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 "X86TargetObjectFile.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalAlias.h"
24 #include "llvm/GlobalVariable.h"
25 #include "llvm/Function.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/Intrinsics.h"
28 #include "llvm/LLVMContext.h"
29 #include "llvm/ADT/BitVector.h"
30 #include "llvm/ADT/VectorExtras.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/ErrorHandling.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 // Disable16Bit - 16-bit operations typically have a larger encoding than
51 // corresponding 32-bit instructions, and 16-bit code is slow on some
52 // processors. This is an experimental flag to disable 16-bit operations
53 // (which forces them to be Legalized to 32-bit operations).
55 Disable16Bit("disable-16bit", cl::Hidden,
56 cl::desc("Disable use of 16-bit instructions"));
58 // Forward declarations.
59 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
62 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
63 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
64 default: llvm_unreachable("unknown subtarget type");
65 case X86Subtarget::isDarwin:
66 if (TM.getSubtarget<X86Subtarget>().is64Bit())
67 return new X8664_MachoTargetObjectFile();
68 return new X8632_MachoTargetObjectFile();
69 case X86Subtarget::isELF:
70 return new TargetLoweringObjectFileELF();
71 case X86Subtarget::isMingw:
72 case X86Subtarget::isCygwin:
73 case X86Subtarget::isWindows:
74 return new TargetLoweringObjectFileCOFF();
79 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
80 : TargetLowering(TM, createTLOF(TM)) {
81 Subtarget = &TM.getSubtarget<X86Subtarget>();
82 X86ScalarSSEf64 = Subtarget->hasSSE2();
83 X86ScalarSSEf32 = Subtarget->hasSSE1();
84 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
86 RegInfo = TM.getRegisterInfo();
89 // Set up the TargetLowering object.
91 // X86 is weird, it always uses i8 for shift amounts and setcc results.
92 setShiftAmountType(MVT::i8);
93 setBooleanContents(ZeroOrOneBooleanContent);
94 setSchedulingPreference(SchedulingForRegPressure);
95 setStackPointerRegisterToSaveRestore(X86StackPtr);
97 if (Subtarget->isTargetDarwin()) {
98 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
99 setUseUnderscoreSetJmp(false);
100 setUseUnderscoreLongJmp(false);
101 } else if (Subtarget->isTargetMingw()) {
102 // MS runtime is weird: it exports _setjmp, but longjmp!
103 setUseUnderscoreSetJmp(true);
104 setUseUnderscoreLongJmp(false);
106 setUseUnderscoreSetJmp(true);
107 setUseUnderscoreLongJmp(true);
110 // Set up the register classes.
111 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
113 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
114 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
115 if (Subtarget->is64Bit())
116 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
118 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
120 // We don't accept any truncstore of integer registers.
121 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
123 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
124 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
126 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
127 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
128 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
130 // SETOEQ and SETUNE require checking two conditions.
131 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
132 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
133 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
135 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
136 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
138 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
140 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
141 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
142 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
144 if (Subtarget->is64Bit()) {
145 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
146 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
147 } else if (!UseSoftFloat) {
148 if (X86ScalarSSEf64) {
149 // We have an impenetrably clever algorithm for ui64->double only.
150 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
152 // We have an algorithm for SSE2, and we turn this into a 64-bit
153 // FILD for other targets.
154 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
157 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
159 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
160 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
163 // SSE has no i16 to fp conversion, only i32
164 if (X86ScalarSSEf32) {
165 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
166 // f32 and f64 cases are Legal, f80 case is not
167 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
169 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
170 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
173 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
174 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
177 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
178 // are Legal, f80 is custom lowered.
179 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
180 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
182 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
184 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
185 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
187 if (X86ScalarSSEf32) {
188 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
189 // f32 and f64 cases are Legal, f80 case is not
190 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
192 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
193 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
196 // Handle FP_TO_UINT by promoting the destination to a larger signed
198 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
199 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
200 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
202 if (Subtarget->is64Bit()) {
203 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
204 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
205 } else if (!UseSoftFloat) {
206 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
207 // Expand FP_TO_UINT into a select.
208 // FIXME: We would like to use a Custom expander here eventually to do
209 // the optimal thing for SSE vs. the default expansion in the legalizer.
210 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
212 // With SSE3 we can use fisttpll to convert to a signed i64; without
213 // SSE, we're stuck with a fistpll.
214 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
217 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
218 if (!X86ScalarSSEf64) {
219 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
220 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
223 // Scalar integer divide and remainder are lowered to use operations that
224 // produce two results, to match the available instructions. This exposes
225 // the two-result form to trivial CSE, which is able to combine x/y and x%y
226 // into a single instruction.
228 // Scalar integer multiply-high is also lowered to use two-result
229 // operations, to match the available instructions. However, plain multiply
230 // (low) operations are left as Legal, as there are single-result
231 // instructions for this in x86. Using the two-result multiply instructions
232 // when both high and low results are needed must be arranged by dagcombine.
233 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
234 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
235 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
236 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
237 setOperationAction(ISD::SREM , MVT::i8 , Expand);
238 setOperationAction(ISD::UREM , MVT::i8 , Expand);
239 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
240 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
241 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
242 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
243 setOperationAction(ISD::SREM , MVT::i16 , Expand);
244 setOperationAction(ISD::UREM , MVT::i16 , Expand);
245 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
246 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
247 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
248 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
249 setOperationAction(ISD::SREM , MVT::i32 , Expand);
250 setOperationAction(ISD::UREM , MVT::i32 , Expand);
251 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
252 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
253 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
254 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
255 setOperationAction(ISD::SREM , MVT::i64 , Expand);
256 setOperationAction(ISD::UREM , MVT::i64 , Expand);
258 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
259 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
260 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
261 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
262 if (Subtarget->is64Bit())
263 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
264 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
265 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
266 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
267 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
268 setOperationAction(ISD::FREM , MVT::f32 , Expand);
269 setOperationAction(ISD::FREM , MVT::f64 , Expand);
270 setOperationAction(ISD::FREM , MVT::f80 , Expand);
271 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
273 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
274 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
275 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
276 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
278 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
279 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
281 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
282 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
285 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
286 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
287 if (Subtarget->is64Bit()) {
288 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
293 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
294 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
296 // These should be promoted to a larger select which is supported.
297 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
298 // X86 wants to expand cmov itself.
