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 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86TargetMachine.h"
19 #include "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/GlobalAlias.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Function.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/Intrinsics.h"
27 #include "llvm/LLVMContext.h"
28 #include "llvm/ADT/BitVector.h"
29 #include "llvm/ADT/VectorExtras.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/PseudoSourceValue.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Target/TargetLoweringObjectFile.h"
40 #include "llvm/Target/TargetOptions.h"
41 #include "llvm/ADT/SmallSet.h"
42 #include "llvm/ADT/StringExtras.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/raw_ostream.h"
48 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
50 // Forward declarations.
51 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
54 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
55 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
56 default: llvm_unreachable("unknown subtarget type");
57 case X86Subtarget::isDarwin:
58 return new TargetLoweringObjectFileMachO();
59 case X86Subtarget::isELF:
60 return new TargetLoweringObjectFileELF();
61 case X86Subtarget::isMingw:
62 case X86Subtarget::isCygwin:
63 case X86Subtarget::isWindows:
64 return new TargetLoweringObjectFileCOFF();
69 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
70 : TargetLowering(TM, createTLOF(TM)) {
71 Subtarget = &TM.getSubtarget<X86Subtarget>();
72 X86ScalarSSEf64 = Subtarget->hasSSE2();
73 X86ScalarSSEf32 = Subtarget->hasSSE1();
74 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
76 RegInfo = TM.getRegisterInfo();
79 // Set up the TargetLowering object.
81 // X86 is weird, it always uses i8 for shift amounts and setcc results.
82 setShiftAmountType(MVT::i8);
83 setBooleanContents(ZeroOrOneBooleanContent);
84 setSchedulingPreference(SchedulingForRegPressure);
85 setStackPointerRegisterToSaveRestore(X86StackPtr);
87 if (Subtarget->isTargetDarwin()) {
88 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
89 setUseUnderscoreSetJmp(false);
90 setUseUnderscoreLongJmp(false);
91 } else if (Subtarget->isTargetMingw()) {
92 // MS runtime is weird: it exports _setjmp, but longjmp!
93 setUseUnderscoreSetJmp(true);
94 setUseUnderscoreLongJmp(false);
96 setUseUnderscoreSetJmp(true);
97 setUseUnderscoreLongJmp(true);
100 // Set up the register classes.
101 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
102 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
103 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
104 if (Subtarget->is64Bit())
105 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
107 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
109 // We don't accept any truncstore of integer registers.
110 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
111 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
112 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
113 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
114 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
115 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
117 // SETOEQ and SETUNE require checking two conditions.
118 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
119 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
120 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
121 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
122 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
123 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
125 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
127 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
128 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
129 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
131 if (Subtarget->is64Bit()) {
132 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
133 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
134 } else if (!UseSoftFloat) {
135 if (X86ScalarSSEf64) {
136 // We have an impenetrably clever algorithm for ui64->double only.
137 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
139 // We have an algorithm for SSE2, and we turn this into a 64-bit
140 // FILD for other targets.
141 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
144 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
146 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
147 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
150 // SSE has no i16 to fp conversion, only i32
151 if (X86ScalarSSEf32) {
152 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
153 // f32 and f64 cases are Legal, f80 case is not
154 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
156 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
157 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
160 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
161 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
164 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
165 // are Legal, f80 is custom lowered.
166 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
167 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
169 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
171 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
172 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
174 if (X86ScalarSSEf32) {
175 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
176 // f32 and f64 cases are Legal, f80 case is not
177 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
179 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
180 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
183 // Handle FP_TO_UINT by promoting the destination to a larger signed
185 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
186 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
187 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
189 if (Subtarget->is64Bit()) {
190 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
191 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
192 } else if (!UseSoftFloat) {
193 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
194 // Expand FP_TO_UINT into a select.
195 // FIXME: We would like to use a Custom expander here eventually to do
196 // the optimal thing for SSE vs. the default expansion in the legalizer.
197 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
199 // With SSE3 we can use fisttpll to convert to a signed i64; without
200 // SSE, we're stuck with a fistpll.
201 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
204 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
205 if (!X86ScalarSSEf64) {
206 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
207 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
210 // Scalar integer divide and remainder are lowered to use operations that
211 // produce two results, to match the available instructions. This exposes
212 // the two-result form to trivial CSE, which is able to combine x/y and x%y
213 // into a single instruction.
215 // Scalar integer multiply-high is also lowered to use two-result
216 // operations, to match the available instructions. However, plain multiply
217 // (low) operations are left as Legal, as there are single-result
218 // instructions for this in x86. Using the two-result multiply instructions
219 // when both high and low results are needed must be arranged by dagcombine.
220 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
221 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
222 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
223 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
224 setOperationAction(ISD::SREM , MVT::i8 , Expand);
225 setOperationAction(ISD::UREM , MVT::i8 , Expand);
226 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
227 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
228 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
229 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
230 setOperationAction(ISD::SREM , MVT::i16 , Expand);
231 setOperationAction(ISD::UREM , MVT::i16 , Expand);
232 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
233 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
234 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
235 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
236 setOperationAction(ISD::SREM , MVT::i32 , Expand);
237 setOperationAction(ISD::UREM , MVT::i32 , Expand);
238 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
239 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
240 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
241 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
242 setOperationAction(ISD::SREM , MVT::i64 , Expand);
243 setOperationAction(ISD::UREM , MVT::i64 , Expand);
245 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
246 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
247 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
248 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
249 if (Subtarget->is64Bit())
250 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
251 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
252 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
253 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
254 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
255 setOperationAction(ISD::FREM , MVT::f32 , Expand);
256 setOperationAction(ISD::FREM , MVT::f64 , Expand);
257 setOperationAction(ISD::FREM , MVT::f80 , Expand);
258 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
260 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
261 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
262 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
263 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
264 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
265 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
266 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
267 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
268 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
269 if (Subtarget->is64Bit()) {
270 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
271 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
272 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
275 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
276 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
278 // These should be promoted to a larger select which is supported.
279 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
280 setOperationAction(ISD::SELECT , MVT::i8 , Promote);
281 // X86 wants to expand cmov itself.
282 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
283 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
284 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
285 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
286 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
287 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
288 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
289 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
290 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
291 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
292 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
293 if (Subtarget->is64Bit()) {
294 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
295 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
297 // X86 ret instruction may pop stack.
298 setOperationAction(ISD::RET , MVT::Other, Custom);
299 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
302 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
303 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
304 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
305 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
306 if (Subtarget->is64Bit())
307 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
308 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
309 if (Subtarget->is64Bit()) {
310 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
311 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
312 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
313 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
315 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
316 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
317 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
318 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
319 if (Subtarget->is64Bit()) {
320 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
321 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
322 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
325 if (Subtarget->hasSSE1())
326 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
328 if (!Subtarget->hasSSE2())
329 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
331 // Expand certain atomics
332 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
333 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
334 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
335 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
337 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
338 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
339 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
340 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
342 if (!Subtarget->is64Bit()) {
343 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
344 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
345 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
346 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
347 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
348 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
349 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
352 // Use the default ISD::DBG_STOPPOINT, ISD::DECLARE expansion.
353 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
354 // FIXME - use subtarget debug flags
355 if (!Subtarget->isTargetDarwin() &&
356 !Subtarget->isTargetELF() &&
357 !Subtarget->isTargetCygMing()) {
358 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
359 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
362 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
363 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
364 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
365 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
366 if (Subtarget->is64Bit()) {
367 setExceptionPointerRegister(X86::RAX);
368 setExceptionSelectorRegister(X86::RDX);
370 setExceptionPointerRegister(X86::EAX);
371 setExceptionSelectorRegister(X86::EDX);
373 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
374 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
376 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
378 setOperationAction(ISD::TRAP, MVT::Other, Legal);
380 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
381 setOperationAction(ISD::VASTART , MVT::Other, Custom);
382 setOperationAction(ISD::VAEND , MVT::Other, Expand);
383 if (Subtarget->is64Bit()) {
384 setOperationAction(ISD::VAARG , MVT::Other, Custom);
385 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
387 setOperationAction(ISD::VAARG , MVT::Other, Expand);
388 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
391 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
392 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
393 if (Subtarget->is64Bit())
394 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
395 if (Subtarget->isTargetCygMing())
396 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
398 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
400 if (!UseSoftFloat && X86ScalarSSEf64) {
401 // f32 and f64 use SSE.
402 // Set up the FP register classes.
403 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
404 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
406 // Use ANDPD to simulate FABS.
407 setOperationAction(ISD::FABS , MVT::f64, Custom);
408 setOperationAction(ISD::FABS , MVT::f32, Custom);
410 // Use XORP to simulate FNEG.
411 setOperationAction(ISD::FNEG , MVT::f64, Custom);
412 setOperationAction(ISD::FNEG , MVT::f32, Custom);
414 // Use ANDPD and ORPD to simulate FCOPYSIGN.
415 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
416 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
418 // We don't support sin/cos/fmod
419 setOperationAction(ISD::FSIN , MVT::f64, Expand);
420 setOperationAction(ISD::FCOS , MVT::f64, Expand);
421 setOperationAction(ISD::FSIN , MVT::f32, Expand);
422 setOperationAction(ISD::FCOS , MVT::f32, Expand);
424 // Expand FP immediates into loads from the stack, except for the special
426 addLegalFPImmediate(APFloat(+0.0)); // xorpd
427 addLegalFPImmediate(APFloat(+0.0f)); // xorps
428 } else if (!UseSoftFloat && X86ScalarSSEf32) {
429 // Use SSE for f32, x87 for f64.
430 // Set up the FP register classes.
431 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
432 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
434 // Use ANDPS to simulate FABS.
435 setOperationAction(ISD::FABS , MVT::f32, Custom);
437 // Use XORP to simulate FNEG.
438 setOperationAction(ISD::FNEG , MVT::f32, Custom);
440 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
442 // Use ANDPS and ORPS to simulate FCOPYSIGN.
443 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
444 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
446 // We don't support sin/cos/fmod
447 setOperationAction(ISD::FSIN , MVT::f32, Expand);
448 setOperationAction(ISD::FCOS , MVT::f32, Expand);
450 // Special cases we handle for FP constants.
451 addLegalFPImmediate(APFloat(+0.0f)); // xorps
452 addLegalFPImmediate(APFloat(+0.0)); // FLD0
453 addLegalFPImmediate(APFloat(+1.0)); // FLD1
454 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
455 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
458 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
459 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
461 } else if (!UseSoftFloat) {
462 // f32 and f64 in x87.
463 // Set up the FP register classes.
464 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
465 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
467 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
468 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
469 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
470 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
473 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
474 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
476 addLegalFPImmediate(APFloat(+0.0)); // FLD0
477 addLegalFPImmediate(APFloat(+1.0)); // FLD1
478 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
479 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
480 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
481 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
482 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
483 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
486 // Long double always uses X87.
488 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
489 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
490 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
493 APFloat TmpFlt(+0.0);
494 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
496 addLegalFPImmediate(TmpFlt); // FLD0
498 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
499 APFloat TmpFlt2(+1.0);
500 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
502 addLegalFPImmediate(TmpFlt2); // FLD1
503 TmpFlt2.changeSign();
504 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
508 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
509 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
513 // Always use a library call for pow.
514 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
515 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
516 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
518 setOperationAction(ISD::FLOG, MVT::f80, Expand);
519 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
520 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
521 setOperationAction(ISD::FEXP, MVT::f80, Expand);
522 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
524 // First set operation action for all vector types to either promote
525 // (for widening) or expand (for scalarization). Then we will selectively
526 // turn on ones that can be effectively codegen'd.
527 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
528 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
529 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
530 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
531 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
532 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
533 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
534 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
535 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
536 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
538 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
540 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
541 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
542 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
543 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
544 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
545 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
579 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
580 // with -msoft-float, disable use of MMX as well.
581 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
582 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
583 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
584 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
585 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
586 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
588 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
589 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
590 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
591 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
593 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
594 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
595 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
596 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
598 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
599 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
601 setOperationAction(ISD::AND, MVT::v8i8, Promote);
602 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
603 setOperationAction(ISD::AND, MVT::v4i16, Promote);
604 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
605 setOperationAction(ISD::AND, MVT::v2i32, Promote);
606 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
607 setOperationAction(ISD::AND, MVT::v1i64, Legal);
609 setOperationAction(ISD::OR, MVT::v8i8, Promote);
610 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
611 setOperationAction(ISD::OR, MVT::v4i16, Promote);
612 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
613 setOperationAction(ISD::OR, MVT::v2i32, Promote);
614 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
615 setOperationAction(ISD::OR, MVT::v1i64, Legal);
617 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
618 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
619 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
620 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
621 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
622 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
623 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
625 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
626 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
627 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
628 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
629 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
630 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
631 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
632 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
633 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
635 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
636 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
637 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
638 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
639 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
641 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
642 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
643 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
644 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
646 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
647 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
648 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
649 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
651 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
653 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
654 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
655 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
656 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
657 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
658 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
659 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
660 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
661 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
664 if (!UseSoftFloat && Subtarget->hasSSE1()) {
665 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
667 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
668 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
669 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
670 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
671 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
672 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
673 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
674 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
675 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
676 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
677 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
678 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
681 if (!UseSoftFloat && Subtarget->hasSSE2()) {
682 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
684 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
685 // registers cannot be used even for integer operations.
686 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
687 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
688 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
689 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
691 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
692 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
693 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
694 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
695 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
696 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
697 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
698 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
699 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
700 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
701 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
702 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
703 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
704 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
705 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
706 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
708 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
709 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
710 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
711 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
713 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
714 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
715 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
716 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
717 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
719 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
720 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
721 MVT VT = (MVT::SimpleValueType)i;
722 // Do not attempt to custom lower non-power-of-2 vectors
723 if (!isPowerOf2_32(VT.getVectorNumElements()))
725 // Do not attempt to custom lower non-128-bit vectors
726 if (!VT.is128BitVector())
728 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
729 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
730 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
733 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
734 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
735 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
736 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
737 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
738 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
740 if (Subtarget->is64Bit()) {
741 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
742 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
745 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
746 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
747 MVT VT = (MVT::SimpleValueType)i;
749 // Do not attempt to promote non-128-bit vectors
750 if (!VT.is128BitVector()) {
753 setOperationAction(ISD::AND, VT, Promote);
754 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
755 setOperationAction(ISD::OR, VT, Promote);
756 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
757 setOperationAction(ISD::XOR, VT, Promote);
758 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
759 setOperationAction(ISD::LOAD, VT, Promote);
760 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
761 setOperationAction(ISD::SELECT, VT, Promote);
762 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
765 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
767 // Custom lower v2i64 and v2f64 selects.
768 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
769 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
770 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
771 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
773 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
774 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
775 if (!DisableMMX && Subtarget->hasMMX()) {
776 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
777 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
781 if (Subtarget->hasSSE41()) {
782 // FIXME: Do we need to handle scalar-to-vector here?
783 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
785 // i8 and i16 vectors are custom , because the source register and source
786 // source memory operand types are not the same width. f32 vectors are
787 // custom since the immediate controlling the insert encodes additional
789 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
790 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
791 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
792 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
794 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
795 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
796 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
797 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
799 if (Subtarget->is64Bit()) {
800 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
801 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
805 if (Subtarget->hasSSE42()) {
806 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
809 if (!UseSoftFloat && Subtarget->hasAVX()) {
810 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
811 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
812 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
813 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
815 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
816 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
817 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
818 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
819 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
820 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
821 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
822 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
823 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
824 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
825 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
826 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
827 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
828 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
829 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
831 // Operations to consider commented out -v16i16 v32i8
832 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
833 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
834 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
835 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
836 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
837 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
838 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
839 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
840 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
841 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
842 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
843 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
844 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
845 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
847 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
848 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
849 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
850 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
852 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
853 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
854 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
855 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
856 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
858 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
859 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
860 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
861 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
862 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
863 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
866 // Not sure we want to do this since there are no 256-bit integer
869 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
870 // This includes 256-bit vectors
871 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
872 MVT VT = (MVT::SimpleValueType)i;
874 // Do not attempt to custom lower non-power-of-2 vectors
875 if (!isPowerOf2_32(VT.getVectorNumElements()))
878 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
879 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
880 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
883 if (Subtarget->is64Bit()) {
884 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
885 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
890 // Not sure we want to do this since there are no 256-bit integer
893 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
894 // Including 256-bit vectors
895 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
896 MVT VT = (MVT::SimpleValueType)i;
898 if (!VT.is256BitVector()) {
901 setOperationAction(ISD::AND, VT, Promote);
902 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
903 setOperationAction(ISD::OR, VT, Promote);
904 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
905 setOperationAction(ISD::XOR, VT, Promote);
906 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
907 setOperationAction(ISD::LOAD, VT, Promote);
908 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
909 setOperationAction(ISD::SELECT, VT, Promote);
910 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
913 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
917 // We want to custom lower some of our intrinsics.
918 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
920 // Add/Sub/Mul with overflow operations are custom lowered.
921 setOperationAction(ISD::SADDO, MVT::i32, Custom);
922 setOperationAction(ISD::SADDO, MVT::i64, Custom);
923 setOperationAction(ISD::UADDO, MVT::i32, Custom);
924 setOperationAction(ISD::UADDO, MVT::i64, Custom);
925 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
926 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
927 setOperationAction(ISD::USUBO, MVT::i32, Custom);
928 setOperationAction(ISD::USUBO, MVT::i64, Custom);
929 setOperationAction(ISD::SMULO, MVT::i32, Custom);
930 setOperationAction(ISD::SMULO, MVT::i64, Custom);
932 if (!Subtarget->is64Bit()) {
933 // These libcalls are not available in 32-bit.
934 setLibcallName(RTLIB::SHL_I128, 0);
935 setLibcallName(RTLIB::SRL_I128, 0);
936 setLibcallName(RTLIB::SRA_I128, 0);
939 // We have target-specific dag combine patterns for the following nodes:
940 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
941 setTargetDAGCombine(ISD::BUILD_VECTOR);
942 setTargetDAGCombine(ISD::SELECT);
943 setTargetDAGCombine(ISD::SHL);
944 setTargetDAGCombine(ISD::SRA);
945 setTargetDAGCombine(ISD::SRL);
946 setTargetDAGCombine(ISD::STORE);
947 setTargetDAGCombine(ISD::MEMBARRIER);
948 if (Subtarget->is64Bit())
949 setTargetDAGCombine(ISD::MUL);
951 computeRegisterProperties();
953 // FIXME: These should be based on subtarget info. Plus, the values should
954 // be smaller when we are in optimizing for size mode.
955 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
956 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
957 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
958 allowUnalignedMemoryAccesses = true; // x86 supports it!
959 setPrefLoopAlignment(16);
960 benefitFromCodePlacementOpt = true;
964 MVT X86TargetLowering::getSetCCResultType(MVT VT) const {
969 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
970 /// the desired ByVal argument alignment.
971 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
974 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
975 if (VTy->getBitWidth() == 128)
977 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
978 unsigned EltAlign = 0;
979 getMaxByValAlign(ATy->getElementType(), EltAlign);
980 if (EltAlign > MaxAlign)
982 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
983 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
984 unsigned EltAlign = 0;
985 getMaxByValAlign(STy->getElementType(i), EltAlign);
986 if (EltAlign > MaxAlign)
995 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
996 /// function arguments in the caller parameter area. For X86, aggregates
997 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
998 /// are at 4-byte boundaries.
999 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1000 if (Subtarget->is64Bit()) {
1001 // Max of 8 and alignment of type.
1002 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1009 if (Subtarget->hasSSE1())
1010 getMaxByValAlign(Ty, Align);
1014 /// getOptimalMemOpType - Returns the target specific optimal type for load
1015 /// and store operations as a result of memset, memcpy, and memmove
1016 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1019 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1020 bool isSrcConst, bool isSrcStr,
1021 SelectionDAG &DAG) const {
1022 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1023 // linux. This is because the stack realignment code can't handle certain
1024 // cases like PR2962. This should be removed when PR2962 is fixed.
1025 const Function *F = DAG.getMachineFunction().getFunction();
1026 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1027 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1028 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1030 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1033 if (Subtarget->is64Bit() && Size >= 8)
1038 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1040 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1041 SelectionDAG &DAG) const {
1042 if (usesGlobalOffsetTable())
1043 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1044 if (!Subtarget->is64Bit())
1045 // This doesn't have DebugLoc associated with it, but is not really the
1046 // same as a Register.
1047 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1052 /// getFunctionAlignment - Return the Log2 alignment of this function.
1053 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1054 return F->hasFnAttr(Attribute::OptimizeForSize) ? 1 : 4;
1057 //===----------------------------------------------------------------------===//
1058 // Return Value Calling Convention Implementation
1059 //===----------------------------------------------------------------------===//
1061 #include "X86GenCallingConv.inc"
1063 /// LowerRET - Lower an ISD::RET node.
1064 SDValue X86TargetLowering::LowerRET(SDValue Op, SelectionDAG &DAG) {
1065 DebugLoc dl = Op.getDebugLoc();
1066 assert((Op.getNumOperands() & 1) == 1 && "ISD::RET should have odd # args");
1068 SmallVector<CCValAssign, 16> RVLocs;
1069 unsigned CC = DAG.getMachineFunction().getFunction()->getCallingConv();
1070 bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
1071 CCState CCInfo(CC, isVarArg, getTargetMachine(), RVLocs, *DAG.getContext());
1072 CCInfo.AnalyzeReturn(Op.getNode(), RetCC_X86);
1074 // If this is the first return lowered for this function, add the regs to the
1075 // liveout set for the function.
1076 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1077 for (unsigned i = 0; i != RVLocs.size(); ++i)
1078 if (RVLocs[i].isRegLoc())
1079 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1081 SDValue Chain = Op.getOperand(0);
1083 // Handle tail call return.
1084 Chain = GetPossiblePreceedingTailCall(Chain, X86ISD::TAILCALL);
1085 if (Chain.getOpcode() == X86ISD::TAILCALL) {
1086 SDValue TailCall = Chain;
1087 SDValue TargetAddress = TailCall.getOperand(1);
1088 SDValue StackAdjustment = TailCall.getOperand(2);
1089 assert(((TargetAddress.getOpcode() == ISD::Register &&
1090 (cast<RegisterSDNode>(TargetAddress)->getReg() == X86::EAX ||
1091 cast<RegisterSDNode>(TargetAddress)->getReg() == X86::R11)) ||
1092 TargetAddress.getOpcode() == ISD::TargetExternalSymbol ||
1093 TargetAddress.getOpcode() == ISD::TargetGlobalAddress) &&
1094 "Expecting an global address, external symbol, or register");
1095 assert(StackAdjustment.getOpcode() == ISD::Constant &&
1096 "Expecting a const value");
1098 SmallVector<SDValue,8> Operands;
1099 Operands.push_back(Chain.getOperand(0));
1100 Operands.push_back(TargetAddress);
1101 Operands.push_back(StackAdjustment);
1102 // Copy registers used by the call. Last operand is a flag so it is not
1104 for (unsigned i=3; i < TailCall.getNumOperands()-1; i++) {
1105 Operands.push_back(Chain.getOperand(i));
1107 return DAG.getNode(X86ISD::TC_RETURN, dl, MVT::Other, &Operands[0],
1114 SmallVector<SDValue, 6> RetOps;
1115 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1116 // Operand #1 = Bytes To Pop
1117 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
1119 // Copy the result values into the output registers.
1120 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1121 CCValAssign &VA = RVLocs[i];
1122 assert(VA.isRegLoc() && "Can only return in registers!");
1123 SDValue ValToCopy = Op.getOperand(i*2+1);
1125 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1126 // the RET instruction and handled by the FP Stackifier.
1127 if (VA.getLocReg() == X86::ST0 ||
1128 VA.getLocReg() == X86::ST1) {
1129 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1130 // change the value to the FP stack register class.
1131 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1132 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1133 RetOps.push_back(ValToCopy);
1134 // Don't emit a copytoreg.
1138 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1139 // which is returned in RAX / RDX.
1140 if (Subtarget->is64Bit()) {
1141 MVT ValVT = ValToCopy.getValueType();
1142 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1143 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1144 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1145 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1149 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1150 Flag = Chain.getValue(1);
1153 // The x86-64 ABI for returning structs by value requires that we copy
1154 // the sret argument into %rax for the return. We saved the argument into
1155 // a virtual register in the entry block, so now we copy the value out
1157 if (Subtarget->is64Bit() &&
1158 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1159 MachineFunction &MF = DAG.getMachineFunction();
1160 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1161 unsigned Reg = FuncInfo->getSRetReturnReg();
1163 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1164 FuncInfo->setSRetReturnReg(Reg);
1166 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1168 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1169 Flag = Chain.getValue(1);
1172 RetOps[0] = Chain; // Update chain.
1174 // Add the flag if we have it.
1176 RetOps.push_back(Flag);
1178 return DAG.getNode(X86ISD::RET_FLAG, dl,
1179 MVT::Other, &RetOps[0], RetOps.size());
1183 /// LowerCallResult - Lower the result values of an ISD::CALL into the
1184 /// appropriate copies out of appropriate physical registers. This assumes that
1185 /// Chain/InFlag are the input chain/flag to use, and that TheCall is the call
1186 /// being lowered. The returns a SDNode with the same number of values as the
1188 SDNode *X86TargetLowering::
1189 LowerCallResult(SDValue Chain, SDValue InFlag, CallSDNode *TheCall,
1190 unsigned CallingConv, SelectionDAG &DAG) {
1192 DebugLoc dl = TheCall->getDebugLoc();
1193 // Assign locations to each value returned by this call.
1194 SmallVector<CCValAssign, 16> RVLocs;
1195 bool isVarArg = TheCall->isVarArg();
1196 bool Is64Bit = Subtarget->is64Bit();
1197 CCState CCInfo(CallingConv, isVarArg, getTargetMachine(),
1198 RVLocs, *DAG.getContext());
1199 CCInfo.AnalyzeCallResult(TheCall, RetCC_X86);
1201 SmallVector<SDValue, 8> ResultVals;
1203 // Copy all of the result registers out of their specified physreg.
1204 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1205 CCValAssign &VA = RVLocs[i];
1206 MVT CopyVT = VA.getValVT();
1208 // If this is x86-64, and we disabled SSE, we can't return FP values
1209 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1210 ((Is64Bit || TheCall->isInreg()) && !Subtarget->hasSSE1())) {
1211 llvm_report_error("SSE register return with SSE disabled");
1214 // If this is a call to a function that returns an fp value on the floating
1215 // point stack, but where we prefer to use the value in xmm registers, copy
1216 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1217 if ((VA.getLocReg() == X86::ST0 ||
1218 VA.getLocReg() == X86::ST1) &&
1219 isScalarFPTypeInSSEReg(VA.getValVT())) {
1224 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1225 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1226 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1227 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1228 MVT::v2i64, InFlag).getValue(1);
1229 Val = Chain.getValue(0);
1230 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1231 Val, DAG.getConstant(0, MVT::i64));
1233 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1234 MVT::i64, InFlag).getValue(1);
1235 Val = Chain.getValue(0);
1237 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1239 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1240 CopyVT, InFlag).getValue(1);
1241 Val = Chain.getValue(0);
1243 InFlag = Chain.getValue(2);
1245 if (CopyVT != VA.getValVT()) {
1246 // Round the F80 the right size, which also moves to the appropriate xmm
1248 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1249 // This truncation won't change the value.
1250 DAG.getIntPtrConstant(1));
1253 ResultVals.push_back(Val);
1256 // Merge everything together with a MERGE_VALUES node.
1257 ResultVals.push_back(Chain);
1258 return DAG.getNode(ISD::MERGE_VALUES, dl, TheCall->getVTList(),
1259 &ResultVals[0], ResultVals.size()).getNode();
1263 //===----------------------------------------------------------------------===//
1264 // C & StdCall & Fast Calling Convention implementation
1265 //===----------------------------------------------------------------------===//
1266 // StdCall calling convention seems to be standard for many Windows' API
1267 // routines and around. It differs from C calling convention just a little:
1268 // callee should clean up the stack, not caller. Symbols should be also
1269 // decorated in some fancy way :) It doesn't support any vector arguments.
1270 // For info on fast calling convention see Fast Calling Convention (tail call)
1271 // implementation LowerX86_32FastCCCallTo.
1273 /// CallIsStructReturn - Determines whether a CALL node uses struct return
1275 static bool CallIsStructReturn(CallSDNode *TheCall) {
1276 unsigned NumOps = TheCall->getNumArgs();
1280 return TheCall->getArgFlags(0).isSRet();
1283 /// ArgsAreStructReturn - Determines whether a function uses struct
1284 /// return semantics.
