Intrinsic::x86_sse2_ucomile_sd:
6779 Opc = X86ISD::UCOMI;
6782 case Intrinsic::x86_sse_ucomigt_ss:
6783 case Intrinsic::x86_sse2_ucomigt_sd:
6784 Opc = X86ISD::UCOMI;
6787 case Intrinsic::x86_sse_ucomige_ss:
6788 case Intrinsic::x86_sse2_ucomige_sd:
6789 Opc = X86ISD::UCOMI;
6792 case Intrinsic::x86_sse_ucomineq_ss:
6793 case Intrinsic::x86_sse2_ucomineq_sd:
6794 Opc = X86ISD::UCOMI;
6799 SDValue LHS = Op.getOperand(1);
6800 SDValue RHS = Op.getOperand(2);
6801 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6802 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6803 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6804 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6805 DAG.getConstant(X86CC, MVT::i8), Cond);
6806 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6808 // ptest intrinsics. The intrinsic these come from are designed to return
6809 // an integer value, not just an instruction so lower it to the ptest
6810 // pattern and a setcc for the result.
6811 case Intrinsic::x86_sse41_ptestz:
6812 case Intrinsic::x86_sse41_ptestc:
6813 case Intrinsic::x86_sse41_ptestnzc:{
6816 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6817 case Intrinsic::x86_sse41_ptestz:
6819 X86CC = X86::COND_E;
6821 case Intrinsic::x86_sse41_ptestc:
6823 X86CC = X86::COND_B;
6825 case Intrinsic::x86_sse41_ptestnzc:
6827 X86CC = X86::COND_A;
6831 SDValue LHS = Op.getOperand(1);
6832 SDValue RHS = Op.getOperand(2);
6833 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6834 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6835 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6836 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6839 // Fix vector shift instructions where the last operand is a non-immediate
6841 case Intrinsic::x86_sse2_pslli_w:
6842 case Intrinsic::x86_sse2_pslli_d:
6843 case Intrinsic::x86_sse2_pslli_q:
6844 case Intrinsic::x86_sse2_psrli_w:
6845 case Intrinsic::x86_sse2_psrli_d:
6846 case Intrinsic::x86_sse2_psrli_q:
6847 case Intrinsic::x86_sse2_psrai_w:
6848 case Intrinsic::x86_sse2_psrai_d:
6849 case Intrinsic::x86_mmx_pslli_w:
6850 case Intrinsic::x86_mmx_pslli_d:
6851 case Intrinsic::x86_mmx_pslli_q:
6852 case Intrinsic::x86_mmx_psrli_w:
6853 case Intrinsic::x86_mmx_psrli_d:
6854 case Intrinsic::x86_mmx_psrli_q:
6855 case Intrinsic::x86_mmx_psrai_w:
6856 case Intrinsic::x86_mmx_psrai_d: {
6857 SDValue ShAmt = Op.getOperand(2);
6858 if (isa<ConstantSDNode>(ShAmt))
6861 unsigned NewIntNo = 0;
6862 EVT ShAmtVT = MVT::v4i32;
6864 case Intrinsic::x86_sse2_pslli_w:
6865 NewIntNo = Intrinsic::x86_sse2_psll_w;
6867 case Intrinsic::x86_sse2_pslli_d:
6868 NewIntNo = Intrinsic::x86_sse2_psll_d;
6870 case Intrinsic::x86_sse2_pslli_q:
6871 NewIntNo = Intrinsic::x86_sse2_psll_q;
6873 case Intrinsic::x86_sse2_psrli_w:
6874 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6876 case Intrinsic::x86_sse2_psrli_d:
6877 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6879 case Intrinsic::x86_sse2_psrli_q:
6880 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6882 case Intrinsic::x86_sse2_psrai_w:
6883 NewIntNo = Intrinsic::x86_sse2_psra_w;
6885 case Intrinsic::x86_sse2_psrai_d:
6886 NewIntNo = Intrinsic::x86_sse2_psra_d;
6889 ShAmtVT = MVT::v2i32;
6891 case Intrinsic::x86_mmx_pslli_w:
6892 NewIntNo = Intrinsic::x86_mmx_psll_w;
6894 case Intrinsic::x86_mmx_pslli_d:
6895 NewIntNo = Intrinsic::x86_mmx_psll_d;
6897 case Intrinsic::x86_mmx_pslli_q:
6898 NewIntNo = Intrinsic::x86_mmx_psll_q;
6900 case Intrinsic::x86_mmx_psrli_w:
6901 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6903 case Intrinsic::x86_mmx_psrli_d:
6904 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6906 case Intrinsic::x86_mmx_psrli_q:
6907 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6909 case Intrinsic::x86_mmx_psrai_w:
6910 NewIntNo = Intrinsic::x86_mmx_psra_w;
6912 case Intrinsic::x86_mmx_psrai_d:
6913 NewIntNo = Intrinsic::x86_mmx_psra_d;
6915 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6921 // The vector shift intrinsics with scalars uses 32b shift amounts but
6922 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6926 ShOps[1] = DAG.getConstant(0, MVT::i32);
6927 if (ShAmtVT == MVT::v4i32) {
6928 ShOps[2] = DAG.getUNDEF(MVT::i32);
6929 ShOps[3] = DAG.getUNDEF(MVT::i32);
6930 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6932 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6935 EVT VT = Op.getValueType();
6936 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6937 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6938 DAG.getConstant(NewIntNo, MVT::i32),
6939 Op.getOperand(1), ShAmt);
6944 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6945 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6946 DebugLoc dl = Op.getDebugLoc();
6949 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6951 DAG.getConstant(TD->getPointerSize(),
6952 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6953 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6954 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6956 NULL, 0, false, false, 0);
6959 // Just load the return address.
6960 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6961 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6962 RetAddrFI, NULL, 0, false, false, 0);
6965 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6966 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6967 MFI->setFrameAddressIsTaken(true);
6968 EVT VT = Op.getValueType();
6969 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6970 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6971 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6972 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6974 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
6979 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6980 SelectionDAG &DAG) {
6981 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6984 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6986 MachineFunction &MF = DAG.getMachineFunction();
6987 SDValue Chain = Op.getOperand(0);
6988 SDValue Offset = Op.getOperand(1);
6989 SDValue Handler = Op.getOperand(2);
6990 DebugLoc dl = Op.getDebugLoc();
6992 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6994 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6996 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6997 DAG.getIntPtrConstant(-TD->getPointerSize()));
6998 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6999 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
7000 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7001 MF.getRegInfo().addLiveOut(StoreAddrReg);
7003 return DAG.getNode(X86ISD::EH_RETURN, dl,
7005 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7008 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7009 SelectionDAG &DAG) {
7010 SDValue Root = Op.getOperand(0);
7011 SDValue Trmp = Op.getOperand(1); // trampoline
7012 SDValue FPtr = Op.getOperand(2); // nested function
7013 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7014 DebugLoc dl = Op.getDebugLoc();
7016 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7018 if (Subtarget->is64Bit()) {
7019 SDValue OutChains[6];
7021 // Large code-model.
7022 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7023 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7025 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7026 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7028 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7030 // Load the pointer to the nested function into R11.
7031 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7032 SDValue Addr = Trmp;
7033 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7034 Addr, TrmpAddr, 0, false, false, 0);
7036 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7037 DAG.getConstant(2, MVT::i64));
7038 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7041 // Load the 'nest' parameter value into R10.
7042 // R10 is specified in X86CallingConv.td
7043 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7044 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7045 DAG.getConstant(10, MVT::i64));
7046 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7047 Addr, TrmpAddr, 10, false, false, 0);
7049 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7050 DAG.getConstant(12, MVT::i64));
7051 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7054 // Jump to the nested function.
7055 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7056 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7057 DAG.getConstant(20, MVT::i64));
7058 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7059 Addr, TrmpAddr, 20, false, false, 0);
7061 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7062 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7063 DAG.getConstant(22, MVT::i64));
7064 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7065 TrmpAddr, 22, false, false, 0);
7068 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7069 return DAG.getMergeValues(Ops, 2, dl);
7071 const Function *Func =
7072 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7073 CallingConv::ID CC = Func->getCallingConv();
7078 llvm_unreachable("Unsupported calling convention");
7079 case CallingConv::C:
7080 case CallingConv::X86_StdCall: {
7081 // Pass 'nest' parameter in ECX.
7082 // Must be kept in sync with X86CallingConv.td
7085 // Check that ECX wasn't needed by an 'inreg' parameter.
7086 const FunctionType *FTy = Func->getFunctionType();
7087 const AttrListPtr &Attrs = Func->getAttributes();
7089 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7090 unsigned InRegCount = 0;
7093 for (FunctionType::param_iterator I = FTy->param_begin(),
7094 E = FTy->param_end(); I != E; ++I, ++Idx)
7095 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7096 // FIXME: should only count parameters that are lowered to integers.
7097 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7099 if (InRegCount > 2) {
7100 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7105 case CallingConv::X86_FastCall:
7106 case CallingConv::Fast:
7107 // Pass 'nest' parameter in EAX.
7108 // Must be kept in sync with X86CallingConv.td
7113 SDValue OutChains[4];
7116 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7117 DAG.getConstant(10, MVT::i32));
7118 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7120 // This is storing the opcode for MOV32ri.
7121 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7122 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7123 OutChains[0] = DAG.getStore(Root, dl,
7124 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7125 Trmp, TrmpAddr, 0, false, false, 0);
7127 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7128 DAG.getConstant(1, MVT::i32));
7129 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7132 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7133 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7134 DAG.getConstant(5, MVT::i32));
7135 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7136 TrmpAddr, 5, false, false, 1);
7138 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7139 DAG.getConstant(6, MVT::i32));
7140 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7144 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7145 return DAG.getMergeValues(Ops, 2, dl);
7149 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7151 The rounding mode is in bits 11:10 of FPSR, and has the following
7158 FLT_ROUNDS, on the other hand, expects the following:
7165 To perform the conversion, we do:
7166 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7169 MachineFunction &MF = DAG.getMachineFunction();
7170 const TargetMachine &TM = MF.getTarget();
7171 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7172 unsigned StackAlignment = TFI.getStackAlignment();
7173 EVT VT = Op.getValueType();
7174 DebugLoc dl = Op.getDebugLoc();
7176 // Save FP Control Word to stack slot
7177 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7178 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7180 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7181 DAG.getEntryNode(), StackSlot);
7183 // Load FP Control Word from stack slot
7184 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7187 // Transform as necessary
7189 DAG.getNode(ISD::SRL, dl, MVT::i16,
7190 DAG.getNode(ISD::AND, dl, MVT::i16,
7191 CWD, DAG.getConstant(0x800, MVT::i16)),
7192 DAG.getConstant(11, MVT::i8));
7194 DAG.getNode(ISD::SRL, dl, MVT::i16,
7195 DAG.getNode(ISD::AND, dl, MVT::i16,
7196 CWD, DAG.getConstant(0x400, MVT::i16)),
7197 DAG.getConstant(9, MVT::i8));
7200 DAG.getNode(ISD::AND, dl, MVT::i16,
7201 DAG.getNode(ISD::ADD, dl, MVT::i16,
7202 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7203 DAG.getConstant(1, MVT::i16)),
7204 DAG.getConstant(3, MVT::i16));
7207 return DAG.getNode((VT.getSizeInBits() < 16 ?