299 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
301 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
303 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
304 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
305 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
306 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
307 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
308 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
310 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
312 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
313 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
314 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
315 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
316 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
317 if (Subtarget->is64Bit()) {
318 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
319 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
321 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
324 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
325 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
326 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
327 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
328 if (Subtarget->is64Bit())
329 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
330 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
331 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
332 if (Subtarget->is64Bit()) {
333 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
334 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
335 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
336 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
337 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
339 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
340 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
341 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
342 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
343 if (Subtarget->is64Bit()) {
344 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
345 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
346 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
349 if (Subtarget->hasSSE1())
350 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
352 if (!Subtarget->hasSSE2())
353 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
355 // Expand certain atomics
356 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
357 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
361 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
362 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
366 if (!Subtarget->is64Bit()) {
367 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
376 // FIXME - use subtarget debug flags
377 if (!Subtarget->isTargetDarwin() &&
378 !Subtarget->isTargetELF() &&
379 !Subtarget->isTargetCygMing()) {
380 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
383 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
384 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
385 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
386 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
387 if (Subtarget->is64Bit()) {
388 setExceptionPointerRegister(X86::RAX);
389 setExceptionSelectorRegister(X86::RDX);
391 setExceptionPointerRegister(X86::EAX);
392 setExceptionSelectorRegister(X86::EDX);
394 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
395 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
397 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
399 setOperationAction(ISD::TRAP, MVT::Other, Legal);
401 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
402 setOperationAction(ISD::VASTART , MVT::Other, Custom);
403 setOperationAction(ISD::VAEND , MVT::Other, Expand);
404 if (Subtarget->is64Bit()) {
405 setOperationAction(ISD::VAARG , MVT::Other, Custom);
406 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
408 setOperationAction(ISD::VAARG , MVT::Other, Expand);
409 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
412 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
413 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
414 if (Subtarget->is64Bit())
415 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
416 if (Subtarget->isTargetCygMing())
417 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
419 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
421 if (!UseSoftFloat && X86ScalarSSEf64) {
422 // f32 and f64 use SSE.
423 // Set up the FP register classes.
424 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
425 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
427 // Use ANDPD to simulate FABS.
428 setOperationAction(ISD::FABS , MVT::f64, Custom);
429 setOperationAction(ISD::FABS , MVT::f32, Custom);
431 // Use XORP to simulate FNEG.
432 setOperationAction(ISD::FNEG , MVT::f64, Custom);
433 setOperationAction(ISD::FNEG , MVT::f32, Custom);
435 // Use ANDPD and ORPD to simulate FCOPYSIGN.
436 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
437 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
439 // We don't support sin/cos/fmod
440 setOperationAction(ISD::FSIN , MVT::f64, Expand);
441 setOperationAction(ISD::FCOS , MVT::f64, Expand);
442 setOperationAction(ISD::FSIN , MVT::f32, Expand);
443 setOperationAction(ISD::FCOS , MVT::f32, Expand);
445 // Expand FP immediates into loads from the stack, except for the special
447 addLegalFPImmediate(APFloat(+0.0)); // xorpd
448 addLegalFPImmediate(APFloat(+0.0f)); // xorps
449 } else if (!UseSoftFloat && X86ScalarSSEf32) {
450 // Use SSE for f32, x87 for f64.
451 // Set up the FP register classes.
452 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
453 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
455 // Use ANDPS to simulate FABS.
456 setOperationAction(ISD::FABS , MVT::f32, Custom);
458 // Use XORP to simulate FNEG.
459 setOperationAction(ISD::FNEG , MVT::f32, Custom);
461 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
463 // Use ANDPS and ORPS to simulate FCOPYSIGN.
464 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
465 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
467 // We don't support sin/cos/fmod
468 setOperationAction(ISD::FSIN , MVT::f32, Expand);
469 setOperationAction(ISD::FCOS , MVT::f32, Expand);
471 // Special cases we handle for FP constants.
472 addLegalFPImmediate(APFloat(+0.0f)); // xorps
473 addLegalFPImmediate(APFloat(+0.0)); // FLD0
474 addLegalFPImmediate(APFloat(+1.0)); // FLD1
475 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
476 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
479 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
480 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
482 } else if (!UseSoftFloat) {
483 // f32 and f64 in x87.
484 // Set up the FP register classes.
485 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
486 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
488 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
489 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
490 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
491 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
494 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
495 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
497 addLegalFPImmediate(APFloat(+0.0)); // FLD0
498 addLegalFPImmediate(APFloat(+1.0)); // FLD1
499 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
500 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
501 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
502 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
503 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
504 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
507 // Long double always uses X87.
509 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
510 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
511 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
514 APFloat TmpFlt(+0.0);
515 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
517 addLegalFPImmediate(TmpFlt); // FLD0
519 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
520 APFloat TmpFlt2(+1.0);
521 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
523 addLegalFPImmediate(TmpFlt2); // FLD1
524 TmpFlt2.changeSign();
525 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
529 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
530 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
534 // Always use a library call for pow.
535 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
536 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
537 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
539 setOperationAction(ISD::FLOG, MVT::f80, Expand);
540 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
541 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
542 setOperationAction(ISD::FEXP, MVT::f80, Expand);
543 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
545 // First set operation action for all vector types to either promote
546 // (for widening) or expand (for scalarization). Then we will selectively
547 // turn on ones that can be effectively codegen'd.
548 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
549 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
550 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
565 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
566 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
599 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
603 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
604 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
605 setTruncStoreAction((MVT::SimpleValueType)VT,
606 (MVT::SimpleValueType)InnerVT, Expand);
607 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
608 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
609 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
612 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
613 // with -msoft-float, disable use of MMX as well.
614 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
615 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
616 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
617 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
618 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
619 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
621 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
622 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
623 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
624 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
626 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
627 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
628 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
629 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
631 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
632 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
634 setOperationAction(ISD::AND, MVT::v8i8, Promote);
635 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
636 setOperationAction(ISD::AND, MVT::v4i16, Promote);
637 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
638 setOperationAction(ISD::AND, MVT::v2i32, Promote);
639 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
640 setOperationAction(ISD::AND, MVT::v1i64, Legal);
642 setOperationAction(ISD::OR, MVT::v8i8, Promote);
643 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
644 setOperationAction(ISD::OR, MVT::v4i16, Promote);
645 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
646 setOperationAction(ISD::OR, MVT::v2i32, Promote);
647 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
648 setOperationAction(ISD::OR, MVT::v1i64, Legal);
650 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
651 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
652 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
653 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
654 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
655 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
656 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
658 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
659 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
660 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
661 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
662 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
663 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
664 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
665 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
666 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
668 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
669 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
670 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
671 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
672 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
674 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
675 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
676 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
677 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
679 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
680 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
681 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
682 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
684 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
686 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
687 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
688 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
689 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
690 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
691 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
692 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
695 if (!UseSoftFloat && Subtarget->hasSSE1()) {
696 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
698 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
699 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
700 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
701 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
702 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
703 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
704 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
705 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
706 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
707 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
708 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
709 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
712 if (!UseSoftFloat && Subtarget->hasSSE2()) {
713 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
715 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
716 // registers cannot be used even for integer operations.