1285 static bool ArgsAreStructReturn(SDValue Op) {
1286 unsigned NumArgs = Op.getNode()->getNumValues() - 1;
1290 return cast<ARG_FLAGSSDNode>(Op.getOperand(3))->getArgFlags().isSRet();
1293 /// IsCalleePop - Determines whether the callee is required to pop its
1294 /// own arguments. Callee pop is necessary to support tail calls.
1295 bool X86TargetLowering::IsCalleePop(bool IsVarArg, unsigned CallingConv) {
1299 switch (CallingConv) {
1302 case CallingConv::X86_StdCall:
1303 return !Subtarget->is64Bit();
1304 case CallingConv::X86_FastCall:
1305 return !Subtarget->is64Bit();
1306 case CallingConv::Fast:
1307 return PerformTailCallOpt;
1311 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1312 /// given CallingConvention value.
1313 CCAssignFn *X86TargetLowering::CCAssignFnForNode(unsigned CC) const {
1314 if (Subtarget->is64Bit()) {
1315 if (Subtarget->isTargetWin64())
1316 return CC_X86_Win64_C;
1321 if (CC == CallingConv::X86_FastCall)
1322 return CC_X86_32_FastCall;
1323 else if (CC == CallingConv::Fast)
1324 return CC_X86_32_FastCC;
1329 /// NameDecorationForFORMAL_ARGUMENTS - Selects the appropriate decoration to
1330 /// apply to a MachineFunction containing a given FORMAL_ARGUMENTS node.
1332 X86TargetLowering::NameDecorationForFORMAL_ARGUMENTS(SDValue Op) {
1333 unsigned CC = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
1334 if (CC == CallingConv::X86_FastCall)
1336 else if (CC == CallingConv::X86_StdCall)
1342 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1343 /// by "Src" to address "Dst" with size and alignment information specified by
1344 /// the specific parameter attribute. The copy will be passed as a byval
1345 /// function parameter.
1347 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1348 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1350 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1351 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1352 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1355 SDValue X86TargetLowering::LowerMemArgument(SDValue Op, SelectionDAG &DAG,
1356 const CCValAssign &VA,
1357 MachineFrameInfo *MFI,
1359 SDValue Root, unsigned i) {
1360 // Create the nodes corresponding to a load from this parameter slot.
1361 ISD::ArgFlagsTy Flags =
1362 cast<ARG_FLAGSSDNode>(Op.getOperand(3 + i))->getArgFlags();
1363 bool AlwaysUseMutable = (CC==CallingConv::Fast) && PerformTailCallOpt;
1364 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1366 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1367 // changed with more analysis.
1368 // In case of tail call optimization mark all arguments mutable. Since they
1369 // could be overwritten by lowering of arguments in case of a tail call.
1370 int FI = MFI->CreateFixedObject(VA.getValVT().getSizeInBits()/8,
1371 VA.getLocMemOffset(), isImmutable);
1372 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1373 if (Flags.isByVal())
1375 return DAG.getLoad(VA.getValVT(), Op.getDebugLoc(), Root, FIN,
1376 PseudoSourceValue::getFixedStack(FI), 0);
1380 X86TargetLowering::LowerFORMAL_ARGUMENTS(SDValue Op, SelectionDAG &DAG) {
1381 MachineFunction &MF = DAG.getMachineFunction();
1382 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1383 DebugLoc dl = Op.getDebugLoc();
1385 const Function* Fn = MF.getFunction();
1386 if (Fn->hasExternalLinkage() &&
1387 Subtarget->isTargetCygMing() &&
1388 Fn->getName() == "main")
1389 FuncInfo->setForceFramePointer(true);
1391 // Decorate the function name.
1392 FuncInfo->setDecorationStyle(NameDecorationForFORMAL_ARGUMENTS(Op));
1394 MachineFrameInfo *MFI = MF.getFrameInfo();
1395 SDValue Root = Op.getOperand(0);
1396 bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() != 0;
1397 unsigned CC = MF.getFunction()->getCallingConv();
1398 bool Is64Bit = Subtarget->is64Bit();
1399 bool IsWin64 = Subtarget->isTargetWin64();
1401 assert(!(isVarArg && CC == CallingConv::Fast) &&
1402 "Var args not supported with calling convention fastcc");
1404 // Assign locations to all of the incoming arguments.
1405 SmallVector<CCValAssign, 16> ArgLocs;
1406 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs, *DAG.getContext());
1407 CCInfo.AnalyzeFormalArguments(Op.getNode(), CCAssignFnForNode(CC));
1409 SmallVector<SDValue, 8> ArgValues;
1410 unsigned LastVal = ~0U;
1411 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1412 CCValAssign &VA = ArgLocs[i];
1413 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1415 assert(VA.getValNo() != LastVal &&
1416 "Don't support value assigned to multiple locs yet");
1417 LastVal = VA.getValNo();
1419 if (VA.isRegLoc()) {
1420 MVT RegVT = VA.getLocVT();
1421 TargetRegisterClass *RC = NULL;
1422 if (RegVT == MVT::i32)
1423 RC = X86::GR32RegisterClass;
1424 else if (Is64Bit && RegVT == MVT::i64)
1425 RC = X86::GR64RegisterClass;
1426 else if (RegVT == MVT::f32)
1427 RC = X86::FR32RegisterClass;
1428 else if (RegVT == MVT::f64)
1429 RC = X86::FR64RegisterClass;
1430 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1431 RC = X86::VR128RegisterClass;
1432 else if (RegVT.isVector()) {
1433 assert(RegVT.getSizeInBits() == 64);
1435 RC = X86::VR64RegisterClass; // MMX values are passed in MMXs.
1437 // Darwin calling convention passes MMX values in either GPRs or
1438 // XMMs in x86-64. Other targets pass them in memory.
1439 if (RegVT != MVT::v1i64 && Subtarget->hasSSE2()) {
1440 RC = X86::VR128RegisterClass; // MMX values are passed in XMMs.
1443 RC = X86::GR64RegisterClass; // v1i64 values are passed in GPRs.
1448 llvm_unreachable("Unknown argument type!");
1451 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1452 SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, RegVT);
1454 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1455 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1457 if (VA.getLocInfo() == CCValAssign::SExt)
1458 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1459 DAG.getValueType(VA.getValVT()));
1460 else if (VA.getLocInfo() == CCValAssign::ZExt)
1461 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1462 DAG.getValueType(VA.getValVT()));
1464 if (VA.getLocInfo() != CCValAssign::Full)
1465 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1467 // Handle MMX values passed in GPRs.
1468 if (Is64Bit && RegVT != VA.getLocVT()) {
1469 if (RegVT.getSizeInBits() == 64 && RC == X86::GR64RegisterClass)
1470 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1471 else if (RC == X86::VR128RegisterClass) {
1472 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1473 ArgValue, DAG.getConstant(0, MVT::i64));
1474 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getLocVT(), ArgValue);
1478 ArgValues.push_back(ArgValue);
1480 assert(VA.isMemLoc());
1481 ArgValues.push_back(LowerMemArgument(Op, DAG, VA, MFI, CC, Root, i));
1485 // The x86-64 ABI for returning structs by value requires that we copy
1486 // the sret argument into %rax for the return. Save the argument into
1487 // a virtual register so that we can access it from the return points.
1488 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1489 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1490 unsigned Reg = FuncInfo->getSRetReturnReg();
1492 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1493 FuncInfo->setSRetReturnReg(Reg);
1495 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, ArgValues[0]);
1496 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Root);
1499 unsigned StackSize = CCInfo.getNextStackOffset();
1500 // align stack specially for tail calls
1501 if (PerformTailCallOpt && CC == CallingConv::Fast)
1502 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1504 // If the function takes variable number of arguments, make a frame index for
1505 // the start of the first vararg value... for expansion of llvm.va_start.
1507 if (Is64Bit || CC != CallingConv::X86_FastCall) {
1508 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1511 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1513 // FIXME: We should really autogenerate these arrays
1514 static const unsigned GPR64ArgRegsWin64[] = {
1515 X86::RCX, X86::RDX, X86::R8, X86::R9
1517 static const unsigned XMMArgRegsWin64[] = {
1518 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1520 static const unsigned GPR64ArgRegs64Bit[] = {
1521 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1523 static const unsigned XMMArgRegs64Bit[] = {
1524 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1525 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1527 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1530 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1531 GPR64ArgRegs = GPR64ArgRegsWin64;
1532 XMMArgRegs = XMMArgRegsWin64;
1534 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1535 GPR64ArgRegs = GPR64ArgRegs64Bit;
1536 XMMArgRegs = XMMArgRegs64Bit;
1538 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1540 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1543 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1544 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1545 "SSE register cannot be used when SSE is disabled!");
1546 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1547 "SSE register cannot be used when SSE is disabled!");
1548 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1549 // Kernel mode asks for SSE to be disabled, so don't push them
1551 TotalNumXMMRegs = 0;
1553 // For X86-64, if there are vararg parameters that are passed via
1554 // registers, then we must store them to their spots on the stack so they
1555 // may be loaded by deferencing the result of va_next.
1556 VarArgsGPOffset = NumIntRegs * 8;
1557 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1558 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1559 TotalNumXMMRegs * 16, 16);
1561 // Store the integer parameter registers.
1562 SmallVector<SDValue, 8> MemOps;
1563 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1564 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1565 DAG.getIntPtrConstant(VarArgsGPOffset));
1566 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1567 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1568 X86::GR64RegisterClass);
1569 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::i64);
1571 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1572 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1573 MemOps.push_back(Store);
1574 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1575 DAG.getIntPtrConstant(8));
1578 // Now store the XMM (fp + vector) parameter registers.
1579 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1580 DAG.getIntPtrConstant(VarArgsFPOffset));
1581 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1582 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1583 X86::VR128RegisterClass);
1584 SDValue Val = DAG.getCopyFromReg(Root, dl, VReg, MVT::v4f32);
1586 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1587 PseudoSourceValue::getFixedStack(RegSaveFrameIndex), 0);
1588 MemOps.push_back(Store);
1589 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
1590 DAG.getIntPtrConstant(16));
1592 if (!MemOps.empty())
1593 Root = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1594 &MemOps[0], MemOps.size());
1598 ArgValues.push_back(Root);
1600 // Some CCs need callee pop.
1601 if (IsCalleePop(isVarArg, CC)) {
1602 BytesToPopOnReturn = StackSize; // Callee pops everything.
1603 BytesCallerReserves = 0;
1605 BytesToPopOnReturn = 0; // Callee pops nothing.
1606 // If this is an sret function, the return should pop the hidden pointer.
1607 if (!Is64Bit && CC != CallingConv::Fast && ArgsAreStructReturn(Op))
1608 BytesToPopOnReturn = 4;
1609 BytesCallerReserves = StackSize;
1613 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1614 if (CC == CallingConv::X86_FastCall)
1615 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1618 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1620 // Return the new list of results.
1621 return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getNode()->getVTList(),
1622 &ArgValues[0], ArgValues.size()).getValue(Op.getResNo());
1626 X86TargetLowering::LowerMemOpCallTo(CallSDNode *TheCall, SelectionDAG &DAG,
1627 const SDValue &StackPtr,
1628 const CCValAssign &VA,
1630 SDValue Arg, ISD::ArgFlagsTy Flags) {
1631 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1632 DebugLoc dl = TheCall->getDebugLoc();
1633 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1634 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1635 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1636 if (Flags.isByVal()) {
1637 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1639 return DAG.getStore(Chain, dl, Arg, PtrOff,
1640 PseudoSourceValue::getStack(), LocMemOffset);
1643 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1644 /// optimization is performed and it is required.
1646 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1647 SDValue &OutRetAddr,
1653 if (!IsTailCall || FPDiff==0) return Chain;
1655 // Adjust the Return address stack slot.
1656 MVT VT = getPointerTy();
1657 OutRetAddr = getReturnAddressFrameIndex(DAG);
1659 // Load the "old" Return address.
1660 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1661 return SDValue(OutRetAddr.getNode(), 1);
1664 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1665 /// optimization is performed and it is required (FPDiff!=0).
1667 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1668 SDValue Chain, SDValue RetAddrFrIdx,
1669 bool Is64Bit, int FPDiff, DebugLoc dl) {
1670 // Store the return address to the appropriate stack slot.
1671 if (!FPDiff) return Chain;
1672 // Calculate the new stack slot for the return address.
1673 int SlotSize = Is64Bit ? 8 : 4;
1674 int NewReturnAddrFI =
1675 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1676 MVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1677 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1678 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1679 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1683 SDValue X86TargetLowering::LowerCALL(SDValue Op, SelectionDAG &DAG) {
1684 MachineFunction &MF = DAG.getMachineFunction();
1685 CallSDNode *TheCall = cast<CallSDNode>(Op.getNode());
1686 SDValue Chain = TheCall->getChain();
1687 unsigned CC = TheCall->getCallingConv();
1688 bool isVarArg = TheCall->isVarArg();
1689 bool IsTailCall = TheCall->isTailCall() &&
1690 CC == CallingConv::Fast && PerformTailCallOpt;
1691 SDValue Callee = TheCall->getCallee();
1692 bool Is64Bit = Subtarget->is64Bit();
1693 bool IsStructRet = CallIsStructReturn(TheCall);
1694 DebugLoc dl = TheCall->getDebugLoc();
1696 assert(!(isVarArg && CC == CallingConv::Fast) &&
1697 "Var args not supported with calling convention fastcc");
1699 // Analyze operands of the call, assigning locations to each operand.
1700 SmallVector<CCValAssign, 16> ArgLocs;
1701 CCState CCInfo(CC, isVarArg, getTargetMachine(), ArgLocs, *DAG.getContext());
1702 CCInfo.AnalyzeCallOperands(TheCall, CCAssignFnForNode(CC));
1704 // Get a count of how many bytes are to be pushed on the stack.
1705 unsigned NumBytes = CCInfo.getNextStackOffset();
1706 if (PerformTailCallOpt && CC == CallingConv::Fast)
1707 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1711 // Lower arguments at fp - stackoffset + fpdiff.
1712 unsigned NumBytesCallerPushed =
1713 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1714 FPDiff = NumBytesCallerPushed - NumBytes;
1716 // Set the delta of movement of the returnaddr stackslot.
1717 // But only set if delta is greater than previous delta.
1718 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1719 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1722 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1724 SDValue RetAddrFrIdx;
1725 // Load return adress for tail calls.
1726 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, IsTailCall, Is64Bit,
1729 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1730 SmallVector<SDValue, 8> MemOpChains;
1733 // Walk the register/memloc assignments, inserting copies/loads. In the case
1734 // of tail call optimization arguments are handle later.
1735 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1736 CCValAssign &VA = ArgLocs[i];
1737 SDValue Arg = TheCall->getArg(i);
1738 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1739 bool isByVal = Flags.isByVal();
1741 // Promote the value if needed.
1742 switch (VA.getLocInfo()) {
1743 default: llvm_unreachable("Unknown loc info!");
1744 case CCValAssign::Full: break;
1745 case CCValAssign::SExt:
1746 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
1748 case CCValAssign::ZExt:
1749 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
1751 case CCValAssign::AExt:
1752 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
1756 if (VA.isRegLoc()) {
1758 MVT RegVT = VA.getLocVT();
1759 if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1760 switch (VA.getLocReg()) {
1763 case X86::RDI: case X86::RSI: case X86::RDX: case X86::RCX:
1765 // Special case: passing MMX values in GPR registers.
1766 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1769 case X86::XMM0: case X86::XMM1: case X86::XMM2: case X86::XMM3:
1770 case X86::XMM4: case X86::XMM5: case X86::XMM6: case X86::XMM7: {
1771 // Special case: passing MMX values in XMM registers.
1772 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1773 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1774 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1779 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1781 if (!IsTailCall || (IsTailCall && isByVal)) {
1782 assert(VA.isMemLoc());
1783 if (StackPtr.getNode() == 0)
1784 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1786 MemOpChains.push_back(LowerMemOpCallTo(TheCall, DAG, StackPtr, VA,
1787 Chain, Arg, Flags));
1792 if (!MemOpChains.empty())
1793 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1794 &MemOpChains[0], MemOpChains.size());
1796 // Build a sequence of copy-to-reg nodes chained together with token chain
1797 // and flag operands which copy the outgoing args into registers.
1799 // Tail call byval lowering might overwrite argument registers so in case of
1800 // tail call optimization the copies to registers are lowered later.
1802 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1803 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1804 RegsToPass[i].second, InFlag);
1805 InFlag = Chain.getValue(1);
1809 if (Subtarget->isPICStyleGOT()) {
1810 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1813 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1814 DAG.getNode(X86ISD::GlobalBaseReg,
1815 DebugLoc::getUnknownLoc(),
1818 InFlag = Chain.getValue(1);
1820 // If we are tail calling and generating PIC/GOT style code load the
1821 // address of the callee into ECX. The value in ecx is used as target of
1822 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1823 // for tail calls on PIC/GOT architectures. Normally we would just put the
1824 // address of GOT into ebx and then call target@PLT. But for tail calls
1825 // ebx would be restored (since ebx is callee saved) before jumping to the
1828 // Note: The actual moving to ECX is done further down.
1829 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1830 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1831 !G->getGlobal()->hasProtectedVisibility())
1832 Callee = LowerGlobalAddress(Callee, DAG);
1833 else if (isa<ExternalSymbolSDNode>(Callee))
1834 Callee = LowerExternalSymbol(Callee, DAG);
1838 if (Is64Bit && isVarArg) {
1839 // From AMD64 ABI document:
1840 // For calls that may call functions that use varargs or stdargs
1841 // (prototype-less calls or calls to functions containing ellipsis (...) in
1842 // the declaration) %al is used as hidden argument to specify the number
1843 // of SSE registers used. The contents of %al do not need to match exactly
1844 // the number of registers, but must be an ubound on the number of SSE
1845 // registers used and is in the range 0 - 8 inclusive.
1847 // FIXME: Verify this on Win64
1848 // Count the number of XMM registers allocated.
1849 static const unsigned XMMArgRegs[] = {
1850 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1851 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1853 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1854 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1855 && "SSE registers cannot be used when SSE is disabled");
1857 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1858 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1859 InFlag = Chain.getValue(1);
1863 // For tail calls lower the arguments to the 'real' stack slot.
1865 SmallVector<SDValue, 8> MemOpChains2;
1868 // Do not flag preceeding copytoreg stuff together with the following stuff.
1870 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1871 CCValAssign &VA = ArgLocs[i];
1872 if (!VA.isRegLoc()) {
1873 assert(VA.isMemLoc());
1874 SDValue Arg = TheCall->getArg(i);
1875 ISD::ArgFlagsTy Flags = TheCall->getArgFlags(i);
1876 // Create frame index.
1877 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1878 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1879 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1880 FIN = DAG.getFrameIndex(FI, getPointerTy());
1882 if (Flags.isByVal()) {
1883 // Copy relative to framepointer.
1884 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1885 if (StackPtr.getNode() == 0)
1886 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1888 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1890 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN, Chain,
1893 // Store relative to framepointer.
1894 MemOpChains2.push_back(
1895 DAG.getStore(Chain, dl, Arg, FIN,
1896 PseudoSourceValue::getFixedStack(FI), 0));
1901 if (!MemOpChains2.empty())
1902 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1903 &MemOpChains2[0], MemOpChains2.size());
1905 // Copy arguments to their registers.
1906 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1907 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1908 RegsToPass[i].second, InFlag);
1909 InFlag = Chain.getValue(1);
1913 // Store the return address to the appropriate stack slot.
1914 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1918 // If the callee is a GlobalAddress node (quite common, every direct call is)
1919 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1920 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1921 // We should use extra load for direct calls to dllimported functions in
1923 GlobalValue *GV = G->getGlobal();
1924 if (!GV->hasDLLImportLinkage()) {
1925 unsigned char OpFlags = 0;
1927 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1928 // external symbols most go through the PLT in PIC mode. If the symbol
1929 // has hidden or protected visibility, or if it is static or local, then
1930 // we don't need to use the PLT - we can directly call it.
1931 if (Subtarget->isTargetELF() &&
1932 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1933 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1934 OpFlags = X86II::MO_PLT;
1935 } else if (Subtarget->isPICStyleStubAny() &&
1936 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1937 Subtarget->getDarwinVers() < 9) {
1938 // PC-relative references to external symbols should go through $stub,
1939 // unless we're building with the leopard linker or later, which
1940 // automatically synthesizes these stubs.
1941 OpFlags = X86II::MO_DARWIN_STUB;
1944 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
1945 G->getOffset(), OpFlags);
1947 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1948 unsigned char OpFlags = 0;
1950 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
1951 // symbols should go through the PLT.
1952 if (Subtarget->isTargetELF() &&
1953 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1954 OpFlags = X86II::MO_PLT;
1955 } else if (Subtarget->isPICStyleStubAny() &&
1956 Subtarget->getDarwinVers() < 9) {
1957 // PC-relative references to external symbols should go through $stub,
1958 // unless we're building with the leopard linker or later, which
1959 // automatically synthesizes these stubs.
1960 OpFlags = X86II::MO_DARWIN_STUB;
1963 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
1965 } else if (IsTailCall) {
1966 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
1968 Chain = DAG.getCopyToReg(Chain, dl,
1969 DAG.getRegister(Opc, getPointerTy()),
1971 Callee = DAG.getRegister(Opc, getPointerTy());
1972 // Add register as live out.
1973 MF.getRegInfo().addLiveOut(Opc);
1976 // Returns a chain & a flag for retval copy to use.
1977 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1978 SmallVector<SDValue, 8> Ops;
1981 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1982 DAG.getIntPtrConstant(0, true), InFlag);
1983 InFlag = Chain.getValue(1);
1985 // Returns a chain & a flag for retval copy to use.
1986 NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1990 Ops.push_back(Chain);
1991 Ops.push_back(Callee);
1994 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1996 // Add argument registers to the end of the list so that they are known live
1998 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1999 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2000 RegsToPass[i].second.getValueType()));
2002 // Add an implicit use GOT pointer in EBX.
2003 if (!IsTailCall && Subtarget->isPICStyleGOT())
2004 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2006 // Add an implicit use of AL for x86 vararg functions.
2007 if (Is64Bit && isVarArg)
2008 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2010 if (InFlag.getNode())
2011 Ops.push_back(InFlag);
2014 assert(InFlag.getNode() &&
2015 "Flag must be set. Depend on flag being set in LowerRET");
2016 Chain = DAG.getNode(X86ISD::TAILCALL, dl,
2017 TheCall->getVTList(), &Ops[0], Ops.size());
2019 return SDValue(Chain.getNode(), Op.getResNo());
2022 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2023 InFlag = Chain.getValue(1);
2025 // Create the CALLSEQ_END node.
2026 unsigned NumBytesForCalleeToPush;
2027 if (IsCalleePop(isVarArg, CC))
2028 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2029 else if (!Is64Bit && CC != CallingConv::Fast && IsStructRet)
2030 // If this is is a call to a struct-return function, the callee
2031 // pops the hidden struct pointer, so we have to push it back.
2032 // This is common for Darwin/X86, Linux & Mingw32 targets.
2033 NumBytesForCalleeToPush = 4;
2035 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2037 // Returns a flag for retval copy to use.
2038 Chain = DAG.getCALLSEQ_END(Chain,
2039 DAG.getIntPtrConstant(NumBytes, true),
2040 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2043 InFlag = Chain.getValue(1);
2045 // Handle result values, copying them out of physregs into vregs that we
2047 return SDValue(LowerCallResult(Chain, InFlag, TheCall, CC, DAG),
2052 //===----------------------------------------------------------------------===//
2053 // Fast Calling Convention (tail call) implementation
2054 //===----------------------------------------------------------------------===//
2056 // Like std call, callee cleans arguments, convention except that ECX is
2057 // reserved for storing the tail called function address. Only 2 registers are
2058 // free for argument passing (inreg). Tail call optimization is performed
2060 // * tailcallopt is enabled
2061 // * caller/callee are fastcc
2062 // On X86_64 architecture with GOT-style position independent code only local
2063 // (within module) calls are supported at the moment.
2064 // To keep the stack aligned according to platform abi the function
2065 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2066 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2067 // If a tail called function callee has more arguments than the caller the
2068 // caller needs to make sure that there is room to move the RETADDR to. This is
2069 // achieved by reserving an area the size of the argument delta right after the
2070 // original REtADDR, but before the saved framepointer or the spilled registers
2071 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2083 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2084 /// for a 16 byte align requirement.
2085 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2086 SelectionDAG& DAG) {
2087 MachineFunction &MF = DAG.getMachineFunction();
2088 const TargetMachine &TM = MF.getTarget();
2089 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2090 unsigned StackAlignment = TFI.getStackAlignment();
2091 uint64_t AlignMask = StackAlignment - 1;
2092 int64_t Offset = StackSize;
2093 uint64_t SlotSize = TD->getPointerSize();
2094 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2095 // Number smaller than 12 so just add the difference.
2096 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2098 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2099 Offset = ((~AlignMask) & Offset) + StackAlignment +
2100 (StackAlignment-SlotSize);
2105 /// IsEligibleForTailCallElimination - Check to see whether the next instruction
2106 /// following the call is a return. A function is eligible if caller/callee
2107 /// calling conventions match, currently only fastcc supports tail calls, and
2108 /// the function CALL is immediatly followed by a RET.
2109 bool X86TargetLowering::IsEligibleForTailCallOptimization(CallSDNode *TheCall,
2111 SelectionDAG& DAG) const {
2112 if (!PerformTailCallOpt)
2115 if (CheckTailCallReturnConstraints(TheCall, Ret)) {
2117 DAG.getMachineFunction().getFunction()->getCallingConv();
2118 unsigned CalleeCC = TheCall->getCallingConv();
2119 if (CalleeCC == CallingConv::Fast && CallerCC == CalleeCC)
2127 X86TargetLowering::createFastISel(MachineFunction &mf,
2128 MachineModuleInfo *mmo,
2130 DenseMap<const Value *, unsigned> &vm,
2131 DenseMap<const BasicBlock *,
2132 MachineBasicBlock *> &bm,
2133 DenseMap<const AllocaInst *, int> &am
2135 , SmallSet<Instruction*, 8> &cil
2138 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2146 //===----------------------------------------------------------------------===//
2147 // Other Lowering Hooks
2148 //===----------------------------------------------------------------------===//
2151 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2152 MachineFunction &MF = DAG.getMachineFunction();
2153 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2154 int ReturnAddrIndex = FuncInfo->getRAIndex();
2156 if (ReturnAddrIndex == 0) {
2157 // Set up a frame object for the return address.
2158 uint64_t SlotSize = TD->getPointerSize();
2159 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2160 FuncInfo->setRAIndex(ReturnAddrIndex);
2163 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2167 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2168 /// specific condition code, returning the condition code and the LHS/RHS of the
2169 /// comparison to make.
2170 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2171 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2173 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2174 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2175 // X > -1 -> X == 0, jump !sign.
2176 RHS = DAG.getConstant(0, RHS.getValueType());
2177 return X86::COND_NS;
2178 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2179 // X < 0 -> X == 0, jump on sign.
2181 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2183 RHS = DAG.getConstant(0, RHS.getValueType());
2184 return X86::COND_LE;
2188 switch (SetCCOpcode) {
2189 default: llvm_unreachable("Invalid integer condition!");
2190 case ISD::SETEQ: return X86::COND_E;
2191 case ISD::SETGT: return X86::COND_G;
2192 case ISD::SETGE: return X86::COND_GE;
2193 case ISD::SETLT: return X86::COND_L;
2194 case ISD::SETLE: return X86::COND_LE;
2195 case ISD::SETNE: return X86::COND_NE;
2196 case ISD::SETULT: return X86::COND_B;
2197 case ISD::SETUGT: return X86::COND_A;
2198 case ISD::SETULE: return X86::COND_BE;
2199 case ISD::SETUGE: return X86::COND_AE;
2203 // First determine if it is required or is profitable to flip the operands.
2205 // If LHS is a foldable load, but RHS is not, flip the condition.