7208 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7211 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7212 EVT VT = Op.getValueType();
7214 unsigned NumBits = VT.getSizeInBits();
7215 DebugLoc dl = Op.getDebugLoc();
7217 Op = Op.getOperand(0);
7218 if (VT == MVT::i8) {
7219 // Zero extend to i32 since there is not an i8 bsr.
7221 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7224 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7225 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7226 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7228 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7231 DAG.getConstant(NumBits+NumBits-1, OpVT),
7232 DAG.getConstant(X86::COND_E, MVT::i8),
7235 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7237 // Finally xor with NumBits-1.
7238 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7241 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7245 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7246 EVT VT = Op.getValueType();
7248 unsigned NumBits = VT.getSizeInBits();
7249 DebugLoc dl = Op.getDebugLoc();
7251 Op = Op.getOperand(0);
7252 if (VT == MVT::i8) {
7254 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7257 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7258 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7259 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7261 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7264 DAG.getConstant(NumBits, OpVT),
7265 DAG.getConstant(X86::COND_E, MVT::i8),
7268 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7271 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7275 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7276 EVT VT = Op.getValueType();
7277 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7278 DebugLoc dl = Op.getDebugLoc();
7280 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7281 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7282 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7283 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7284 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7286 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7287 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7288 // return AloBlo + AloBhi + AhiBlo;
7290 SDValue A = Op.getOperand(0);
7291 SDValue B = Op.getOperand(1);
7293 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7294 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7295 A, DAG.getConstant(32, MVT::i32));
7296 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7297 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7298 B, DAG.getConstant(32, MVT::i32));
7299 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7300 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7302 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7303 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7305 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7306 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7308 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7309 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7310 AloBhi, DAG.getConstant(32, MVT::i32));
7311 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7312 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7313 AhiBlo, DAG.getConstant(32, MVT::i32));
7314 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7315 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7320 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7321 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7322 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7323 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7324 // has only one use.
7325 SDNode *N = Op.getNode();
7326 SDValue LHS = N->getOperand(0);
7327 SDValue RHS = N->getOperand(1);
7328 unsigned BaseOp = 0;
7330 DebugLoc dl = Op.getDebugLoc();
7332 switch (Op.getOpcode()) {
7333 default: llvm_unreachable("Unknown ovf instruction!");
7335 // A subtract of one will be selected as a INC. Note that INC doesn't
7336 // set CF, so we can't do this for UADDO.
7337 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7338 if (C->getAPIntValue() == 1) {
7339 BaseOp = X86ISD::INC;
7343 BaseOp = X86ISD::ADD;
7347 BaseOp = X86ISD::ADD;
7351 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7352 // set CF, so we can't do this for USUBO.
7353 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7354 if (C->getAPIntValue() == 1) {
7355 BaseOp = X86ISD::DEC;
7359 BaseOp = X86ISD::SUB;
7363 BaseOp = X86ISD::SUB;
7367 BaseOp = X86ISD::SMUL;
7371 BaseOp = X86ISD::UMUL;
7376 // Also sets EFLAGS.
7377 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7378 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7381 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7382 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7384 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7388 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7389 EVT T = Op.getValueType();
7390 DebugLoc dl = Op.getDebugLoc();
7393 switch(T.getSimpleVT().SimpleTy) {
7395 assert(false && "Invalid value type!");
7396 case MVT::i8: Reg = X86::AL; size = 1; break;
7397 case MVT::i16: Reg = X86::AX; size = 2; break;
7398 case MVT::i32: Reg = X86::EAX; size = 4; break;
7400 assert(Subtarget->is64Bit() && "Node not type legal!");
7401 Reg = X86::RAX; size = 8;
7404 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7405 Op.getOperand(2), SDValue());
7406 SDValue Ops[] = { cpIn.getValue(0),
7409 DAG.getTargetConstant(size, MVT::i8),
7411 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7412 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7414 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7418 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7419 SelectionDAG &DAG) {
7420 assert(Subtarget->is64Bit() && "Result not type legalized?");
7421 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7422 SDValue TheChain = Op.getOperand(0);
7423 DebugLoc dl = Op.getDebugLoc();
7424 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7425 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7426 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7428 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7429 DAG.getConstant(32, MVT::i8));
7431 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7434 return DAG.getMergeValues(Ops, 2, dl);
7437 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7438 SDNode *Node = Op.getNode();
7439 DebugLoc dl = Node->getDebugLoc();
7440 EVT T = Node->getValueType(0);
7441 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7442 DAG.getConstant(0, T), Node->getOperand(2));
7443 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7444 cast<AtomicSDNode>(Node)->getMemoryVT(),
7445 Node->getOperand(0),
7446 Node->getOperand(1), negOp,
7447 cast<AtomicSDNode>(Node)->getSrcValue(),
7448 cast<AtomicSDNode>(Node)->getAlignment());
7451 /// LowerOperation - Provide custom lowering hooks for some operations.
7453 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7454 switch (Op.getOpcode()) {
7455 default: llvm_unreachable("Should not custom lower this!");
7456 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7457 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7458 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7459 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7460 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7461 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7462 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7463 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7464 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7465 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7466 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7467 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7468 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7469 case ISD::SHL_PARTS:
7470 case ISD::SRA_PARTS:
7471 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7472 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7473 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7474 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7475 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7476 case ISD::FABS: return LowerFABS(Op, DAG);
7477 case ISD::FNEG: return LowerFNEG(Op, DAG);
7478 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7479 case ISD::SETCC: return LowerSETCC(Op, DAG);
7480 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7481 case ISD::SELECT: return LowerSELECT(Op, DAG);
7482 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7483 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7484 case ISD::VASTART: return LowerVASTART(Op, DAG);
7485 case ISD::VAARG: return LowerVAARG(Op, DAG);
7486 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7487 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7488 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7489 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7490 case ISD::FRAME_TO_ARGS_OFFSET:
7491 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7492 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7493 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7494 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7495 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7496 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7497 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7498 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7504 case ISD::UMULO: return LowerXALUO(Op, DAG);
7505 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7509 void X86TargetLowering::
7510 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7511 SelectionDAG &DAG, unsigned NewOp) {
7512 EVT T = Node->getValueType(0);
7513 DebugLoc dl = Node->getDebugLoc();
7514 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7516 SDValue Chain = Node->getOperand(0);
7517 SDValue In1 = Node->getOperand(1);
7518 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7519 Node->getOperand(2), DAG.getIntPtrConstant(0));
7520 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7521 Node->getOperand(2), DAG.getIntPtrConstant(1));
7522 SDValue Ops[] = { Chain, In1, In2L, In2H };
7523 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7525 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7526 cast<MemSDNode>(Node)->getMemOperand());
7527 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7528 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7529 Results.push_back(Result.getValue(2));
7532 /// ReplaceNodeResults - Replace a node with an illegal result type
7533 /// with a new node built out of custom code.
7534 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7535 SmallVectorImpl<SDValue>&Results,
7536 SelectionDAG &DAG) {
7537 DebugLoc dl = N->getDebugLoc();
7538 switch (N->getOpcode()) {
7540 assert(false && "Do not know how to custom type legalize this operation!");
7542 case ISD::FP_TO_SINT: {
7543 std::pair<SDValue,SDValue> Vals =
7544 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7545 SDValue FIST = Vals.first, StackSlot = Vals.second;
7546 if (FIST.getNode() != 0) {
7547 EVT VT = N->getValueType(0);
7548 // Return a load from the stack slot.
7549 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7554 case ISD::READCYCLECOUNTER: {
7555 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7556 SDValue TheChain = N->getOperand(0);
7557 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7558 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7560 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7562 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7563 SDValue Ops[] = { eax, edx };
7564 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7565 Results.push_back(edx.getValue(1));
7568 case ISD::ATOMIC_CMP_SWAP: {
7569 EVT T = N->getValueType(0);
7570 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7571 SDValue cpInL, cpInH;
7572 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7573 DAG.getConstant(0, MVT::i32));
7574 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7575 DAG.getConstant(1, MVT::i32));
7576 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7577 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7579 SDValue swapInL, swapInH;
7580 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7581 DAG.getConstant(0, MVT::i32));
7582 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7583 DAG.getConstant(1, MVT::i32));
7584 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7586 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7587 swapInL.getValue(1));
7588 SDValue Ops[] = { swapInH.getValue(0),
7590 swapInH.getValue(1) };
7591 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7592 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7593 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7594 MVT::i32, Result.getValue(1));
7595 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7596 MVT::i32, cpOutL.getValue(2));
7597 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7598 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7599 Results.push_back(cpOutH.getValue(1));
7602 case ISD::ATOMIC_LOAD_ADD:
7603 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7605 case ISD::ATOMIC_LOAD_AND:
7606 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7608 case ISD::ATOMIC_LOAD_NAND:
7609 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7611 case ISD::ATOMIC_LOAD_OR:
7612 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7614 case ISD::ATOMIC_LOAD_SUB:
7615 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7617 case ISD::ATOMIC_LOAD_XOR:
7618 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7620 case ISD::ATOMIC_SWAP:
7621 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7626 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7628 default: return NULL;
7629 case X86ISD::BSF: return "X86ISD::BSF";
7630 case X86ISD::BSR: return "X86ISD::BSR";
7631 case X86ISD::SHLD: return "X86ISD::SHLD";
7632 case X86ISD::SHRD: return "X86ISD::SHRD";
7633 case X86ISD::FAND: return "X86ISD::FAND";
7634 case X86ISD::FOR: return "X86ISD::FOR";
7635 case X86ISD::FXOR: return "X86ISD::FXOR";
7636 case X86ISD::FSRL: return "X86ISD::FSRL";
7637 case X86ISD::FILD: return "X86ISD::FILD";
7638 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7639 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7640 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7641 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7642 case X86ISD::FLD: return "X86ISD::FLD";
7643 case X86ISD::FST: return "X86ISD::FST";
7644 case X86ISD::CALL: return "X86ISD::CALL";
7645 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7646 case X86ISD::BT: return "X86ISD::BT";
7647 case X86ISD::CMP: return "X86ISD::CMP";
7648 case X86ISD::COMI: return "X86ISD::COMI";
7649 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7650 case X86ISD::SETCC: return "X86ISD::SETCC";
7651 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7652 case X86ISD::CMOV: return "X86ISD::CMOV";
7653 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7654 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7655 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7656 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7657 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7658 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7659 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7660 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7661 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7662 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7663 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7664 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7665 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7666 case X86ISD::FMAX: return "X86ISD::FMAX";
7667 case X86ISD::FMIN: return "X86ISD::FMIN";
7668 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7669 case X86ISD::FRCP: return "X86ISD::FRCP";
7670 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7671 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7672 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7673 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7674 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7675 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7676 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7677 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7678 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7679 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7680 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7681 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7682 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7683 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7684 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7685 case X86ISD::VSHL: return "X86ISD::VSHL";
7686 case X86ISD::VSRL: return "X86ISD::VSRL";
7687 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7688 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7689 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7690 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7691 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7692 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7693 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7694 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7695 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7696 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7697 case X86ISD::ADD: return "X86ISD::ADD";
7698 case X86ISD::SUB: return "X86ISD::SUB";
7699 case X86ISD::SMUL: return "X86ISD::SMUL";
7700 case X86ISD::UMUL: return "X86ISD::UMUL";
7701 case X86ISD::INC: return "X86ISD::INC";
7702 case X86ISD::DEC: return "X86ISD::DEC";
7703 case X86ISD::OR: return "X86ISD::OR";
7704 case X86ISD::XOR: return "X86ISD::XOR";
7705 case X86ISD::AND: return "X86ISD::AND";
7706 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7707 case X86ISD::PTEST: return "X86ISD::PTEST";
7708 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7712 // isLegalAddressingMode - Return true if the addressing mode represented
7713 // by AM is legal for this target, for a load/store of the specified type.