717 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
718 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
719 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
720 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
722 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
723 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
724 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
725 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
726 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
727 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
728 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
729 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
730 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
731 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
732 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
733 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
734 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
735 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
736 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
737 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
739 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
740 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
741 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
742 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
744 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
745 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
746 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
747 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
748 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
750 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
751 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
752 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
753 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
754 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
756 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
757 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
758 EVT VT = (MVT::SimpleValueType)i;
759 // Do not attempt to custom lower non-power-of-2 vectors
760 if (!isPowerOf2_32(VT.getVectorNumElements()))
762 // Do not attempt to custom lower non-128-bit vectors
763 if (!VT.is128BitVector())
765 setOperationAction(ISD::BUILD_VECTOR,
766 VT.getSimpleVT().SimpleTy, Custom);
767 setOperationAction(ISD::VECTOR_SHUFFLE,
768 VT.getSimpleVT().SimpleTy, Custom);
769 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
770 VT.getSimpleVT().SimpleTy, Custom);
773 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
774 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
775 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
776 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
777 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
778 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
780 if (Subtarget->is64Bit()) {
781 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
782 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
785 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
786 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
787 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
790 // Do not attempt to promote non-128-bit vectors
791 if (!VT.is128BitVector()) {
794 setOperationAction(ISD::AND, SVT, Promote);
795 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
796 setOperationAction(ISD::OR, SVT, Promote);
797 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
798 setOperationAction(ISD::XOR, SVT, Promote);
799 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
800 setOperationAction(ISD::LOAD, SVT, Promote);
801 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
802 setOperationAction(ISD::SELECT, SVT, Promote);
803 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
806 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
808 // Custom lower v2i64 and v2f64 selects.
809 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
810 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
811 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
812 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
814 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
815 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
816 if (!DisableMMX && Subtarget->hasMMX()) {
817 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
818 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
822 if (Subtarget->hasSSE41()) {
823 // FIXME: Do we need to handle scalar-to-vector here?
824 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
826 // i8 and i16 vectors are custom , because the source register and source
827 // source memory operand types are not the same width. f32 vectors are
828 // custom since the immediate controlling the insert encodes additional
830 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
831 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
832 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
833 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
835 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
836 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
837 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
838 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
840 if (Subtarget->is64Bit()) {
841 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
842 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
846 if (Subtarget->hasSSE42()) {
847 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
850 if (!UseSoftFloat && Subtarget->hasAVX()) {
851 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
852 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
853 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
854 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
856 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
857 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
858 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
859 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
860 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
861 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
862 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
863 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
864 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
865 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
866 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
867 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
868 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
869 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
870 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
872 // Operations to consider commented out -v16i16 v32i8
873 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
874 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
875 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
876 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
877 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
878 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
879 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
880 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
881 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
882 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
883 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
884 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
885 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
886 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
888 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
889 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
890 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
891 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
893 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
894 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
895 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
896 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
897 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
899 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
900 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
901 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
902 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
903 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
904 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
907 // Not sure we want to do this since there are no 256-bit integer
910 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
911 // This includes 256-bit vectors
912 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
913 EVT VT = (MVT::SimpleValueType)i;
915 // Do not attempt to custom lower non-power-of-2 vectors
916 if (!isPowerOf2_32(VT.getVectorNumElements()))
919 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
920 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
921 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
924 if (Subtarget->is64Bit()) {
925 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
926 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
931 // Not sure we want to do this since there are no 256-bit integer
934 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
935 // Including 256-bit vectors
936 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
937 EVT VT = (MVT::SimpleValueType)i;
939 if (!VT.is256BitVector()) {
942 setOperationAction(ISD::AND, VT, Promote);
943 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
944 setOperationAction(ISD::OR, VT, Promote);
945 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
946 setOperationAction(ISD::XOR, VT, Promote);
947 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
948 setOperationAction(ISD::LOAD, VT, Promote);
949 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
950 setOperationAction(ISD::SELECT, VT, Promote);
951 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
954 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
958 // We want to custom lower some of our intrinsics.
959 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
961 // Add/Sub/Mul with overflow operations are custom lowered.
962 setOperationAction(ISD::SADDO, MVT::i32, Custom);
963 setOperationAction(ISD::SADDO, MVT::i64, Custom);
964 setOperationAction(ISD::UADDO, MVT::i32, Custom);
965 setOperationAction(ISD::UADDO, MVT::i64, Custom);
966 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
967 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
968 setOperationAction(ISD::USUBO, MVT::i32, Custom);
969 setOperationAction(ISD::USUBO, MVT::i64, Custom);
970 setOperationAction(ISD::SMULO, MVT::i32, Custom);
971 setOperationAction(ISD::SMULO, MVT::i64, Custom);
973 if (!Subtarget->is64Bit()) {
974 // These libcalls are not available in 32-bit.
975 setLibcallName(RTLIB::SHL_I128, 0);
976 setLibcallName(RTLIB::SRL_I128, 0);
977 setLibcallName(RTLIB::SRA_I128, 0);
980 // We have target-specific dag combine patterns for the following nodes:
981 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
982 setTargetDAGCombine(ISD::BUILD_VECTOR);
983 setTargetDAGCombine(ISD::SELECT);
984 setTargetDAGCombine(ISD::SHL);
985 setTargetDAGCombine(ISD::SRA);
986 setTargetDAGCombine(ISD::SRL);
987 setTargetDAGCombine(ISD::OR);
988 setTargetDAGCombine(ISD::STORE);
989 setTargetDAGCombine(ISD::MEMBARRIER);
990 setTargetDAGCombine(ISD::ZERO_EXTEND);
991 if (Subtarget->is64Bit())
992 setTargetDAGCombine(ISD::MUL);
994 computeRegisterProperties();
996 // Divide and reminder operations have no vector equivalent and can
997 // trap. Do a custom widening for these operations in which we never
998 // generate more divides/remainder than the original vector width.
999 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1000 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
1001 if (!isTypeLegal((MVT::SimpleValueType)VT)) {
1002 setOperationAction(ISD::SDIV, (MVT::SimpleValueType) VT, Custom);
1003 setOperationAction(ISD::UDIV, (MVT::SimpleValueType) VT, Custom);
1004 setOperationAction(ISD::SREM, (MVT::SimpleValueType) VT, Custom);
1005 setOperationAction(ISD::UREM, (MVT::SimpleValueType) VT, Custom);
1009 // FIXME: These should be based on subtarget info. Plus, the values should
1010 // be smaller when we are in optimizing for size mode.
1011 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1012 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1013 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1014 setPrefLoopAlignment(16);
1015 benefitFromCodePlacementOpt = true;
1019 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1024 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1025 /// the desired ByVal argument alignment.
1026 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1029 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1030 if (VTy->getBitWidth() == 128)
1032 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1033 unsigned EltAlign = 0;
1034 getMaxByValAlign(ATy->getElementType(), EltAlign);
1035 if (EltAlign > MaxAlign)
1036 MaxAlign = EltAlign;
1037 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1038 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1039 unsigned EltAlign = 0;
1040 getMaxByValAlign(STy->getElementType(i), EltAlign);
1041 if (EltAlign > MaxAlign)
1042 MaxAlign = EltAlign;
1050 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1051 /// function arguments in the caller parameter area. For X86, aggregates
1052 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1053 /// are at 4-byte boundaries.
1054 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1055 if (Subtarget->is64Bit()) {
1056 // Max of 8 and alignment of type.
1057 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1064 if (Subtarget->hasSSE1())
1065 getMaxByValAlign(Ty, Align);
1069 /// getOptimalMemOpType - Returns the target specific optimal type for load
1070 /// and store operations as a result of memset, memcpy, and memmove
1071 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1074 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1075 bool isSrcConst, bool isSrcStr,
1076 SelectionDAG &DAG) const {
1077 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1078 // linux. This is because the stack realignment code can't handle certain
1079 // cases like PR2962. This should be removed when PR2962 is fixed.