2206 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2207 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2208 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2209 std::swap(LHS, RHS);
2212 switch (SetCCOpcode) {
2218 std::swap(LHS, RHS);
2222 // On a floating point condition, the flags are set as follows:
2224 // 0 | 0 | 0 | X > Y
2225 // 0 | 0 | 1 | X < Y
2226 // 1 | 0 | 0 | X == Y
2227 // 1 | 1 | 1 | unordered
2228 switch (SetCCOpcode) {
2229 default: llvm_unreachable("Condcode should be pre-legalized away");
2231 case ISD::SETEQ: return X86::COND_E;
2232 case ISD::SETOLT: // flipped
2234 case ISD::SETGT: return X86::COND_A;
2235 case ISD::SETOLE: // flipped
2237 case ISD::SETGE: return X86::COND_AE;
2238 case ISD::SETUGT: // flipped
2240 case ISD::SETLT: return X86::COND_B;
2241 case ISD::SETUGE: // flipped
2243 case ISD::SETLE: return X86::COND_BE;
2245 case ISD::SETNE: return X86::COND_NE;
2246 case ISD::SETUO: return X86::COND_P;
2247 case ISD::SETO: return X86::COND_NP;
2251 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2252 /// code. Current x86 isa includes the following FP cmov instructions:
2253 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2254 static bool hasFPCMov(unsigned X86CC) {
2270 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2271 /// the specified range (L, H].
2272 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2273 return (Val < 0) || (Val >= Low && Val < Hi);
2276 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2277 /// specified value.
2278 static bool isUndefOrEqual(int Val, int CmpVal) {
2279 if (Val < 0 || Val == CmpVal)
2284 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2285 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2286 /// the second operand.
2287 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2288 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2289 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2290 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2291 return (Mask[0] < 2 && Mask[1] < 2);
2295 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2296 SmallVector<int, 8> M;
2298 return ::isPSHUFDMask(M, N->getValueType(0));
2301 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2302 /// is suitable for input to PSHUFHW.
2303 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2304 if (VT != MVT::v8i16)
2307 // Lower quadword copied in order or undef.
2308 for (int i = 0; i != 4; ++i)
2309 if (Mask[i] >= 0 && Mask[i] != i)
2312 // Upper quadword shuffled.
2313 for (int i = 4; i != 8; ++i)
2314 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2320 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2321 SmallVector<int, 8> M;
2323 return ::isPSHUFHWMask(M, N->getValueType(0));
2326 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2327 /// is suitable for input to PSHUFLW.
2328 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2329 if (VT != MVT::v8i16)
2332 // Upper quadword copied in order.
2333 for (int i = 4; i != 8; ++i)
2334 if (Mask[i] >= 0 && Mask[i] != i)
2337 // Lower quadword shuffled.
2338 for (int i = 0; i != 4; ++i)
2345 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2346 SmallVector<int, 8> M;
2348 return ::isPSHUFLWMask(M, N->getValueType(0));
2351 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2352 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2353 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2354 int NumElems = VT.getVectorNumElements();
2355 if (NumElems != 2 && NumElems != 4)
2358 int Half = NumElems / 2;
2359 for (int i = 0; i < Half; ++i)
2360 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2362 for (int i = Half; i < NumElems; ++i)
2363 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2369 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2370 SmallVector<int, 8> M;
2372 return ::isSHUFPMask(M, N->getValueType(0));
2375 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2376 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2377 /// half elements to come from vector 1 (which would equal the dest.) and
2378 /// the upper half to come from vector 2.
2379 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2380 int NumElems = VT.getVectorNumElements();
2382 if (NumElems != 2 && NumElems != 4)
2385 int Half = NumElems / 2;
2386 for (int i = 0; i < Half; ++i)
2387 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2389 for (int i = Half; i < NumElems; ++i)
2390 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2395 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2396 SmallVector<int, 8> M;
2398 return isCommutedSHUFPMask(M, N->getValueType(0));
2401 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2402 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2403 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2404 if (N->getValueType(0).getVectorNumElements() != 4)
2407 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2408 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2409 isUndefOrEqual(N->getMaskElt(1), 7) &&
2410 isUndefOrEqual(N->getMaskElt(2), 2) &&
2411 isUndefOrEqual(N->getMaskElt(3), 3);
2414 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2415 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2416 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2417 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2419 if (NumElems != 2 && NumElems != 4)
2422 for (unsigned i = 0; i < NumElems/2; ++i)
2423 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2426 for (unsigned i = NumElems/2; i < NumElems; ++i)
2427 if (!isUndefOrEqual(N->getMaskElt(i), i))
2433 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2434 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2436 bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
2437 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2439 if (NumElems != 2 && NumElems != 4)
2442 for (unsigned i = 0; i < NumElems/2; ++i)
2443 if (!isUndefOrEqual(N->getMaskElt(i), i))
2446 for (unsigned i = 0; i < NumElems/2; ++i)
2447 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2453 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2454 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2456 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2457 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2462 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2463 isUndefOrEqual(N->getMaskElt(1), 3) &&
2464 isUndefOrEqual(N->getMaskElt(2), 2) &&
2465 isUndefOrEqual(N->getMaskElt(3), 3);
2468 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2469 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2470 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2471 bool V2IsSplat = false) {
2472 int NumElts = VT.getVectorNumElements();
2473 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2476 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2478 int BitI1 = Mask[i+1];
2479 if (!isUndefOrEqual(BitI, j))
2482 if (!isUndefOrEqual(BitI1, NumElts))
2485 if (!isUndefOrEqual(BitI1, j + NumElts))
2492 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2493 SmallVector<int, 8> M;
2495 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2498 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2499 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2500 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, MVT VT,
2501 bool V2IsSplat = false) {
2502 int NumElts = VT.getVectorNumElements();
2503 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2506 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2508 int BitI1 = Mask[i+1];
2509 if (!isUndefOrEqual(BitI, j + NumElts/2))
2512 if (isUndefOrEqual(BitI1, NumElts))
2515 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2522 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2523 SmallVector<int, 8> M;
2525 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2528 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2529 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2531 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2532 int NumElems = VT.getVectorNumElements();
2533 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2536 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2538 int BitI1 = Mask[i+1];
2539 if (!isUndefOrEqual(BitI, j))
2541 if (!isUndefOrEqual(BitI1, j))
2547 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2548 SmallVector<int, 8> M;
2550 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2553 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2554 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2556 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, MVT VT) {
2557 int NumElems = VT.getVectorNumElements();
2558 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2561 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2563 int BitI1 = Mask[i+1];
2564 if (!isUndefOrEqual(BitI, j))
2566 if (!isUndefOrEqual(BitI1, j))
2572 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2573 SmallVector<int, 8> M;
2575 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2578 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2579 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2580 /// MOVSD, and MOVD, i.e. setting the lowest element.
2581 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT) {
2582 if (VT.getVectorElementType().getSizeInBits() < 32)
2585 int NumElts = VT.getVectorNumElements();
2587 if (!isUndefOrEqual(Mask[0], NumElts))
2590 for (int i = 1; i < NumElts; ++i)
2591 if (!isUndefOrEqual(Mask[i], i))
2597 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2598 SmallVector<int, 8> M;
2600 return ::isMOVLMask(M, N->getValueType(0));
2603 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2604 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2605 /// element of vector 2 and the other elements to come from vector 1 in order.
2606 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, MVT VT,
2607 bool V2IsSplat = false, bool V2IsUndef = false) {
2608 int NumOps = VT.getVectorNumElements();
2609 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2612 if (!isUndefOrEqual(Mask[0], 0))
2615 for (int i = 1; i < NumOps; ++i)
2616 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2617 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2618 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2624 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2625 bool V2IsUndef = false) {
2626 SmallVector<int, 8> M;
2628 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2631 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2632 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2633 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2634 if (N->getValueType(0).getVectorNumElements() != 4)
2637 // Expect 1, 1, 3, 3
2638 for (unsigned i = 0; i < 2; ++i) {
2639 int Elt = N->getMaskElt(i);
2640 if (Elt >= 0 && Elt != 1)
2645 for (unsigned i = 2; i < 4; ++i) {
2646 int Elt = N->getMaskElt(i);
2647 if (Elt >= 0 && Elt != 3)
2652 // Don't use movshdup if it can be done with a shufps.
2653 // FIXME: verify that matching u, u, 3, 3 is what we want.
2657 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2658 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2659 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2660 if (N->getValueType(0).getVectorNumElements() != 4)
2663 // Expect 0, 0, 2, 2
2664 for (unsigned i = 0; i < 2; ++i)
2665 if (N->getMaskElt(i) > 0)
2669 for (unsigned i = 2; i < 4; ++i) {
2670 int Elt = N->getMaskElt(i);
2671 if (Elt >= 0 && Elt != 2)
2676 // Don't use movsldup if it can be done with a shufps.
2680 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2681 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2682 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2683 int e = N->getValueType(0).getVectorNumElements() / 2;
2685 for (int i = 0; i < e; ++i)
2686 if (!isUndefOrEqual(N->getMaskElt(i), i))
2688 for (int i = 0; i < e; ++i)
2689 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2694 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2695 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2697 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2698 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2699 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2701 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2703 for (int i = 0; i < NumOperands; ++i) {
2704 int Val = SVOp->getMaskElt(NumOperands-i-1);
2705 if (Val < 0) Val = 0;
2706 if (Val >= NumOperands) Val -= NumOperands;
2708 if (i != NumOperands - 1)
2714 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2715 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2717 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2718 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2720 // 8 nodes, but we only care about the last 4.
2721 for (unsigned i = 7; i >= 4; --i) {
2722 int Val = SVOp->getMaskElt(i);
2731 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2732 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2734 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2735 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2737 // 8 nodes, but we only care about the first 4.
2738 for (int i = 3; i >= 0; --i) {
2739 int Val = SVOp->getMaskElt(i);
2748 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2750 bool X86::isZeroNode(SDValue Elt) {
2751 return ((isa<ConstantSDNode>(Elt) &&
2752 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2753 (isa<ConstantFPSDNode>(Elt) &&
2754 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2757 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2758 /// their permute mask.
2759 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2760 SelectionDAG &DAG) {
2761 MVT VT = SVOp->getValueType(0);
2762 unsigned NumElems = VT.getVectorNumElements();
2763 SmallVector<int, 8> MaskVec;
2765 for (unsigned i = 0; i != NumElems; ++i) {
2766 int idx = SVOp->getMaskElt(i);
2768 MaskVec.push_back(idx);
2769 else if (idx < (int)NumElems)
2770 MaskVec.push_back(idx + NumElems);
2772 MaskVec.push_back(idx - NumElems);
2774 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2775 SVOp->getOperand(0), &MaskVec[0]);
2778 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2779 /// the two vector operands have swapped position.
2780 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, MVT VT) {
2781 unsigned NumElems = VT.getVectorNumElements();
2782 for (unsigned i = 0; i != NumElems; ++i) {
2786 else if (idx < (int)NumElems)
2787 Mask[i] = idx + NumElems;
2789 Mask[i] = idx - NumElems;
2793 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2794 /// match movhlps. The lower half elements should come from upper half of
2795 /// V1 (and in order), and the upper half elements should come from the upper
2796 /// half of V2 (and in order).
2797 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2798 if (Op->getValueType(0).getVectorNumElements() != 4)
2800 for (unsigned i = 0, e = 2; i != e; ++i)
2801 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2803 for (unsigned i = 2; i != 4; ++i)
2804 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2809 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2810 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2812 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2813 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2815 N = N->getOperand(0).getNode();
2816 if (!ISD::isNON_EXTLoad(N))
2819 *LD = cast<LoadSDNode>(N);
2823 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2824 /// match movlp{s|d}. The lower half elements should come from lower half of
2825 /// V1 (and in order), and the upper half elements should come from the upper
2826 /// half of V2 (and in order). And since V1 will become the source of the
2827 /// MOVLP, it must be either a vector load or a scalar load to vector.
2828 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2829 ShuffleVectorSDNode *Op) {
2830 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2832 // Is V2 is a vector load, don't do this transformation. We will try to use
2833 // load folding shufps op.
2834 if (ISD::isNON_EXTLoad(V2))
2837 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2839 if (NumElems != 2 && NumElems != 4)
2841 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2842 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2844 for (unsigned i = NumElems/2; i != NumElems; ++i)
2845 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2850 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2852 static bool isSplatVector(SDNode *N) {
2853 if (N->getOpcode() != ISD::BUILD_VECTOR)
2856 SDValue SplatValue = N->getOperand(0);
2857 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2858 if (N->getOperand(i) != SplatValue)
2863 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2864 /// to an zero vector.
2865 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
2866 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
2867 SDValue V1 = N->getOperand(0);
2868 SDValue V2 = N->getOperand(1);
2869 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2870 for (unsigned i = 0; i != NumElems; ++i) {
2871 int Idx = N->getMaskElt(i);
2872 if (Idx >= (int)NumElems) {
2873 unsigned Opc = V2.getOpcode();
2874 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2876 if (Opc != ISD::BUILD_VECTOR ||
2877 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
2879 } else if (Idx >= 0) {
2880 unsigned Opc = V1.getOpcode();
2881 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2883 if (Opc != ISD::BUILD_VECTOR ||
2884 !X86::isZeroNode(V1.getOperand(Idx)))
2891 /// getZeroVector - Returns a vector of specified type with all zero elements.
2893 static SDValue getZeroVector(MVT VT, bool HasSSE2, SelectionDAG &DAG,
2895 assert(VT.isVector() && "Expected a vector type");
2897 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2898 // type. This ensures they get CSE'd.
2900 if (VT.getSizeInBits() == 64) { // MMX
2901 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2902 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2903 } else if (HasSSE2) { // SSE2
2904 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2905 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2907 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2908 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2910 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2913 /// getOnesVector - Returns a vector of specified type with all bits set.
2915 static SDValue getOnesVector(MVT VT, SelectionDAG &DAG, DebugLoc dl) {
2916 assert(VT.isVector() && "Expected a vector type");
2918 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2919 // type. This ensures they get CSE'd.
2920 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2922 if (VT.getSizeInBits() == 64) // MMX
2923 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2925 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2926 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2930 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2931 /// that point to V2 points to its first element.
2932 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
2933 MVT VT = SVOp->getValueType(0);
2934 unsigned NumElems = VT.getVectorNumElements();
2936 bool Changed = false;
2937 SmallVector<int, 8> MaskVec;
2938 SVOp->getMask(MaskVec);
2940 for (unsigned i = 0; i != NumElems; ++i) {
2941 if (MaskVec[i] > (int)NumElems) {
2942 MaskVec[i] = NumElems;
2947 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
2948 SVOp->getOperand(1), &MaskVec[0]);
2949 return SDValue(SVOp, 0);
2952 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2953 /// operation of specified width.
2954 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2956 unsigned NumElems = VT.getVectorNumElements();
2957 SmallVector<int, 8> Mask;
2958 Mask.push_back(NumElems);
2959 for (unsigned i = 1; i != NumElems; ++i)
2961 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2964 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
2965 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2967 unsigned NumElems = VT.getVectorNumElements();
2968 SmallVector<int, 8> Mask;
2969 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2971 Mask.push_back(i + NumElems);
2973 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2976 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
2977 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, MVT VT, SDValue V1,
2979 unsigned NumElems = VT.getVectorNumElements();
2980 unsigned Half = NumElems/2;
2981 SmallVector<int, 8> Mask;
2982 for (unsigned i = 0; i != Half; ++i) {
2983 Mask.push_back(i + Half);
2984 Mask.push_back(i + NumElems + Half);
2986 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2989 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
2990 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
2992 if (SV->getValueType(0).getVectorNumElements() <= 4)
2993 return SDValue(SV, 0);
2995 MVT PVT = MVT::v4f32;
2996 MVT VT = SV->getValueType(0);
2997 DebugLoc dl = SV->getDebugLoc();
2998 SDValue V1 = SV->getOperand(0);
2999 int NumElems = VT.getVectorNumElements();
3000 int EltNo = SV->getSplatIndex();
3002 // unpack elements to the correct location
3003 while (NumElems > 4) {
3004 if (EltNo < NumElems/2) {
3005 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3007 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3008 EltNo -= NumElems/2;
3013 // Perform the splat.
3014 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3015 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3016 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3017 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3020 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3021 /// vector of zero or undef vector. This produces a shuffle where the low
3022 /// element of V2 is swizzled into the zero/undef vector, landing at element
3023 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3024 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3025 bool isZero, bool HasSSE2,
3026 SelectionDAG &DAG) {
3027 MVT VT = V2.getValueType();
3029 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3030 unsigned NumElems = VT.getVectorNumElements();
3031 SmallVector<int, 16> MaskVec;
3032 for (unsigned i = 0; i != NumElems; ++i)
3033 // If this is the insertion idx, put the low elt of V2 here.
3034 MaskVec.push_back(i == Idx ? NumElems : i);
3035 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3038 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3039 /// a shuffle that is zero.
3041 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3042 bool Low, SelectionDAG &DAG) {
3043 unsigned NumZeros = 0;
3044 for (int i = 0; i < NumElems; ++i) {
3045 unsigned Index = Low ? i : NumElems-i-1;
3046 int Idx = SVOp->getMaskElt(Index);
3051 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3052 if (Elt.getNode() && X86::isZeroNode(Elt))
3060 /// isVectorShift - Returns true if the shuffle can be implemented as a
3061 /// logical left or right shift of a vector.
3062 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3063 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3064 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3065 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3068 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3071 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3075 bool SeenV1 = false;
3076 bool SeenV2 = false;
3077 for (int i = NumZeros; i < NumElems; ++i) {
3078 int Val = isLeft ? (i - NumZeros) : i;
3079 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3091 if (SeenV1 && SeenV2)
3094 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3100 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3102 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3103 unsigned NumNonZero, unsigned NumZero,
3104 SelectionDAG &DAG, TargetLowering &TLI) {
3108 DebugLoc dl = Op.getDebugLoc();
3111 for (unsigned i = 0; i < 16; ++i) {
3112 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3113 if (ThisIsNonZero && First) {
3115 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3117 V = DAG.getUNDEF(MVT::v8i16);
3122 SDValue ThisElt(0, 0), LastElt(0, 0);
3123 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3124 if (LastIsNonZero) {
3125 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3126 MVT::i16, Op.getOperand(i-1));
3128 if (ThisIsNonZero) {
3129 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3130 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3131 ThisElt, DAG.getConstant(8, MVT::i8));
3133 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3137 if (ThisElt.getNode())
3138 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3139 DAG.getIntPtrConstant(i/2));
3143 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3146 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3148 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3149 unsigned NumNonZero, unsigned NumZero,
3150 SelectionDAG &DAG, TargetLowering &TLI) {
3154 DebugLoc dl = Op.getDebugLoc();
3157 for (unsigned i = 0; i < 8; ++i) {
3158 bool isNonZero = (NonZeros & (1 << i)) != 0;
3162 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3164 V = DAG.getUNDEF(MVT::v8i16);
3167 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3168 MVT::v8i16, V, Op.getOperand(i),
3169 DAG.getIntPtrConstant(i));
3176 /// getVShift - Return a vector logical shift node.
3178 static SDValue getVShift(bool isLeft, MVT VT, SDValue SrcOp,
3179 unsigned NumBits, SelectionDAG &DAG,
3180 const TargetLowering &TLI, DebugLoc dl) {
3181 bool isMMX = VT.getSizeInBits() == 64;
3182 MVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3183 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3184 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3185 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3186 DAG.getNode(Opc, dl, ShVT, SrcOp,
3187 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3191 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3192 DebugLoc dl = Op.getDebugLoc();
3193 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3194 if (ISD::isBuildVectorAllZeros(Op.getNode())
3195 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3196 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3197 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3198 // eliminated on x86-32 hosts.
3199 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3202 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3203 return getOnesVector(Op.getValueType(), DAG, dl);
3204 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3207 MVT VT = Op.getValueType();
3208 MVT EVT = VT.getVectorElementType();
3209 unsigned EVTBits = EVT.getSizeInBits();
3211 unsigned NumElems = Op.getNumOperands();
3212 unsigned NumZero = 0;
3213 unsigned NumNonZero = 0;
3214 unsigned NonZeros = 0;
3215 bool IsAllConstants = true;
3216 SmallSet<SDValue, 8> Values;
3217 for (unsigned i = 0; i < NumElems; ++i) {
3218 SDValue Elt = Op.getOperand(i);
3219 if (Elt.getOpcode() == ISD::UNDEF)
3222 if (Elt.getOpcode() != ISD::Constant &&
3223 Elt.getOpcode() != ISD::ConstantFP)
3224 IsAllConstants = false;
3225 if (X86::isZeroNode(Elt))
3228 NonZeros |= (1 << i);
3233 if (NumNonZero == 0) {
3234 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3235 return DAG.getUNDEF(VT);
3238 // Special case for single non-zero, non-undef, element.
3239 if (NumNonZero == 1) {
3240 unsigned Idx = CountTrailingZeros_32(NonZeros);
3241 SDValue Item = Op.getOperand(Idx);
3243 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3244 // the value are obviously zero, truncate the value to i32 and do the
3245 // insertion that way. Only do this if the value is non-constant or if the
3246 // value is a constant being inserted into element 0. It is cheaper to do
3247 // a constant pool load than it is to do a movd + shuffle.
3248 if (EVT == MVT::i64 && !Subtarget->is64Bit() &&
3249 (!IsAllConstants || Idx == 0)) {
3250 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3251 // Handle MMX and SSE both.
3252 MVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3253 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3255 // Truncate the value (which may itself be a constant) to i32, and
3256 // convert it to a vector with movd (S2V+shuffle to zero extend).
3257 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3258 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3259 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3260 Subtarget->hasSSE2(), DAG);
3262 // Now we have our 32-bit value zero extended in the low element of
3263 // a vector. If Idx != 0, swizzle it into place.
3265 SmallVector<int, 4> Mask;
3266 Mask.push_back(Idx);
3267 for (unsigned i = 1; i != VecElts; ++i)
3269 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3270 DAG.getUNDEF(Item.getValueType()),
3273 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3277 // If we have a constant or non-constant insertion into the low element of
3278 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3279 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3280 // depending on what the source datatype is.
3283 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3284 } else if (EVT == MVT::i32 || EVT == MVT::f32 || EVT == MVT::f64 ||
3285 (EVT == MVT::i64 && Subtarget->is64Bit())) {
3286 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3287 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3288 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3290 } else if (EVT == MVT::i16 || EVT == MVT::i8) {
3291 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3292 MVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3293 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3294 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3295 Subtarget->hasSSE2(), DAG);
3296 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3300 // Is it a vector logical left shift?
3301 if (NumElems == 2 && Idx == 1 &&
3302 X86::isZeroNode(Op.getOperand(0)) &&
3303 !X86::isZeroNode(Op.getOperand(1))) {
3304 unsigned NumBits = VT.getSizeInBits();
3305 return getVShift(true, VT,
3306 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3307 VT, Op.getOperand(1)),
3308 NumBits/2, DAG, *this, dl);
3311 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3314 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3315 // is a non-constant being inserted into an element other than the low one,
3316 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3317 // movd/movss) to move this into the low element, then shuffle it into
3319 if (EVTBits == 32) {
3320 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3322 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3323 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3324 Subtarget->hasSSE2(), DAG);
3325 SmallVector<int, 8> MaskVec;
3326 for (unsigned i = 0; i < NumElems; i++)
3327 MaskVec.push_back(i == Idx ? 0 : 1);
3328 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3332 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3333 if (Values.size() == 1)
3336 // A vector full of immediates; various special cases are already
3337 // handled, so this is best done with a single constant-pool load.
3341 // Let legalizer expand 2-wide build_vectors.
3342 if (EVTBits == 64) {
3343 if (NumNonZero == 1) {
3344 // One half is zero or undef.
3345 unsigned Idx = CountTrailingZeros_32(NonZeros);
3346 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3347 Op.getOperand(Idx));
3348 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3349 Subtarget->hasSSE2(), DAG);
3354 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3355 if (EVTBits == 8 && NumElems == 16) {
3356 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3358 if (V.getNode()) return V;
3361 if (EVTBits == 16 && NumElems == 8) {
3362 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3364 if (V.getNode()) return V;
3367 // If element VT is == 32 bits, turn it into a number of shuffles.
3368 SmallVector<SDValue, 8> V;
3370 if (NumElems == 4 && NumZero > 0) {
3371 for (unsigned i = 0; i < 4; ++i) {
3372 bool isZero = !(NonZeros & (1 << i));
3374 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3376 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3379 for (unsigned i = 0; i < 2; ++i) {
3380 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3383 V[i] = V[i*2]; // Must be a zero vector.
3386 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3389 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3392 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3397 SmallVector<int, 8> MaskVec;
3398 bool Reverse = (NonZeros & 0x3) == 2;
3399 for (unsigned i = 0; i < 2; ++i)
3400 MaskVec.push_back(Reverse ? 1-i : i);
3401 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3402 for (unsigned i = 0; i < 2; ++i)
3403 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3404 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3407 if (Values.size() > 2) {
3408 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3409 // values to be inserted is equal to the number of elements, in which case
3410 // use the unpack code below in the hopes of matching the consecutive elts
3411 // load merge pattern for shuffles.
3412 // FIXME: We could probably just check that here directly.
3413 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3414 getSubtarget()->hasSSE41()) {
3415 V[0] = DAG.getUNDEF(VT);
3416 for (unsigned i = 0; i < NumElems; ++i)
3417 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3418 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3419 Op.getOperand(i), DAG.getIntPtrConstant(i));
3422 // Expand into a number of unpckl*.
3424 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3425 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3426 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3427 for (unsigned i = 0; i < NumElems; ++i)
3428 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3430 while (NumElems != 0) {
3431 for (unsigned i = 0; i < NumElems; ++i)
3432 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3441 // v8i16 shuffles - Prefer shuffles in the following order:
3442 // 1. [all] pshuflw, pshufhw, optional move
3443 // 2. [ssse3] 1 x pshufb
3444 // 3. [ssse3] 2 x pshufb + 1 x por
3445 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3447 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3448 SelectionDAG &DAG, X86TargetLowering &TLI) {
3449 SDValue V1 = SVOp->getOperand(0);
3450 SDValue V2 = SVOp->getOperand(1);
3451 DebugLoc dl = SVOp->getDebugLoc();
3452 SmallVector<int, 8> MaskVals;
3454 // Determine if more than 1 of the words in each of the low and high quadwords
3455 // of the result come from the same quadword of one of the two inputs. Undef
3456 // mask values count as coming from any quadword, for better codegen.
3457 SmallVector<unsigned, 4> LoQuad(4);
3458 SmallVector<unsigned, 4> HiQuad(4);
3459 BitVector InputQuads(4);
3460 for (unsigned i = 0; i < 8; ++i) {
3461 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3462 int EltIdx = SVOp->getMaskElt(i);
3463 MaskVals.push_back(EltIdx);
3472 InputQuads.set(EltIdx / 4);
3475 int BestLoQuad = -1;
3476 unsigned MaxQuad = 1;
3477 for (unsigned i = 0; i < 4; ++i) {
3478 if (LoQuad[i] > MaxQuad) {
3480 MaxQuad = LoQuad[i];
3484 int BestHiQuad = -1;
3486 for (unsigned i = 0; i < 4; ++i) {
3487 if (HiQuad[i] > MaxQuad) {
3489 MaxQuad = HiQuad[i];
3493 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3494 // of the two input vectors, shuffle them into one input vector so only a
3495 // single pshufb instruction is necessary. If There are more than 2 input
3496 // quads, disable the next transformation since it does not help SSSE3.
3497 bool V1Used = InputQuads[0] || InputQuads[1];
3498 bool V2Used = InputQuads[2] || InputQuads[3];
3499 if (TLI.getSubtarget()->hasSSSE3()) {
3500 if (InputQuads.count() == 2 && V1Used && V2Used) {
3501 BestLoQuad = InputQuads.find_first();
3502 BestHiQuad = InputQuads.find_next(BestLoQuad);
3504 if (InputQuads.count() > 2) {
3510 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3511 // the shuffle mask. If a quad is scored as -1, that means that it contains
3512 // words from all 4 input quadwords.
3514 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3515 SmallVector<int, 8> MaskV;
3516 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3517 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3518 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3519 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3520 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3521 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3523 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3524 // source words for the shuffle, to aid later transformations.
3525 bool AllWordsInNewV = true;
3526 bool InOrder[2] = { true, true };
3527 for (unsigned i = 0; i != 8; ++i) {
3528 int idx = MaskVals[i];
3530 InOrder[i/4] = false;
3531 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3533 AllWordsInNewV = false;
3537 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3538 if (AllWordsInNewV) {
3539 for (int i = 0; i != 8; ++i) {
3540 int idx = MaskVals[i];
3543 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3544 if ((idx != i) && idx < 4)
3546 if ((idx != i) && idx > 3)
3555 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3556 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3557 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3558 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3559 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3563 // If we have SSSE3, and all words of the result are from 1 input vector,
3564 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3565 // is present, fall back to case 4.