7714 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7715 const Type *Ty) const {
7716 // X86 supports extremely general addressing modes.
7717 CodeModel::Model M = getTargetMachine().getCodeModel();
7719 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7720 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7725 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7727 // If a reference to this global requires an extra load, we can't fold it.
7728 if (isGlobalStubReference(GVFlags))
7731 // If BaseGV requires a register for the PIC base, we cannot also have a
7732 // BaseReg specified.
7733 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7736 // If lower 4G is not available, then we must use rip-relative addressing.
7737 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7747 // These scales always work.
7752 // These scales are formed with basereg+scalereg. Only accept if there is
7757 default: // Other stuff never works.
7765 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7766 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7768 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7769 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7770 if (NumBits1 <= NumBits2)
7772 return Subtarget->is64Bit() || NumBits1 < 64;
7775 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7776 if (!VT1.isInteger() || !VT2.isInteger())
7778 unsigned NumBits1 = VT1.getSizeInBits();
7779 unsigned NumBits2 = VT2.getSizeInBits();
7780 if (NumBits1 <= NumBits2)
7782 return Subtarget->is64Bit() || NumBits1 < 64;
7785 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7786 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7787 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7790 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7791 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7792 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7795 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7796 // i16 instructions are longer (0x66 prefix) and potentially slower.
7797 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7800 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7801 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7802 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7803 /// are assumed to be legal.
7805 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7807 // Only do shuffles on 128-bit vector types for now.
7808 if (VT.getSizeInBits() == 64)
7811 // FIXME: pshufb, blends, shifts.
7812 return (VT.getVectorNumElements() == 2 ||
7813 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7814 isMOVLMask(M, VT) ||
7815 isSHUFPMask(M, VT) ||
7816 isPSHUFDMask(M, VT) ||
7817 isPSHUFHWMask(M, VT) ||
7818 isPSHUFLWMask(M, VT) ||
7819 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7820 isUNPCKLMask(M, VT) ||
7821 isUNPCKHMask(M, VT) ||
7822 isUNPCKL_v_undef_Mask(M, VT) ||
7823 isUNPCKH_v_undef_Mask(M, VT));
7827 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7829 unsigned NumElts = VT.getVectorNumElements();
7830 // FIXME: This collection of masks seems suspect.
7833 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7834 return (isMOVLMask(Mask, VT) ||
7835 isCommutedMOVLMask(Mask, VT, true) ||
7836 isSHUFPMask(Mask, VT) ||
7837 isCommutedSHUFPMask(Mask, VT));
7842 //===----------------------------------------------------------------------===//
7843 // X86 Scheduler Hooks
7844 //===----------------------------------------------------------------------===//
7846 // private utility function
7848 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7849 MachineBasicBlock *MBB,
7857 TargetRegisterClass *RC,
7858 bool invSrc) const {
7859 // For the atomic bitwise operator, we generate
7862 // ld t1 = [bitinstr.addr]
7863 // op t2 = t1, [bitinstr.val]
7865 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7867 // fallthrough -->nextMBB
7868 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7869 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7870 MachineFunction::iterator MBBIter = MBB;
7873 /// First build the CFG
7874 MachineFunction *F = MBB->getParent();
7875 MachineBasicBlock *thisMBB = MBB;
7876 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7877 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7878 F->insert(MBBIter, newMBB);
7879 F->insert(MBBIter, nextMBB);
7881 // Move all successors to thisMBB to nextMBB
7882 nextMBB->transferSuccessors(thisMBB);
7884 // Update thisMBB to fall through to newMBB
7885 thisMBB->addSuccessor(newMBB);
7887 // newMBB jumps to itself and fall through to nextMBB
7888 newMBB->addSuccessor(nextMBB);
7889 newMBB->addSuccessor(newMBB);
7891 // Insert instructions into newMBB based on incoming instruction
7892 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7893 "unexpected number of operands");
7894 DebugLoc dl = bInstr->getDebugLoc();
7895 MachineOperand& destOper = bInstr->getOperand(0);
7896 MachineOperand* argOpers[2 + X86AddrNumOperands];
7897 int numArgs = bInstr->getNumOperands() - 1;
7898 for (int i=0; i < numArgs; ++i)
7899 argOpers[i] = &bInstr->getOperand(i+1);
7901 // x86 address has 4 operands: base, index, scale, and displacement
7902 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7903 int valArgIndx = lastAddrIndx + 1;
7905 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7906 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7907 for (int i=0; i <= lastAddrIndx; ++i)
7908 (*MIB).addOperand(*argOpers[i]);
7910 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7912 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7917 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7918 assert((argOpers[valArgIndx]->isReg() ||
7919 argOpers[valArgIndx]->isImm()) &&
7921 if (argOpers[valArgIndx]->isReg())
7922 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7924 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7926 (*MIB).addOperand(*argOpers[valArgIndx]);
7928 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7931 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7932 for (int i=0; i <= lastAddrIndx; ++i)
7933 (*MIB).addOperand(*argOpers[i]);
7935 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7936 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7937 bInstr->memoperands_end());
7939 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7943 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
7945 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7949 // private utility function: 64 bit atomics on 32 bit host.
7951 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7952 MachineBasicBlock *MBB,
7957 bool invSrc) const {
7958 // For the atomic bitwise operator, we generate
7959 // thisMBB (instructions are in pairs, except cmpxchg8b)
7960 // ld t1,t2 = [bitinstr.addr]
7962 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7963 // op t5, t6 <- out1, out2, [bitinstr.val]
7964 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7965 // mov ECX, EBX <- t5, t6
7966 // mov EAX, EDX <- t1, t2
7967 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7968 // mov t3, t4 <- EAX, EDX
7970 // result in out1, out2
7971 // fallthrough -->nextMBB
7973 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7974 const unsigned LoadOpc = X86::MOV32rm;
7975 const unsigned copyOpc = X86::MOV32rr;
7976 const unsigned NotOpc = X86::NOT32r;
7977 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7978 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7979 MachineFunction::iterator MBBIter = MBB;
7982 /// First build the CFG
7983 MachineFunction *F = MBB->getParent();
7984 MachineBasicBlock *thisMBB = MBB;
7985 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7986 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7987 F->insert(MBBIter, newMBB);
7988 F->insert(MBBIter, nextMBB);
7990 // Move all successors to thisMBB to nextMBB
7991 nextMBB->transferSuccessors(thisMBB);
7993 // Update thisMBB to fall through to newMBB
7994 thisMBB->addSuccessor(newMBB);
7996 // newMBB jumps to itself and fall through to nextMBB
7997 newMBB->addSuccessor(nextMBB);
7998 newMBB->addSuccessor(newMBB);
8000 DebugLoc dl = bInstr->getDebugLoc();
8001 // Insert instructions into newMBB based on incoming instruction
8002 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8003 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8004 "unexpected number of operands");
8005 MachineOperand& dest1Oper = bInstr->getOperand(0);
8006 MachineOperand& dest2Oper = bInstr->getOperand(1);
8007 MachineOperand* argOpers[2 + X86AddrNumOperands];
8008 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8009 argOpers[i] = &bInstr->getOperand(i+2);
8011 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8012 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8014 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8015 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8016 for (int i=0; i <= lastAddrIndx; ++i)
8017 (*MIB).addOperand(*argOpers[i]);
8018 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8019 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8020 // add 4 to displacement.
8021 for (int i=0; i <= lastAddrIndx-2; ++i)
8022 (*MIB).addOperand(*argOpers[i]);
8023 MachineOperand newOp3 = *(argOpers[3]);
8025 newOp3.setImm(newOp3.getImm()+4);
8027 newOp3.setOffset(newOp3.getOffset()+4);
8028 (*MIB).addOperand(newOp3);
8029 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8031 // t3/4 are defined later, at the bottom of the loop
8032 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8033 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8034 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8035 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8036 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8037 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8039 // The subsequent operations should be using the destination registers of
8040 //the PHI instructions.
8042 t1 = F->getRegInfo().createVirtualRegister(RC);
8043 t2 = F->getRegInfo().createVirtualRegister(RC);
8044 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8045 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8047 t1 = dest1Oper.getReg();
8048 t2 = dest2Oper.getReg();
8051 int valArgIndx = lastAddrIndx + 1;
8052 assert((argOpers[valArgIndx]->isReg() ||
8053 argOpers[valArgIndx]->isImm()) &&
8055 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8056 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8057 if (argOpers[valArgIndx]->isReg())
8058 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8060 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8061 if (regOpcL != X86::MOV32rr)
8063 (*MIB).addOperand(*argOpers[valArgIndx]);
8064 assert(argOpers[valArgIndx + 1]->isReg() ==
8065 argOpers[valArgIndx]->isReg());
8066 assert(argOpers[valArgIndx + 1]->isImm() ==
8067 argOpers[valArgIndx]->isImm());
8068 if (argOpers[valArgIndx + 1]->isReg())
8069 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8071 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8072 if (regOpcH != X86::MOV32rr)
8074 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8076 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8078 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8081 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8083 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8086 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8087 for (int i=0; i <= lastAddrIndx; ++i)
8088 (*MIB).addOperand(*argOpers[i]);
8090 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8091 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8092 bInstr->memoperands_end());
8094 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8095 MIB.addReg(X86::EAX);
8096 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8097 MIB.addReg(X86::EDX);
8100 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8102 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8106 // private utility function
8108 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8109 MachineBasicBlock *MBB,
8110 unsigned cmovOpc) const {
8111 // For the atomic min/max operator, we generate
8114 // ld t1 = [min/max.addr]
8115 // mov t2 = [min/max.val]
8117 // cmov[cond] t2 = t1
8119 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8121 // fallthrough -->nextMBB
8123 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8124 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8125 MachineFunction::iterator MBBIter = MBB;
8128 /// First build the CFG
8129 MachineFunction *F = MBB->getParent();
8130 MachineBasicBlock *thisMBB = MBB;
8131 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8132 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8133 F->insert(MBBIter, newMBB);
8134 F->insert(MBBIter, nextMBB);
8136 // Move all successors of thisMBB to nextMBB
8137 nextMBB->transferSuccessors(thisMBB);
8139 // Update thisMBB to fall through to newMBB
8140 thisMBB->addSuccessor(newMBB);
8142 // newMBB jumps to newMBB and fall through to nextMBB
8143 newMBB->addSuccessor(nextMBB);
8144 newMBB->addSuccessor(newMBB);
8146 DebugLoc dl = mInstr->getDebugLoc();
8147 // Insert instructions into newMBB based on incoming instruction
8148 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8149 "unexpected number of operands");
8150 MachineOperand& destOper = mInstr->getOperand(0);
8151 MachineOperand* argOpers[2 + X86AddrNumOperands];
8152 int numArgs = mInstr->getNumOperands() - 1;
8153 for (int i=0; i < numArgs; ++i)
8154 argOpers[i] = &mInstr->getOperand(i+1);
8156 // x86 address has 4 operands: base, index, scale, and displacement
8157 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8158 int valArgIndx = lastAddrIndx + 1;
8160 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8161 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8162 for (int i=0; i <= lastAddrIndx; ++i)
8163 (*MIB).addOperand(*argOpers[i]);
8165 // We only support register and immediate values
8166 assert((argOpers[valArgIndx]->isReg() ||
8167 argOpers[valArgIndx]->isImm()) &&
8170 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8171 if (argOpers[valArgIndx]->isReg())
8172 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8174 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8175 (*MIB).addOperand(*argOpers[valArgIndx]);
8177 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8180 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8185 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8186 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8190 // Cmp and exchange if none has modified the memory location
8191 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8192 for (int i=0; i <= lastAddrIndx; ++i)
8193 (*MIB).addOperand(*argOpers[i]);
8195 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8196 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8197 mInstr->memoperands_end());
8199 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8200 MIB.addReg(X86::EAX);
8203 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8205 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8209 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8210 // all of this code can be replaced with that in the .td file.