1080 const Function *F = DAG.getMachineFunction().getFunction();
1081 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1082 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1083 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1085 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1088 if (Subtarget->is64Bit() && Size >= 8)
1093 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1095 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1096 SelectionDAG &DAG) const {
1097 if (!Subtarget->is64Bit())
1098 // This doesn't have DebugLoc associated with it, but is not really the
1099 // same as a Register.
1100 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1105 /// getFunctionAlignment - Return the Log2 alignment of this function.
1106 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1107 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1110 //===----------------------------------------------------------------------===//
1111 // Return Value Calling Convention Implementation
1112 //===----------------------------------------------------------------------===//
1114 #include "X86GenCallingConv.inc"
1117 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1118 const SmallVectorImpl<EVT> &OutTys,
1119 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1120 SelectionDAG &DAG) {
1121 SmallVector<CCValAssign, 16> RVLocs;
1122 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1123 RVLocs, *DAG.getContext());
1124 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1128 X86TargetLowering::LowerReturn(SDValue Chain,
1129 CallingConv::ID CallConv, bool isVarArg,
1130 const SmallVectorImpl<ISD::OutputArg> &Outs,
1131 DebugLoc dl, SelectionDAG &DAG) {
1133 SmallVector<CCValAssign, 16> RVLocs;
1134 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1135 RVLocs, *DAG.getContext());
1136 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1138 // If this is the first return lowered for this function, add the regs to the
1139 // liveout set for the function.
1140 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1141 for (unsigned i = 0; i != RVLocs.size(); ++i)
1142 if (RVLocs[i].isRegLoc())
1143 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1148 SmallVector<SDValue, 6> RetOps;
1149 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1150 // Operand #1 = Bytes To Pop
1151 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1153 // Copy the result values into the output registers.
1154 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1155 CCValAssign &VA = RVLocs[i];
1156 assert(VA.isRegLoc() && "Can only return in registers!");
1157 SDValue ValToCopy = Outs[i].Val;
1159 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1160 // the RET instruction and handled by the FP Stackifier.
1161 if (VA.getLocReg() == X86::ST0 ||
1162 VA.getLocReg() == X86::ST1) {
1163 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1164 // change the value to the FP stack register class.
1165 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1166 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1167 RetOps.push_back(ValToCopy);
1168 // Don't emit a copytoreg.
1172 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1173 // which is returned in RAX / RDX.
1174 if (Subtarget->is64Bit()) {
1175 EVT ValVT = ValToCopy.getValueType();
1176 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1177 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1178 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1179 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1183 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1184 Flag = Chain.getValue(1);
1187 // The x86-64 ABI for returning structs by value requires that we copy
1188 // the sret argument into %rax for the return. We saved the argument into
1189 // a virtual register in the entry block, so now we copy the value out
1191 if (Subtarget->is64Bit() &&
1192 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1193 MachineFunction &MF = DAG.getMachineFunction();
1194 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1195 unsigned Reg = FuncInfo->getSRetReturnReg();
1197 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1198 FuncInfo->setSRetReturnReg(Reg);
1200 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1202 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1203 Flag = Chain.getValue(1);
1205 // RAX now acts like a return value.
1206 MF.getRegInfo().addLiveOut(X86::RAX);
1209 RetOps[0] = Chain; // Update chain.
1211 // Add the flag if we have it.
1213 RetOps.push_back(Flag);
1215 return DAG.getNode(X86ISD::RET_FLAG, dl,
1216 MVT::Other, &RetOps[0], RetOps.size());
1219 /// LowerCallResult - Lower the result values of a call into the
1220 /// appropriate copies out of appropriate physical registers.
1223 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1224 CallingConv::ID CallConv, bool isVarArg,
1225 const SmallVectorImpl<ISD::InputArg> &Ins,
1226 DebugLoc dl, SelectionDAG &DAG,
1227 SmallVectorImpl<SDValue> &InVals) {
1229 // Assign locations to each value returned by this call.
1230 SmallVector<CCValAssign, 16> RVLocs;
1231 bool Is64Bit = Subtarget->is64Bit();
1232 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1233 RVLocs, *DAG.getContext());
1234 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1236 // Copy all of the result registers out of their specified physreg.
1237 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1238 CCValAssign &VA = RVLocs[i];
1239 EVT CopyVT = VA.getValVT();
1241 // If this is x86-64, and we disabled SSE, we can't return FP values
1242 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1243 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1244 llvm_report_error("SSE register return with SSE disabled");
1247 // If this is a call to a function that returns an fp value on the floating
1248 // point stack, but where we prefer to use the value in xmm registers, copy
1249 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1250 if ((VA.getLocReg() == X86::ST0 ||
1251 VA.getLocReg() == X86::ST1) &&
1252 isScalarFPTypeInSSEReg(VA.getValVT())) {
1257 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1258 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1259 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1260 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1261 MVT::v2i64, InFlag).getValue(1);
1262 Val = Chain.getValue(0);
1263 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1264 Val, DAG.getConstant(0, MVT::i64));
1266 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1267 MVT::i64, InFlag).getValue(1);
1268 Val = Chain.getValue(0);
1270 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1272 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1273 CopyVT, InFlag).getValue(1);
1274 Val = Chain.getValue(0);
1276 InFlag = Chain.getValue(2);
1278 if (CopyVT != VA.getValVT()) {
1279 // Round the F80 the right size, which also moves to the appropriate xmm
1281 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1282 // This truncation won't change the value.
1283 DAG.getIntPtrConstant(1));
1286 InVals.push_back(Val);
1293 //===----------------------------------------------------------------------===//
1294 // C & StdCall & Fast Calling Convention implementation
1295 //===----------------------------------------------------------------------===//
1296 // StdCall calling convention seems to be standard for many Windows' API
1297 // routines and around. It differs from C calling convention just a little:
1298 // callee should clean up the stack, not caller. Symbols should be also
1299 // decorated in some fancy way :) It doesn't support any vector arguments.
1300 // For info on fast calling convention see Fast Calling Convention (tail call)
1301 // implementation LowerX86_32FastCCCallTo.
1303 /// CallIsStructReturn - Determines whether a call uses struct return
1305 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1309 return Outs[0].Flags.isSRet();
1312 /// ArgsAreStructReturn - Determines whether a function uses struct
1313 /// return semantics.
1315 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1319 return Ins[0].Flags.isSRet();
1322 /// IsCalleePop - Determines whether the callee is required to pop its
1323 /// own arguments. Callee pop is necessary to support tail calls.
1324 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1328 switch (CallingConv) {
1331 case CallingConv::X86_StdCall:
1332 return !Subtarget->is64Bit();
1333 case CallingConv::X86_FastCall:
1334 return !Subtarget->is64Bit();
1335 case CallingConv::Fast:
1336 return PerformTailCallOpt;
1340 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1341 /// given CallingConvention value.