3566 if (TLI.getSubtarget()->hasSSSE3()) {
3567 SmallVector<SDValue,16> pshufbMask;
3569 // If we have elements from both input vectors, set the high bit of the
3570 // shuffle mask element to zero out elements that come from V2 in the V1
3571 // mask, and elements that come from V1 in the V2 mask, so that the two
3572 // results can be OR'd together.
3573 bool TwoInputs = V1Used && V2Used;
3574 for (unsigned i = 0; i != 8; ++i) {
3575 int EltIdx = MaskVals[i] * 2;
3576 if (TwoInputs && (EltIdx >= 16)) {
3577 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3578 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3581 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3582 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3584 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3585 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3586 DAG.getNode(ISD::BUILD_VECTOR, dl,
3587 MVT::v16i8, &pshufbMask[0], 16));
3589 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3591 // Calculate the shuffle mask for the second input, shuffle it, and
3592 // OR it with the first shuffled input.
3594 for (unsigned i = 0; i != 8; ++i) {
3595 int EltIdx = MaskVals[i] * 2;
3597 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3598 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3601 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3602 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3604 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3605 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3606 DAG.getNode(ISD::BUILD_VECTOR, dl,
3607 MVT::v16i8, &pshufbMask[0], 16));
3608 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3609 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3612 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3613 // and update MaskVals with new element order.
3614 BitVector InOrder(8);
3615 if (BestLoQuad >= 0) {
3616 SmallVector<int, 8> MaskV;
3617 for (int i = 0; i != 4; ++i) {
3618 int idx = MaskVals[i];
3620 MaskV.push_back(-1);
3622 } else if ((idx / 4) == BestLoQuad) {
3623 MaskV.push_back(idx & 3);
3626 MaskV.push_back(-1);
3629 for (unsigned i = 4; i != 8; ++i)
3631 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3635 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3636 // and update MaskVals with the new element order.
3637 if (BestHiQuad >= 0) {
3638 SmallVector<int, 8> MaskV;
3639 for (unsigned i = 0; i != 4; ++i)
3641 for (unsigned i = 4; i != 8; ++i) {
3642 int idx = MaskVals[i];
3644 MaskV.push_back(-1);
3646 } else if ((idx / 4) == BestHiQuad) {
3647 MaskV.push_back((idx & 3) + 4);
3650 MaskV.push_back(-1);
3653 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3657 // In case BestHi & BestLo were both -1, which means each quadword has a word
3658 // from each of the four input quadwords, calculate the InOrder bitvector now
3659 // before falling through to the insert/extract cleanup.
3660 if (BestLoQuad == -1 && BestHiQuad == -1) {
3662 for (int i = 0; i != 8; ++i)
3663 if (MaskVals[i] < 0 || MaskVals[i] == i)
3667 // The other elements are put in the right place using pextrw and pinsrw.
3668 for (unsigned i = 0; i != 8; ++i) {
3671 int EltIdx = MaskVals[i];
3674 SDValue ExtOp = (EltIdx < 8)
3675 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3676 DAG.getIntPtrConstant(EltIdx))
3677 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3678 DAG.getIntPtrConstant(EltIdx - 8));
3679 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3680 DAG.getIntPtrConstant(i));
3685 // v16i8 shuffles - Prefer shuffles in the following order:
3686 // 1. [ssse3] 1 x pshufb
3687 // 2. [ssse3] 2 x pshufb + 1 x por
3688 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3690 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3691 SelectionDAG &DAG, X86TargetLowering &TLI) {
3692 SDValue V1 = SVOp->getOperand(0);
3693 SDValue V2 = SVOp->getOperand(1);
3694 DebugLoc dl = SVOp->getDebugLoc();
3695 SmallVector<int, 16> MaskVals;
3696 SVOp->getMask(MaskVals);
3698 // If we have SSSE3, case 1 is generated when all result bytes come from
3699 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3700 // present, fall back to case 3.
3701 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3704 for (unsigned i = 0; i < 16; ++i) {
3705 int EltIdx = MaskVals[i];
3714 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3715 if (TLI.getSubtarget()->hasSSSE3()) {
3716 SmallVector<SDValue,16> pshufbMask;
3718 // If all result elements are from one input vector, then only translate
3719 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3721 // Otherwise, we have elements from both input vectors, and must zero out
3722 // elements that come from V2 in the first mask, and V1 in the second mask
3723 // so that we can OR them together.
3724 bool TwoInputs = !(V1Only || V2Only);
3725 for (unsigned i = 0; i != 16; ++i) {
3726 int EltIdx = MaskVals[i];
3727 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3728 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3731 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3733 // If all the elements are from V2, assign it to V1 and return after
3734 // building the first pshufb.
3737 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3738 DAG.getNode(ISD::BUILD_VECTOR, dl,
3739 MVT::v16i8, &pshufbMask[0], 16));
3743 // Calculate the shuffle mask for the second input, shuffle it, and
3744 // OR it with the first shuffled input.
3746 for (unsigned i = 0; i != 16; ++i) {
3747 int EltIdx = MaskVals[i];
3749 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3752 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3754 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3755 DAG.getNode(ISD::BUILD_VECTOR, dl,
3756 MVT::v16i8, &pshufbMask[0], 16));
3757 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3760 // No SSSE3 - Calculate in place words and then fix all out of place words
3761 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3762 // the 16 different words that comprise the two doublequadword input vectors.
3763 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3764 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3765 SDValue NewV = V2Only ? V2 : V1;
3766 for (int i = 0; i != 8; ++i) {
3767 int Elt0 = MaskVals[i*2];
3768 int Elt1 = MaskVals[i*2+1];
3770 // This word of the result is all undef, skip it.
3771 if (Elt0 < 0 && Elt1 < 0)
3774 // This word of the result is already in the correct place, skip it.
3775 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3777 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3780 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3781 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3784 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3785 // using a single extract together, load it and store it.
3786 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3787 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3788 DAG.getIntPtrConstant(Elt1 / 2));
3789 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3790 DAG.getIntPtrConstant(i));
3794 // If Elt1 is defined, extract it from the appropriate source. If the
3795 // source byte is not also odd, shift the extracted word left 8 bits
3796 // otherwise clear the bottom 8 bits if we need to do an or.
3798 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3799 DAG.getIntPtrConstant(Elt1 / 2));
3800 if ((Elt1 & 1) == 0)
3801 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3802 DAG.getConstant(8, TLI.getShiftAmountTy()));
3804 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3805 DAG.getConstant(0xFF00, MVT::i16));
3807 // If Elt0 is defined, extract it from the appropriate source. If the
3808 // source byte is not also even, shift the extracted word right 8 bits. If
3809 // Elt1 was also defined, OR the extracted values together before
3810 // inserting them in the result.
3812 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3813 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3814 if ((Elt0 & 1) != 0)
3815 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3816 DAG.getConstant(8, TLI.getShiftAmountTy()));
3818 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3819 DAG.getConstant(0x00FF, MVT::i16));
3820 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3823 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3824 DAG.getIntPtrConstant(i));
3826 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3829 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3830 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3831 /// done when every pair / quad of shuffle mask elements point to elements in
3832 /// the right sequence. e.g.
3833 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3835 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3837 TargetLowering &TLI, DebugLoc dl) {
3838 MVT VT = SVOp->getValueType(0);
3839 SDValue V1 = SVOp->getOperand(0);
3840 SDValue V2 = SVOp->getOperand(1);
3841 unsigned NumElems = VT.getVectorNumElements();
3842 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3843 MVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3844 MVT MaskEltVT = MaskVT.getVectorElementType();
3846 switch (VT.getSimpleVT()) {
3847 default: assert(false && "Unexpected!");
3848 case MVT::v4f32: NewVT = MVT::v2f64; break;
3849 case MVT::v4i32: NewVT = MVT::v2i64; break;
3850 case MVT::v8i16: NewVT = MVT::v4i32; break;
3851 case MVT::v16i8: NewVT = MVT::v4i32; break;
3854 if (NewWidth == 2) {
3860 int Scale = NumElems / NewWidth;
3861 SmallVector<int, 8> MaskVec;
3862 for (unsigned i = 0; i < NumElems; i += Scale) {
3864 for (int j = 0; j < Scale; ++j) {
3865 int EltIdx = SVOp->getMaskElt(i+j);
3869 StartIdx = EltIdx - (EltIdx % Scale);
3870 if (EltIdx != StartIdx + j)
3874 MaskVec.push_back(-1);
3876 MaskVec.push_back(StartIdx / Scale);
3879 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3880 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3881 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
3884 /// getVZextMovL - Return a zero-extending vector move low node.
3886 static SDValue getVZextMovL(MVT VT, MVT OpVT,
3887 SDValue SrcOp, SelectionDAG &DAG,
3888 const X86Subtarget *Subtarget, DebugLoc dl) {
3889 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3890 LoadSDNode *LD = NULL;
3891 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3892 LD = dyn_cast<LoadSDNode>(SrcOp);
3894 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3896 MVT EVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3897 if ((EVT != MVT::i64 || Subtarget->is64Bit()) &&
3898 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3899 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3900 SrcOp.getOperand(0).getOperand(0).getValueType() == EVT) {
3902 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3903 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3904 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3905 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3913 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3914 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3915 DAG.getNode(ISD::BIT_CONVERT, dl,
3919 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3922 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3923 SDValue V1 = SVOp->getOperand(0);
3924 SDValue V2 = SVOp->getOperand(1);
3925 DebugLoc dl = SVOp->getDebugLoc();
3926 MVT VT = SVOp->getValueType(0);
3928 SmallVector<std::pair<int, int>, 8> Locs;
3930 SmallVector<int, 8> Mask1(4U, -1);
3931 SmallVector<int, 8> PermMask;
3932 SVOp->getMask(PermMask);
3936 for (unsigned i = 0; i != 4; ++i) {
3937 int Idx = PermMask[i];
3939 Locs[i] = std::make_pair(-1, -1);
3941 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
3943 Locs[i] = std::make_pair(0, NumLo);
3947 Locs[i] = std::make_pair(1, NumHi);
3949 Mask1[2+NumHi] = Idx;
3955 if (NumLo <= 2 && NumHi <= 2) {
3956 // If no more than two elements come from either vector. This can be
3957 // implemented with two shuffles. First shuffle gather the elements.
3958 // The second shuffle, which takes the first shuffle as both of its
3959 // vector operands, put the elements into the right order.
3960 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3962 SmallVector<int, 8> Mask2(4U, -1);
3964 for (unsigned i = 0; i != 4; ++i) {
3965 if (Locs[i].first == -1)
3968 unsigned Idx = (i < 2) ? 0 : 4;
3969 Idx += Locs[i].first * 2 + Locs[i].second;
3974 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
3975 } else if (NumLo == 3 || NumHi == 3) {
3976 // Otherwise, we must have three elements from one vector, call it X, and
3977 // one element from the other, call it Y. First, use a shufps to build an
3978 // intermediate vector with the one element from Y and the element from X
3979 // that will be in the same half in the final destination (the indexes don't
3980 // matter). Then, use a shufps to build the final vector, taking the half
3981 // containing the element from Y from the intermediate, and the other half
3984 // Normalize it so the 3 elements come from V1.
3985 CommuteVectorShuffleMask(PermMask, VT);
3989 // Find the element from V2.
3991 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
3992 int Val = PermMask[HiIndex];
3999 Mask1[0] = PermMask[HiIndex];
4001 Mask1[2] = PermMask[HiIndex^1];
4003 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4006 Mask1[0] = PermMask[0];
4007 Mask1[1] = PermMask[1];
4008 Mask1[2] = HiIndex & 1 ? 6 : 4;
4009 Mask1[3] = HiIndex & 1 ? 4 : 6;
4010 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4012 Mask1[0] = HiIndex & 1 ? 2 : 0;
4013 Mask1[1] = HiIndex & 1 ? 0 : 2;
4014 Mask1[2] = PermMask[2];
4015 Mask1[3] = PermMask[3];
4020 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4024 // Break it into (shuffle shuffle_hi, shuffle_lo).
4026 SmallVector<int,8> LoMask(4U, -1);
4027 SmallVector<int,8> HiMask(4U, -1);
4029 SmallVector<int,8> *MaskPtr = &LoMask;
4030 unsigned MaskIdx = 0;
4033 for (unsigned i = 0; i != 4; ++i) {
4040 int Idx = PermMask[i];
4042 Locs[i] = std::make_pair(-1, -1);
4043 } else if (Idx < 4) {
4044 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4045 (*MaskPtr)[LoIdx] = Idx;
4048 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4049 (*MaskPtr)[HiIdx] = Idx;
4054 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4055 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4056 SmallVector<int, 8> MaskOps;
4057 for (unsigned i = 0; i != 4; ++i) {
4058 if (Locs[i].first == -1) {
4059 MaskOps.push_back(-1);
4061 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4062 MaskOps.push_back(Idx);
4065 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4069 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4070 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4071 SDValue V1 = Op.getOperand(0);
4072 SDValue V2 = Op.getOperand(1);
4073 MVT VT = Op.getValueType();
4074 DebugLoc dl = Op.getDebugLoc();
4075 unsigned NumElems = VT.getVectorNumElements();
4076 bool isMMX = VT.getSizeInBits() == 64;
4077 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4078 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4079 bool V1IsSplat = false;
4080 bool V2IsSplat = false;
4082 if (isZeroShuffle(SVOp))
4083 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4085 // Promote splats to v4f32.
4086 if (SVOp->isSplat()) {
4087 if (isMMX || NumElems < 4)
4089 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4092 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4094 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4095 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4096 if (NewOp.getNode())
4097 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4098 LowerVECTOR_SHUFFLE(NewOp, DAG));
4099 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4100 // FIXME: Figure out a cleaner way to do this.
4101 // Try to make use of movq to zero out the top part.
4102 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4103 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4104 if (NewOp.getNode()) {
4105 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4106 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4107 DAG, Subtarget, dl);
4109 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4110 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4111 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4112 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4113 DAG, Subtarget, dl);
4117 if (X86::isPSHUFDMask(SVOp))
4120 // Check if this can be converted into a logical shift.
4121 bool isLeft = false;
4124 bool isShift = getSubtarget()->hasSSE2() &&
4125 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4126 if (isShift && ShVal.hasOneUse()) {
4127 // If the shifted value has multiple uses, it may be cheaper to use
4128 // v_set0 + movlhps or movhlps, etc.
4129 MVT EVT = VT.getVectorElementType();
4130 ShAmt *= EVT.getSizeInBits();
4131 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4134 if (X86::isMOVLMask(SVOp)) {
4137 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4138 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4143 // FIXME: fold these into legal mask.
4144 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4145 X86::isMOVSLDUPMask(SVOp) ||
4146 X86::isMOVHLPSMask(SVOp) ||
4147 X86::isMOVHPMask(SVOp) ||
4148 X86::isMOVLPMask(SVOp)))
4151 if (ShouldXformToMOVHLPS(SVOp) ||
4152 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4153 return CommuteVectorShuffle(SVOp, DAG);
4156 // No better options. Use a vshl / vsrl.
4157 MVT EVT = VT.getVectorElementType();
4158 ShAmt *= EVT.getSizeInBits();
4159 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4162 bool Commuted = false;
4163 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4164 // 1,1,1,1 -> v8i16 though.
4165 V1IsSplat = isSplatVector(V1.getNode());
4166 V2IsSplat = isSplatVector(V2.getNode());
4168 // Canonicalize the splat or undef, if present, to be on the RHS.
4169 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4170 Op = CommuteVectorShuffle(SVOp, DAG);
4171 SVOp = cast<ShuffleVectorSDNode>(Op);
4172 V1 = SVOp->getOperand(0);
4173 V2 = SVOp->getOperand(1);
4174 std::swap(V1IsSplat, V2IsSplat);
4175 std::swap(V1IsUndef, V2IsUndef);
4179 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4180 // Shuffling low element of v1 into undef, just return v1.
4183 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4184 // the instruction selector will not match, so get a canonical MOVL with
4185 // swapped operands to undo the commute.
4186 return getMOVL(DAG, dl, VT, V2, V1);
4189 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4190 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4191 X86::isUNPCKLMask(SVOp) ||
4192 X86::isUNPCKHMask(SVOp))
4196 // Normalize mask so all entries that point to V2 points to its first
4197 // element then try to match unpck{h|l} again. If match, return a
4198 // new vector_shuffle with the corrected mask.
4199 SDValue NewMask = NormalizeMask(SVOp, DAG);
4200 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4201 if (NSVOp != SVOp) {
4202 if (X86::isUNPCKLMask(NSVOp, true)) {
4204 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4211 // Commute is back and try unpck* again.
4212 // FIXME: this seems wrong.
4213 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4214 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4215 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4216 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4217 X86::isUNPCKLMask(NewSVOp) ||
4218 X86::isUNPCKHMask(NewSVOp))
4222 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4224 // Normalize the node to match x86 shuffle ops if needed
4225 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4226 return CommuteVectorShuffle(SVOp, DAG);
4228 // Check for legal shuffle and return?
4229 SmallVector<int, 16> PermMask;
4230 SVOp->getMask(PermMask);
4231 if (isShuffleMaskLegal(PermMask, VT))
4234 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4235 if (VT == MVT::v8i16) {
4236 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4237 if (NewOp.getNode())
4241 if (VT == MVT::v16i8) {
4242 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4243 if (NewOp.getNode())
4247 // Handle all 4 wide cases with a number of shuffles except for MMX.
4248 if (NumElems == 4 && !isMMX)
4249 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4255 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4256 SelectionDAG &DAG) {
4257 MVT VT = Op.getValueType();
4258 DebugLoc dl = Op.getDebugLoc();
4259 if (VT.getSizeInBits() == 8) {
4260 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4261 Op.getOperand(0), Op.getOperand(1));
4262 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4263 DAG.getValueType(VT));
4264 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4265 } else if (VT.getSizeInBits() == 16) {
4266 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4267 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4269 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4270 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4271 DAG.getNode(ISD::BIT_CONVERT, dl,
4275 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4276 Op.getOperand(0), Op.getOperand(1));
4277 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4278 DAG.getValueType(VT));
4279 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4280 } else if (VT == MVT::f32) {
4281 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4282 // the result back to FR32 register. It's only worth matching if the
4283 // result has a single use which is a store or a bitcast to i32. And in
4284 // the case of a store, it's not worth it if the index is a constant 0,
4285 // because a MOVSSmr can be used instead, which is smaller and faster.
4286 if (!Op.hasOneUse())
4288 SDNode *User = *Op.getNode()->use_begin();
4289 if ((User->getOpcode() != ISD::STORE ||
4290 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4291 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4292 (User->getOpcode() != ISD::BIT_CONVERT ||
4293 User->getValueType(0) != MVT::i32))
4295 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4296 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4299 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4300 } else if (VT == MVT::i32) {
4301 // ExtractPS works with constant index.
4302 if (isa<ConstantSDNode>(Op.getOperand(1)))
4310 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4311 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4314 if (Subtarget->hasSSE41()) {
4315 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4320 MVT VT = Op.getValueType();
4321 DebugLoc dl = Op.getDebugLoc();
4322 // TODO: handle v16i8.
4323 if (VT.getSizeInBits() == 16) {
4324 SDValue Vec = Op.getOperand(0);
4325 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4327 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4328 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4329 DAG.getNode(ISD::BIT_CONVERT, dl,
4332 // Transform it so it match pextrw which produces a 32-bit result.
4333 MVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT()+1);
4334 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4335 Op.getOperand(0), Op.getOperand(1));
4336 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4337 DAG.getValueType(VT));
4338 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4339 } else if (VT.getSizeInBits() == 32) {
4340 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4344 // SHUFPS the element to the lowest double word, then movss.
4345 int Mask[4] = { Idx, -1, -1, -1 };
4346 MVT VVT = Op.getOperand(0).getValueType();
4347 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4348 DAG.getUNDEF(VVT), Mask);
4349 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4350 DAG.getIntPtrConstant(0));
4351 } else if (VT.getSizeInBits() == 64) {
4352 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4353 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4354 // to match extract_elt for f64.
4355 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4359 // UNPCKHPD the element to the lowest double word, then movsd.
4360 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4361 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4362 int Mask[2] = { 1, -1 };
4363 MVT VVT = Op.getOperand(0).getValueType();
4364 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4365 DAG.getUNDEF(VVT), Mask);
4366 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4367 DAG.getIntPtrConstant(0));
4374 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4375 MVT VT = Op.getValueType();
4376 MVT EVT = VT.getVectorElementType();
4377 DebugLoc dl = Op.getDebugLoc();
4379 SDValue N0 = Op.getOperand(0);
4380 SDValue N1 = Op.getOperand(1);
4381 SDValue N2 = Op.getOperand(2);
4383 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4384 isa<ConstantSDNode>(N2)) {
4385 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4387 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4389 if (N1.getValueType() != MVT::i32)
4390 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4391 if (N2.getValueType() != MVT::i32)
4392 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4393 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4394 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4395 // Bits [7:6] of the constant are the source select. This will always be
4396 // zero here. The DAG Combiner may combine an extract_elt index into these
4397 // bits. For example (insert (extract, 3), 2) could be matched by putting
4398 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4399 // Bits [5:4] of the constant are the destination select. This is the
4400 // value of the incoming immediate.
4401 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4402 // combine either bitwise AND or insert of float 0.0 to set these bits.
4403 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4404 // Create this as a scalar to vector..
4405 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4406 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4407 } else if (EVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4408 // PINSR* works with constant index.
4415 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4416 MVT VT = Op.getValueType();
4417 MVT EVT = VT.getVectorElementType();
4419 if (Subtarget->hasSSE41())
4420 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4425 DebugLoc dl = Op.getDebugLoc();
4426 SDValue N0 = Op.getOperand(0);
4427 SDValue N1 = Op.getOperand(1);
4428 SDValue N2 = Op.getOperand(2);
4430 if (EVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4431 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4432 // as its second argument.
4433 if (N1.getValueType() != MVT::i32)
4434 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4435 if (N2.getValueType() != MVT::i32)
4436 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4437 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4443 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4444 DebugLoc dl = Op.getDebugLoc();
4445 if (Op.getValueType() == MVT::v2f32)
4446 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4447 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4448 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4449 Op.getOperand(0))));
4451 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4452 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4454 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4455 MVT VT = MVT::v2i32;
4456 switch (Op.getValueType().getSimpleVT()) {
4463 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4464 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4467 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4468 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4469 // one of the above mentioned nodes. It has to be wrapped because otherwise
4470 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4471 // be used to form addressing mode. These wrapped nodes will be selected
4474 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4475 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4477 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4479 unsigned char OpFlag = 0;
4480 unsigned WrapperKind = X86ISD::Wrapper;
4482 if (Subtarget->isPICStyleRIPRel() &&
4483 getTargetMachine().getCodeModel() == CodeModel::Small)
4484 WrapperKind = X86ISD::WrapperRIP;
4485 else if (Subtarget->isPICStyleGOT())
4486 OpFlag = X86II::MO_GOTOFF;
4487 else if (Subtarget->isPICStyleStubPIC())
4488 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4490 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4492 CP->getOffset(), OpFlag);
4493 DebugLoc DL = CP->getDebugLoc();
4494 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4495 // With PIC, the address is actually $g + Offset.
4497 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4498 DAG.getNode(X86ISD::GlobalBaseReg,
4499 DebugLoc::getUnknownLoc(), getPointerTy()),
4506 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4507 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4509 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4511 unsigned char OpFlag = 0;
4512 unsigned WrapperKind = X86ISD::Wrapper;
4514 if (Subtarget->isPICStyleRIPRel() &&
4515 getTargetMachine().getCodeModel() == CodeModel::Small)
4516 WrapperKind = X86ISD::WrapperRIP;
4517 else if (Subtarget->isPICStyleGOT())
4518 OpFlag = X86II::MO_GOTOFF;
4519 else if (Subtarget->isPICStyleStubPIC())
4520 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4522 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4524 DebugLoc DL = JT->getDebugLoc();
4525 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4527 // With PIC, the address is actually $g + Offset.
4529 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4530 DAG.getNode(X86ISD::GlobalBaseReg,
4531 DebugLoc::getUnknownLoc(), getPointerTy()),
4539 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4540 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4542 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4544 unsigned char OpFlag = 0;
4545 unsigned WrapperKind = X86ISD::Wrapper;
4546 if (Subtarget->isPICStyleRIPRel() &&
4547 getTargetMachine().getCodeModel() == CodeModel::Small)
4548 WrapperKind = X86ISD::WrapperRIP;
4549 else if (Subtarget->isPICStyleGOT())
4550 OpFlag = X86II::MO_GOTOFF;
4551 else if (Subtarget->isPICStyleStubPIC())
4552 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4554 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4556 DebugLoc DL = Op.getDebugLoc();
4557 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4560 // With PIC, the address is actually $g + Offset.
4561 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4562 !Subtarget->is64Bit()) {
4563 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4564 DAG.getNode(X86ISD::GlobalBaseReg,
4565 DebugLoc::getUnknownLoc(),
4574 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4576 SelectionDAG &DAG) const {
4577 // Create the TargetGlobalAddress node, folding in the constant
4578 // offset if it is legal.
4579 unsigned char OpFlags =
4580 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4582 if (OpFlags == X86II::MO_NO_FLAG && isInt32(Offset)) {
4583 // A direct static reference to a global.
4584 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4587 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4590 if (Subtarget->isPICStyleRIPRel() &&
4591 getTargetMachine().getCodeModel() == CodeModel::Small)
4592 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4594 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4596 // With PIC, the address is actually $g + Offset.
4597 if (isGlobalRelativeToPICBase(OpFlags)) {
4598 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4599 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4603 // For globals that require a load from a stub to get the address, emit the
4605 if (isGlobalStubReference(OpFlags))
4606 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4607 PseudoSourceValue::getGOT(), 0);
4609 // If there was a non-zero offset that we didn't fold, create an explicit
4612 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4613 DAG.getConstant(Offset, getPointerTy()));
4619 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4620 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4621 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4622 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4626 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4627 SDValue *InFlag, const MVT PtrVT, unsigned ReturnReg,
4628 unsigned char OperandFlags) {
4629 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4630 DebugLoc dl = GA->getDebugLoc();
4631 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4632 GA->getValueType(0),
4636 SDValue Ops[] = { Chain, TGA, *InFlag };
4637 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4639 SDValue Ops[] = { Chain, TGA };
4640 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4642 SDValue Flag = Chain.getValue(1);
4643 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4646 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4648 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4651 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4652 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4653 DAG.getNode(X86ISD::GlobalBaseReg,
4654 DebugLoc::getUnknownLoc(),
4656 InFlag = Chain.getValue(1);
4658 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4661 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4663 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4665 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4666 X86::RAX, X86II::MO_TLSGD);
4669 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4670 // "local exec" model.
4671 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4672 const MVT PtrVT, TLSModel::Model model,
4674 DebugLoc dl = GA->getDebugLoc();
4675 // Get the Thread Pointer
4676 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4677 DebugLoc::getUnknownLoc(), PtrVT,
4678 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4681 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4684 unsigned char OperandFlags = 0;
4685 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
4687 unsigned WrapperKind = X86ISD::Wrapper;
4688 if (model == TLSModel::LocalExec) {
4689 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
4690 } else if (is64Bit) {
4691 assert(model == TLSModel::InitialExec);
4692 OperandFlags = X86II::MO_GOTTPOFF;
4693 WrapperKind = X86ISD::WrapperRIP;
4695 assert(model == TLSModel::InitialExec);
4696 OperandFlags = X86II::MO_INDNTPOFF;
4699 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4701 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4702 GA->getOffset(), OperandFlags);
4703 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
4705 if (model == TLSModel::InitialExec)
4706 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4707 PseudoSourceValue::getGOT(), 0);
4709 // The address of the thread local variable is the add of the thread
4710 // pointer with the offset of the variable.
4711 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4715 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4716 // TODO: implement the "local dynamic" model
4717 // TODO: implement the "initial exec"model for pic executables
4718 assert(Subtarget->isTargetELF() &&
4719 "TLS not implemented for non-ELF targets");
4720 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4721 const GlobalValue *GV = GA->getGlobal();
4723 // If GV is an alias then use the aliasee for determining
4724 // thread-localness.
4725 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
4726 GV = GA->resolveAliasedGlobal(false);
4728 TLSModel::Model model = getTLSModel(GV,
4729 getTargetMachine().getRelocationModel());
4732 case TLSModel::GeneralDynamic:
4733 case TLSModel::LocalDynamic: // not implemented
4734 if (Subtarget->is64Bit())
4735 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4736 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4738 case TLSModel::InitialExec:
4739 case TLSModel::LocalExec:
4740 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
4741 Subtarget->is64Bit());
4744 llvm_unreachable("Unreachable");
4749 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4750 /// take a 2 x i32 value to shift plus a shift amount.