8212 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8213 unsigned numArgs, bool memArg) const {
8215 MachineFunction *F = BB->getParent();
8216 DebugLoc dl = MI->getDebugLoc();
8217 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8221 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8223 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8225 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8227 for (unsigned i = 0; i < numArgs; ++i) {
8228 MachineOperand &Op = MI->getOperand(i+1);
8230 if (!(Op.isReg() && Op.isImplicit()))
8234 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8237 F->DeleteMachineInstr(MI);
8243 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8245 MachineBasicBlock *MBB) const {
8246 // Emit code to save XMM registers to the stack. The ABI says that the
8247 // number of registers to save is given in %al, so it's theoretically
8248 // possible to do an indirect jump trick to avoid saving all of them,
8249 // however this code takes a simpler approach and just executes all
8250 // of the stores if %al is non-zero. It's less code, and it's probably
8251 // easier on the hardware branch predictor, and stores aren't all that
8252 // expensive anyway.
8254 // Create the new basic blocks. One block contains all the XMM stores,
8255 // and one block is the final destination regardless of whether any
8256 // stores were performed.
8257 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8258 MachineFunction *F = MBB->getParent();
8259 MachineFunction::iterator MBBIter = MBB;
8261 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8262 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8263 F->insert(MBBIter, XMMSaveMBB);
8264 F->insert(MBBIter, EndMBB);
8267 // Move any original successors of MBB to the end block.
8268 EndMBB->transferSuccessors(MBB);
8269 // The original block will now fall through to the XMM save block.
8270 MBB->addSuccessor(XMMSaveMBB);
8271 // The XMMSaveMBB will fall through to the end block.
8272 XMMSaveMBB->addSuccessor(EndMBB);
8274 // Now add the instructions.
8275 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8276 DebugLoc DL = MI->getDebugLoc();
8278 unsigned CountReg = MI->getOperand(0).getReg();
8279 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8280 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8282 if (!Subtarget->isTargetWin64()) {
8283 // If %al is 0, branch around the XMM save block.
8284 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8285 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8286 MBB->addSuccessor(EndMBB);
8289 // In the XMM save block, save all the XMM argument registers.
8290 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8291 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8292 MachineMemOperand *MMO =
8293 F->getMachineMemOperand(
8294 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8295 MachineMemOperand::MOStore, Offset,
8296 /*Size=*/16, /*Align=*/16);
8297 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8298 .addFrameIndex(RegSaveFrameIndex)
8299 .addImm(/*Scale=*/1)
8300 .addReg(/*IndexReg=*/0)
8301 .addImm(/*Disp=*/Offset)
8302 .addReg(/*Segment=*/0)
8303 .addReg(MI->getOperand(i).getReg())
8304 .addMemOperand(MMO);
8307 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8313 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8314 MachineBasicBlock *BB,
8315 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8316 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8317 DebugLoc DL = MI->getDebugLoc();
8319 // To "insert" a SELECT_CC instruction, we actually have to insert the
8320 // diamond control-flow pattern. The incoming instruction knows the
8321 // destination vreg to set, the condition code register to branch on, the
8322 // true/false values to select between, and a branch opcode to use.
8323 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8324 MachineFunction::iterator It = BB;
8330 // cmpTY ccX, r1, r2
8332 // fallthrough --> copy0MBB
8333 MachineBasicBlock *thisMBB = BB;
8334 MachineFunction *F = BB->getParent();
8335 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8336 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8338 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8339 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8340 F->insert(It, copy0MBB);
8341 F->insert(It, sinkMBB);
8342 // Update machine-CFG edges by first adding all successors of the current
8343 // block to the new block which will contain the Phi node for the select.
8344 // Also inform sdisel of the edge changes.
8345 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8346 E = BB->succ_end(); I != E; ++I) {
8347 EM->insert(std::make_pair(*I, sinkMBB));
8348 sinkMBB->addSuccessor(*I);
8350 // Next, remove all successors of the current block, and add the true
8351 // and fallthrough blocks as its successors.
8352 while (!BB->succ_empty())
8353 BB->removeSuccessor(BB->succ_begin());
8354 // Add the true and fallthrough blocks as its successors.
8355 BB->addSuccessor(copy0MBB);
8356 BB->addSuccessor(sinkMBB);
8359 // %FalseValue = ...
8360 // # fallthrough to sinkMBB
8363 // Update machine-CFG edges
8364 BB->addSuccessor(sinkMBB);
8367 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8370 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8371 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8372 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8374 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8380 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8381 MachineBasicBlock *BB,
8382 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8383 switch (MI->getOpcode()) {
8384 default: assert(false && "Unexpected instr type to insert");
8386 case X86::CMOV_V1I64:
8387 case X86::CMOV_FR32:
8388 case X86::CMOV_FR64:
8389 case X86::CMOV_V4F32:
8390 case X86::CMOV_V2F64:
8391 case X86::CMOV_V2I64:
8392 return EmitLoweredSelect(MI, BB, EM);
8394 case X86::FP32_TO_INT16_IN_MEM:
8395 case X86::FP32_TO_INT32_IN_MEM:
8396 case X86::FP32_TO_INT64_IN_MEM:
8397 case X86::FP64_TO_INT16_IN_MEM:
8398 case X86::FP64_TO_INT32_IN_MEM:
8399 case X86::FP64_TO_INT64_IN_MEM:
8400 case X86::FP80_TO_INT16_IN_MEM:
8401 case X86::FP80_TO_INT32_IN_MEM:
8402 case X86::FP80_TO_INT64_IN_MEM: {
8403 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8404 DebugLoc DL = MI->getDebugLoc();
8406 // Change the floating point control register to use "round towards zero"
8407 // mode when truncating to an integer value.
8408 MachineFunction *F = BB->getParent();
8409 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8410 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8412 // Load the old value of the high byte of the control word...
8414 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8415 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8418 // Set the high part to be round to zero...
8419 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8422 // Reload the modified control word now...
8423 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8425 // Restore the memory image of control word to original value
8426 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8429 // Get the X86 opcode to use.
8431 switch (MI->getOpcode()) {
8432 default: llvm_unreachable("illegal opcode!");
8433 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8434 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8435 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8436 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8437 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8438 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8439 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8440 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8441 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8445 MachineOperand &Op = MI->getOperand(0);
8447 AM.BaseType = X86AddressMode::RegBase;
8448 AM.Base.Reg = Op.getReg();
8450 AM.BaseType = X86AddressMode::FrameIndexBase;
8451 AM.Base.FrameIndex = Op.getIndex();
8453 Op = MI->getOperand(1);
8455 AM.Scale = Op.getImm();
8456 Op = MI->getOperand(2);
8458 AM.IndexReg = Op.getImm();
8459 Op = MI->getOperand(3);
8460 if (Op.isGlobal()) {
8461 AM.GV = Op.getGlobal();
8463 AM.Disp = Op.getImm();
8465 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8466 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8468 // Reload the original control word now.
8469 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8471 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8474 // String/text processing lowering.
8475 case X86::PCMPISTRM128REG:
8476 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8477 case X86::PCMPISTRM128MEM:
8478 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8479 case X86::PCMPESTRM128REG:
8480 return EmitPCMP(MI, BB, 5, false /* in mem */);
8481 case X86::PCMPESTRM128MEM:
8482 return EmitPCMP(MI, BB, 5, true /* in mem */);
8485 case X86::ATOMAND32:
8486 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8487 X86::AND32ri, X86::MOV32rm,
8488 X86::LCMPXCHG32, X86::MOV32rr,
8489 X86::NOT32r, X86::EAX,
8490 X86::GR32RegisterClass);
8492 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8493 X86::OR32ri, X86::MOV32rm,
8494 X86::LCMPXCHG32, X86::MOV32rr,
8495 X86::NOT32r, X86::EAX,
8496 X86::GR32RegisterClass);
8497 case X86::ATOMXOR32:
8498 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8499 X86::XOR32ri, X86::MOV32rm,
8500 X86::LCMPXCHG32, X86::MOV32rr,
8501 X86::NOT32r, X86::EAX,
8502 X86::GR32RegisterClass);
8503 case X86::ATOMNAND32:
8504 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8505 X86::AND32ri, X86::MOV32rm,
8506 X86::LCMPXCHG32, X86::MOV32rr,
8507 X86::NOT32r, X86::EAX,
8508 X86::GR32RegisterClass, true);
8509 case X86::ATOMMIN32:
8510 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8511 case X86::ATOMMAX32:
8512 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8513 case X86::ATOMUMIN32:
8514 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8515 case X86::ATOMUMAX32:
8516 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8518 case X86::ATOMAND16:
8519 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8520 X86::AND16ri, X86::MOV16rm,
8521 X86::LCMPXCHG16, X86::MOV16rr,
8522 X86::NOT16r, X86::AX,
8523 X86::GR16RegisterClass);
8525 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8526 X86::OR16ri, X86::MOV16rm,
8527 X86::LCMPXCHG16, X86::MOV16rr,
8528 X86::NOT16r, X86::AX,
8529 X86::GR16RegisterClass);
8530 case X86::ATOMXOR16:
8531 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8532 X86::XOR16ri, X86::MOV16rm,
8533 X86::LCMPXCHG16, X86::MOV16rr,
8534 X86::NOT16r, X86::AX,
8535 X86::GR16RegisterClass);
8536 case X86::ATOMNAND16:
8537 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8538 X86::AND16ri, X86::MOV16rm,
8539 X86::LCMPXCHG16, X86::MOV16rr,
8540 X86::NOT16r, X86::AX,
8541 X86::GR16RegisterClass, true);
8542 case X86::ATOMMIN16:
8543 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8544 case X86::ATOMMAX16:
8545 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8546 case X86::ATOMUMIN16:
8547 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8548 case X86::ATOMUMAX16:
8549 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8552 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8553 X86::AND8ri, X86::MOV8rm,
8554 X86::LCMPXCHG8, X86::MOV8rr,
8555 X86::NOT8r, X86::AL,
8556 X86::GR8RegisterClass);
8558 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8559 X86::OR8ri, X86::MOV8rm,
8560 X86::LCMPXCHG8, X86::MOV8rr,
8561 X86::NOT8r, X86::AL,
8562 X86::GR8RegisterClass);
8564 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8565 X86::XOR8ri, X86::MOV8rm,
8566 X86::LCMPXCHG8, X86::MOV8rr,
8567 X86::NOT8r, X86::AL,
8568 X86::GR8RegisterClass);
8569 case X86::ATOMNAND8:
8570 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8571 X86::AND8ri, X86::MOV8rm,
8572 X86::LCMPXCHG8, X86::MOV8rr,
8573 X86::NOT8r, X86::AL,
8574 X86::GR8RegisterClass, true);
8575 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8576 // This group is for 64-bit host.