1342 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1343 if (Subtarget->is64Bit()) {
1344 if (Subtarget->isTargetWin64())
1345 return CC_X86_Win64_C;
1350 if (CC == CallingConv::X86_FastCall)
1351 return CC_X86_32_FastCall;
1352 else if (CC == CallingConv::Fast)
1353 return CC_X86_32_FastCC;
1358 /// NameDecorationForCallConv - Selects the appropriate decoration to
1359 /// apply to a MachineFunction containing a given calling convention.
1361 X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
1362 if (CallConv == CallingConv::X86_FastCall)
1364 else if (CallConv == CallingConv::X86_StdCall)
1370 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1371 /// by "Src" to address "Dst" with size and alignment information specified by
1372 /// the specific parameter attribute. The copy will be passed as a byval
1373 /// function parameter.
1375 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1376 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1378 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1379 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1380 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1384 X86TargetLowering::LowerMemArgument(SDValue Chain,
1385 CallingConv::ID CallConv,
1386 const SmallVectorImpl<ISD::InputArg> &Ins,
1387 DebugLoc dl, SelectionDAG &DAG,
1388 const CCValAssign &VA,
1389 MachineFrameInfo *MFI,
1392 // Create the nodes corresponding to a load from this parameter slot.
1393 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1394 bool AlwaysUseMutable = (CallConv==CallingConv::Fast) && PerformTailCallOpt;
1395 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1398 // If value is passed by pointer we have address passed instead of the value
1400 if (VA.getLocInfo() == CCValAssign::Indirect)
1401 ValVT = VA.getLocVT();
1403 ValVT = VA.getValVT();
1405 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1406 // changed with more analysis.
1407 // In case of tail call optimization mark all arguments mutable. Since they
1408 // could be overwritten by lowering of arguments in case of a tail call.
1409 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1410 VA.getLocMemOffset(), isImmutable, false);
1411 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1412 if (Flags.isByVal())
1414 return DAG.getLoad(ValVT, dl, Chain, FIN,
1415 PseudoSourceValue::getFixedStack(FI), 0);
1419 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1420 CallingConv::ID CallConv,
1422 const SmallVectorImpl<ISD::InputArg> &Ins,
1425 SmallVectorImpl<SDValue> &InVals) {
1427 MachineFunction &MF = DAG.getMachineFunction();
1428 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1430 const Function* Fn = MF.getFunction();
1431 if (Fn->hasExternalLinkage() &&
1432 Subtarget->isTargetCygMing() &&
1433 Fn->getName() == "main")
1434 FuncInfo->setForceFramePointer(true);
1436 // Decorate the function name.
1437 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1439 MachineFrameInfo *MFI = MF.getFrameInfo();
1440 bool Is64Bit = Subtarget->is64Bit();
1441 bool IsWin64 = Subtarget->isTargetWin64();
1443 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1444 "Var args not supported with calling convention fastcc");
1446 // Assign locations to all of the incoming arguments.
1447 SmallVector<CCValAssign, 16> ArgLocs;
1448 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1449 ArgLocs, *DAG.getContext());
1450 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1452 unsigned LastVal = ~0U;
1454 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1455 CCValAssign &VA = ArgLocs[i];
1456 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1458 assert(VA.getValNo() != LastVal &&
1459 "Don't support value assigned to multiple locs yet");
1460 LastVal = VA.getValNo();
1462 if (VA.isRegLoc()) {
1463 EVT RegVT = VA.getLocVT();
1464 TargetRegisterClass *RC = NULL;
1465 if (RegVT == MVT::i32)
1466 RC = X86::GR32RegisterClass;
1467 else if (Is64Bit && RegVT == MVT::i64)
1468 RC = X86::GR64RegisterClass;
1469 else if (RegVT == MVT::f32)
1470 RC = X86::FR32RegisterClass;
1471 else if (RegVT == MVT::f64)
1472 RC = X86::FR64RegisterClass;
1473 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1474 RC = X86::VR128RegisterClass;
1475 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1476 RC = X86::VR64RegisterClass;
1478 llvm_unreachable("Unknown argument type!");
1480 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1481 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1483 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1484 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1486 if (VA.getLocInfo() == CCValAssign::SExt)
1487 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1488 DAG.getValueType(VA.getValVT()));
1489 else if (VA.getLocInfo() == CCValAssign::ZExt)
1490 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1491 DAG.getValueType(VA.getValVT()));
1492 else if (VA.getLocInfo() == CCValAssign::BCvt)
1493 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1495 if (VA.isExtInLoc()) {
1496 // Handle MMX values passed in XMM regs.
1497 if (RegVT.isVector()) {
1498 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1499 ArgValue, DAG.getConstant(0, MVT::i64));
1500 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1502 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1505 assert(VA.isMemLoc());
1506 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1509 // If value is passed via pointer - do a load.
1510 if (VA.getLocInfo() == CCValAssign::Indirect)
1511 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1513 InVals.push_back(ArgValue);
1516 // The x86-64 ABI for returning structs by value requires that we copy
1517 // the sret argument into %rax for the return. Save the argument into
1518 // a virtual register so that we can access it from the return points.
1519 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1520 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1521 unsigned Reg = FuncInfo->getSRetReturnReg();
1523 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1524 FuncInfo->setSRetReturnReg(Reg);
1526 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1527 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1530 unsigned StackSize = CCInfo.getNextStackOffset();
1531 // align stack specially for tail calls
1532 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1533 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1535 // If the function takes variable number of arguments, make a frame index for
1536 // the start of the first vararg value... for expansion of llvm.va_start.
1538 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1539 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1542 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1544 // FIXME: We should really autogenerate these arrays
1545 static const unsigned GPR64ArgRegsWin64[] = {
1546 X86::RCX, X86::RDX, X86::R8, X86::R9
1548 static const unsigned XMMArgRegsWin64[] = {
1549 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1551 static const unsigned GPR64ArgRegs64Bit[] = {
1552 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1554 static const unsigned XMMArgRegs64Bit[] = {
1555 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1556 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1558 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1561 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1562 GPR64ArgRegs = GPR64ArgRegsWin64;
1563 XMMArgRegs = XMMArgRegsWin64;
1565 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1566 GPR64ArgRegs = GPR64ArgRegs64Bit;
1567 XMMArgRegs = XMMArgRegs64Bit;
1569 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1571 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1574 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1575 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1576 "SSE register cannot be used when SSE is disabled!");
1577 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1578 "SSE register cannot be used when SSE is disabled!");
1579 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1580 // Kernel mode asks for SSE to be disabled, so don't push them
1582 TotalNumXMMRegs = 0;
1584 // For X86-64, if there are vararg parameters that are passed via
1585 // registers, then we must store them to their spots on the stack so they
1586 // may be loaded by deferencing the result of va_next.
1587 VarArgsGPOffset = NumIntRegs * 8;
1588 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1589 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1590 TotalNumXMMRegs * 16, 16,
1593 // Store the integer parameter registers.
1594 SmallVector<SDValue, 8> MemOps;
1595 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1596 unsigned Offset = VarArgsGPOffset;
1597 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1598 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1599 DAG.getIntPtrConstant(Offset));
1600 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1601 X86::GR64RegisterClass);
1602 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1604 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1605 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1607 MemOps.push_back(Store);
1611 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1612 // Now store the XMM (fp + vector) parameter registers.