4751 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4752 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4753 MVT VT = Op.getValueType();
4754 unsigned VTBits = VT.getSizeInBits();
4755 DebugLoc dl = Op.getDebugLoc();
4756 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4757 SDValue ShOpLo = Op.getOperand(0);
4758 SDValue ShOpHi = Op.getOperand(1);
4759 SDValue ShAmt = Op.getOperand(2);
4760 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4761 DAG.getConstant(VTBits - 1, MVT::i8))
4762 : DAG.getConstant(0, VT);
4765 if (Op.getOpcode() == ISD::SHL_PARTS) {
4766 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4767 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4769 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4770 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4773 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4774 DAG.getConstant(VTBits, MVT::i8));
4775 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4776 AndNode, DAG.getConstant(0, MVT::i8));
4779 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4780 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4781 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4783 if (Op.getOpcode() == ISD::SHL_PARTS) {
4784 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4785 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4787 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4788 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4791 SDValue Ops[2] = { Lo, Hi };
4792 return DAG.getMergeValues(Ops, 2, dl);
4795 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4796 MVT SrcVT = Op.getOperand(0).getValueType();
4798 if (SrcVT.isVector()) {
4799 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
4805 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4806 "Unknown SINT_TO_FP to lower!");
4808 // These are really Legal; return the operand so the caller accepts it as
4810 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4812 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
4813 Subtarget->is64Bit()) {
4817 DebugLoc dl = Op.getDebugLoc();
4818 unsigned Size = SrcVT.getSizeInBits()/8;
4819 MachineFunction &MF = DAG.getMachineFunction();
4820 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4821 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4822 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4824 PseudoSourceValue::getFixedStack(SSFI), 0);
4825 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
4828 SDValue X86TargetLowering::BuildFILD(SDValue Op, MVT SrcVT, SDValue Chain,
4830 SelectionDAG &DAG) {
4832 DebugLoc dl = Op.getDebugLoc();
4834 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4836 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4838 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4839 SmallVector<SDValue, 8> Ops;
4840 Ops.push_back(Chain);
4841 Ops.push_back(StackSlot);
4842 Ops.push_back(DAG.getValueType(SrcVT));
4843 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4844 Tys, &Ops[0], Ops.size());
4847 Chain = Result.getValue(1);
4848 SDValue InFlag = Result.getValue(2);
4850 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4851 // shouldn't be necessary except that RFP cannot be live across
4852 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4853 MachineFunction &MF = DAG.getMachineFunction();
4854 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4855 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4856 Tys = DAG.getVTList(MVT::Other);
4857 SmallVector<SDValue, 8> Ops;
4858 Ops.push_back(Chain);
4859 Ops.push_back(Result);
4860 Ops.push_back(StackSlot);
4861 Ops.push_back(DAG.getValueType(Op.getValueType()));
4862 Ops.push_back(InFlag);
4863 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4864 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4865 PseudoSourceValue::getFixedStack(SSFI), 0);
4871 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4872 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4873 // This algorithm is not obvious. Here it is in C code, more or less:
4875 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4876 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4877 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4879 // Copy ints to xmm registers.
4880 __m128i xh = _mm_cvtsi32_si128( hi );
4881 __m128i xl = _mm_cvtsi32_si128( lo );
4883 // Combine into low half of a single xmm register.
4884 __m128i x = _mm_unpacklo_epi32( xh, xl );
4888 // Merge in appropriate exponents to give the integer bits the right
4890 x = _mm_unpacklo_epi32( x, exp );
4892 // Subtract away the biases to deal with the IEEE-754 double precision
4894 d = _mm_sub_pd( (__m128d) x, bias );
4896 // All conversions up to here are exact. The correctly rounded result is
4897 // calculated using the current rounding mode using the following
4899 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4900 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4901 // store doesn't really need to be here (except
4902 // maybe to zero the other double)
4907 DebugLoc dl = Op.getDebugLoc();
4908 LLVMContext *Context = DAG.getContext();
4910 // Build some magic constants.
4911 std::vector<Constant*> CV0;
4912 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
4913 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
4914 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
4915 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
4916 Constant *C0 = ConstantVector::get(CV0);
4917 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
4919 std::vector<Constant*> CV1;
4921 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
4923 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
4924 Constant *C1 = ConstantVector::get(CV1);
4925 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
4927 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4928 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4930 DAG.getIntPtrConstant(1)));
4931 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4932 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4934 DAG.getIntPtrConstant(0)));
4935 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
4936 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
4937 PseudoSourceValue::getConstantPool(), 0,
4939 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
4940 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
4941 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
4942 PseudoSourceValue::getConstantPool(), 0,
4944 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
4946 // Add the halves; easiest way is to swap them into another reg first.
4947 int ShufMask[2] = { 1, -1 };
4948 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
4949 DAG.getUNDEF(MVT::v2f64), ShufMask);
4950 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
4951 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
4952 DAG.getIntPtrConstant(0));
4955 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
4956 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
4957 DebugLoc dl = Op.getDebugLoc();
4958 // FP constant to bias correct the final result.
4959 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
4962 // Load the 32-bit value into an XMM register.
4963 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4964 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4966 DAG.getIntPtrConstant(0)));
4968 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4969 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
4970 DAG.getIntPtrConstant(0));
4972 // Or the load with the bias.
4973 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
4974 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4975 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4977 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4978 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4979 MVT::v2f64, Bias)));
4980 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4981 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
4982 DAG.getIntPtrConstant(0));
4984 // Subtract the bias.
4985 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
4987 // Handle final rounding.
4988 MVT DestVT = Op.getValueType();
4990 if (DestVT.bitsLT(MVT::f64)) {
4991 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
4992 DAG.getIntPtrConstant(0));
4993 } else if (DestVT.bitsGT(MVT::f64)) {
4994 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
4997 // Handle final rounding.
5001 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5002 SDValue N0 = Op.getOperand(0);
5003 DebugLoc dl = Op.getDebugLoc();
5005 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5006 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5007 // the optimization here.
5008 if (DAG.SignBitIsZero(N0))
5009 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5011 MVT SrcVT = N0.getValueType();
5012 if (SrcVT == MVT::i64) {
5013 // We only handle SSE2 f64 target here; caller can expand the rest.
5014 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5017 return LowerUINT_TO_FP_i64(Op, DAG);
5018 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5019 return LowerUINT_TO_FP_i32(Op, DAG);
5022 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5024 // Make a 64-bit buffer, and use it to build an FILD.
5025 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5026 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5027 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5028 getPointerTy(), StackSlot, WordOff);
5029 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5030 StackSlot, NULL, 0);
5031 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5032 OffsetSlot, NULL, 0);
5033 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5036 std::pair<SDValue,SDValue> X86TargetLowering::
5037 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5038 DebugLoc dl = Op.getDebugLoc();
5040 MVT DstTy = Op.getValueType();
5043 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5047 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5048 DstTy.getSimpleVT() >= MVT::i16 &&
5049 "Unknown FP_TO_SINT to lower!");
5051 // These are really Legal.
5052 if (DstTy == MVT::i32 &&
5053 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5054 return std::make_pair(SDValue(), SDValue());
5055 if (Subtarget->is64Bit() &&
5056 DstTy == MVT::i64 &&
5057 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5058 return std::make_pair(SDValue(), SDValue());
5060 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5062 MachineFunction &MF = DAG.getMachineFunction();
5063 unsigned MemSize = DstTy.getSizeInBits()/8;
5064 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5065 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5068 switch (DstTy.getSimpleVT()) {
5069 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5070 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5071 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5072 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5075 SDValue Chain = DAG.getEntryNode();
5076 SDValue Value = Op.getOperand(0);
5077 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5078 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5079 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5080 PseudoSourceValue::getFixedStack(SSFI), 0);
5081 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5083 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5085 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5086 Chain = Value.getValue(1);
5087 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5088 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5091 // Build the FP_TO_INT*_IN_MEM
5092 SDValue Ops[] = { Chain, Value, StackSlot };
5093 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5095 return std::make_pair(FIST, StackSlot);
5098 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5099 if (Op.getValueType().isVector()) {
5100 if (Op.getValueType() == MVT::v2i32 &&
5101 Op.getOperand(0).getValueType() == MVT::v2f64) {
5107 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5108 SDValue FIST = Vals.first, StackSlot = Vals.second;
5109 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5110 if (FIST.getNode() == 0) return Op;
5113 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5114 FIST, StackSlot, NULL, 0);
5117 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5118 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5119 SDValue FIST = Vals.first, StackSlot = Vals.second;
5120 assert(FIST.getNode() && "Unexpected failure");
5123 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5124 FIST, StackSlot, NULL, 0);
5127 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5128 LLVMContext *Context = DAG.getContext();
5129 DebugLoc dl = Op.getDebugLoc();
5130 MVT VT = Op.getValueType();
5133 EltVT = VT.getVectorElementType();
5134 std::vector<Constant*> CV;
5135 if (EltVT == MVT::f64) {
5136 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5140 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5146 Constant *C = ConstantVector::get(CV);
5147 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5148 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5149 PseudoSourceValue::getConstantPool(), 0,
5151 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5154 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5155 LLVMContext *Context = DAG.getContext();
5156 DebugLoc dl = Op.getDebugLoc();
5157 MVT VT = Op.getValueType();
5159 unsigned EltNum = 1;
5160 if (VT.isVector()) {
5161 EltVT = VT.getVectorElementType();
5162 EltNum = VT.getVectorNumElements();
5164 std::vector<Constant*> CV;
5165 if (EltVT == MVT::f64) {
5166 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5170 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5176 Constant *C = ConstantVector::get(CV);
5177 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5178 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5179 PseudoSourceValue::getConstantPool(), 0,
5181 if (VT.isVector()) {
5182 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5183 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5184 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5186 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5188 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5192 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5193 LLVMContext *Context = DAG.getContext();
5194 SDValue Op0 = Op.getOperand(0);
5195 SDValue Op1 = Op.getOperand(1);
5196 DebugLoc dl = Op.getDebugLoc();
5197 MVT VT = Op.getValueType();
5198 MVT SrcVT = Op1.getValueType();
5200 // If second operand is smaller, extend it first.
5201 if (SrcVT.bitsLT(VT)) {
5202 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5205 // And if it is bigger, shrink it first.
5206 if (SrcVT.bitsGT(VT)) {
5207 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5211 // At this point the operands and the result should have the same
5212 // type, and that won't be f80 since that is not custom lowered.
5214 // First get the sign bit of second operand.
5215 std::vector<Constant*> CV;
5216 if (SrcVT == MVT::f64) {
5217 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5218 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5220 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5221 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5222 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5223 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5225 Constant *C = ConstantVector::get(CV);
5226 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5227 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5228 PseudoSourceValue::getConstantPool(), 0,
5230 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5232 // Shift sign bit right or left if the two operands have different types.
5233 if (SrcVT.bitsGT(VT)) {
5234 // Op0 is MVT::f32, Op1 is MVT::f64.
5235 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5236 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5237 DAG.getConstant(32, MVT::i32));
5238 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5239 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5240 DAG.getIntPtrConstant(0));
5243 // Clear first operand sign bit.
5245 if (VT == MVT::f64) {
5246 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5247 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5249 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5250 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5251 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5252 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5254 C = ConstantVector::get(CV);
5255 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5256 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5257 PseudoSourceValue::getConstantPool(), 0,
5259 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5261 // Or the value with the sign bit.
5262 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5265 /// Emit nodes that will be selected as "test Op0,Op0", or something
5267 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5268 SelectionDAG &DAG) {
5269 DebugLoc dl = Op.getDebugLoc();
5271 // CF and OF aren't always set the way we want. Determine which
5272 // of these we need.
5273 bool NeedCF = false;
5274 bool NeedOF = false;
5276 case X86::COND_A: case X86::COND_AE:
5277 case X86::COND_B: case X86::COND_BE:
5280 case X86::COND_G: case X86::COND_GE:
5281 case X86::COND_L: case X86::COND_LE:
5282 case X86::COND_O: case X86::COND_NO:
5288 // See if we can use the EFLAGS value from the operand instead of
5289 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5290 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5291 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5292 unsigned Opcode = 0;
5293 unsigned NumOperands = 0;
5294 switch (Op.getNode()->getOpcode()) {
5296 // Due to an isel shortcoming, be conservative if this add is likely to
5297 // be selected as part of a load-modify-store instruction. When the root
5298 // node in a match is a store, isel doesn't know how to remap non-chain
5299 // non-flag uses of other nodes in the match, such as the ADD in this
5300 // case. This leads to the ADD being left around and reselected, with
5301 // the result being two adds in the output.
5302 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5303 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5304 if (UI->getOpcode() == ISD::STORE)
5306 if (ConstantSDNode *C =
5307 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5308 // An add of one will be selected as an INC.
5309 if (C->getAPIntValue() == 1) {
5310 Opcode = X86ISD::INC;
5314 // An add of negative one (subtract of one) will be selected as a DEC.
5315 if (C->getAPIntValue().isAllOnesValue()) {
5316 Opcode = X86ISD::DEC;
5321 // Otherwise use a regular EFLAGS-setting add.
5322 Opcode = X86ISD::ADD;
5326 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5327 // likely to be selected as part of a load-modify-store instruction.
5328 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5329 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5330 if (UI->getOpcode() == ISD::STORE)
5332 // Otherwise use a regular EFLAGS-setting sub.
5333 Opcode = X86ISD::SUB;
5340 return SDValue(Op.getNode(), 1);
5346 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5347 SmallVector<SDValue, 4> Ops;
5348 for (unsigned i = 0; i != NumOperands; ++i)
5349 Ops.push_back(Op.getOperand(i));
5350 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5351 DAG.ReplaceAllUsesWith(Op, New);
5352 return SDValue(New.getNode(), 1);
5356 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5357 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5358 DAG.getConstant(0, Op.getValueType()));
5361 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5363 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5364 SelectionDAG &DAG) {
5365 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5366 if (C->getAPIntValue() == 0)
5367 return EmitTest(Op0, X86CC, DAG);
5369 DebugLoc dl = Op0.getDebugLoc();
5370 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5373 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5374 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5375 SDValue Op0 = Op.getOperand(0);
5376 SDValue Op1 = Op.getOperand(1);
5377 DebugLoc dl = Op.getDebugLoc();
5378 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5380 // Lower (X & (1 << N)) == 0 to BT(X, N).
5381 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5382 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5383 if (Op0.getOpcode() == ISD::AND &&
5385 Op1.getOpcode() == ISD::Constant &&
5386 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5387 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5389 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5390 if (ConstantSDNode *Op010C =
5391 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5392 if (Op010C->getZExtValue() == 1) {
5393 LHS = Op0.getOperand(0);
5394 RHS = Op0.getOperand(1).getOperand(1);
5396 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5397 if (ConstantSDNode *Op000C =
5398 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5399 if (Op000C->getZExtValue() == 1) {
5400 LHS = Op0.getOperand(1);
5401 RHS = Op0.getOperand(0).getOperand(1);
5403 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5404 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5405 SDValue AndLHS = Op0.getOperand(0);
5406 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5407 LHS = AndLHS.getOperand(0);
5408 RHS = AndLHS.getOperand(1);
5412 if (LHS.getNode()) {
5413 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5414 // instruction. Since the shift amount is in-range-or-undefined, we know
5415 // that doing a bittest on the i16 value is ok. We extend to i32 because
5416 // the encoding for the i16 version is larger than the i32 version.
5417 if (LHS.getValueType() == MVT::i8)
5418 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5420 // If the operand types disagree, extend the shift amount to match. Since
5421 // BT ignores high bits (like shifts) we can use anyextend.
5422 if (LHS.getValueType() != RHS.getValueType())
5423 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5425 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5426 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5427 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5428 DAG.getConstant(Cond, MVT::i8), BT);
5432 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5433 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5435 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5436 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5437 DAG.getConstant(X86CC, MVT::i8), Cond);
5440 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5442 SDValue Op0 = Op.getOperand(0);
5443 SDValue Op1 = Op.getOperand(1);
5444 SDValue CC = Op.getOperand(2);
5445 MVT VT = Op.getValueType();
5446 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5447 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5448 DebugLoc dl = Op.getDebugLoc();
5452 MVT VT0 = Op0.getValueType();
5453 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5454 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5457 switch (SetCCOpcode) {
5460 case ISD::SETEQ: SSECC = 0; break;
5462 case ISD::SETGT: Swap = true; // Fallthrough
5464 case ISD::SETOLT: SSECC = 1; break;
5466 case ISD::SETGE: Swap = true; // Fallthrough
5468 case ISD::SETOLE: SSECC = 2; break;
5469 case ISD::SETUO: SSECC = 3; break;
5471 case ISD::SETNE: SSECC = 4; break;
5472 case ISD::SETULE: Swap = true;
5473 case ISD::SETUGE: SSECC = 5; break;
5474 case ISD::SETULT: Swap = true;
5475 case ISD::SETUGT: SSECC = 6; break;
5476 case ISD::SETO: SSECC = 7; break;
5479 std::swap(Op0, Op1);
5481 // In the two special cases we can't handle, emit two comparisons.
5483 if (SetCCOpcode == ISD::SETUEQ) {
5485 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5486 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5487 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5489 else if (SetCCOpcode == ISD::SETONE) {
5491 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5492 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5493 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5495 llvm_unreachable("Illegal FP comparison");
5497 // Handle all other FP comparisons here.
5498 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5501 // We are handling one of the integer comparisons here. Since SSE only has
5502 // GT and EQ comparisons for integer, swapping operands and multiple
5503 // operations may be required for some comparisons.
5504 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5505 bool Swap = false, Invert = false, FlipSigns = false;
5507 switch (VT.getSimpleVT()) {
5510 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5512 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5514 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5515 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5518 switch (SetCCOpcode) {
5520 case ISD::SETNE: Invert = true;
5521 case ISD::SETEQ: Opc = EQOpc; break;
5522 case ISD::SETLT: Swap = true;
5523 case ISD::SETGT: Opc = GTOpc; break;
5524 case ISD::SETGE: Swap = true;
5525 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5526 case ISD::SETULT: Swap = true;
5527 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5528 case ISD::SETUGE: Swap = true;
5529 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5532 std::swap(Op0, Op1);
5534 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5535 // bits of the inputs before performing those operations.
5537 MVT EltVT = VT.getVectorElementType();
5538 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5540 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5541 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5543 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5544 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5547 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5549 // If the logical-not of the result is required, perform that now.
5551 Result = DAG.getNOT(dl, Result, VT);
5556 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5557 static bool isX86LogicalCmp(SDValue Op) {
5558 unsigned Opc = Op.getNode()->getOpcode();
5559 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5561 if (Op.getResNo() == 1 &&
5562 (Opc == X86ISD::ADD ||
5563 Opc == X86ISD::SUB ||
5564 Opc == X86ISD::SMUL ||
5565 Opc == X86ISD::UMUL ||
5566 Opc == X86ISD::INC ||
5567 Opc == X86ISD::DEC))
5573 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5574 bool addTest = true;
5575 SDValue Cond = Op.getOperand(0);
5576 DebugLoc dl = Op.getDebugLoc();
5579 if (Cond.getOpcode() == ISD::SETCC)
5580 Cond = LowerSETCC(Cond, DAG);
5582 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5583 // setting operand in place of the X86ISD::SETCC.
5584 if (Cond.getOpcode() == X86ISD::SETCC) {
5585 CC = Cond.getOperand(0);
5587 SDValue Cmp = Cond.getOperand(1);
5588 unsigned Opc = Cmp.getOpcode();
5589 MVT VT = Op.getValueType();
5591 bool IllegalFPCMov = false;
5592 if (VT.isFloatingPoint() && !VT.isVector() &&
5593 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5594 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5596 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5597 Opc == X86ISD::BT) { // FIXME
5604 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5605 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5608 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5609 SmallVector<SDValue, 4> Ops;
5610 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5611 // condition is true.
5612 Ops.push_back(Op.getOperand(2));
5613 Ops.push_back(Op.getOperand(1));
5615 Ops.push_back(Cond);
5616 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5619 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5620 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5621 // from the AND / OR.
5622 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5623 Opc = Op.getOpcode();
5624 if (Opc != ISD::OR && Opc != ISD::AND)
5626 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5627 Op.getOperand(0).hasOneUse() &&
5628 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5629 Op.getOperand(1).hasOneUse());
5632 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5633 // 1 and that the SETCC node has a single use.
5634 static bool isXor1OfSetCC(SDValue Op) {
5635 if (Op.getOpcode() != ISD::XOR)
5637 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5638 if (N1C && N1C->getAPIntValue() == 1) {
5639 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5640 Op.getOperand(0).hasOneUse();
5645 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5646 bool addTest = true;
5647 SDValue Chain = Op.getOperand(0);
5648 SDValue Cond = Op.getOperand(1);
5649 SDValue Dest = Op.getOperand(2);
5650 DebugLoc dl = Op.getDebugLoc();
5653 if (Cond.getOpcode() == ISD::SETCC)
5654 Cond = LowerSETCC(Cond, DAG);
5656 // FIXME: LowerXALUO doesn't handle these!!
5657 else if (Cond.getOpcode() == X86ISD::ADD ||
5658 Cond.getOpcode() == X86ISD::SUB ||
5659 Cond.getOpcode() == X86ISD::SMUL ||
5660 Cond.getOpcode() == X86ISD::UMUL)
5661 Cond = LowerXALUO(Cond, DAG);
5664 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5665 // setting operand in place of the X86ISD::SETCC.
5666 if (Cond.getOpcode() == X86ISD::SETCC) {
5667 CC = Cond.getOperand(0);
5669 SDValue Cmp = Cond.getOperand(1);
5670 unsigned Opc = Cmp.getOpcode();
5671 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5672 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5676 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5680 // These can only come from an arithmetic instruction with overflow,
5681 // e.g. SADDO, UADDO.
5682 Cond = Cond.getNode()->getOperand(1);
5689 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5690 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5691 if (CondOpc == ISD::OR) {
5692 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5693 // two branches instead of an explicit OR instruction with a
5695 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5696 isX86LogicalCmp(Cmp)) {
5697 CC = Cond.getOperand(0).getOperand(0);
5698 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5699 Chain, Dest, CC, Cmp);
5700 CC = Cond.getOperand(1).getOperand(0);
5704 } else { // ISD::AND
5705 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5706 // two branches instead of an explicit AND instruction with a
5707 // separate test. However, we only do this if this block doesn't
5708 // have a fall-through edge, because this requires an explicit
5709 // jmp when the condition is false.
5710 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5711 isX86LogicalCmp(Cmp) &&
5712 Op.getNode()->hasOneUse()) {
5713 X86::CondCode CCode =
5714 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5715 CCode = X86::GetOppositeBranchCondition(CCode);
5716 CC = DAG.getConstant(CCode, MVT::i8);
5717 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5718 // Look for an unconditional branch following this conditional branch.
5719 // We need this because we need to reverse the successors in order
5720 // to implement FCMP_OEQ.
5721 if (User.getOpcode() == ISD::BR) {
5722 SDValue FalseBB = User.getOperand(1);
5724 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5725 assert(NewBR == User);
5728 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5729 Chain, Dest, CC, Cmp);
5730 X86::CondCode CCode =
5731 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5732 CCode = X86::GetOppositeBranchCondition(CCode);
5733 CC = DAG.getConstant(CCode, MVT::i8);
5739 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5740 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5741 // It should be transformed during dag combiner except when the condition
5742 // is set by a arithmetics with overflow node.
5743 X86::CondCode CCode =
5744 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5745 CCode = X86::GetOppositeBranchCondition(CCode);
5746 CC = DAG.getConstant(CCode, MVT::i8);
5747 Cond = Cond.getOperand(0).getOperand(1);
5753 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5754 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5756 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5757 Chain, Dest, CC, Cond);
5761 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5762 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5763 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5764 // that the guard pages used by the OS virtual memory manager are allocated in
5765 // correct sequence.
5767 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5768 SelectionDAG &DAG) {
5769 assert(Subtarget->isTargetCygMing() &&
5770 "This should be used only on Cygwin/Mingw targets");
5771 DebugLoc dl = Op.getDebugLoc();
5774 SDValue Chain = Op.getOperand(0);
5775 SDValue Size = Op.getOperand(1);
5776 // FIXME: Ensure alignment here
5780 MVT IntPtr = getPointerTy();
5781 MVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5783 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5785 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5786 Flag = Chain.getValue(1);
5788 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5789 SDValue Ops[] = { Chain,
5790 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5791 DAG.getRegister(X86::EAX, IntPtr),
5792 DAG.getRegister(X86StackPtr, SPTy),
5794 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5795 Flag = Chain.getValue(1);
5797 Chain = DAG.getCALLSEQ_END(Chain,
5798 DAG.getIntPtrConstant(0, true),
5799 DAG.getIntPtrConstant(0, true),
5802 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5804 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5805 return DAG.getMergeValues(Ops1, 2, dl);
5809 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5811 SDValue Dst, SDValue Src,
5812 SDValue Size, unsigned Align,
5814 uint64_t DstSVOff) {
5815 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5817 // If not DWORD aligned or size is more than the threshold, call the library.
5818 // The libc version is likely to be faster for these cases. It can use the
5819 // address value and run time information about the CPU.
5820 if ((Align & 3) != 0 ||
5822 ConstantSize->getZExtValue() >
5823 getSubtarget()->getMaxInlineSizeThreshold()) {
5824 SDValue InFlag(0, 0);
5826 // Check to see if there is a specialized entry-point for memory zeroing.
5827 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5829 if (const char *bzeroEntry = V &&
5830 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5831 MVT IntPtr = getPointerTy();
5832 const Type *IntPtrTy = TD->getIntPtrType();
5833 TargetLowering::ArgListTy Args;
5834 TargetLowering::ArgListEntry Entry;
5836 Entry.Ty = IntPtrTy;
5837 Args.push_back(Entry);
5839 Args.push_back(Entry);
5840 std::pair<SDValue,SDValue> CallResult =
5841 LowerCallTo(Chain, Type::VoidTy, false, false, false, false,
5842 0, CallingConv::C, false,
5843 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5844 return CallResult.second;
5847 // Otherwise have the target-independent code call memset.
5851 uint64_t SizeVal = ConstantSize->getZExtValue();
5852 SDValue InFlag(0, 0);
5855 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5856 unsigned BytesLeft = 0;
5857 bool TwoRepStos = false;
5860 uint64_t Val = ValC->getZExtValue() & 255;
5862 // If the value is a constant, then we can potentially use larger sets.
5863 switch (Align & 3) {
5864 case 2: // WORD aligned
5867 Val = (Val << 8) | Val;
5869 case 0: // DWORD aligned
5872 Val = (Val << 8) | Val;
5873 Val = (Val << 16) | Val;
5874 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5877 Val = (Val << 32) | Val;
5880 default: // Byte aligned
5883 Count = DAG.getIntPtrConstant(SizeVal);
5887 if (AVT.bitsGT(MVT::i8)) {
5888 unsigned UBytes = AVT.getSizeInBits() / 8;
5889 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5890 BytesLeft = SizeVal % UBytes;
5893 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
5895 InFlag = Chain.getValue(1);
5898 Count = DAG.getIntPtrConstant(SizeVal);
5899 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
5900 InFlag = Chain.getValue(1);
5903 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5906 InFlag = Chain.getValue(1);
5907 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5910 InFlag = Chain.getValue(1);
5912 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5913 SmallVector<SDValue, 8> Ops;
5914 Ops.push_back(Chain);
5915 Ops.push_back(DAG.getValueType(AVT));
5916 Ops.push_back(InFlag);
5917 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5920 InFlag = Chain.getValue(1);
5922 MVT CVT = Count.getValueType();
5923 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
5924 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5925 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
5928 InFlag = Chain.getValue(1);
5929 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5931 Ops.push_back(Chain);
5932 Ops.push_back(DAG.getValueType(MVT::i8));
5933 Ops.push_back(InFlag);
5934 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5935 } else if (BytesLeft) {
5936 // Handle the last 1 - 7 bytes.