8577 case X86::ATOMAND64:
8578 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8579 X86::AND64ri32, X86::MOV64rm,
8580 X86::LCMPXCHG64, X86::MOV64rr,
8581 X86::NOT64r, X86::RAX,
8582 X86::GR64RegisterClass);
8584 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8585 X86::OR64ri32, X86::MOV64rm,
8586 X86::LCMPXCHG64, X86::MOV64rr,
8587 X86::NOT64r, X86::RAX,
8588 X86::GR64RegisterClass);
8589 case X86::ATOMXOR64:
8590 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8591 X86::XOR64ri32, X86::MOV64rm,
8592 X86::LCMPXCHG64, X86::MOV64rr,
8593 X86::NOT64r, X86::RAX,
8594 X86::GR64RegisterClass);
8595 case X86::ATOMNAND64:
8596 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8597 X86::AND64ri32, X86::MOV64rm,
8598 X86::LCMPXCHG64, X86::MOV64rr,
8599 X86::NOT64r, X86::RAX,
8600 X86::GR64RegisterClass, true);
8601 case X86::ATOMMIN64:
8602 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8603 case X86::ATOMMAX64:
8604 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8605 case X86::ATOMUMIN64:
8606 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8607 case X86::ATOMUMAX64:
8608 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8610 // This group does 64-bit operations on a 32-bit host.
8611 case X86::ATOMAND6432:
8612 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8613 X86::AND32rr, X86::AND32rr,
8614 X86::AND32ri, X86::AND32ri,
8616 case X86::ATOMOR6432:
8617 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8618 X86::OR32rr, X86::OR32rr,
8619 X86::OR32ri, X86::OR32ri,
8621 case X86::ATOMXOR6432:
8622 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8623 X86::XOR32rr, X86::XOR32rr,
8624 X86::XOR32ri, X86::XOR32ri,
8626 case X86::ATOMNAND6432:
8627 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8628 X86::AND32rr, X86::AND32rr,
8629 X86::AND32ri, X86::AND32ri,
8631 case X86::ATOMADD6432:
8632 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8633 X86::ADD32rr, X86::ADC32rr,
8634 X86::ADD32ri, X86::ADC32ri,
8636 case X86::ATOMSUB6432:
8637 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8638 X86::SUB32rr, X86::SBB32rr,
8639 X86::SUB32ri, X86::SBB32ri,
8641 case X86::ATOMSWAP6432:
8642 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8643 X86::MOV32rr, X86::MOV32rr,
8644 X86::MOV32ri, X86::MOV32ri,
8646 case X86::VASTART_SAVE_XMM_REGS:
8647 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8651 //===----------------------------------------------------------------------===//
8652 // X86 Optimization Hooks
8653 //===----------------------------------------------------------------------===//
8655 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8659 const SelectionDAG &DAG,
8660 unsigned Depth) const {
8661 unsigned Opc = Op.getOpcode();
8662 assert((Opc >= ISD::BUILTIN_OP_END ||
8663 Opc == ISD::INTRINSIC_WO_CHAIN ||
8664 Opc == ISD::INTRINSIC_W_CHAIN ||
8665 Opc == ISD::INTRINSIC_VOID) &&
8666 "Should use MaskedValueIsZero if you don't know whether Op"
8667 " is a target node!");
8669 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8681 // These nodes' second result is a boolean.
8682 if (Op.getResNo() == 0)
8686 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8687 Mask.getBitWidth() - 1);
8692 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8693 /// node is a GlobalAddress + offset.
8694 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8695 GlobalValue* &GA, int64_t &Offset) const{
8696 if (N->getOpcode() == X86ISD::Wrapper) {
8697 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8698 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8699 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8703 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8706 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8707 EVT EltVT, LoadSDNode *&LDBase,
8708 unsigned &LastLoadedElt,
8709 SelectionDAG &DAG, MachineFrameInfo *MFI,
8710 const TargetLowering &TLI) {
8712 LastLoadedElt = -1U;
8713 for (unsigned i = 0; i < NumElems; ++i) {
8714 if (N->getMaskElt(i) < 0) {
8720 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8721 if (!Elt.getNode() ||
8722 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8725 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8727 LDBase = cast<LoadSDNode>(Elt.getNode());
8731 if (Elt.getOpcode() == ISD::UNDEF)
8734 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8735 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8742 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8743 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8744 /// if the load addresses are consecutive, non-overlapping, and in the right
8745 /// order. In the case of v2i64, it will see if it can rewrite the
8746 /// shuffle to be an appropriate build vector so it can take advantage of
8747 // performBuildVectorCombine.
8748 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8749 const TargetLowering &TLI) {
8750 DebugLoc dl = N->getDebugLoc();
8751 EVT VT = N->getValueType(0);
8752 EVT EltVT = VT.getVectorElementType();
8753 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8754 unsigned NumElems = VT.getVectorNumElements();
8756 if (VT.getSizeInBits() != 128)
8759 // Try to combine a vector_shuffle into a 128-bit load.
8760 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8761 LoadSDNode *LD = NULL;
8762 unsigned LastLoadedElt;
8763 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8767 if (LastLoadedElt == NumElems - 1) {
8768 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8769 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8770 LD->getSrcValue(), LD->getSrcValueOffset(),
8771 LD->isVolatile(), LD->isNonTemporal(), 0);
8772 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8773 LD->getSrcValue(), LD->getSrcValueOffset(),
8774 LD->isVolatile(), LD->isNonTemporal(),
8775 LD->getAlignment());
8776 } else if (NumElems == 4 && LastLoadedElt == 1) {
8777 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8778 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8779 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8780 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8785 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8786 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8787 const X86Subtarget *Subtarget) {
8788 DebugLoc DL = N->getDebugLoc();
8789 SDValue Cond = N->getOperand(0);
8790 // Get the LHS/RHS of the select.
8791 SDValue LHS = N->getOperand(1);
8792 SDValue RHS = N->getOperand(2);
8794 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8795 // instructions have the peculiarity that if either operand is a NaN,
8796 // they chose what we call the RHS operand (and as such are not symmetric).
8797 // It happens that this matches the semantics of the common C idiom
8798 // x<y?x:y and related forms, so we can recognize these cases.
8799 if (Subtarget->hasSSE2() &&
8800 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8801 Cond.getOpcode() == ISD::SETCC) {
8802 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8804 unsigned Opcode = 0;
8805 // Check for x CC y ? x : y.
8806 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8810 // This can be a min if we can prove that at least one of the operands
8812 if (!FiniteOnlyFPMath()) {
8813 if (DAG.isKnownNeverNaN(RHS)) {
8814 // Put the potential NaN in the RHS so that SSE will preserve it.
8815 std::swap(LHS, RHS);
8816 } else if (!DAG.isKnownNeverNaN(LHS))
8819 Opcode = X86ISD::FMIN;
8822 // This can be a min if we can prove that at least one of the operands
8824 if (!FiniteOnlyFPMath()) {
8825 if (DAG.isKnownNeverNaN(LHS)) {
8826 // Put the potential NaN in the RHS so that SSE will preserve it.
8827 std::swap(LHS, RHS);
8828 } else if (!DAG.isKnownNeverNaN(RHS))
8831 Opcode = X86ISD::FMIN;
8834 // This can be a min, but if either operand is a NaN we need it to
8835 // preserve the original LHS.
8836 std::swap(LHS, RHS);
8840 Opcode = X86ISD::FMIN;
8844 // This can be a max if we can prove that at least one of the operands
8846 if (!FiniteOnlyFPMath()) {
8847 if (DAG.isKnownNeverNaN(LHS)) {
8848 // Put the potential NaN in the RHS so that SSE will preserve it.
8849 std::swap(LHS, RHS);
8850 } else if (!DAG.isKnownNeverNaN(RHS))
8853 Opcode = X86ISD::FMAX;
8856 // This can be a max if we can prove that at least one of the operands
8858 if (!FiniteOnlyFPMath()) {
8859 if (DAG.isKnownNeverNaN(RHS)) {
8860 // Put the potential NaN in the RHS so that SSE will preserve it.
8861 std::swap(LHS, RHS);
8862 } else if (!DAG.isKnownNeverNaN(LHS))
8865 Opcode = X86ISD::FMAX;
8868 // This can be a max, but if either operand is a NaN we need it to
8869 // preserve the original LHS.
8870 std::swap(LHS, RHS);
8874 Opcode = X86ISD::FMAX;
8877 // Check for x CC y ? y : x -- a min/max with reversed arms.
8878 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8882 // This can be a min if we can prove that at least one of the operands
8884 if (!FiniteOnlyFPMath()) {
8885 if (DAG.isKnownNeverNaN(RHS)) {
8886 // Put the potential NaN in the RHS so that SSE will preserve it.
8887 std::swap(LHS, RHS);
8888 } else if (!DAG.isKnownNeverNaN(LHS))
8891 Opcode = X86ISD::FMIN;
8894 // This can be a min if we can prove that at least one of the operands
8896 if (!FiniteOnlyFPMath()) {
8897 if (DAG.isKnownNeverNaN(LHS)) {
8898 // Put the potential NaN in the RHS so that SSE will preserve it.
8899 std::swap(LHS, RHS);
8900 } else if (!DAG.isKnownNeverNaN(RHS))
8903 Opcode = X86ISD::FMIN;
8906 // This can be a min, but if either operand is a NaN we need it to
8907 // preserve the original LHS.
8908 std::swap(LHS, RHS);
8912 Opcode = X86ISD::FMIN;
8916 // This can be a max if we can prove that at least one of the operands
8918 if (!FiniteOnlyFPMath()) {
8919 if (DAG.isKnownNeverNaN(LHS)) {
8920 // Put the potential NaN in the RHS so that SSE will preserve it.
8921 std::swap(LHS, RHS);
8922 } else if (!DAG.isKnownNeverNaN(RHS))
8925 Opcode = X86ISD::FMAX;
8928 // This can be a max if we can prove that at least one of the operands
8930 if (!FiniteOnlyFPMath()) {
8931 if (DAG.isKnownNeverNaN(RHS)) {
8932 // Put the potential NaN in the RHS so that SSE will preserve it.
8933 std::swap(LHS, RHS);
8934 } else if (!DAG.isKnownNeverNaN(LHS))
8937 Opcode = X86ISD::FMAX;
8940 // This can be a max, but if either operand is a NaN we need it to
8941 // preserve the original LHS.
8942 std::swap(LHS, RHS);
8946 Opcode = X86ISD::FMAX;
8952 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8955 // If this is a select between two integer constants, try to do some
8957 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8958 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8959 // Don't do this for crazy integer types.
8960 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8961 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8962 // so that TrueC (the true value) is larger than FalseC.
8963 bool NeedsCondInvert = false;
8965 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8966 // Efficiently invertible.
8967 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8968 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8969 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8970 NeedsCondInvert = true;
8971 std::swap(TrueC, FalseC);
8974 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8975 if (FalseC->getAPIntValue() == 0 &&
8976 TrueC->getAPIntValue().isPowerOf2()) {
8977 if (NeedsCondInvert) // Invert the condition if needed.