1613 SmallVector<SDValue, 11> SaveXMMOps;
1614 SaveXMMOps.push_back(Chain);
1616 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1617 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1618 SaveXMMOps.push_back(ALVal);
1620 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1621 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1623 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1624 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1625 X86::VR128RegisterClass);
1626 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1627 SaveXMMOps.push_back(Val);
1629 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1631 &SaveXMMOps[0], SaveXMMOps.size()));
1634 if (!MemOps.empty())
1635 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1636 &MemOps[0], MemOps.size());
1640 // Some CCs need callee pop.
1641 if (IsCalleePop(isVarArg, CallConv)) {
1642 BytesToPopOnReturn = StackSize; // Callee pops everything.
1643 BytesCallerReserves = 0;
1645 BytesToPopOnReturn = 0; // Callee pops nothing.
1646 // If this is an sret function, the return should pop the hidden pointer.
1647 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1648 BytesToPopOnReturn = 4;
1649 BytesCallerReserves = StackSize;
1653 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1654 if (CallConv == CallingConv::X86_FastCall)
1655 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1658 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1664 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1665 SDValue StackPtr, SDValue Arg,
1666 DebugLoc dl, SelectionDAG &DAG,
1667 const CCValAssign &VA,
1668 ISD::ArgFlagsTy Flags) {
1669 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1670 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1671 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1672 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1673 if (Flags.isByVal()) {
1674 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1676 return DAG.getStore(Chain, dl, Arg, PtrOff,
1677 PseudoSourceValue::getStack(), LocMemOffset);
1680 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1681 /// optimization is performed and it is required.
1683 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1684 SDValue &OutRetAddr,
1690 if (!IsTailCall || FPDiff==0) return Chain;
1692 // Adjust the Return address stack slot.
1693 EVT VT = getPointerTy();
1694 OutRetAddr = getReturnAddressFrameIndex(DAG);
1696 // Load the "old" Return address.
1697 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1698 return SDValue(OutRetAddr.getNode(), 1);
1701 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1702 /// optimization is performed and it is required (FPDiff!=0).
1704 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1705 SDValue Chain, SDValue RetAddrFrIdx,
1706 bool Is64Bit, int FPDiff, DebugLoc dl) {
1707 // Store the return address to the appropriate stack slot.
1708 if (!FPDiff) return Chain;
1709 // Calculate the new stack slot for the return address.
1710 int SlotSize = Is64Bit ? 8 : 4;
1711 int NewReturnAddrFI =
1712 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize,
1714 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1715 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1716 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1717 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1722 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1723 CallingConv::ID CallConv, bool isVarArg,
1725 const SmallVectorImpl<ISD::OutputArg> &Outs,
1726 const SmallVectorImpl<ISD::InputArg> &Ins,
1727 DebugLoc dl, SelectionDAG &DAG,
1728 SmallVectorImpl<SDValue> &InVals) {
1730 MachineFunction &MF = DAG.getMachineFunction();
1731 bool Is64Bit = Subtarget->is64Bit();
1732 bool IsStructRet = CallIsStructReturn(Outs);
1734 assert((!isTailCall ||
1735 (CallConv == CallingConv::Fast && PerformTailCallOpt)) &&
1736 "IsEligibleForTailCallOptimization missed a case!");
1737 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1738 "Var args not supported with calling convention fastcc");
1740 // Analyze operands of the call, assigning locations to each operand.
1741 SmallVector<CCValAssign, 16> ArgLocs;
1742 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1743 ArgLocs, *DAG.getContext());
1744 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1746 // Get a count of how many bytes are to be pushed on the stack.
1747 unsigned NumBytes = CCInfo.getNextStackOffset();
1748 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1749 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1753 // Lower arguments at fp - stackoffset + fpdiff.
1754 unsigned NumBytesCallerPushed =
1755 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1756 FPDiff = NumBytesCallerPushed - NumBytes;
1758 // Set the delta of movement of the returnaddr stackslot.
1759 // But only set if delta is greater than previous delta.
1760 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1761 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1764 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1766 SDValue RetAddrFrIdx;
1767 // Load return adress for tail calls.
1768 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
1771 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1772 SmallVector<SDValue, 8> MemOpChains;
1775 // Walk the register/memloc assignments, inserting copies/loads. In the case
1776 // of tail call optimization arguments are handle later.
1777 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1778 CCValAssign &VA = ArgLocs[i];
1779 EVT RegVT = VA.getLocVT();
1780 SDValue Arg = Outs[i].Val;
1781 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1782 bool isByVal = Flags.isByVal();
1784 // Promote the value if needed.
1785 switch (VA.getLocInfo()) {
1786 default: llvm_unreachable("Unknown loc info!");
1787 case CCValAssign::Full: break;
1788 case CCValAssign::SExt:
1789 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1791 case CCValAssign::ZExt:
1792 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1794 case CCValAssign::AExt:
1795 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1796 // Special case: passing MMX values in XMM registers.
1797 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1798 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1799 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1801 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1803 case CCValAssign::BCvt:
1804 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1806 case CCValAssign::Indirect: {
1807 // Store the argument.
1808 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1809 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1810 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1811 PseudoSourceValue::getFixedStack(FI), 0);
1817 if (VA.isRegLoc()) {
1818 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1820 if (!isTailCall || (isTailCall && isByVal)) {
1821 assert(VA.isMemLoc());
1822 if (StackPtr.getNode() == 0)
1823 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1825 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1826 dl, DAG, VA, Flags));
1831 if (!MemOpChains.empty())
1832 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1833 &MemOpChains[0], MemOpChains.size());
1835 // Build a sequence of copy-to-reg nodes chained together with token chain
1836 // and flag operands which copy the outgoing args into registers.
1838 // Tail call byval lowering might overwrite argument registers so in case of
1839 // tail call optimization the copies to registers are lowered later.
1841 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1842 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1843 RegsToPass[i].second, InFlag);
1844 InFlag = Chain.getValue(1);
1848 if (Subtarget->isPICStyleGOT()) {
1849 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1852 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1853 DAG.getNode(X86ISD::GlobalBaseReg,
1854 DebugLoc::getUnknownLoc(),
1857 InFlag = Chain.getValue(1);
1859 // If we are tail calling and generating PIC/GOT style code load the
1860 // address of the callee into ECX. The value in ecx is used as target of
1861 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1862 // for tail calls on PIC/GOT architectures. Normally we would just put the
1863 // address of GOT into ebx and then call target@PLT. But for tail calls
1864 // ebx would be restored (since ebx is callee saved) before jumping to the
1867 // Note: The actual moving to ECX is done further down.