5937 unsigned Offset = SizeVal - BytesLeft;
5938 MVT AddrVT = Dst.getValueType();
5939 MVT SizeVT = Size.getValueType();
5941 Chain = DAG.getMemset(Chain, dl,
5942 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
5943 DAG.getConstant(Offset, AddrVT)),
5945 DAG.getConstant(BytesLeft, SizeVT),
5946 Align, DstSV, DstSVOff + Offset);
5949 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5954 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
5955 SDValue Chain, SDValue Dst, SDValue Src,
5956 SDValue Size, unsigned Align,
5958 const Value *DstSV, uint64_t DstSVOff,
5959 const Value *SrcSV, uint64_t SrcSVOff) {
5960 // This requires the copy size to be a constant, preferrably
5961 // within a subtarget-specific limit.
5962 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5965 uint64_t SizeVal = ConstantSize->getZExtValue();
5966 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5969 /// If not DWORD aligned, call the library.
5970 if ((Align & 3) != 0)
5975 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5978 unsigned UBytes = AVT.getSizeInBits() / 8;
5979 unsigned CountVal = SizeVal / UBytes;
5980 SDValue Count = DAG.getIntPtrConstant(CountVal);
5981 unsigned BytesLeft = SizeVal % UBytes;
5983 SDValue InFlag(0, 0);
5984 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5987 InFlag = Chain.getValue(1);
5988 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5991 InFlag = Chain.getValue(1);
5992 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
5995 InFlag = Chain.getValue(1);
5997 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5998 SmallVector<SDValue, 8> Ops;
5999 Ops.push_back(Chain);
6000 Ops.push_back(DAG.getValueType(AVT));
6001 Ops.push_back(InFlag);
6002 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
6004 SmallVector<SDValue, 4> Results;
6005 Results.push_back(RepMovs);
6007 // Handle the last 1 - 7 bytes.
6008 unsigned Offset = SizeVal - BytesLeft;
6009 MVT DstVT = Dst.getValueType();
6010 MVT SrcVT = Src.getValueType();
6011 MVT SizeVT = Size.getValueType();
6012 Results.push_back(DAG.getMemcpy(Chain, dl,
6013 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6014 DAG.getConstant(Offset, DstVT)),
6015 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6016 DAG.getConstant(Offset, SrcVT)),
6017 DAG.getConstant(BytesLeft, SizeVT),
6018 Align, AlwaysInline,
6019 DstSV, DstSVOff + Offset,
6020 SrcSV, SrcSVOff + Offset));
6023 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6024 &Results[0], Results.size());
6027 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6028 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6029 DebugLoc dl = Op.getDebugLoc();
6031 if (!Subtarget->is64Bit()) {
6032 // vastart just stores the address of the VarArgsFrameIndex slot into the
6033 // memory location argument.
6034 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6035 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6039 // gp_offset (0 - 6 * 8)
6040 // fp_offset (48 - 48 + 8 * 16)
6041 // overflow_arg_area (point to parameters coming in memory).
6043 SmallVector<SDValue, 8> MemOps;
6044 SDValue FIN = Op.getOperand(1);
6046 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6047 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6049 MemOps.push_back(Store);
6052 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6053 FIN, DAG.getIntPtrConstant(4));
6054 Store = DAG.getStore(Op.getOperand(0), dl,
6055 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6057 MemOps.push_back(Store);
6059 // Store ptr to overflow_arg_area
6060 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6061 FIN, DAG.getIntPtrConstant(4));
6062 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6063 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6064 MemOps.push_back(Store);
6066 // Store ptr to reg_save_area.
6067 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6068 FIN, DAG.getIntPtrConstant(8));
6069 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6070 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6071 MemOps.push_back(Store);
6072 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6073 &MemOps[0], MemOps.size());
6076 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6077 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6078 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6079 SDValue Chain = Op.getOperand(0);
6080 SDValue SrcPtr = Op.getOperand(1);
6081 SDValue SrcSV = Op.getOperand(2);
6083 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6087 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6088 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6089 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6090 SDValue Chain = Op.getOperand(0);
6091 SDValue DstPtr = Op.getOperand(1);
6092 SDValue SrcPtr = Op.getOperand(2);
6093 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6094 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6095 DebugLoc dl = Op.getDebugLoc();
6097 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6098 DAG.getIntPtrConstant(24), 8, false,
6099 DstSV, 0, SrcSV, 0);
6103 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6104 DebugLoc dl = Op.getDebugLoc();
6105 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6107 default: return SDValue(); // Don't custom lower most intrinsics.
6108 // Comparison intrinsics.
6109 case Intrinsic::x86_sse_comieq_ss:
6110 case Intrinsic::x86_sse_comilt_ss:
6111 case Intrinsic::x86_sse_comile_ss:
6112 case Intrinsic::x86_sse_comigt_ss:
6113 case Intrinsic::x86_sse_comige_ss:
6114 case Intrinsic::x86_sse_comineq_ss:
6115 case Intrinsic::x86_sse_ucomieq_ss:
6116 case Intrinsic::x86_sse_ucomilt_ss:
6117 case Intrinsic::x86_sse_ucomile_ss:
6118 case Intrinsic::x86_sse_ucomigt_ss:
6119 case Intrinsic::x86_sse_ucomige_ss:
6120 case Intrinsic::x86_sse_ucomineq_ss:
6121 case Intrinsic::x86_sse2_comieq_sd:
6122 case Intrinsic::x86_sse2_comilt_sd:
6123 case Intrinsic::x86_sse2_comile_sd:
6124 case Intrinsic::x86_sse2_comigt_sd:
6125 case Intrinsic::x86_sse2_comige_sd:
6126 case Intrinsic::x86_sse2_comineq_sd:
6127 case Intrinsic::x86_sse2_ucomieq_sd:
6128 case Intrinsic::x86_sse2_ucomilt_sd:
6129 case Intrinsic::x86_sse2_ucomile_sd:
6130 case Intrinsic::x86_sse2_ucomigt_sd:
6131 case Intrinsic::x86_sse2_ucomige_sd:
6132 case Intrinsic::x86_sse2_ucomineq_sd: {
6134 ISD::CondCode CC = ISD::SETCC_INVALID;
6137 case Intrinsic::x86_sse_comieq_ss:
6138 case Intrinsic::x86_sse2_comieq_sd:
6142 case Intrinsic::x86_sse_comilt_ss:
6143 case Intrinsic::x86_sse2_comilt_sd:
6147 case Intrinsic::x86_sse_comile_ss:
6148 case Intrinsic::x86_sse2_comile_sd:
6152 case Intrinsic::x86_sse_comigt_ss:
6153 case Intrinsic::x86_sse2_comigt_sd:
6157 case Intrinsic::x86_sse_comige_ss:
6158 case Intrinsic::x86_sse2_comige_sd:
6162 case Intrinsic::x86_sse_comineq_ss:
6163 case Intrinsic::x86_sse2_comineq_sd:
6167 case Intrinsic::x86_sse_ucomieq_ss:
6168 case Intrinsic::x86_sse2_ucomieq_sd:
6169 Opc = X86ISD::UCOMI;
6172 case Intrinsic::x86_sse_ucomilt_ss:
6173 case Intrinsic::x86_sse2_ucomilt_sd:
6174 Opc = X86ISD::UCOMI;
6177 case Intrinsic::x86_sse_ucomile_ss:
6178 case Intrinsic::x86_sse2_ucomile_sd:
6179 Opc = X86ISD::UCOMI;
6182 case Intrinsic::x86_sse_ucomigt_ss:
6183 case Intrinsic::x86_sse2_ucomigt_sd:
6184 Opc = X86ISD::UCOMI;
6187 case Intrinsic::x86_sse_ucomige_ss:
6188 case Intrinsic::x86_sse2_ucomige_sd:
6189 Opc = X86ISD::UCOMI;
6192 case Intrinsic::x86_sse_ucomineq_ss:
6193 case Intrinsic::x86_sse2_ucomineq_sd:
6194 Opc = X86ISD::UCOMI;
6199 SDValue LHS = Op.getOperand(1);
6200 SDValue RHS = Op.getOperand(2);
6201 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6202 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6203 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6204 DAG.getConstant(X86CC, MVT::i8), Cond);
6205 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6207 // ptest intrinsics. The intrinsic these come from are designed to return
6208 // an integer value, not just an instruction so lower it to the ptest
6209 // pattern and a setcc for the result.
6210 case Intrinsic::x86_sse41_ptestz:
6211 case Intrinsic::x86_sse41_ptestc:
6212 case Intrinsic::x86_sse41_ptestnzc:{
6215 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6216 case Intrinsic::x86_sse41_ptestz:
6218 X86CC = X86::COND_E;
6220 case Intrinsic::x86_sse41_ptestc:
6222 X86CC = X86::COND_B;
6224 case Intrinsic::x86_sse41_ptestnzc:
6226 X86CC = X86::COND_A;
6230 SDValue LHS = Op.getOperand(1);
6231 SDValue RHS = Op.getOperand(2);
6232 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6233 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6234 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6235 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6238 // Fix vector shift instructions where the last operand is a non-immediate
6240 case Intrinsic::x86_sse2_pslli_w:
6241 case Intrinsic::x86_sse2_pslli_d:
6242 case Intrinsic::x86_sse2_pslli_q:
6243 case Intrinsic::x86_sse2_psrli_w:
6244 case Intrinsic::x86_sse2_psrli_d:
6245 case Intrinsic::x86_sse2_psrli_q:
6246 case Intrinsic::x86_sse2_psrai_w:
6247 case Intrinsic::x86_sse2_psrai_d:
6248 case Intrinsic::x86_mmx_pslli_w:
6249 case Intrinsic::x86_mmx_pslli_d:
6250 case Intrinsic::x86_mmx_pslli_q:
6251 case Intrinsic::x86_mmx_psrli_w:
6252 case Intrinsic::x86_mmx_psrli_d:
6253 case Intrinsic::x86_mmx_psrli_q:
6254 case Intrinsic::x86_mmx_psrai_w:
6255 case Intrinsic::x86_mmx_psrai_d: {
6256 SDValue ShAmt = Op.getOperand(2);
6257 if (isa<ConstantSDNode>(ShAmt))
6260 unsigned NewIntNo = 0;
6261 MVT ShAmtVT = MVT::v4i32;
6263 case Intrinsic::x86_sse2_pslli_w:
6264 NewIntNo = Intrinsic::x86_sse2_psll_w;
6266 case Intrinsic::x86_sse2_pslli_d:
6267 NewIntNo = Intrinsic::x86_sse2_psll_d;
6269 case Intrinsic::x86_sse2_pslli_q:
6270 NewIntNo = Intrinsic::x86_sse2_psll_q;
6272 case Intrinsic::x86_sse2_psrli_w:
6273 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6275 case Intrinsic::x86_sse2_psrli_d:
6276 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6278 case Intrinsic::x86_sse2_psrli_q:
6279 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6281 case Intrinsic::x86_sse2_psrai_w:
6282 NewIntNo = Intrinsic::x86_sse2_psra_w;
6284 case Intrinsic::x86_sse2_psrai_d:
6285 NewIntNo = Intrinsic::x86_sse2_psra_d;
6288 ShAmtVT = MVT::v2i32;
6290 case Intrinsic::x86_mmx_pslli_w:
6291 NewIntNo = Intrinsic::x86_mmx_psll_w;
6293 case Intrinsic::x86_mmx_pslli_d:
6294 NewIntNo = Intrinsic::x86_mmx_psll_d;
6296 case Intrinsic::x86_mmx_pslli_q:
6297 NewIntNo = Intrinsic::x86_mmx_psll_q;
6299 case Intrinsic::x86_mmx_psrli_w:
6300 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6302 case Intrinsic::x86_mmx_psrli_d:
6303 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6305 case Intrinsic::x86_mmx_psrli_q:
6306 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6308 case Intrinsic::x86_mmx_psrai_w:
6309 NewIntNo = Intrinsic::x86_mmx_psra_w;
6311 case Intrinsic::x86_mmx_psrai_d:
6312 NewIntNo = Intrinsic::x86_mmx_psra_d;
6314 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6319 MVT VT = Op.getValueType();
6320 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6321 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6322 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6323 DAG.getConstant(NewIntNo, MVT::i32),
6324 Op.getOperand(1), ShAmt);
6329 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6330 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6331 DebugLoc dl = Op.getDebugLoc();
6334 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6336 DAG.getConstant(TD->getPointerSize(),
6337 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6338 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6339 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6344 // Just load the return address.
6345 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6346 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6347 RetAddrFI, NULL, 0);
6350 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6351 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6352 MFI->setFrameAddressIsTaken(true);
6353 MVT VT = Op.getValueType();
6354 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6355 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6356 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6357 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6359 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6363 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6364 SelectionDAG &DAG) {
6365 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6368 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6370 MachineFunction &MF = DAG.getMachineFunction();
6371 SDValue Chain = Op.getOperand(0);
6372 SDValue Offset = Op.getOperand(1);
6373 SDValue Handler = Op.getOperand(2);
6374 DebugLoc dl = Op.getDebugLoc();
6376 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6378 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6380 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6381 DAG.getIntPtrConstant(-TD->getPointerSize()));
6382 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6383 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6384 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6385 MF.getRegInfo().addLiveOut(StoreAddrReg);
6387 return DAG.getNode(X86ISD::EH_RETURN, dl,
6389 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6392 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6393 SelectionDAG &DAG) {
6394 SDValue Root = Op.getOperand(0);
6395 SDValue Trmp = Op.getOperand(1); // trampoline
6396 SDValue FPtr = Op.getOperand(2); // nested function
6397 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6398 DebugLoc dl = Op.getDebugLoc();
6400 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6402 const X86InstrInfo *TII =
6403 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6405 if (Subtarget->is64Bit()) {
6406 SDValue OutChains[6];
6408 // Large code-model.
6410 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6411 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6413 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6414 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6416 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6418 // Load the pointer to the nested function into R11.
6419 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6420 SDValue Addr = Trmp;
6421 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6424 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6425 DAG.getConstant(2, MVT::i64));
6426 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6428 // Load the 'nest' parameter value into R10.
6429 // R10 is specified in X86CallingConv.td
6430 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6431 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6432 DAG.getConstant(10, MVT::i64));
6433 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6434 Addr, TrmpAddr, 10);
6436 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6437 DAG.getConstant(12, MVT::i64));
6438 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6440 // Jump to the nested function.
6441 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6442 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6443 DAG.getConstant(20, MVT::i64));
6444 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6445 Addr, TrmpAddr, 20);
6447 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6448 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6449 DAG.getConstant(22, MVT::i64));
6450 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6454 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6455 return DAG.getMergeValues(Ops, 2, dl);
6457 const Function *Func =
6458 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6459 unsigned CC = Func->getCallingConv();
6464 llvm_unreachable("Unsupported calling convention");
6465 case CallingConv::C:
6466 case CallingConv::X86_StdCall: {
6467 // Pass 'nest' parameter in ECX.
6468 // Must be kept in sync with X86CallingConv.td
6471 // Check that ECX wasn't needed by an 'inreg' parameter.
6472 const FunctionType *FTy = Func->getFunctionType();
6473 const AttrListPtr &Attrs = Func->getAttributes();
6475 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6476 unsigned InRegCount = 0;
6479 for (FunctionType::param_iterator I = FTy->param_begin(),
6480 E = FTy->param_end(); I != E; ++I, ++Idx)
6481 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6482 // FIXME: should only count parameters that are lowered to integers.
6483 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6485 if (InRegCount > 2) {
6486 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6491 case CallingConv::X86_FastCall:
6492 case CallingConv::Fast:
6493 // Pass 'nest' parameter in EAX.
6494 // Must be kept in sync with X86CallingConv.td
6499 SDValue OutChains[4];
6502 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6503 DAG.getConstant(10, MVT::i32));
6504 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6506 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6507 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6508 OutChains[0] = DAG.getStore(Root, dl,
6509 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6512 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6513 DAG.getConstant(1, MVT::i32));
6514 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6516 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6517 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6518 DAG.getConstant(5, MVT::i32));
6519 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6520 TrmpAddr, 5, false, 1);
6522 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6523 DAG.getConstant(6, MVT::i32));
6524 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6527 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6528 return DAG.getMergeValues(Ops, 2, dl);
6532 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6534 The rounding mode is in bits 11:10 of FPSR, and has the following
6541 FLT_ROUNDS, on the other hand, expects the following:
6548 To perform the conversion, we do:
6549 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6552 MachineFunction &MF = DAG.getMachineFunction();
6553 const TargetMachine &TM = MF.getTarget();
6554 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6555 unsigned StackAlignment = TFI.getStackAlignment();
6556 MVT VT = Op.getValueType();
6557 DebugLoc dl = Op.getDebugLoc();
6559 // Save FP Control Word to stack slot
6560 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6561 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6563 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6564 DAG.getEntryNode(), StackSlot);
6566 // Load FP Control Word from stack slot
6567 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6569 // Transform as necessary
6571 DAG.getNode(ISD::SRL, dl, MVT::i16,
6572 DAG.getNode(ISD::AND, dl, MVT::i16,
6573 CWD, DAG.getConstant(0x800, MVT::i16)),
6574 DAG.getConstant(11, MVT::i8));
6576 DAG.getNode(ISD::SRL, dl, MVT::i16,
6577 DAG.getNode(ISD::AND, dl, MVT::i16,
6578 CWD, DAG.getConstant(0x400, MVT::i16)),
6579 DAG.getConstant(9, MVT::i8));
6582 DAG.getNode(ISD::AND, dl, MVT::i16,
6583 DAG.getNode(ISD::ADD, dl, MVT::i16,
6584 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6585 DAG.getConstant(1, MVT::i16)),
6586 DAG.getConstant(3, MVT::i16));
6589 return DAG.getNode((VT.getSizeInBits() < 16 ?
6590 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6593 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6594 MVT VT = Op.getValueType();
6596 unsigned NumBits = VT.getSizeInBits();
6597 DebugLoc dl = Op.getDebugLoc();
6599 Op = Op.getOperand(0);
6600 if (VT == MVT::i8) {
6601 // Zero extend to i32 since there is not an i8 bsr.
6603 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6606 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6607 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6608 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6610 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6611 SmallVector<SDValue, 4> Ops;
6613 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6614 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6615 Ops.push_back(Op.getValue(1));
6616 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6618 // Finally xor with NumBits-1.
6619 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6622 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6626 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6627 MVT VT = Op.getValueType();
6629 unsigned NumBits = VT.getSizeInBits();
6630 DebugLoc dl = Op.getDebugLoc();
6632 Op = Op.getOperand(0);
6633 if (VT == MVT::i8) {
6635 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6638 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6639 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6640 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6642 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6643 SmallVector<SDValue, 4> Ops;
6645 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6646 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6647 Ops.push_back(Op.getValue(1));
6648 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6651 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6655 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6656 MVT VT = Op.getValueType();
6657 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6658 DebugLoc dl = Op.getDebugLoc();
6660 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6661 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6662 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6663 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6664 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6666 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6667 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6668 // return AloBlo + AloBhi + AhiBlo;
6670 SDValue A = Op.getOperand(0);
6671 SDValue B = Op.getOperand(1);
6673 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6674 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6675 A, DAG.getConstant(32, MVT::i32));
6676 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6677 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6678 B, DAG.getConstant(32, MVT::i32));
6679 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6680 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6682 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6683 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6685 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6686 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6688 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6689 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6690 AloBhi, DAG.getConstant(32, MVT::i32));
6691 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6692 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6693 AhiBlo, DAG.getConstant(32, MVT::i32));
6694 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6695 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6700 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6701 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6702 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6703 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6704 // has only one use.
6705 SDNode *N = Op.getNode();
6706 SDValue LHS = N->getOperand(0);
6707 SDValue RHS = N->getOperand(1);
6708 unsigned BaseOp = 0;
6710 DebugLoc dl = Op.getDebugLoc();
6712 switch (Op.getOpcode()) {
6713 default: llvm_unreachable("Unknown ovf instruction!");
6715 // A subtract of one will be selected as a INC. Note that INC doesn't
6716 // set CF, so we can't do this for UADDO.
6717 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6718 if (C->getAPIntValue() == 1) {
6719 BaseOp = X86ISD::INC;
6723 BaseOp = X86ISD::ADD;
6727 BaseOp = X86ISD::ADD;
6731 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6732 // set CF, so we can't do this for USUBO.
6733 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6734 if (C->getAPIntValue() == 1) {
6735 BaseOp = X86ISD::DEC;
6739 BaseOp = X86ISD::SUB;
6743 BaseOp = X86ISD::SUB;
6747 BaseOp = X86ISD::SMUL;
6751 BaseOp = X86ISD::UMUL;
6756 // Also sets EFLAGS.
6757 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6758 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6761 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6762 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6764 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6768 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6769 MVT T = Op.getValueType();
6770 DebugLoc dl = Op.getDebugLoc();
6773 switch(T.getSimpleVT()) {
6775 assert(false && "Invalid value type!");
6776 case MVT::i8: Reg = X86::AL; size = 1; break;
6777 case MVT::i16: Reg = X86::AX; size = 2; break;
6778 case MVT::i32: Reg = X86::EAX; size = 4; break;
6780 assert(Subtarget->is64Bit() && "Node not type legal!");
6781 Reg = X86::RAX; size = 8;
6784 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6785 Op.getOperand(2), SDValue());
6786 SDValue Ops[] = { cpIn.getValue(0),
6789 DAG.getTargetConstant(size, MVT::i8),
6791 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6792 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6794 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6798 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6799 SelectionDAG &DAG) {
6800 assert(Subtarget->is64Bit() && "Result not type legalized?");
6801 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6802 SDValue TheChain = Op.getOperand(0);
6803 DebugLoc dl = Op.getDebugLoc();
6804 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6805 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6806 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6808 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6809 DAG.getConstant(32, MVT::i8));
6811 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6814 return DAG.getMergeValues(Ops, 2, dl);
6817 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6818 SDNode *Node = Op.getNode();
6819 DebugLoc dl = Node->getDebugLoc();
6820 MVT T = Node->getValueType(0);
6821 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6822 DAG.getConstant(0, T), Node->getOperand(2));
6823 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6824 cast<AtomicSDNode>(Node)->getMemoryVT(),
6825 Node->getOperand(0),
6826 Node->getOperand(1), negOp,
6827 cast<AtomicSDNode>(Node)->getSrcValue(),
6828 cast<AtomicSDNode>(Node)->getAlignment());
6831 /// LowerOperation - Provide custom lowering hooks for some operations.
6833 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6834 switch (Op.getOpcode()) {
6835 default: llvm_unreachable("Should not custom lower this!");
6836 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6837 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6838 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6839 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6840 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6841 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6842 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6843 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6844 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6845 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6846 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6847 case ISD::SHL_PARTS:
6848 case ISD::SRA_PARTS:
6849 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6850 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6851 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6852 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6853 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
6854 case ISD::FABS: return LowerFABS(Op, DAG);
6855 case ISD::FNEG: return LowerFNEG(Op, DAG);
6856 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6857 case ISD::SETCC: return LowerSETCC(Op, DAG);
6858 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6859 case ISD::SELECT: return LowerSELECT(Op, DAG);
6860 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6861 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6862 case ISD::CALL: return LowerCALL(Op, DAG);
6863 case ISD::RET: return LowerRET(Op, DAG);
6864 case ISD::FORMAL_ARGUMENTS: return LowerFORMAL_ARGUMENTS(Op, DAG);
6865 case ISD::VASTART: return LowerVASTART(Op, DAG);
6866 case ISD::VAARG: return LowerVAARG(Op, DAG);
6867 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6868 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6869 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6870 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6871 case ISD::FRAME_TO_ARGS_OFFSET:
6872 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6873 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6874 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6875 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6876 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6877 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6878 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6879 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6885 case ISD::UMULO: return LowerXALUO(Op, DAG);
6886 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6890 void X86TargetLowering::
6891 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6892 SelectionDAG &DAG, unsigned NewOp) {
6893 MVT T = Node->getValueType(0);
6894 DebugLoc dl = Node->getDebugLoc();
6895 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6897 SDValue Chain = Node->getOperand(0);
6898 SDValue In1 = Node->getOperand(1);
6899 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6900 Node->getOperand(2), DAG.getIntPtrConstant(0));
6901 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6902 Node->getOperand(2), DAG.getIntPtrConstant(1));
6903 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6904 // have a MemOperand. Pass the info through as a normal operand.
6905 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6906 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6907 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6908 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
6909 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6910 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6911 Results.push_back(Result.getValue(2));
6914 /// ReplaceNodeResults - Replace a node with an illegal result type
6915 /// with a new node built out of custom code.
6916 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6917 SmallVectorImpl<SDValue>&Results,
6918 SelectionDAG &DAG) {
6919 DebugLoc dl = N->getDebugLoc();
6920 switch (N->getOpcode()) {
6922 assert(false && "Do not know how to custom type legalize this operation!");
6924 case ISD::FP_TO_SINT: {
6925 std::pair<SDValue,SDValue> Vals =
6926 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
6927 SDValue FIST = Vals.first, StackSlot = Vals.second;
6928 if (FIST.getNode() != 0) {
6929 MVT VT = N->getValueType(0);
6930 // Return a load from the stack slot.
6931 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
6935 case ISD::READCYCLECOUNTER: {
6936 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6937 SDValue TheChain = N->getOperand(0);
6938 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6939 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
6941 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
6943 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
6944 SDValue Ops[] = { eax, edx };
6945 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
6946 Results.push_back(edx.getValue(1));
6949 case ISD::ATOMIC_CMP_SWAP: {
6950 MVT T = N->getValueType(0);
6951 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
6952 SDValue cpInL, cpInH;
6953 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6954 DAG.getConstant(0, MVT::i32));
6955 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6956 DAG.getConstant(1, MVT::i32));
6957 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
6958 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
6960 SDValue swapInL, swapInH;
6961 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6962 DAG.getConstant(0, MVT::i32));
6963 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6964 DAG.getConstant(1, MVT::i32));
6965 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
6967 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
6968 swapInL.getValue(1));
6969 SDValue Ops[] = { swapInH.getValue(0),
6971 swapInH.getValue(1) };
6972 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6973 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
6974 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
6975 MVT::i32, Result.getValue(1));
6976 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
6977 MVT::i32, cpOutL.getValue(2));
6978 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6979 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6980 Results.push_back(cpOutH.getValue(1));
6983 case ISD::ATOMIC_LOAD_ADD:
6984 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
6986 case ISD::ATOMIC_LOAD_AND:
6987 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
6989 case ISD::ATOMIC_LOAD_NAND:
6990 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
6992 case ISD::ATOMIC_LOAD_OR:
6993 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
6995 case ISD::ATOMIC_LOAD_SUB:
6996 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
6998 case ISD::ATOMIC_LOAD_XOR:
6999 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7001 case ISD::ATOMIC_SWAP:
7002 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7007 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7009 default: return NULL;
7010 case X86ISD::BSF: return "X86ISD::BSF";
7011 case X86ISD::BSR: return "X86ISD::BSR";
7012 case X86ISD::SHLD: return "X86ISD::SHLD";
7013 case X86ISD::SHRD: return "X86ISD::SHRD";
7014 case X86ISD::FAND: return "X86ISD::FAND";
7015 case X86ISD::FOR: return "X86ISD::FOR";
7016 case X86ISD::FXOR: return "X86ISD::FXOR";
7017 case X86ISD::FSRL: return "X86ISD::FSRL";
7018 case X86ISD::FILD: return "X86ISD::FILD";
7019 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7020 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7021 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7022 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7023 case X86ISD::FLD: return "X86ISD::FLD";
7024 case X86ISD::FST: return "X86ISD::FST";
7025 case X86ISD::CALL: return "X86ISD::CALL";
7026 case X86ISD::TAILCALL: return "X86ISD::TAILCALL";
7027 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7028 case X86ISD::BT: return "X86ISD::BT";
7029 case X86ISD::CMP: return "X86ISD::CMP";
7030 case X86ISD::COMI: return "X86ISD::COMI";
7031 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7032 case X86ISD::SETCC: return "X86ISD::SETCC";
7033 case X86ISD::CMOV: return "X86ISD::CMOV";
7034 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7035 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7036 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7037 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7038 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7039 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7040 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7041 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7042 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7043 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7044 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7045 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7046 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7047 case X86ISD::FMAX: return "X86ISD::FMAX";
7048 case X86ISD::FMIN: return "X86ISD::FMIN";
7049 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7050 case X86ISD::FRCP: return "X86ISD::FRCP";
7051 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7052 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7053 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7054 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7055 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7056 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7057 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7058 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7059 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7060 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7061 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7062 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7063 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7064 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7065 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7066 case X86ISD::VSHL: return "X86ISD::VSHL";
7067 case X86ISD::VSRL: return "X86ISD::VSRL";
7068 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7069 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7070 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7071 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7072 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7073 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7074 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7075 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7076 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7077 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7078 case X86ISD::ADD: return "X86ISD::ADD";
7079 case X86ISD::SUB: return "X86ISD::SUB";
7080 case X86ISD::SMUL: return "X86ISD::SMUL";
7081 case X86ISD::UMUL: return "X86ISD::UMUL";
7082 case X86ISD::INC: return "X86ISD::INC";
7083 case X86ISD::DEC: return "X86ISD::DEC";
7084 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7085 case X86ISD::PTEST: return "X86ISD::PTEST";
7089 // isLegalAddressingMode - Return true if the addressing mode represented
7090 // by AM is legal for this target, for a load/store of the specified type.