8978 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8979 DAG.getConstant(1, Cond.getValueType()));
8981 // Zero extend the condition if needed.
8982 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8984 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8985 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8986 DAG.getConstant(ShAmt, MVT::i8));
8989 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8990 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8991 if (NeedsCondInvert) // Invert the condition if needed.
8992 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8993 DAG.getConstant(1, Cond.getValueType()));
8995 // Zero extend the condition if needed.
8996 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8997 FalseC->getValueType(0), Cond);
8998 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8999 SDValue(FalseC, 0));
9002 // Optimize cases that will turn into an LEA instruction. This requires
9003 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9004 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9005 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9006 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9008 bool isFastMultiplier = false;
9010 switch ((unsigned char)Diff) {
9012 case 1: // result = add base, cond
9013 case 2: // result = lea base( , cond*2)
9014 case 3: // result = lea base(cond, cond*2)
9015 case 4: // result = lea base( , cond*4)
9016 case 5: // result = lea base(cond, cond*4)
9017 case 8: // result = lea base( , cond*8)
9018 case 9: // result = lea base(cond, cond*8)
9019 isFastMultiplier = true;
9024 if (isFastMultiplier) {
9025 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9026 if (NeedsCondInvert) // Invert the condition if needed.
9027 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9028 DAG.getConstant(1, Cond.getValueType()));
9030 // Zero extend the condition if needed.
9031 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9033 // Scale the condition by the difference.
9035 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9036 DAG.getConstant(Diff, Cond.getValueType()));
9038 // Add the base if non-zero.
9039 if (FalseC->getAPIntValue() != 0)
9040 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9041 SDValue(FalseC, 0));
9051 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9052 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9053 TargetLowering::DAGCombinerInfo &DCI) {
9054 DebugLoc DL = N->getDebugLoc();
9056 // If the flag operand isn't dead, don't touch this CMOV.
9057 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9060 // If this is a select between two integer constants, try to do some
9061 // optimizations. Note that the operands are ordered the opposite of SELECT
9063 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9064 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9065 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9066 // larger than FalseC (the false value).
9067 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9069 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9070 CC = X86::GetOppositeBranchCondition(CC);
9071 std::swap(TrueC, FalseC);
9074 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9075 // This is efficient for any integer data type (including i8/i16) and
9077 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9078 SDValue Cond = N->getOperand(3);
9079 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9080 DAG.getConstant(CC, MVT::i8), Cond);
9082 // Zero extend the condition if needed.
9083 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9085 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9086 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9087 DAG.getConstant(ShAmt, MVT::i8));
9088 if (N->getNumValues() == 2) // Dead flag value?
9089 return DCI.CombineTo(N, Cond, SDValue());
9093 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9094 // for any integer data type, including i8/i16.
9095 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9096 SDValue Cond = N->getOperand(3);
9097 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9098 DAG.getConstant(CC, MVT::i8), Cond);
9100 // Zero extend the condition if needed.
9101 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9102 FalseC->getValueType(0), Cond);
9103 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9104 SDValue(FalseC, 0));
9106 if (N->getNumValues() == 2) // Dead flag value?
9107 return DCI.CombineTo(N, Cond, SDValue());
9111 // Optimize cases that will turn into an LEA instruction. This requires
9112 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9113 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9114 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9115 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9117 bool isFastMultiplier = false;
9119 switch ((unsigned char)Diff) {
9121 case 1: // result = add base, cond
9122 case 2: // result = lea base( , cond*2)
9123 case 3: // result = lea base(cond, cond*2)
9124 case 4: // result = lea base( , cond*4)
9125 case 5: // result = lea base(cond, cond*4)
9126 case 8: // result = lea base( , cond*8)
9127 case 9: // result = lea base(cond, cond*8)
9128 isFastMultiplier = true;
9133 if (isFastMultiplier) {
9134 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9135 SDValue Cond = N->getOperand(3);
9136 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9137 DAG.getConstant(CC, MVT::i8), Cond);
9138 // Zero extend the condition if needed.
9139 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9141 // Scale the condition by the difference.
9143 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9144 DAG.getConstant(Diff, Cond.getValueType()));
9146 // Add the base if non-zero.
9147 if (FalseC->getAPIntValue() != 0)
9148 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9149 SDValue(FalseC, 0));
9150 if (N->getNumValues() == 2) // Dead flag value?
9151 return DCI.CombineTo(N, Cond, SDValue());
9161 /// PerformMulCombine - Optimize a single multiply with constant into two
9162 /// in order to implement it with two cheaper instructions, e.g.
9163 /// LEA + SHL, LEA + LEA.
9164 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9165 TargetLowering::DAGCombinerInfo &DCI) {
9166 if (DAG.getMachineFunction().
9167 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
9170 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9173 EVT VT = N->getValueType(0);
9177 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9180 uint64_t MulAmt = C->getZExtValue();
9181 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9184 uint64_t MulAmt1 = 0;
9185 uint64_t MulAmt2 = 0;
9186 if ((MulAmt % 9) == 0) {
9188 MulAmt2 = MulAmt / 9;
9189 } else if ((MulAmt % 5) == 0) {
9191 MulAmt2 = MulAmt / 5;
9192 } else if ((MulAmt % 3) == 0) {
9194 MulAmt2 = MulAmt / 3;
9197 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9198 DebugLoc DL = N->getDebugLoc();
9200 if (isPowerOf2_64(MulAmt2) &&
9201 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9202 // If second multiplifer is pow2, issue it first. We want the multiply by
9203 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9205 std::swap(MulAmt1, MulAmt2);
9208 if (isPowerOf2_64(MulAmt1))
9209 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9210 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9212 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9213 DAG.getConstant(MulAmt1, VT));
9215 if (isPowerOf2_64(MulAmt2))
9216 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9217 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9219 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9220 DAG.getConstant(MulAmt2, VT));
9222 // Do not add new nodes to DAG combiner worklist.
9223 DCI.CombineTo(N, NewMul, false);
9228 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9229 SDValue N0 = N->getOperand(0);
9230 SDValue N1 = N->getOperand(1);
9231 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9232 EVT VT = N0.getValueType();
9234 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9235 // since the result of setcc_c is all zero's or all ones.
9236 if (N1C && N0.getOpcode() == ISD::AND &&
9237 N0.getOperand(1).getOpcode() == ISD::Constant) {
9238 SDValue N00 = N0.getOperand(0);
9239 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9240 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9241 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9242 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9243 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9244 APInt ShAmt = N1C->getAPIntValue();
9245 Mask = Mask.shl(ShAmt);
9247 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9248 N00, DAG.getConstant(Mask, VT));
9255 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9257 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9258 const X86Subtarget *Subtarget) {
9259 EVT VT = N->getValueType(0);
9260 if (!VT.isVector() && VT.isInteger() &&
9261 N->getOpcode() == ISD::SHL)
9262 return PerformSHLCombine(N, DAG);
9264 // On X86 with SSE2 support, we can transform this to a vector shift if
9265 // all elements are shifted by the same amount. We can't do this in legalize
9266 // because the a constant vector is typically transformed to a constant pool
9267 // so we have no knowledge of the shift amount.
9268 if (!Subtarget->hasSSE2())
9271 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9274 SDValue ShAmtOp = N->getOperand(1);
9275 EVT EltVT = VT.getVectorElementType();
9276 DebugLoc DL = N->getDebugLoc();
9277 SDValue BaseShAmt = SDValue();
9278 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9279 unsigned NumElts = VT.getVectorNumElements();
9281 for (; i != NumElts; ++i) {
9282 SDValue Arg = ShAmtOp.getOperand(i);
9283 if (Arg.getOpcode() == ISD::UNDEF) continue;
9287 for (; i != NumElts; ++i) {
9288 SDValue Arg = ShAmtOp.getOperand(i);
9289 if (Arg.getOpcode() == ISD::UNDEF) continue;
9290 if (Arg != BaseShAmt) {
9294 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9295 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9296 SDValue InVec = ShAmtOp.getOperand(0);
9297 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9298 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9300 for (; i != NumElts; ++i) {
9301 SDValue Arg = InVec.getOperand(i);
9302 if (Arg.getOpcode() == ISD::UNDEF) continue;
9306 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9307 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9308 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9309 if (C->getZExtValue() == SplatIdx)
9310 BaseShAmt = InVec.getOperand(1);
9313 if (BaseShAmt.getNode() == 0)
9314 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9315 DAG.getIntPtrConstant(0));
9319 // The shift amount is an i32.
9320 if (EltVT.bitsGT(MVT::i32))
9321 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9322 else if (EltVT.bitsLT(MVT::i32))
9323 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9325 // The shift amount is identical so we can do a vector shift.
9326 SDValue ValOp = N->getOperand(0);
9327 switch (N->getOpcode()) {
9329 llvm_unreachable("Unknown shift opcode!");
9332 if (VT == MVT::v2i64)
9333 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9334 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9336 if (VT == MVT::v4i32)
9337 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9338 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9340 if (VT == MVT::v8i16)
9341 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9342 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9346 if (VT == MVT::v4i32)
9347 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9348 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9350 if (VT == MVT::v8i16)
9351 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9352 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9356 if (VT == MVT::v2i64)
9357 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9358 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9360 if (VT == MVT::v4i32)
9361 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9362 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9364 if (VT == MVT::v8i16)
9365 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9366 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9373 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9374 const X86Subtarget *Subtarget) {
9375 EVT VT = N->getValueType(0);
9376 if (VT != MVT::i64 || !Subtarget->is64Bit())
9379 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9380 SDValue N0 = N->getOperand(0);
9381 SDValue N1 = N->getOperand(1);
9382 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9384 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9387 SDValue ShAmt0 = N0.getOperand(1);
9388 if (ShAmt0.getValueType() != MVT::i8)
9390 SDValue ShAmt1 = N1.getOperand(1);
9391 if (ShAmt1.getValueType() != MVT::i8)
9393 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9394 ShAmt0 = ShAmt0.getOperand(0);
9395 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9396 ShAmt1 = ShAmt1.getOperand(0);
9398 DebugLoc DL = N->getDebugLoc();
9399 unsigned Opc = X86ISD::SHLD;
9400 SDValue Op0 = N0.getOperand(0);
9401 SDValue Op1 = N1.getOperand(0);
9402 if (ShAmt0.getOpcode() == ISD::SUB) {
9404 std::swap(Op0, Op1);
9405 std::swap(ShAmt0, ShAmt1);
9408 if (ShAmt1.getOpcode() == ISD::SUB) {
9409 SDValue Sum = ShAmt1.getOperand(0);
9410 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9411 if (SumC->getSExtValue() == 64 &&
9412 ShAmt1.getOperand(1) == ShAmt0)
9413 return DAG.getNode(Opc, DL, VT,
9415 DAG.getNode(ISD::TRUNCATE, DL,
9418 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9419 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9421 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9422 return DAG.getNode(Opc, DL, VT,
9423 N0.getOperand(0), N1.getOperand(0),
9424 DAG.getNode(ISD::TRUNCATE, DL,
9431 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9432 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9433 const X86Subtarget *Subtarget) {
9434 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9435 // the FP state in cases where an emms may be missing.
9436 // A preferable solution to the general problem is to figure out the right
9437 // places to insert EMMS. This qualifies as a quick hack.