1868 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1869 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1870 !G->getGlobal()->hasProtectedVisibility())
1871 Callee = LowerGlobalAddress(Callee, DAG);
1872 else if (isa<ExternalSymbolSDNode>(Callee))
1873 Callee = LowerExternalSymbol(Callee, DAG);
1877 if (Is64Bit && isVarArg) {
1878 // From AMD64 ABI document:
1879 // For calls that may call functions that use varargs or stdargs
1880 // (prototype-less calls or calls to functions containing ellipsis (...) in
1881 // the declaration) %al is used as hidden argument to specify the number
1882 // of SSE registers used. The contents of %al do not need to match exactly
1883 // the number of registers, but must be an ubound on the number of SSE
1884 // registers used and is in the range 0 - 8 inclusive.
1886 // FIXME: Verify this on Win64
1887 // Count the number of XMM registers allocated.
1888 static const unsigned XMMArgRegs[] = {
1889 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1890 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1892 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1893 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1894 && "SSE registers cannot be used when SSE is disabled");
1896 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1897 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1898 InFlag = Chain.getValue(1);
1902 // For tail calls lower the arguments to the 'real' stack slot.
1904 // Force all the incoming stack arguments to be loaded from the stack
1905 // before any new outgoing arguments are stored to the stack, because the
1906 // outgoing stack slots may alias the incoming argument stack slots, and
1907 // the alias isn't otherwise explicit. This is slightly more conservative
1908 // than necessary, because it means that each store effectively depends
1909 // on every argument instead of just those arguments it would clobber.
1910 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1912 SmallVector<SDValue, 8> MemOpChains2;
1915 // Do not flag preceeding copytoreg stuff together with the following stuff.
1917 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1918 CCValAssign &VA = ArgLocs[i];
1919 if (!VA.isRegLoc()) {
1920 assert(VA.isMemLoc());
1921 SDValue Arg = Outs[i].Val;
1922 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1923 // Create frame index.
1924 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1925 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1926 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1927 FIN = DAG.getFrameIndex(FI, getPointerTy());
1929 if (Flags.isByVal()) {
1930 // Copy relative to framepointer.
1931 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1932 if (StackPtr.getNode() == 0)
1933 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1935 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1937 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1941 // Store relative to framepointer.
1942 MemOpChains2.push_back(
1943 DAG.getStore(ArgChain, dl, Arg, FIN,
1944 PseudoSourceValue::getFixedStack(FI), 0));
1949 if (!MemOpChains2.empty())
1950 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1951 &MemOpChains2[0], MemOpChains2.size());
1953 // Copy arguments to their registers.
1954 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1955 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1956 RegsToPass[i].second, InFlag);
1957 InFlag = Chain.getValue(1);
1961 // Store the return address to the appropriate stack slot.
1962 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1966 bool WasGlobalOrExternal = false;
1967 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
1968 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
1969 // In the 64-bit large code model, we have to make all calls
1970 // through a register, since the call instruction's 32-bit
1971 // pc-relative offset may not be large enough to hold the whole
1973 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1974 WasGlobalOrExternal = true;
1975 // If the callee is a GlobalAddress node (quite common, every direct call
1976 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
1979 // We should use extra load for direct calls to dllimported functions in
1981 GlobalValue *GV = G->getGlobal();
1982 if (!GV->hasDLLImportLinkage()) {
1983 unsigned char OpFlags = 0;
1985 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1986 // external symbols most go through the PLT in PIC mode. If the symbol
1987 // has hidden or protected visibility, or if it is static or local, then
1988 // we don't need to use the PLT - we can directly call it.
1989 if (Subtarget->isTargetELF() &&
1990 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1991 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1992 OpFlags = X86II::MO_PLT;
1993 } else if (Subtarget->isPICStyleStubAny() &&
1994 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1995 Subtarget->getDarwinVers() < 9) {
1996 // PC-relative references to external symbols should go through $stub,
1997 // unless we're building with the leopard linker or later, which
1998 // automatically synthesizes these stubs.
1999 OpFlags = X86II::MO_DARWIN_STUB;
2002 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2003 G->getOffset(), OpFlags);
2005 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2006 WasGlobalOrExternal = true;
2007 unsigned char OpFlags = 0;
2009 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2010 // symbols should go through the PLT.
2011 if (Subtarget->isTargetELF() &&
2012 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2013 OpFlags = X86II::MO_PLT;
2014 } else if (Subtarget->isPICStyleStubAny() &&
2015 Subtarget->getDarwinVers() < 9) {
2016 // PC-relative references to external symbols should go through $stub,
2017 // unless we're building with the leopard linker or later, which
2018 // automatically synthesizes these stubs.
2019 OpFlags = X86II::MO_DARWIN_STUB;
2022 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2026 if (isTailCall && !WasGlobalOrExternal) {
2027 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
2029 Chain = DAG.getCopyToReg(Chain, dl,
2030 DAG.getRegister(Opc, getPointerTy()),
2032 Callee = DAG.getRegister(Opc, getPointerTy());
2033 // Add register as live out.
2034 MF.getRegInfo().addLiveOut(Opc);
2037 // Returns a chain & a flag for retval copy to use.
2038 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2039 SmallVector<SDValue, 8> Ops;
2042 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2043 DAG.getIntPtrConstant(0, true), InFlag);
2044 InFlag = Chain.getValue(1);
2047 Ops.push_back(Chain);
2048 Ops.push_back(Callee);
2051 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2053 // Add argument registers to the end of the list so that they are known live
2055 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2056 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2057 RegsToPass[i].second.getValueType()));
2059 // Add an implicit use GOT pointer in EBX.
2060 if (!isTailCall && Subtarget->isPICStyleGOT())
2061 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2063 // Add an implicit use of AL for x86 vararg functions.
2064 if (Is64Bit && isVarArg)
2065 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2067 if (InFlag.getNode())
2068 Ops.push_back(InFlag);
2071 // If this is the first return lowered for this function, add the regs
2072 // to the liveout set for the function.
2073 if (MF.getRegInfo().liveout_empty()) {
2074 SmallVector<CCValAssign, 16> RVLocs;
2075 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2077 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2078 for (unsigned i = 0; i != RVLocs.size(); ++i)
2079 if (RVLocs[i].isRegLoc())
2080 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2083 assert(((Callee.getOpcode() == ISD::Register &&
2084 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2085 cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
2086 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2087 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2088 "Expecting a global address, external symbol, or scratch register");
2090 return DAG.getNode(X86ISD::TC_RETURN, dl,
2091 NodeTys, &Ops[0], Ops.size());
2094 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2095 InFlag = Chain.getValue(1);
2097 // Create the CALLSEQ_END node.
2098 unsigned NumBytesForCalleeToPush;
2099 if (IsCalleePop(isVarArg, CallConv))
2100 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2101 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2102 // If this is is a call to a struct-return function, the callee
2103 // pops the hidden struct pointer, so we have to push it back.
2104 // This is common for Darwin/X86, Linux & Mingw32 targets.
2105 NumBytesForCalleeToPush = 4;
2107 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2109 // Returns a flag for retval copy to use.