7091 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7092 const Type *Ty) const {
7093 // X86 supports extremely general addressing modes.
7095 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7096 if (AM.BaseOffs <= -(1LL << 32) || AM.BaseOffs >= (1LL << 32)-1)
7101 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7103 // If a reference to this global requires an extra load, we can't fold it.
7104 if (isGlobalStubReference(GVFlags))
7107 // If BaseGV requires a register for the PIC base, we cannot also have a
7108 // BaseReg specified.
7109 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7112 // X86-64 only supports addr of globals in small code model.
7113 if (Subtarget->is64Bit()) {
7114 if (getTargetMachine().getCodeModel() != CodeModel::Small)
7116 // If lower 4G is not available, then we must use rip-relative addressing.
7117 if (AM.BaseOffs || AM.Scale > 1)
7128 // These scales always work.
7133 // These scales are formed with basereg+scalereg. Only accept if there is
7138 default: // Other stuff never works.
7146 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7147 if (!Ty1->isInteger() || !Ty2->isInteger())
7149 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7150 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7151 if (NumBits1 <= NumBits2)
7153 return Subtarget->is64Bit() || NumBits1 < 64;
7156 bool X86TargetLowering::isTruncateFree(MVT VT1, MVT VT2) const {
7157 if (!VT1.isInteger() || !VT2.isInteger())
7159 unsigned NumBits1 = VT1.getSizeInBits();
7160 unsigned NumBits2 = VT2.getSizeInBits();
7161 if (NumBits1 <= NumBits2)
7163 return Subtarget->is64Bit() || NumBits1 < 64;
7166 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7167 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7168 return Ty1 == Type::Int32Ty && Ty2 == Type::Int64Ty && Subtarget->is64Bit();
7171 bool X86TargetLowering::isZExtFree(MVT VT1, MVT VT2) const {
7172 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7173 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7176 bool X86TargetLowering::isNarrowingProfitable(MVT VT1, MVT VT2) const {
7177 // i16 instructions are longer (0x66 prefix) and potentially slower.
7178 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7181 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7182 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7183 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7184 /// are assumed to be legal.
7186 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7188 // Only do shuffles on 128-bit vector types for now.
7189 if (VT.getSizeInBits() == 64)
7192 // FIXME: pshufb, blends, palignr, shifts.
7193 return (VT.getVectorNumElements() == 2 ||
7194 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7195 isMOVLMask(M, VT) ||
7196 isSHUFPMask(M, VT) ||
7197 isPSHUFDMask(M, VT) ||
7198 isPSHUFHWMask(M, VT) ||
7199 isPSHUFLWMask(M, VT) ||
7200 isUNPCKLMask(M, VT) ||
7201 isUNPCKHMask(M, VT) ||
7202 isUNPCKL_v_undef_Mask(M, VT) ||
7203 isUNPCKH_v_undef_Mask(M, VT));
7207 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7209 unsigned NumElts = VT.getVectorNumElements();
7210 // FIXME: This collection of masks seems suspect.
7213 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7214 return (isMOVLMask(Mask, VT) ||
7215 isCommutedMOVLMask(Mask, VT, true) ||
7216 isSHUFPMask(Mask, VT) ||
7217 isCommutedSHUFPMask(Mask, VT));
7222 //===----------------------------------------------------------------------===//
7223 // X86 Scheduler Hooks
7224 //===----------------------------------------------------------------------===//
7226 // private utility function
7228 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7229 MachineBasicBlock *MBB,
7237 TargetRegisterClass *RC,
7238 bool invSrc) const {
7239 // For the atomic bitwise operator, we generate
7242 // ld t1 = [bitinstr.addr]
7243 // op t2 = t1, [bitinstr.val]
7245 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7247 // fallthrough -->nextMBB
7248 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7249 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7250 MachineFunction::iterator MBBIter = MBB;
7253 /// First build the CFG
7254 MachineFunction *F = MBB->getParent();
7255 MachineBasicBlock *thisMBB = MBB;
7256 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7257 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7258 F->insert(MBBIter, newMBB);
7259 F->insert(MBBIter, nextMBB);
7261 // Move all successors to thisMBB to nextMBB
7262 nextMBB->transferSuccessors(thisMBB);
7264 // Update thisMBB to fall through to newMBB
7265 thisMBB->addSuccessor(newMBB);
7267 // newMBB jumps to itself and fall through to nextMBB
7268 newMBB->addSuccessor(nextMBB);
7269 newMBB->addSuccessor(newMBB);
7271 // Insert instructions into newMBB based on incoming instruction
7272 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7273 "unexpected number of operands");
7274 DebugLoc dl = bInstr->getDebugLoc();
7275 MachineOperand& destOper = bInstr->getOperand(0);
7276 MachineOperand* argOpers[2 + X86AddrNumOperands];
7277 int numArgs = bInstr->getNumOperands() - 1;
7278 for (int i=0; i < numArgs; ++i)
7279 argOpers[i] = &bInstr->getOperand(i+1);
7281 // x86 address has 4 operands: base, index, scale, and displacement
7282 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7283 int valArgIndx = lastAddrIndx + 1;
7285 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7286 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7287 for (int i=0; i <= lastAddrIndx; ++i)
7288 (*MIB).addOperand(*argOpers[i]);
7290 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7292 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7297 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7298 assert((argOpers[valArgIndx]->isReg() ||
7299 argOpers[valArgIndx]->isImm()) &&
7301 if (argOpers[valArgIndx]->isReg())
7302 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7304 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7306 (*MIB).addOperand(*argOpers[valArgIndx]);
7308 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7311 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7312 for (int i=0; i <= lastAddrIndx; ++i)
7313 (*MIB).addOperand(*argOpers[i]);
7315 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7316 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7318 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7322 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7324 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7328 // private utility function: 64 bit atomics on 32 bit host.
7330 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7331 MachineBasicBlock *MBB,
7336 bool invSrc) const {
7337 // For the atomic bitwise operator, we generate
7338 // thisMBB (instructions are in pairs, except cmpxchg8b)
7339 // ld t1,t2 = [bitinstr.addr]
7341 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7342 // op t5, t6 <- out1, out2, [bitinstr.val]
7343 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7344 // mov ECX, EBX <- t5, t6
7345 // mov EAX, EDX <- t1, t2
7346 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7347 // mov t3, t4 <- EAX, EDX
7349 // result in out1, out2
7350 // fallthrough -->nextMBB
7352 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7353 const unsigned LoadOpc = X86::MOV32rm;
7354 const unsigned copyOpc = X86::MOV32rr;
7355 const unsigned NotOpc = X86::NOT32r;
7356 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7357 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7358 MachineFunction::iterator MBBIter = MBB;
7361 /// First build the CFG
7362 MachineFunction *F = MBB->getParent();
7363 MachineBasicBlock *thisMBB = MBB;
7364 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7365 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7366 F->insert(MBBIter, newMBB);
7367 F->insert(MBBIter, nextMBB);
7369 // Move all successors to thisMBB to nextMBB
7370 nextMBB->transferSuccessors(thisMBB);
7372 // Update thisMBB to fall through to newMBB
7373 thisMBB->addSuccessor(newMBB);
7375 // newMBB jumps to itself and fall through to nextMBB
7376 newMBB->addSuccessor(nextMBB);
7377 newMBB->addSuccessor(newMBB);
7379 DebugLoc dl = bInstr->getDebugLoc();
7380 // Insert instructions into newMBB based on incoming instruction
7381 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7382 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7383 "unexpected number of operands");
7384 MachineOperand& dest1Oper = bInstr->getOperand(0);
7385 MachineOperand& dest2Oper = bInstr->getOperand(1);
7386 MachineOperand* argOpers[2 + X86AddrNumOperands];
7387 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7388 argOpers[i] = &bInstr->getOperand(i+2);
7390 // x86 address has 4 operands: base, index, scale, and displacement
7391 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7393 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7394 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7395 for (int i=0; i <= lastAddrIndx; ++i)
7396 (*MIB).addOperand(*argOpers[i]);
7397 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7398 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7399 // add 4 to displacement.
7400 for (int i=0; i <= lastAddrIndx-2; ++i)
7401 (*MIB).addOperand(*argOpers[i]);
7402 MachineOperand newOp3 = *(argOpers[3]);
7404 newOp3.setImm(newOp3.getImm()+4);
7406 newOp3.setOffset(newOp3.getOffset()+4);
7407 (*MIB).addOperand(newOp3);
7408 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7410 // t3/4 are defined later, at the bottom of the loop
7411 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7412 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7413 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7414 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7415 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7416 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7418 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7419 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7421 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7422 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7428 int valArgIndx = lastAddrIndx + 1;
7429 assert((argOpers[valArgIndx]->isReg() ||
7430 argOpers[valArgIndx]->isImm()) &&
7432 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7433 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7434 if (argOpers[valArgIndx]->isReg())
7435 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7437 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7438 if (regOpcL != X86::MOV32rr)
7440 (*MIB).addOperand(*argOpers[valArgIndx]);
7441 assert(argOpers[valArgIndx + 1]->isReg() ==
7442 argOpers[valArgIndx]->isReg());
7443 assert(argOpers[valArgIndx + 1]->isImm() ==
7444 argOpers[valArgIndx]->isImm());
7445 if (argOpers[valArgIndx + 1]->isReg())
7446 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7448 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7449 if (regOpcH != X86::MOV32rr)
7451 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7453 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7455 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7458 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7460 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7463 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7464 for (int i=0; i <= lastAddrIndx; ++i)
7465 (*MIB).addOperand(*argOpers[i]);
7467 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7468 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7470 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7471 MIB.addReg(X86::EAX);
7472 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7473 MIB.addReg(X86::EDX);
7476 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7478 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7482 // private utility function
7484 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7485 MachineBasicBlock *MBB,
7486 unsigned cmovOpc) const {
7487 // For the atomic min/max operator, we generate
7490 // ld t1 = [min/max.addr]
7491 // mov t2 = [min/max.val]
7493 // cmov[cond] t2 = t1
7495 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7497 // fallthrough -->nextMBB
7499 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7500 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7501 MachineFunction::iterator MBBIter = MBB;
7504 /// First build the CFG
7505 MachineFunction *F = MBB->getParent();
7506 MachineBasicBlock *thisMBB = MBB;
7507 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7508 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7509 F->insert(MBBIter, newMBB);
7510 F->insert(MBBIter, nextMBB);
7512 // Move all successors to thisMBB to nextMBB
7513 nextMBB->transferSuccessors(thisMBB);
7515 // Update thisMBB to fall through to newMBB
7516 thisMBB->addSuccessor(newMBB);
7518 // newMBB jumps to newMBB and fall through to nextMBB
7519 newMBB->addSuccessor(nextMBB);
7520 newMBB->addSuccessor(newMBB);
7522 DebugLoc dl = mInstr->getDebugLoc();
7523 // Insert instructions into newMBB based on incoming instruction
7524 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7525 "unexpected number of operands");
7526 MachineOperand& destOper = mInstr->getOperand(0);
7527 MachineOperand* argOpers[2 + X86AddrNumOperands];
7528 int numArgs = mInstr->getNumOperands() - 1;
7529 for (int i=0; i < numArgs; ++i)
7530 argOpers[i] = &mInstr->getOperand(i+1);
7532 // x86 address has 4 operands: base, index, scale, and displacement
7533 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7534 int valArgIndx = lastAddrIndx + 1;
7536 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7537 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7538 for (int i=0; i <= lastAddrIndx; ++i)
7539 (*MIB).addOperand(*argOpers[i]);
7541 // We only support register and immediate values
7542 assert((argOpers[valArgIndx]->isReg() ||
7543 argOpers[valArgIndx]->isImm()) &&
7546 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7547 if (argOpers[valArgIndx]->isReg())
7548 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7550 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7551 (*MIB).addOperand(*argOpers[valArgIndx]);
7553 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7556 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7561 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7562 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7566 // Cmp and exchange if none has modified the memory location
7567 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7568 for (int i=0; i <= lastAddrIndx; ++i)
7569 (*MIB).addOperand(*argOpers[i]);
7571 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7572 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7574 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7575 MIB.addReg(X86::EAX);
7578 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7580 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7586 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7587 MachineBasicBlock *BB) const {
7588 DebugLoc dl = MI->getDebugLoc();
7589 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7590 switch (MI->getOpcode()) {
7591 default: assert(false && "Unexpected instr type to insert");
7592 case X86::CMOV_V1I64:
7593 case X86::CMOV_FR32:
7594 case X86::CMOV_FR64:
7595 case X86::CMOV_V4F32:
7596 case X86::CMOV_V2F64:
7597 case X86::CMOV_V2I64: {
7598 // To "insert" a SELECT_CC instruction, we actually have to insert the
7599 // diamond control-flow pattern. The incoming instruction knows the
7600 // destination vreg to set, the condition code register to branch on, the
7601 // true/false values to select between, and a branch opcode to use.
7602 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7603 MachineFunction::iterator It = BB;
7609 // cmpTY ccX, r1, r2
7611 // fallthrough --> copy0MBB
7612 MachineBasicBlock *thisMBB = BB;
7613 MachineFunction *F = BB->getParent();
7614 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7615 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7617 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7618 BuildMI(BB, dl, TII->get(Opc)).addMBB(sinkMBB);
7619 F->insert(It, copy0MBB);
7620 F->insert(It, sinkMBB);
7621 // Update machine-CFG edges by transferring all successors of the current
7622 // block to the new block which will contain the Phi node for the select.
7623 sinkMBB->transferSuccessors(BB);
7625 // Add the true and fallthrough blocks as its successors.
7626 BB->addSuccessor(copy0MBB);
7627 BB->addSuccessor(sinkMBB);
7630 // %FalseValue = ...
7631 // # fallthrough to sinkMBB
7634 // Update machine-CFG edges
7635 BB->addSuccessor(sinkMBB);
7638 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7641 BuildMI(BB, dl, TII->get(X86::PHI), MI->getOperand(0).getReg())
7642 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7643 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7645 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7649 case X86::FP32_TO_INT16_IN_MEM:
7650 case X86::FP32_TO_INT32_IN_MEM:
7651 case X86::FP32_TO_INT64_IN_MEM:
7652 case X86::FP64_TO_INT16_IN_MEM:
7653 case X86::FP64_TO_INT32_IN_MEM:
7654 case X86::FP64_TO_INT64_IN_MEM:
7655 case X86::FP80_TO_INT16_IN_MEM:
7656 case X86::FP80_TO_INT32_IN_MEM:
7657 case X86::FP80_TO_INT64_IN_MEM: {
7658 // Change the floating point control register to use "round towards zero"
7659 // mode when truncating to an integer value.
7660 MachineFunction *F = BB->getParent();
7661 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7662 addFrameReference(BuildMI(BB, dl, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7664 // Load the old value of the high byte of the control word...
7666 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7667 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16rm), OldCW),
7670 // Set the high part to be round to zero...
7671 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mi)), CWFrameIdx)
7674 // Reload the modified control word now...
7675 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7677 // Restore the memory image of control word to original value
7678 addFrameReference(BuildMI(BB, dl, TII->get(X86::MOV16mr)), CWFrameIdx)
7681 // Get the X86 opcode to use.
7683 switch (MI->getOpcode()) {
7684 default: llvm_unreachable("illegal opcode!");
7685 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7686 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7687 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7688 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7689 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7690 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7691 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7692 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7693 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7697 MachineOperand &Op = MI->getOperand(0);
7699 AM.BaseType = X86AddressMode::RegBase;
7700 AM.Base.Reg = Op.getReg();
7702 AM.BaseType = X86AddressMode::FrameIndexBase;
7703 AM.Base.FrameIndex = Op.getIndex();
7705 Op = MI->getOperand(1);
7707 AM.Scale = Op.getImm();
7708 Op = MI->getOperand(2);
7710 AM.IndexReg = Op.getImm();
7711 Op = MI->getOperand(3);
7712 if (Op.isGlobal()) {
7713 AM.GV = Op.getGlobal();
7715 AM.Disp = Op.getImm();
7717 addFullAddress(BuildMI(BB, dl, TII->get(Opc)), AM)
7718 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7720 // Reload the original control word now.
7721 addFrameReference(BuildMI(BB, dl, TII->get(X86::FLDCW16m)), CWFrameIdx);
7723 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7726 case X86::ATOMAND32:
7727 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7728 X86::AND32ri, X86::MOV32rm,
7729 X86::LCMPXCHG32, X86::MOV32rr,
7730 X86::NOT32r, X86::EAX,
7731 X86::GR32RegisterClass);
7733 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7734 X86::OR32ri, X86::MOV32rm,
7735 X86::LCMPXCHG32, X86::MOV32rr,
7736 X86::NOT32r, X86::EAX,
7737 X86::GR32RegisterClass);
7738 case X86::ATOMXOR32:
7739 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7740 X86::XOR32ri, X86::MOV32rm,
7741 X86::LCMPXCHG32, X86::MOV32rr,
7742 X86::NOT32r, X86::EAX,
7743 X86::GR32RegisterClass);
7744 case X86::ATOMNAND32:
7745 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7746 X86::AND32ri, X86::MOV32rm,
7747 X86::LCMPXCHG32, X86::MOV32rr,
7748 X86::NOT32r, X86::EAX,
7749 X86::GR32RegisterClass, true);
7750 case X86::ATOMMIN32:
7751 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7752 case X86::ATOMMAX32:
7753 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7754 case X86::ATOMUMIN32:
7755 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7756 case X86::ATOMUMAX32:
7757 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7759 case X86::ATOMAND16:
7760 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7761 X86::AND16ri, X86::MOV16rm,
7762 X86::LCMPXCHG16, X86::MOV16rr,
7763 X86::NOT16r, X86::AX,
7764 X86::GR16RegisterClass);
7766 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7767 X86::OR16ri, X86::MOV16rm,
7768 X86::LCMPXCHG16, X86::MOV16rr,
7769 X86::NOT16r, X86::AX,
7770 X86::GR16RegisterClass);
7771 case X86::ATOMXOR16:
7772 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7773 X86::XOR16ri, X86::MOV16rm,
7774 X86::LCMPXCHG16, X86::MOV16rr,
7775 X86::NOT16r, X86::AX,
7776 X86::GR16RegisterClass);
7777 case X86::ATOMNAND16:
7778 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7779 X86::AND16ri, X86::MOV16rm,
7780 X86::LCMPXCHG16, X86::MOV16rr,
7781 X86::NOT16r, X86::AX,
7782 X86::GR16RegisterClass, true);
7783 case X86::ATOMMIN16:
7784 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7785 case X86::ATOMMAX16:
7786 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7787 case X86::ATOMUMIN16:
7788 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7789 case X86::ATOMUMAX16:
7790 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7793 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7794 X86::AND8ri, X86::MOV8rm,
7795 X86::LCMPXCHG8, X86::MOV8rr,
7796 X86::NOT8r, X86::AL,
7797 X86::GR8RegisterClass);
7799 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7800 X86::OR8ri, X86::MOV8rm,
7801 X86::LCMPXCHG8, X86::MOV8rr,
7802 X86::NOT8r, X86::AL,
7803 X86::GR8RegisterClass);
7805 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7806 X86::XOR8ri, X86::MOV8rm,
7807 X86::LCMPXCHG8, X86::MOV8rr,
7808 X86::NOT8r, X86::AL,
7809 X86::GR8RegisterClass);
7810 case X86::ATOMNAND8:
7811 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7812 X86::AND8ri, X86::MOV8rm,
7813 X86::LCMPXCHG8, X86::MOV8rr,
7814 X86::NOT8r, X86::AL,
7815 X86::GR8RegisterClass, true);
7816 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7817 // This group is for 64-bit host.
7818 case X86::ATOMAND64:
7819 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7820 X86::AND64ri32, X86::MOV64rm,
7821 X86::LCMPXCHG64, X86::MOV64rr,
7822 X86::NOT64r, X86::RAX,
7823 X86::GR64RegisterClass);
7825 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7826 X86::OR64ri32, X86::MOV64rm,
7827 X86::LCMPXCHG64, X86::MOV64rr,
7828 X86::NOT64r, X86::RAX,
7829 X86::GR64RegisterClass);
7830 case X86::ATOMXOR64:
7831 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7832 X86::XOR64ri32, X86::MOV64rm,
7833 X86::LCMPXCHG64, X86::MOV64rr,
7834 X86::NOT64r, X86::RAX,
7835 X86::GR64RegisterClass);
7836 case X86::ATOMNAND64:
7837 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7838 X86::AND64ri32, X86::MOV64rm,
7839 X86::LCMPXCHG64, X86::MOV64rr,
7840 X86::NOT64r, X86::RAX,
7841 X86::GR64RegisterClass, true);
7842 case X86::ATOMMIN64:
7843 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7844 case X86::ATOMMAX64:
7845 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7846 case X86::ATOMUMIN64:
7847 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7848 case X86::ATOMUMAX64:
7849 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7851 // This group does 64-bit operations on a 32-bit host.
7852 case X86::ATOMAND6432:
7853 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7854 X86::AND32rr, X86::AND32rr,
7855 X86::AND32ri, X86::AND32ri,
7857 case X86::ATOMOR6432:
7858 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7859 X86::OR32rr, X86::OR32rr,
7860 X86::OR32ri, X86::OR32ri,
7862 case X86::ATOMXOR6432:
7863 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7864 X86::XOR32rr, X86::XOR32rr,
7865 X86::XOR32ri, X86::XOR32ri,
7867 case X86::ATOMNAND6432:
7868 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7869 X86::AND32rr, X86::AND32rr,
7870 X86::AND32ri, X86::AND32ri,
7872 case X86::ATOMADD6432:
7873 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7874 X86::ADD32rr, X86::ADC32rr,
7875 X86::ADD32ri, X86::ADC32ri,
7877 case X86::ATOMSUB6432:
7878 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7879 X86::SUB32rr, X86::SBB32rr,
7880 X86::SUB32ri, X86::SBB32ri,
7882 case X86::ATOMSWAP6432:
7883 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7884 X86::MOV32rr, X86::MOV32rr,
7885 X86::MOV32ri, X86::MOV32ri,
7890 //===----------------------------------------------------------------------===//
7891 // X86 Optimization Hooks
7892 //===----------------------------------------------------------------------===//
7894 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
7898 const SelectionDAG &DAG,
7899 unsigned Depth) const {
7900 unsigned Opc = Op.getOpcode();
7901 assert((Opc >= ISD::BUILTIN_OP_END ||
7902 Opc == ISD::INTRINSIC_WO_CHAIN ||
7903 Opc == ISD::INTRINSIC_W_CHAIN ||
7904 Opc == ISD::INTRINSIC_VOID) &&
7905 "Should use MaskedValueIsZero if you don't know whether Op"
7906 " is a target node!");
7908 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
7917 // These nodes' second result is a boolean.
7918 if (Op.getResNo() == 0)
7922 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
7923 Mask.getBitWidth() - 1);
7928 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
7929 /// node is a GlobalAddress + offset.
7930 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
7931 GlobalValue* &GA, int64_t &Offset) const{
7932 if (N->getOpcode() == X86ISD::Wrapper) {
7933 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
7934 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
7935 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
7939 return TargetLowering::isGAPlusOffset(N, GA, Offset);
7942 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
7943 const TargetLowering &TLI) {
7946 if (TLI.isGAPlusOffset(Base, GV, Offset))
7947 return (GV->getAlignment() >= N && (Offset % N) == 0);
7948 // DAG combine handles the stack object case.
7952 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
7953 MVT EVT, LoadSDNode *&LDBase,
7954 unsigned &LastLoadedElt,
7955 SelectionDAG &DAG, MachineFrameInfo *MFI,
7956 const TargetLowering &TLI) {
7958 LastLoadedElt = -1U;
7959 for (unsigned i = 0; i < NumElems; ++i) {
7960 if (N->getMaskElt(i) < 0) {
7966 SDValue Elt = DAG.getShuffleScalarElt(N, i);
7967 if (!Elt.getNode() ||
7968 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
7971 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
7973 LDBase = cast<LoadSDNode>(Elt.getNode());
7977 if (Elt.getOpcode() == ISD::UNDEF)
7980 LoadSDNode *LD = cast<LoadSDNode>(Elt);
7981 if (!TLI.isConsecutiveLoad(LD, LDBase, EVT.getSizeInBits()/8, i, MFI))
7988 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
7989 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
7990 /// if the load addresses are consecutive, non-overlapping, and in the right
7991 /// order. In the case of v2i64, it will see if it can rewrite the
7992 /// shuffle to be an appropriate build vector so it can take advantage of
7993 // performBuildVectorCombine.
7994 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
7995 const TargetLowering &TLI) {
7996 DebugLoc dl = N->getDebugLoc();
7997 MVT VT = N->getValueType(0);
7998 MVT EVT = VT.getVectorElementType();
7999 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8000 unsigned NumElems = VT.getVectorNumElements();
8002 if (VT.getSizeInBits() != 128)
8005 // Try to combine a vector_shuffle into a 128-bit load.
8006 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8007 LoadSDNode *LD = NULL;
8008 unsigned LastLoadedElt;
8009 if (!EltsFromConsecutiveLoads(SVN, NumElems, EVT, LD, LastLoadedElt, DAG,
8013 if (LastLoadedElt == NumElems - 1) {
8014 if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
8015 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8016 LD->getSrcValue(), LD->getSrcValueOffset(),
8018 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8019 LD->getSrcValue(), LD->getSrcValueOffset(),
8020 LD->isVolatile(), LD->getAlignment());
8021 } else if (NumElems == 4 && LastLoadedElt == 1) {
8022 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8023 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8024 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8025 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8030 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8031 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8032 const X86Subtarget *Subtarget) {
8033 DebugLoc DL = N->getDebugLoc();
8034 SDValue Cond = N->getOperand(0);
8035 // Get the LHS/RHS of the select.
8036 SDValue LHS = N->getOperand(1);
8037 SDValue RHS = N->getOperand(2);
8039 // If we have SSE[12] support, try to form min/max nodes.
8040 if (Subtarget->hasSSE2() &&
8041 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8042 Cond.getOpcode() == ISD::SETCC) {
8043 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8045 unsigned Opcode = 0;
8046 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8049 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
8052 if (!UnsafeFPMath) break;
8054 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
8056 Opcode = X86ISD::FMIN;
8059 case ISD::SETOGT: // (X > Y) ? X : Y -> max
8062 if (!UnsafeFPMath) break;
8064 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
8066 Opcode = X86ISD::FMAX;
8069 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8072 case ISD::SETOGT: // (X > Y) ? Y : X -> min
8075 if (!UnsafeFPMath) break;
8077 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
8079 Opcode = X86ISD::FMIN;
8082 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
8085 if (!UnsafeFPMath) break;
8087 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
8089 Opcode = X86ISD::FMAX;
8095 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8098 // If this is a select between two integer constants, try to do some
8100 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8101 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8102 // Don't do this for crazy integer types.
8103 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8104 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8105 // so that TrueC (the true value) is larger than FalseC.
8106 bool NeedsCondInvert = false;
8108 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8109 // Efficiently invertible.
8110 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8111 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8112 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8113 NeedsCondInvert = true;
8114 std::swap(TrueC, FalseC);
8117 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8118 if (FalseC->getAPIntValue() == 0 &&
8119 TrueC->getAPIntValue().isPowerOf2()) {
8120 if (NeedsCondInvert) // Invert the condition if needed.
8121 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8122 DAG.getConstant(1, Cond.getValueType()));
8124 // Zero extend the condition if needed.
8125 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8127 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8128 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8129 DAG.getConstant(ShAmt, MVT::i8));
8132 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8133 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8134 if (NeedsCondInvert) // Invert the condition if needed.
8135 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8136 DAG.getConstant(1, Cond.getValueType()));
8138 // Zero extend the condition if needed.