9439 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9440 StoreSDNode *St = cast<StoreSDNode>(N);
9441 EVT VT = St->getValue().getValueType();
9442 if (VT.getSizeInBits() != 64)
9445 const Function *F = DAG.getMachineFunction().getFunction();
9446 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9447 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9448 && Subtarget->hasSSE2();
9449 if ((VT.isVector() ||
9450 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9451 isa<LoadSDNode>(St->getValue()) &&
9452 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9453 St->getChain().hasOneUse() && !St->isVolatile()) {
9454 SDNode* LdVal = St->getValue().getNode();
9456 int TokenFactorIndex = -1;
9457 SmallVector<SDValue, 8> Ops;
9458 SDNode* ChainVal = St->getChain().getNode();
9459 // Must be a store of a load. We currently handle two cases: the load
9460 // is a direct child, and it's under an intervening TokenFactor. It is
9461 // possible to dig deeper under nested TokenFactors.
9462 if (ChainVal == LdVal)
9463 Ld = cast<LoadSDNode>(St->getChain());
9464 else if (St->getValue().hasOneUse() &&
9465 ChainVal->getOpcode() == ISD::TokenFactor) {
9466 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9467 if (ChainVal->getOperand(i).getNode() == LdVal) {
9468 TokenFactorIndex = i;
9469 Ld = cast<LoadSDNode>(St->getValue());
9471 Ops.push_back(ChainVal->getOperand(i));
9475 if (!Ld || !ISD::isNormalLoad(Ld))
9478 // If this is not the MMX case, i.e. we are just turning i64 load/store
9479 // into f64 load/store, avoid the transformation if there are multiple
9480 // uses of the loaded value.
9481 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9484 DebugLoc LdDL = Ld->getDebugLoc();
9485 DebugLoc StDL = N->getDebugLoc();
9486 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9487 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9489 if (Subtarget->is64Bit() || F64IsLegal) {
9490 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9491 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9492 Ld->getBasePtr(), Ld->getSrcValue(),
9493 Ld->getSrcValueOffset(), Ld->isVolatile(),
9494 Ld->isNonTemporal(), Ld->getAlignment());
9495 SDValue NewChain = NewLd.getValue(1);
9496 if (TokenFactorIndex != -1) {
9497 Ops.push_back(NewChain);
9498 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9501 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9502 St->getSrcValue(), St->getSrcValueOffset(),
9503 St->isVolatile(), St->isNonTemporal(),
9504 St->getAlignment());
9507 // Otherwise, lower to two pairs of 32-bit loads / stores.
9508 SDValue LoAddr = Ld->getBasePtr();
9509 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9510 DAG.getConstant(4, MVT::i32));
9512 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9513 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9514 Ld->isVolatile(), Ld->isNonTemporal(),
9515 Ld->getAlignment());
9516 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9517 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9518 Ld->isVolatile(), Ld->isNonTemporal(),
9519 MinAlign(Ld->getAlignment(), 4));
9521 SDValue NewChain = LoLd.getValue(1);
9522 if (TokenFactorIndex != -1) {
9523 Ops.push_back(LoLd);
9524 Ops.push_back(HiLd);
9525 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9529 LoAddr = St->getBasePtr();
9530 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9531 DAG.getConstant(4, MVT::i32));
9533 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9534 St->getSrcValue(), St->getSrcValueOffset(),
9535 St->isVolatile(), St->isNonTemporal(),
9536 St->getAlignment());
9537 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9539 St->getSrcValueOffset() + 4,
9541 St->isNonTemporal(),
9542 MinAlign(St->getAlignment(), 4));
9543 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9548 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9549 /// X86ISD::FXOR nodes.
9550 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9551 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9552 // F[X]OR(0.0, x) -> x
9553 // F[X]OR(x, 0.0) -> x
9554 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9555 if (C->getValueAPF().isPosZero())
9556 return N->getOperand(1);
9557 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9558 if (C->getValueAPF().isPosZero())
9559 return N->getOperand(0);
9563 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9564 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9565 // FAND(0.0, x) -> 0.0
9566 // FAND(x, 0.0) -> 0.0
9567 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9568 if (C->getValueAPF().isPosZero())
9569 return N->getOperand(0);
9570 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9571 if (C->getValueAPF().isPosZero())
9572 return N->getOperand(1);
9576 static SDValue PerformBTCombine(SDNode *N,
9578 TargetLowering::DAGCombinerInfo &DCI) {
9579 // BT ignores high bits in the bit index operand.
9580 SDValue Op1 = N->getOperand(1);
9581 if (Op1.hasOneUse()) {
9582 unsigned BitWidth = Op1.getValueSizeInBits();
9583 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9584 APInt KnownZero, KnownOne;
9585 TargetLowering::TargetLoweringOpt TLO(DAG);
9586 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9587 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9588 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9589 DCI.CommitTargetLoweringOpt(TLO);
9594 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9595 SDValue Op = N->getOperand(0);
9596 if (Op.getOpcode() == ISD::BIT_CONVERT)
9597 Op = Op.getOperand(0);
9598 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9599 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9600 VT.getVectorElementType().getSizeInBits() ==
9601 OpVT.getVectorElementType().getSizeInBits()) {
9602 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9607 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9608 // Locked instructions, in turn, have implicit fence semantics (all memory
9609 // operations are flushed before issuing the locked instruction, and the
9610 // are not buffered), so we can fold away the common pattern of
9611 // fence-atomic-fence.
9612 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9613 SDValue atomic = N->getOperand(0);
9614 switch (atomic.getOpcode()) {
9615 case ISD::ATOMIC_CMP_SWAP:
9616 case ISD::ATOMIC_SWAP:
9617 case ISD::ATOMIC_LOAD_ADD:
9618 case ISD::ATOMIC_LOAD_SUB:
9619 case ISD::ATOMIC_LOAD_AND:
9620 case ISD::ATOMIC_LOAD_OR:
9621 case ISD::ATOMIC_LOAD_XOR:
9622 case ISD::ATOMIC_LOAD_NAND:
9623 case ISD::ATOMIC_LOAD_MIN:
9624 case ISD::ATOMIC_LOAD_MAX:
9625 case ISD::ATOMIC_LOAD_UMIN:
9626 case ISD::ATOMIC_LOAD_UMAX:
9632 SDValue fence = atomic.getOperand(0);
9633 if (fence.getOpcode() != ISD::MEMBARRIER)
9636 switch (atomic.getOpcode()) {
9637 case ISD::ATOMIC_CMP_SWAP:
9638 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9639 atomic.getOperand(1), atomic.getOperand(2),
9640 atomic.getOperand(3));
9641 case ISD::ATOMIC_SWAP:
9642 case ISD::ATOMIC_LOAD_ADD:
9643 case ISD::ATOMIC_LOAD_SUB:
9644 case ISD::ATOMIC_LOAD_AND:
9645 case ISD::ATOMIC_LOAD_OR:
9646 case ISD::ATOMIC_LOAD_XOR:
9647 case ISD::ATOMIC_LOAD_NAND:
9648 case ISD::ATOMIC_LOAD_MIN:
9649 case ISD::ATOMIC_LOAD_MAX:
9650 case ISD::ATOMIC_LOAD_UMIN:
9651 case ISD::ATOMIC_LOAD_UMAX:
9652 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9653 atomic.getOperand(1), atomic.getOperand(2));
9659 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9660 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9661 // (and (i32 x86isd::setcc_carry), 1)
9662 // This eliminates the zext. This transformation is necessary because
9663 // ISD::SETCC is always legalized to i8.
9664 DebugLoc dl = N->getDebugLoc();
9665 SDValue N0 = N->getOperand(0);
9666 EVT VT = N->getValueType(0);
9667 if (N0.getOpcode() == ISD::AND &&
9669 N0.getOperand(0).hasOneUse()) {
9670 SDValue N00 = N0.getOperand(0);
9671 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9673 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9674 if (!C || C->getZExtValue() != 1)
9676 return DAG.getNode(ISD::AND, dl, VT,
9677 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9678 N00.getOperand(0), N00.getOperand(1)),
9679 DAG.getConstant(1, VT));
9685 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9686 DAGCombinerInfo &DCI) const {
9687 SelectionDAG &DAG = DCI.DAG;
9688 switch (N->getOpcode()) {
9690 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9691 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9692 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9693 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9696 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9697 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9698 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9700 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9701 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9702 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9703 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9704 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9705 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9711 //===----------------------------------------------------------------------===//
9712 // X86 Inline Assembly Support
9713 //===----------------------------------------------------------------------===//
9715 static bool LowerToBSwap(CallInst *CI) {
9716 // FIXME: this should verify that we are targetting a 486 or better. If not,
9717 // we will turn this bswap into something that will be lowered to logical ops
9718 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9719 // so don't worry about this.
9721 // Verify this is a simple bswap.
9722 if (CI->getNumOperands() != 2 ||
9723 CI->getType() != CI->getOperand(1)->getType() ||
9724 !CI->getType()->isIntegerTy())
9727 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9728 if (!Ty || Ty->getBitWidth() % 16 != 0)
9731 // Okay, we can do this xform, do so now.
9732 const Type *Tys[] = { Ty };
9733 Module *M = CI->getParent()->getParent()->getParent();
9734 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9736 Value *Op = CI->getOperand(1);
9737 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9739 CI->replaceAllUsesWith(Op);
9740 CI->eraseFromParent();
9744 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9745 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9746 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9748 std::string AsmStr = IA->getAsmString();
9750 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9751 SmallVector<StringRef, 4> AsmPieces;
9752 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9754 switch (AsmPieces.size()) {
9755 default: return false;
9757 AsmStr = AsmPieces[0];
9759 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9762 if (AsmPieces.size() == 2 &&
9763 (AsmPieces[0] == "bswap" ||
9764 AsmPieces[0] == "bswapq" ||
9765 AsmPieces[0] == "bswapl") &&
9766 (AsmPieces[1] == "$0" ||
9767 AsmPieces[1] == "${0:q}")) {
9768 // No need to check constraints, nothing other than the equivalent of
9769 // "=r,0" would be valid here.
9770 return LowerToBSwap(CI);
9772 // rorw $$8, ${0:w} --> llvm.bswap.i16
9773 if (CI->getType()->isIntegerTy(16) &&
9774 AsmPieces.size() == 3 &&
9775 AsmPieces[0] == "rorw" &&
9776 AsmPieces[1] == "$$8," &&
9777 AsmPieces[2] == "${0:w}" &&
9778 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9779 return LowerToBSwap(CI);
9783 if (CI->getType()->isIntegerTy(64) &&
9784 Constraints.size() >= 2 &&
9785 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9786 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9787 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9788 SmallVector<StringRef, 4> Words;
9789 SplitString(AsmPieces[0], Words, " \t");
9790 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9792 SplitString(AsmPieces[1], Words, " \t");
9793 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9795 SplitString(AsmPieces[2], Words, " \t,");
9796 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9797 Words[2] == "%edx") {
9798 return LowerToBSwap(CI);
9810 /// getConstraintType - Given a constraint letter, return the type of
9811 /// constraint it is for this target.
9812 X86TargetLowering::ConstraintType
9813 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9814 if (Constraint.size() == 1) {
9815 switch (Constraint[0]) {
9827 return C_RegisterClass;
9835 return TargetLowering::getConstraintType(Constraint);
9838 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9839 /// with another that has more specific requirements based on the type of the
9840 /// corresponding operand.