8139 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8140 FalseC->getValueType(0), Cond);
8141 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8142 SDValue(FalseC, 0));
8145 // Optimize cases that will turn into an LEA instruction. This requires
8146 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8147 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8148 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8149 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8151 bool isFastMultiplier = false;
8153 switch ((unsigned char)Diff) {
8155 case 1: // result = add base, cond
8156 case 2: // result = lea base( , cond*2)
8157 case 3: // result = lea base(cond, cond*2)
8158 case 4: // result = lea base( , cond*4)
8159 case 5: // result = lea base(cond, cond*4)
8160 case 8: // result = lea base( , cond*8)
8161 case 9: // result = lea base(cond, cond*8)
8162 isFastMultiplier = true;
8167 if (isFastMultiplier) {
8168 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8169 if (NeedsCondInvert) // Invert the condition if needed.
8170 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8171 DAG.getConstant(1, Cond.getValueType()));
8173 // Zero extend the condition if needed.
8174 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8176 // Scale the condition by the difference.
8178 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8179 DAG.getConstant(Diff, Cond.getValueType()));
8181 // Add the base if non-zero.
8182 if (FalseC->getAPIntValue() != 0)
8183 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8184 SDValue(FalseC, 0));
8194 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8195 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8196 TargetLowering::DAGCombinerInfo &DCI) {
8197 DebugLoc DL = N->getDebugLoc();
8199 // If the flag operand isn't dead, don't touch this CMOV.
8200 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8203 // If this is a select between two integer constants, try to do some
8204 // optimizations. Note that the operands are ordered the opposite of SELECT
8206 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8207 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8208 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8209 // larger than FalseC (the false value).
8210 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8212 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8213 CC = X86::GetOppositeBranchCondition(CC);
8214 std::swap(TrueC, FalseC);
8217 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8218 // This is efficient for any integer data type (including i8/i16) and
8220 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8221 SDValue Cond = N->getOperand(3);
8222 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8223 DAG.getConstant(CC, MVT::i8), Cond);
8225 // Zero extend the condition if needed.
8226 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8228 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8229 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8230 DAG.getConstant(ShAmt, MVT::i8));
8231 if (N->getNumValues() == 2) // Dead flag value?
8232 return DCI.CombineTo(N, Cond, SDValue());
8236 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8237 // for any integer data type, including i8/i16.
8238 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8239 SDValue Cond = N->getOperand(3);
8240 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8241 DAG.getConstant(CC, MVT::i8), Cond);
8243 // Zero extend the condition if needed.
8244 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8245 FalseC->getValueType(0), Cond);
8246 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8247 SDValue(FalseC, 0));
8249 if (N->getNumValues() == 2) // Dead flag value?
8250 return DCI.CombineTo(N, Cond, SDValue());
8254 // Optimize cases that will turn into an LEA instruction. This requires
8255 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8256 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8257 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8258 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8260 bool isFastMultiplier = false;
8262 switch ((unsigned char)Diff) {
8264 case 1: // result = add base, cond
8265 case 2: // result = lea base( , cond*2)
8266 case 3: // result = lea base(cond, cond*2)
8267 case 4: // result = lea base( , cond*4)
8268 case 5: // result = lea base(cond, cond*4)
8269 case 8: // result = lea base( , cond*8)
8270 case 9: // result = lea base(cond, cond*8)
8271 isFastMultiplier = true;
8276 if (isFastMultiplier) {
8277 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8278 SDValue Cond = N->getOperand(3);
8279 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8280 DAG.getConstant(CC, MVT::i8), Cond);
8281 // Zero extend the condition if needed.
8282 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8284 // Scale the condition by the difference.
8286 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8287 DAG.getConstant(Diff, Cond.getValueType()));
8289 // Add the base if non-zero.
8290 if (FalseC->getAPIntValue() != 0)
8291 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8292 SDValue(FalseC, 0));
8293 if (N->getNumValues() == 2) // Dead flag value?
8294 return DCI.CombineTo(N, Cond, SDValue());
8304 /// PerformMulCombine - Optimize a single multiply with constant into two
8305 /// in order to implement it with two cheaper instructions, e.g.
8306 /// LEA + SHL, LEA + LEA.
8307 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8308 TargetLowering::DAGCombinerInfo &DCI) {
8309 if (DAG.getMachineFunction().
8310 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8313 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8316 MVT VT = N->getValueType(0);
8320 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8323 uint64_t MulAmt = C->getZExtValue();
8324 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8327 uint64_t MulAmt1 = 0;
8328 uint64_t MulAmt2 = 0;
8329 if ((MulAmt % 9) == 0) {
8331 MulAmt2 = MulAmt / 9;
8332 } else if ((MulAmt % 5) == 0) {
8334 MulAmt2 = MulAmt / 5;
8335 } else if ((MulAmt % 3) == 0) {
8337 MulAmt2 = MulAmt / 3;
8340 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8341 DebugLoc DL = N->getDebugLoc();
8343 if (isPowerOf2_64(MulAmt2) &&
8344 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8345 // If second multiplifer is pow2, issue it first. We want the multiply by
8346 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8348 std::swap(MulAmt1, MulAmt2);
8351 if (isPowerOf2_64(MulAmt1))
8352 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8353 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8355 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8356 DAG.getConstant(MulAmt1, VT));
8358 if (isPowerOf2_64(MulAmt2))
8359 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8360 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8362 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8363 DAG.getConstant(MulAmt2, VT));
8365 // Do not add new nodes to DAG combiner worklist.
8366 DCI.CombineTo(N, NewMul, false);
8372 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8374 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8375 const X86Subtarget *Subtarget) {
8376 // On X86 with SSE2 support, we can transform this to a vector shift if
8377 // all elements are shifted by the same amount. We can't do this in legalize
8378 // because the a constant vector is typically transformed to a constant pool
8379 // so we have no knowledge of the shift amount.
8380 if (!Subtarget->hasSSE2())
8383 MVT VT = N->getValueType(0);
8384 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8387 SDValue ShAmtOp = N->getOperand(1);
8388 MVT EltVT = VT.getVectorElementType();
8389 DebugLoc DL = N->getDebugLoc();
8391 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8392 unsigned NumElts = VT.getVectorNumElements();
8394 for (; i != NumElts; ++i) {
8395 SDValue Arg = ShAmtOp.getOperand(i);
8396 if (Arg.getOpcode() == ISD::UNDEF) continue;
8400 for (; i != NumElts; ++i) {
8401 SDValue Arg = ShAmtOp.getOperand(i);
8402 if (Arg.getOpcode() == ISD::UNDEF) continue;
8403 if (Arg != BaseShAmt) {
8407 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8408 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8409 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8410 DAG.getIntPtrConstant(0));
8414 if (EltVT.bitsGT(MVT::i32))
8415 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8416 else if (EltVT.bitsLT(MVT::i32))
8417 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8419 // The shift amount is identical so we can do a vector shift.
8420 SDValue ValOp = N->getOperand(0);
8421 switch (N->getOpcode()) {
8423 llvm_unreachable("Unknown shift opcode!");
8426 if (VT == MVT::v2i64)
8427 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8428 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8430 if (VT == MVT::v4i32)
8431 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8432 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8434 if (VT == MVT::v8i16)
8435 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8436 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8440 if (VT == MVT::v4i32)
8441 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8442 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8444 if (VT == MVT::v8i16)
8445 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8446 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8450 if (VT == MVT::v2i64)
8451 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8452 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8454 if (VT == MVT::v4i32)
8455 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8456 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8458 if (VT == MVT::v8i16)
8459 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8460 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8467 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8468 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8469 const X86Subtarget *Subtarget) {
8470 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8471 // the FP state in cases where an emms may be missing.
8472 // A preferable solution to the general problem is to figure out the right
8473 // places to insert EMMS. This qualifies as a quick hack.
8475 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8476 StoreSDNode *St = cast<StoreSDNode>(N);
8477 MVT VT = St->getValue().getValueType();
8478 if (VT.getSizeInBits() != 64)
8481 const Function *F = DAG.getMachineFunction().getFunction();
8482 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
8483 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
8484 && Subtarget->hasSSE2();
8485 if ((VT.isVector() ||
8486 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8487 isa<LoadSDNode>(St->getValue()) &&
8488 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8489 St->getChain().hasOneUse() && !St->isVolatile()) {
8490 SDNode* LdVal = St->getValue().getNode();
8492 int TokenFactorIndex = -1;
8493 SmallVector<SDValue, 8> Ops;
8494 SDNode* ChainVal = St->getChain().getNode();
8495 // Must be a store of a load. We currently handle two cases: the load
8496 // is a direct child, and it's under an intervening TokenFactor. It is
8497 // possible to dig deeper under nested TokenFactors.
8498 if (ChainVal == LdVal)
8499 Ld = cast<LoadSDNode>(St->getChain());
8500 else if (St->getValue().hasOneUse() &&
8501 ChainVal->getOpcode() == ISD::TokenFactor) {
8502 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8503 if (ChainVal->getOperand(i).getNode() == LdVal) {
8504 TokenFactorIndex = i;
8505 Ld = cast<LoadSDNode>(St->getValue());
8507 Ops.push_back(ChainVal->getOperand(i));
8511 if (!Ld || !ISD::isNormalLoad(Ld))
8514 // If this is not the MMX case, i.e. we are just turning i64 load/store
8515 // into f64 load/store, avoid the transformation if there are multiple
8516 // uses of the loaded value.
8517 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8520 DebugLoc LdDL = Ld->getDebugLoc();
8521 DebugLoc StDL = N->getDebugLoc();
8522 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8523 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8525 if (Subtarget->is64Bit() || F64IsLegal) {
8526 MVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8527 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8528 Ld->getBasePtr(), Ld->getSrcValue(),
8529 Ld->getSrcValueOffset(), Ld->isVolatile(),
8530 Ld->getAlignment());
8531 SDValue NewChain = NewLd.getValue(1);
8532 if (TokenFactorIndex != -1) {
8533 Ops.push_back(NewChain);
8534 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8537 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8538 St->getSrcValue(), St->getSrcValueOffset(),
8539 St->isVolatile(), St->getAlignment());
8542 // Otherwise, lower to two pairs of 32-bit loads / stores.
8543 SDValue LoAddr = Ld->getBasePtr();
8544 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8545 DAG.getConstant(4, MVT::i32));
8547 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8548 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8549 Ld->isVolatile(), Ld->getAlignment());
8550 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8551 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8553 MinAlign(Ld->getAlignment(), 4));
8555 SDValue NewChain = LoLd.getValue(1);
8556 if (TokenFactorIndex != -1) {
8557 Ops.push_back(LoLd);
8558 Ops.push_back(HiLd);
8559 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8563 LoAddr = St->getBasePtr();
8564 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8565 DAG.getConstant(4, MVT::i32));
8567 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8568 St->getSrcValue(), St->getSrcValueOffset(),
8569 St->isVolatile(), St->getAlignment());
8570 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8572 St->getSrcValueOffset() + 4,
8574 MinAlign(St->getAlignment(), 4));
8575 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8580 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8581 /// X86ISD::FXOR nodes.
8582 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8583 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8584 // F[X]OR(0.0, x) -> x
8585 // F[X]OR(x, 0.0) -> x
8586 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8587 if (C->getValueAPF().isPosZero())
8588 return N->getOperand(1);
8589 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8590 if (C->getValueAPF().isPosZero())
8591 return N->getOperand(0);
8595 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8596 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8597 // FAND(0.0, x) -> 0.0
8598 // FAND(x, 0.0) -> 0.0
8599 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8600 if (C->getValueAPF().isPosZero())
8601 return N->getOperand(0);
8602 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8603 if (C->getValueAPF().isPosZero())
8604 return N->getOperand(1);
8608 static SDValue PerformBTCombine(SDNode *N,
8610 TargetLowering::DAGCombinerInfo &DCI) {
8611 // BT ignores high bits in the bit index operand.
8612 SDValue Op1 = N->getOperand(1);
8613 if (Op1.hasOneUse()) {
8614 unsigned BitWidth = Op1.getValueSizeInBits();
8615 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8616 APInt KnownZero, KnownOne;
8617 TargetLowering::TargetLoweringOpt TLO(DAG);
8618 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8619 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8620 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8621 DCI.CommitTargetLoweringOpt(TLO);
8626 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
8627 SDValue Op = N->getOperand(0);
8628 if (Op.getOpcode() == ISD::BIT_CONVERT)
8629 Op = Op.getOperand(0);
8630 MVT VT = N->getValueType(0), OpVT = Op.getValueType();
8631 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
8632 VT.getVectorElementType().getSizeInBits() ==
8633 OpVT.getVectorElementType().getSizeInBits()) {
8634 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
8639 // On X86 and X86-64, atomic operations are lowered to locked instructions.
8640 // Locked instructions, in turn, have implicit fence semantics (all memory
8641 // operations are flushed before issuing the locked instruction, and the
8642 // are not buffered), so we can fold away the common pattern of
8643 // fence-atomic-fence.
8644 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
8645 SDValue atomic = N->getOperand(0);
8646 switch (atomic.getOpcode()) {
8647 case ISD::ATOMIC_CMP_SWAP:
8648 case ISD::ATOMIC_SWAP:
8649 case ISD::ATOMIC_LOAD_ADD:
8650 case ISD::ATOMIC_LOAD_SUB:
8651 case ISD::ATOMIC_LOAD_AND:
8652 case ISD::ATOMIC_LOAD_OR:
8653 case ISD::ATOMIC_LOAD_XOR:
8654 case ISD::ATOMIC_LOAD_NAND:
8655 case ISD::ATOMIC_LOAD_MIN:
8656 case ISD::ATOMIC_LOAD_MAX:
8657 case ISD::ATOMIC_LOAD_UMIN:
8658 case ISD::ATOMIC_LOAD_UMAX:
8664 SDValue fence = atomic.getOperand(0);
8665 if (fence.getOpcode() != ISD::MEMBARRIER)
8668 switch (atomic.getOpcode()) {
8669 case ISD::ATOMIC_CMP_SWAP:
8670 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8671 atomic.getOperand(1), atomic.getOperand(2),
8672 atomic.getOperand(3));
8673 case ISD::ATOMIC_SWAP:
8674 case ISD::ATOMIC_LOAD_ADD:
8675 case ISD::ATOMIC_LOAD_SUB:
8676 case ISD::ATOMIC_LOAD_AND:
8677 case ISD::ATOMIC_LOAD_OR:
8678 case ISD::ATOMIC_LOAD_XOR:
8679 case ISD::ATOMIC_LOAD_NAND:
8680 case ISD::ATOMIC_LOAD_MIN:
8681 case ISD::ATOMIC_LOAD_MAX:
8682 case ISD::ATOMIC_LOAD_UMIN:
8683 case ISD::ATOMIC_LOAD_UMAX:
8684 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8685 atomic.getOperand(1), atomic.getOperand(2));
8691 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8692 DAGCombinerInfo &DCI) const {
8693 SelectionDAG &DAG = DCI.DAG;
8694 switch (N->getOpcode()) {
8696 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8697 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8698 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8699 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
8702 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8703 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8705 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8706 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8707 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8708 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
8709 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
8715 //===----------------------------------------------------------------------===//
8716 // X86 Inline Assembly Support
8717 //===----------------------------------------------------------------------===//
8719 static bool LowerToBSwap(CallInst *CI) {
8720 // FIXME: this should verify that we are targetting a 486 or better. If not,
8721 // we will turn this bswap into something that will be lowered to logical ops
8722 // instead of emitting the bswap asm. For now, we don't support 486 or lower
8723 // so don't worry about this.
8725 // Verify this is a simple bswap.
8726 if (CI->getNumOperands() != 2 ||
8727 CI->getType() != CI->getOperand(1)->getType() ||
8728 !CI->getType()->isInteger())
8731 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8732 if (!Ty || Ty->getBitWidth() % 16 != 0)
8735 // Okay, we can do this xform, do so now.
8736 const Type *Tys[] = { Ty };
8737 Module *M = CI->getParent()->getParent()->getParent();
8738 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
8740 Value *Op = CI->getOperand(1);
8741 Op = CallInst::Create(Int, Op, CI->getName(), CI);
8743 CI->replaceAllUsesWith(Op);
8744 CI->eraseFromParent();
8748 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
8749 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8750 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
8752 std::string AsmStr = IA->getAsmString();
8754 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
8755 std::vector<std::string> AsmPieces;
8756 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
8758 switch (AsmPieces.size()) {
8759 default: return false;
8761 AsmStr = AsmPieces[0];
8763 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
8766 if (AsmPieces.size() == 2 &&
8767 (AsmPieces[0] == "bswap" ||
8768 AsmPieces[0] == "bswapq" ||
8769 AsmPieces[0] == "bswapl") &&
8770 (AsmPieces[1] == "$0" ||
8771 AsmPieces[1] == "${0:q}")) {
8772 // No need to check constraints, nothing other than the equivalent of
8773 // "=r,0" would be valid here.
8774 return LowerToBSwap(CI);
8776 // rorw $$8, ${0:w} --> llvm.bswap.i16
8777 if (CI->getType() == Type::Int16Ty &&
8778 AsmPieces.size() == 3 &&
8779 AsmPieces[0] == "rorw" &&
8780 AsmPieces[1] == "$$8," &&
8781 AsmPieces[2] == "${0:w}" &&
8782 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
8783 return LowerToBSwap(CI);
8787 if (CI->getType() == Type::Int64Ty && Constraints.size() >= 2 &&
8788 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
8789 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
8790 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
8791 std::vector<std::string> Words;
8792 SplitString(AsmPieces[0], Words, " \t");
8793 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
8795 SplitString(AsmPieces[1], Words, " \t");
8796 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
8798 SplitString(AsmPieces[2], Words, " \t,");
8799 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
8800 Words[2] == "%edx") {
8801 return LowerToBSwap(CI);
8813 /// getConstraintType - Given a constraint letter, return the type of
8814 /// constraint it is for this target.
8815 X86TargetLowering::ConstraintType
8816 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8817 if (Constraint.size() == 1) {
8818 switch (Constraint[0]) {
8830 return C_RegisterClass;
8838 return TargetLowering::getConstraintType(Constraint);
8841 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8842 /// with another that has more specific requirements based on the type of the
8843 /// corresponding operand.
8844 const char *X86TargetLowering::
8845 LowerXConstraint(MVT ConstraintVT) const {
8846 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8847 // 'f' like normal targets.
8848 if (ConstraintVT.isFloatingPoint()) {
8849 if (Subtarget->hasSSE2())
8851 if (Subtarget->hasSSE1())
8855 return TargetLowering::LowerXConstraint(ConstraintVT);
8858 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
8859 /// vector. If it is invalid, don't add anything to Ops.
8860 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
8863 std::vector<SDValue>&Ops,
8864 SelectionDAG &DAG) const {
8865 SDValue Result(0, 0);
8867 switch (Constraint) {
8870 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8871 if (C->getZExtValue() <= 31) {
8872 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8878 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8879 if (C->getZExtValue() <= 63) {
8880 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8886 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8887 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
8888 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8894 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8895 if (C->getZExtValue() <= 255) {
8896 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8902 // 32-bit signed value
8903 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8904 const ConstantInt *CI = C->getConstantIntValue();
8905 if (CI->isValueValidForType(Type::Int32Ty, C->getSExtValue())) {
8906 // Widen to 64 bits here to get it sign extended.
8907 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
8910 // FIXME gcc accepts some relocatable values here too, but only in certain
8911 // memory models; it's complicated.
8916 // 32-bit unsigned value
8917 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
8918 const ConstantInt *CI = C->getConstantIntValue();
8919 if (CI->isValueValidForType(Type::Int32Ty, C->getZExtValue())) {
8920 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
8924 // FIXME gcc accepts some relocatable values here too, but only in certain
8925 // memory models; it's complicated.
8929 // Literal immediates are always ok.
8930 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
8931 // Widen to 64 bits here to get it sign extended.
8932 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
8936 // If we are in non-pic codegen mode, we allow the address of a global (with
8937 // an optional displacement) to be used with 'i'.
8938 GlobalAddressSDNode *GA = 0;
8941 // Match either (GA), (GA+C), (GA+C1+C2), etc.
8943 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
8944 Offset += GA->getOffset();
8946 } else if (Op.getOpcode() == ISD::ADD) {
8947 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8948 Offset += C->getZExtValue();
8949 Op = Op.getOperand(0);
8952 } else if (Op.getOpcode() == ISD::SUB) {
8953 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
8954 Offset += -C->getZExtValue();
8955 Op = Op.getOperand(0);
8960 // Otherwise, this isn't something we can handle, reject it.
8964 GlobalValue *GV = GA->getGlobal();
8965 // If we require an extra load to get this address, as in PIC mode, we
8967 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
8968 getTargetMachine())))
8972 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
8974 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
8980 if (Result.getNode()) {
8981 Ops.push_back(Result);
8984 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
8988 std::vector<unsigned> X86TargetLowering::
8989 getRegClassForInlineAsmConstraint(const std::string &Constraint,
8991 if (Constraint.size() == 1) {
8992 // FIXME: not handling fp-stack yet!
8993 switch (Constraint[0]) { // GCC X86 Constraint Letters
8994 default: break; // Unknown constraint letter
8995 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
8996 if (Subtarget->is64Bit()) {
8998 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
8999 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9000 X86::R10D,X86::R11D,X86::R12D,
9001 X86::R13D,X86::R14D,X86::R15D,
9002 X86::EBP, X86::ESP, 0);
9003 else if (VT == MVT::i16)
9004 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9005 X86::SI, X86::DI, X86::R8W,X86::R9W,
9006 X86::R10W,X86::R11W,X86::R12W,
9007 X86::R13W,X86::R14W,X86::R15W,
9008 X86::BP, X86::SP, 0);
9009 else if (VT == MVT::i8)
9010 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9011 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9012 X86::R10B,X86::R11B,X86::R12B,
9013 X86::R13B,X86::R14B,X86::R15B,
9014 X86::BPL, X86::SPL, 0);
9016 else if (VT == MVT::i64)
9017 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9018 X86::RSI, X86::RDI, X86::R8, X86::R9,
9019 X86::R10, X86::R11, X86::R12,
9020 X86::R13, X86::R14, X86::R15,
9021 X86::RBP, X86::RSP, 0);
9025 // 32-bit fallthrough
9028 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9029 else if (VT == MVT::i16)
9030 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9031 else if (VT == MVT::i8)
9032 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9033 else if (VT == MVT::i64)
9034 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9039 return std::vector<unsigned>();
9042 std::pair<unsigned, const TargetRegisterClass*>
9043 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9045 // First, see if this is a constraint that directly corresponds to an LLVM
9047 if (Constraint.size() == 1) {
9048 // GCC Constraint Letters
9049 switch (Constraint[0]) {
9051 case 'r': // GENERAL_REGS
9052 case 'R': // LEGACY_REGS
9053 case 'l': // INDEX_REGS
9055 return std::make_pair(0U, X86::GR8RegisterClass);
9057 return std::make_pair(0U, X86::GR16RegisterClass);
9058 if (VT == MVT::i32 || !Subtarget->is64Bit())
9059 return std::make_pair(0U, X86::GR32RegisterClass);
9060 return std::make_pair(0U, X86::GR64RegisterClass);
9061 case 'f': // FP Stack registers.
9062 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
9063 // value to the correct fpstack register class.
9064 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9065 return std::make_pair(0U, X86::RFP32RegisterClass);
9066 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9067 return std::make_pair(0U, X86::RFP64RegisterClass);
9068 return std::make_pair(0U, X86::RFP80RegisterClass);
9069 case 'y': // MMX_REGS if MMX allowed.
9070 if (!Subtarget->hasMMX()) break;
9071 return std::make_pair(0U, X86::VR64RegisterClass);
9072 case 'Y': // SSE_REGS if SSE2 allowed
9073 if (!Subtarget->hasSSE2()) break;
9075 case 'x': // SSE_REGS if SSE1 allowed
9076 if (!Subtarget->hasSSE1()) break;
9078 switch (VT.getSimpleVT()) {
9080 // Scalar SSE types.
9083 return std::make_pair(0U, X86::FR32RegisterClass);
9086 return std::make_pair(0U, X86::FR64RegisterClass);
9094 return std::make_pair(0U, X86::VR128RegisterClass);
9100 // Use the default implementation in TargetLowering to convert the register
9101 // constraint into a member of a register class.
9102 std::pair<unsigned, const TargetRegisterClass*> Res;
9103 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9105 // Not found as a standard register?
9106 if (Res.second == 0) {
9107 // GCC calls "st(0)" just plain "st".
9108 if (StringsEqualNoCase("{st}", Constraint)) {
9109 Res.first = X86::ST0;
9110 Res.second = X86::RFP80RegisterClass;
9112 // 'A' means EAX + EDX.
9113 if (Constraint == "A") {
9114 Res.first = X86::EAX;
9115 Res.second = X86::GR32_ADRegisterClass;
9120 // Otherwise, check to see if this is a register class of the wrong value
9121 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
9122 // turn into {ax},{dx}.
9123 if (Res.second->hasType(VT))
9124 return Res; // Correct type already, nothing to do.
9126 // All of the single-register GCC register classes map their values onto
9127 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
9128 // really want an 8-bit or 32-bit register, map to the appropriate register
9129 // class and return the appropriate register.
9130 if (Res.second == X86::GR16RegisterClass) {
9131 if (VT == MVT::i8) {
9132 unsigned DestReg = 0;
9133 switch (Res.first) {
9135 case X86::AX: DestReg = X86::AL; break;
9136 case X86::DX: DestReg = X86::DL; break;
9137 case X86::CX: DestReg = X86::CL; break;
9138 case X86::BX: DestReg = X86::BL; break;
9141 Res.first = DestReg;
9142 Res.second = X86::GR8RegisterClass;
9144 } else if (VT == MVT::i32) {
9145 unsigned DestReg = 0;
9146 switch (Res.first) {
9148 case X86::AX: DestReg = X86::EAX; break;
9149 case X86::DX: DestReg = X86::EDX; break;
9150 case X86::CX: DestReg = X86::ECX; break;
9151 case X86::BX: DestReg = X86::EBX; break;
9152 case X86::SI: DestReg = X86::ESI; break;
9153 case X86::DI: DestReg = X86::EDI; break;
9154 case X86::BP: DestReg = X86::EBP; break;
9155 case X86::SP: DestReg = X86::ESP; break;
9158 Res.first = DestReg;
9159 Res.second = X86::GR32RegisterClass;
9161 } else if (VT == MVT::i64) {
9162 unsigned DestReg = 0;
9163 switch (Res.first) {
9165 case X86::AX: DestReg = X86::RAX; break;
9166 case X86::DX: DestReg = X86::RDX; break;
9167 case X86::CX: DestReg = X86::RCX; break;
9168 case X86::BX: DestReg = X86::RBX; break;
9169 case X86::SI: DestReg = X86::RSI; break;
9170 case X86::DI: DestReg = X86::RDI; break;
9171 case X86::BP: DestReg = X86::RBP; break;
9172 case X86::SP: DestReg = X86::RSP; break;
9175 Res.first = DestReg;
9176 Res.second = X86::GR64RegisterClass;
9179 } else if (Res.second == X86::FR32RegisterClass ||
9180 Res.second == X86::FR64RegisterClass ||
9181 Res.second == X86::VR128RegisterClass) {
9182 // Handle references to XMM physical registers that got mapped into the
9183 // wrong class. This can happen with constraints like {xmm0} where the
9184 // target independent register mapper will just pick the first match it can
9185 // find, ignoring the required type.
9187 Res.second = X86::FR32RegisterClass;
9188 else if (VT == MVT::f64)
9189 Res.second = X86::FR64RegisterClass;
9190 else if (X86::VR128RegisterClass->hasType(VT))
9191 Res.second = X86::VR128RegisterClass;
9197 //===----------------------------------------------------------------------===//
9198 // X86 Widen vector type
9199 //===----------------------------------------------------------------------===//
9201 /// getWidenVectorType: given a vector type, returns the type to widen
9202 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9203 /// If there is no vector type that we want to widen to, returns MVT::Other
9204 /// When and where to widen is target dependent based on the cost of
9205 /// scalarizing vs using the wider vector type.
9207 MVT X86TargetLowering::getWidenVectorType(MVT VT) const {
9208 assert(VT.isVector());
9209 if (isTypeLegal(VT))
9212 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9213 // type based on element type. This would speed up our search (though
9214 // it may not be worth it since the size of the list is relatively
9216 MVT EltVT = VT.getVectorElementType();
9217 unsigned NElts = VT.getVectorNumElements();
9219 // On X86, it make sense to widen any vector wider than 1
9223 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9224 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9225 MVT SVT = (MVT::SimpleValueType)nVT;
9227 if (isTypeLegal(SVT) &&
9228 SVT.getVectorElementType() == EltVT &&
9229 SVT.getVectorNumElements() > NElts)