9841 const char *X86TargetLowering::
9842 LowerXConstraint(EVT ConstraintVT) const {
9843 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9844 // 'f' like normal targets.
9845 if (ConstraintVT.isFloatingPoint()) {
9846 if (Subtarget->hasSSE2())
9848 if (Subtarget->hasSSE1())
9852 return TargetLowering::LowerXConstraint(ConstraintVT);
9855 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9856 /// vector. If it is invalid, don't add anything to Ops.
9857 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9860 std::vector<SDValue>&Ops,
9861 SelectionDAG &DAG) const {
9862 SDValue Result(0, 0);
9864 switch (Constraint) {
9867 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9868 if (C->getZExtValue() <= 31) {
9869 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9875 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9876 if (C->getZExtValue() <= 63) {
9877 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9883 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9884 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9885 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9891 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9892 if (C->getZExtValue() <= 255) {
9893 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9899 // 32-bit signed value
9900 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9901 const ConstantInt *CI = C->getConstantIntValue();
9902 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9903 C->getSExtValue())) {
9904 // Widen to 64 bits here to get it sign extended.
9905 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9908 // FIXME gcc accepts some relocatable values here too, but only in certain
9909 // memory models; it's complicated.
9914 // 32-bit unsigned value
9915 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9916 const ConstantInt *CI = C->getConstantIntValue();
9917 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9918 C->getZExtValue())) {
9919 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9923 // FIXME gcc accepts some relocatable values here too, but only in certain
9924 // memory models; it's complicated.
9928 // Literal immediates are always ok.
9929 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9930 // Widen to 64 bits here to get it sign extended.
9931 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9935 // If we are in non-pic codegen mode, we allow the address of a global (with
9936 // an optional displacement) to be used with 'i'.
9937 GlobalAddressSDNode *GA = 0;
9940 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9942 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9943 Offset += GA->getOffset();
9945 } else if (Op.getOpcode() == ISD::ADD) {
9946 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9947 Offset += C->getZExtValue();
9948 Op = Op.getOperand(0);
9951 } else if (Op.getOpcode() == ISD::SUB) {
9952 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9953 Offset += -C->getZExtValue();
9954 Op = Op.getOperand(0);
9959 // Otherwise, this isn't something we can handle, reject it.
9963 GlobalValue *GV = GA->getGlobal();
9964 // If we require an extra load to get this address, as in PIC mode, we
9966 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9967 getTargetMachine())))
9971 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9973 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9979 if (Result.getNode()) {
9980 Ops.push_back(Result);
9983 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9987 std::vector<unsigned> X86TargetLowering::
9988 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9990 if (Constraint.size() == 1) {
9991 // FIXME: not handling fp-stack yet!
9992 switch (Constraint[0]) { // GCC X86 Constraint Letters
9993 default: break; // Unknown constraint letter
9994 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9995 if (Subtarget->is64Bit()) {
9997 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9998 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9999 X86::R10D,X86::R11D,X86::R12D,
10000 X86::R13D,X86::R14D,X86::R15D,
10001 X86::EBP, X86::ESP, 0);
10002 else if (VT == MVT::i16)
10003 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10004 X86::SI, X86::DI, X86::R8W,X86::R9W,
10005 X86::R10W,X86::R11W,X86::R12W,
10006 X86::R13W,X86::R14W,X86::R15W,
10007 X86::BP, X86::SP, 0);
10008 else if (VT == MVT::i8)
10009 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10010 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10011 X86::R10B,X86::R11B,X86::R12B,
10012 X86::R13B,X86::R14B,X86::R15B,
10013 X86::BPL, X86::SPL, 0);
10015 else if (VT == MVT::i64)
10016 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10017 X86::RSI, X86::RDI, X86::R8, X86::R9,
10018 X86::R10, X86::R11, X86::R12,
10019 X86::R13, X86::R14, X86::R15,
10020 X86::RBP, X86::RSP, 0);
10024 // 32-bit fallthrough
10025 case 'Q': // Q_REGS
10026 if (VT == MVT::i32)
10027 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10028 else if (VT == MVT::i16)
10029 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10030 else if (VT == MVT::i8)
10031 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10032 else if (VT == MVT::i64)
10033 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10038 return std::vector<unsigned>();
10041 std::pair<unsigned, const TargetRegisterClass*>
10042 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10044 // First, see if this is a constraint that directly corresponds to an LLVM
10046 if (Constraint.size() == 1) {
10047 // GCC Constraint Letters
10048 switch (Constraint[0]) {
10050 case 'r': // GENERAL_REGS
10051 case 'l': // INDEX_REGS
10053 return std::make_pair(0U, X86::GR8RegisterClass);
10054 if (VT == MVT::i16)
10055 return std::make_pair(0U, X86::GR16RegisterClass);
10056 if (VT == MVT::i32 || !Subtarget->is64Bit())
10057 return std::make_pair(0U, X86::GR32RegisterClass);
10058 return std::make_pair(0U, X86::GR64RegisterClass);
10059 case 'R': // LEGACY_REGS
10061 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10062 if (VT == MVT::i16)
10063 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10064 if (VT == MVT::i32 || !Subtarget->is64Bit())
10065 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10066 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10067 case 'f': // FP Stack registers.
10068 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10069 // value to the correct fpstack register class.
10070 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10071 return std::make_pair(0U, X86::RFP32RegisterClass);
10072 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10073 return std::make_pair(0U, X86::RFP64RegisterClass);
10074 return std::make_pair(0U, X86::RFP80RegisterClass);
10075 case 'y': // MMX_REGS if MMX allowed.
10076 if (!Subtarget->hasMMX()) break;
10077 return std::make_pair(0U, X86::VR64RegisterClass);
10078 case 'Y': // SSE_REGS if SSE2 allowed
10079 if (!Subtarget->hasSSE2()) break;
10081 case 'x': // SSE_REGS if SSE1 allowed
10082 if (!Subtarget->hasSSE1()) break;
10084 switch (VT.getSimpleVT().SimpleTy) {
10086 // Scalar SSE types.
10089 return std::make_pair(0U, X86::FR32RegisterClass);
10092 return std::make_pair(0U, X86::FR64RegisterClass);
10100 return std::make_pair(0U, X86::VR128RegisterClass);
10106 // Use the default implementation in TargetLowering to convert the register
10107 // constraint into a member of a register class.
10108 std::pair<unsigned, const TargetRegisterClass*> Res;
10109 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10111 // Not found as a standard register?
10112 if (Res.second == 0) {
10113 // Map st(0) -> st(7) -> ST0
10114 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10115 tolower(Constraint[1]) == 's' &&
10116 tolower(Constraint[2]) == 't' &&
10117 Constraint[3] == '(' &&
10118 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10119 Constraint[5] == ')' &&
10120 Constraint[6] == '}') {
10122 Res.first = X86::ST0+Constraint[4]-'0';
10123 Res.second = X86::RFP80RegisterClass;
10127 // GCC allows "st(0)" to be called just plain "st".
10128 if (StringRef("{st}").equals_lower(Constraint)) {
10129 Res.first = X86::ST0;
10130 Res.second = X86::RFP80RegisterClass;
10135 if (StringRef("{flags}").equals_lower(Constraint)) {
10136 Res.first = X86::EFLAGS;
10137 Res.second = X86::CCRRegisterClass;
10141 // 'A' means EAX + EDX.
10142 if (Constraint == "A") {
10143 Res.first = X86::EAX;
10144 Res.second = X86::GR32_ADRegisterClass;
10150 // Otherwise, check to see if this is a register class of the wrong value
10151 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10152 // turn into {ax},{dx}.
10153 if (Res.second->hasType(VT))
10154 return Res; // Correct type already, nothing to do.
10156 // All of the single-register GCC register classes map their values onto
10157 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10158 // really want an 8-bit or 32-bit register, map to the appropriate register
10159 // class and return the appropriate register.
10160 if (Res.second == X86::GR16RegisterClass) {
10161 if (VT == MVT::i8) {
10162 unsigned DestReg = 0;
10163 switch (Res.first) {
10165 case X86::AX: DestReg = X86::AL; break;
10166 case X86::DX: DestReg = X86::DL; break;
10167 case X86::CX: DestReg = X86::CL; break;
10168 case X86::BX: DestReg = X86::BL; break;
10171 Res.first = DestReg;
10172 Res.second = X86::GR8RegisterClass;
10174 } else if (VT == MVT::i32) {
10175 unsigned DestReg = 0;
10176 switch (Res.first) {
10178 case X86::AX: DestReg = X86::EAX; break;
10179 case X86::DX: DestReg = X86::EDX; break;
10180 case X86::CX: DestReg = X86::ECX; break;
10181 case X86::BX: DestReg = X86::EBX; break;
10182 case X86::SI: DestReg = X86::ESI; break;
10183 case X86::DI: DestReg = X86::EDI; break;
10184 case X86::BP: DestReg = X86::EBP; break;
10185 case X86::SP: DestReg = X86::ESP; break;
10188 Res.first = DestReg;
10189 Res.second = X86::GR32RegisterClass;
10191 } else if (VT == MVT::i64) {
10192 unsigned DestReg = 0;
10193 switch (Res.first) {
10195 case X86::AX: DestReg = X86::RAX; break;
10196 case X86::DX: DestReg = X86::RDX; break;
10197 case X86::CX: DestReg = X86::RCX; break;
10198 case X86::BX: DestReg = X86::RBX; break;
10199 case X86::SI: DestReg = X86::RSI; break;
10200 case X86::DI: DestReg = X86::RDI; break;
10201 case X86::BP: DestReg = X86::RBP; break;
10202 case X86::SP: DestReg = X86::RSP; break;
10205 Res.first = DestReg;
10206 Res.second = X86::GR64RegisterClass;
10209 } else if (Res.second == X86::FR32RegisterClass ||
10210 Res.second == X86::FR64RegisterClass ||
10211 Res.second == X86::VR128RegisterClass) {
10212 // Handle references to XMM physical registers that got mapped into the
10213 // wrong class. This can happen with constraints like {xmm0} where the
10214 // target independent register mapper will just pick the first match it can
10215 // find, ignoring the required type.
10216 if (VT == MVT::f32)
10217 Res.second = X86::FR32RegisterClass;
10218 else if (VT == MVT::f64)
10219 Res.second = X86::FR64RegisterClass;
10220 else if (X86::VR128RegisterClass->hasType(VT))
10221 Res.second = X86::VR128RegisterClass;
10227 //===----------------------------------------------------------------------===//
10228 // X86 Widen vector type
10229 //===----------------------------------------------------------------------===//
10231 /// getWidenVectorType: given a vector type, returns the type to widen
10232 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10233 /// If there is no vector type that we want to widen to, returns MVT::Other
10234 /// When and where to widen is target dependent based on the cost of
10235 /// scalarizing vs using the wider vector type.
10237 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10238 assert(VT.isVector());
10239 if (isTypeLegal(VT))
10242 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10243 // type based on element type. This would speed up our search (though
10244 // it may not be worth it since the size of the list is relatively
10246 EVT EltVT = VT.getVectorElementType();
10247 unsigned NElts = VT.getVectorNumElements();
10249 // On X86, it make sense to widen any vector wider than 1
10253 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10254 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10255 EVT SVT = (MVT::SimpleValueType)nVT;
10257 if (isTypeLegal(SVT) &&
10258 SVT.getVectorElementType() == EltVT &&
10259 SVT.getVectorNumElements() > NElts)