2110 Chain = DAG.getCALLSEQ_END(Chain,
2111 DAG.getIntPtrConstant(NumBytes, true),
2112 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2115 InFlag = Chain.getValue(1);
2117 // Handle result values, copying them out of physregs into vregs that we
2119 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2120 Ins, dl, DAG, InVals);
2124 //===----------------------------------------------------------------------===//
2125 // Fast Calling Convention (tail call) implementation
2126 //===----------------------------------------------------------------------===//
2128 // Like std call, callee cleans arguments, convention except that ECX is
2129 // reserved for storing the tail called function address. Only 2 registers are
2130 // free for argument passing (inreg). Tail call optimization is performed
2132 // * tailcallopt is enabled
2133 // * caller/callee are fastcc
2134 // On X86_64 architecture with GOT-style position independent code only local
2135 // (within module) calls are supported at the moment.
2136 // To keep the stack aligned according to platform abi the function
2137 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2138 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2139 // If a tail called function callee has more arguments than the caller the
2140 // caller needs to make sure that there is room to move the RETADDR to. This is
2141 // achieved by reserving an area the size of the argument delta right after the
2142 // original REtADDR, but before the saved framepointer or the spilled registers
2143 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2155 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2156 /// for a 16 byte align requirement.
2157 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2158 SelectionDAG& DAG) {
2159 MachineFunction &MF = DAG.getMachineFunction();
2160 const TargetMachine &TM = MF.getTarget();
2161 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2162 unsigned StackAlignment = TFI.getStackAlignment();
2163 uint64_t AlignMask = StackAlignment - 1;
2164 int64_t Offset = StackSize;
2165 uint64_t SlotSize = TD->getPointerSize();
2166 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2167 // Number smaller than 12 so just add the difference.
2168 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2170 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2171 Offset = ((~AlignMask) & Offset) + StackAlignment +
2172 (StackAlignment-SlotSize);
2177 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2178 /// for tail call optimization. Targets which want to do tail call
2179 /// optimization should implement this function.
2181 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2182 CallingConv::ID CalleeCC,
2184 const SmallVectorImpl<ISD::InputArg> &Ins,
2185 SelectionDAG& DAG) const {
2186 MachineFunction &MF = DAG.getMachineFunction();
2187 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2188 return CalleeCC == CallingConv::Fast && CallerCC == CalleeCC;
2192 X86TargetLowering::createFastISel(MachineFunction &mf,
2193 MachineModuleInfo *mmo,
2195 DenseMap<const Value *, unsigned> &vm,
2196 DenseMap<const BasicBlock *,
2197 MachineBasicBlock *> &bm,
2198 DenseMap<const AllocaInst *, int> &am
2200 , SmallSet<Instruction*, 8> &cil
2203 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2211 //===----------------------------------------------------------------------===//
2212 // Other Lowering Hooks
2213 //===----------------------------------------------------------------------===//
2216 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2217 MachineFunction &MF = DAG.getMachineFunction();
2218 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2219 int ReturnAddrIndex = FuncInfo->getRAIndex();
2221 if (ReturnAddrIndex == 0) {
2222 // Set up a frame object for the return address.
2223 uint64_t SlotSize = TD->getPointerSize();
2224 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2226 FuncInfo->setRAIndex(ReturnAddrIndex);
2229 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2233 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2234 bool hasSymbolicDisplacement) {
2235 // Offset should fit into 32 bit immediate field.
2236 if (!isInt32(Offset))
2239 // If we don't have a symbolic displacement - we don't have any extra
2241 if (!hasSymbolicDisplacement)
2244 // FIXME: Some tweaks might be needed for medium code model.
2245 if (M != CodeModel::Small && M != CodeModel::Kernel)
2248 // For small code model we assume that latest object is 16MB before end of 31
2249 // bits boundary. We may also accept pretty large negative constants knowing
2250 // that all objects are in the positive half of address space.
2251 if (M == CodeModel::Small && Offset < 16*1024*1024)
2254 // For kernel code model we know that all object resist in the negative half
2255 // of 32bits address space. We may not accept negative offsets, since they may
2256 // be just off and we may accept pretty large positive ones.
2257 if (M == CodeModel::Kernel && Offset > 0)
2263 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2264 /// specific condition code, returning the condition code and the LHS/RHS of the
2265 /// comparison to make.
2266 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2267 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2269 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2270 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2271 // X > -1 -> X == 0, jump !sign.
2272 RHS = DAG.getConstant(0, RHS.getValueType());
2273 return X86::COND_NS;
2274 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2275 // X < 0 -> X == 0, jump on sign.
2277 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2279 RHS = DAG.getConstant(0, RHS.getValueType());
2280 return X86::COND_LE;
2284 switch (SetCCOpcode) {
2285 default: llvm_unreachable("Invalid integer condition!");
2286 case ISD::SETEQ: return X86::COND_E;
2287 case ISD::SETGT: return X86::COND_G;
2288 case ISD::SETGE: return X86::COND_GE;
2289 case ISD::SETLT: return X86::COND_L;
2290 case ISD::SETLE: return X86::COND_LE;
2291 case ISD::SETNE: return X86::COND_NE;
2292 case ISD::SETULT: return X86::COND_B;
2293 case ISD::SETUGT: return X86::COND_A;
2294 case ISD::SETULE: return X86::COND_BE;
2295 case ISD::SETUGE: return X86::COND_AE;
2299 // First determine if it is required or is profitable to flip the operands.
2301 // If LHS is a foldable load, but RHS is not, flip the condition.
2302 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2303 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2304 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2305 std::swap(LHS, RHS);
2308 switch (SetCCOpcode) {
2314 std::swap(LHS, RHS);
2318 // On a floating point condition, the flags are set as follows:
2320 // 0 | 0 | 0 | X > Y
2321 // 0 | 0 | 1 | X < Y
2322 // 1 | 0 | 0 | X == Y
2323 // 1 | 1 | 1 | unordered
2324 switch (SetCCOpcode) {
2325 default: llvm_unreachable("Condcode should be pre-legalized away");
2327 case ISD::SETEQ: return X86::COND_E;
2328 case ISD::SETOLT: // flipped
2330 case ISD::SETGT: return X86::COND_A;
2331 case ISD::SETOLE: // flipped
2333 case ISD::SETGE: return X86::COND_AE;
2334 case ISD::SETUGT: // flipped
2336 case ISD::SETLT: return X86::COND_B;
2337 case ISD::SETUGE: // flipped
2339 case ISD::SETLE: return X86::COND_BE;
2341 case ISD::SETNE: return X86::COND_NE;
2342 case ISD::SETUO: return X86::COND_P;
2343 case ISD::SETO: return X86::COND_NP;
2345 case ISD::SETUNE: return X86::COND_INVALID;
2349 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2350 /// code. Current x86 isa includes the following FP cmov instructions:
2351 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2352 static bool hasFPCMov(unsigned X86CC) {
2368 /// isFPImmLegal - Returns true if the target can instruction select the
2369 /// specified FP immediate natively. If false, the legalizer will
2370 /// materialize the FP immediate as a load from a constant pool.
2371 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2372 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2373 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2379 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2380 /// the specified range (L, H].
2381 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2382 return (Val < 0) || (Val >= Low && Val < Hi);
2385 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2386 /// specified value.
2387 static bool isUndefOrEqual(int Val, int CmpVal) {
2388 if (Val < 0 || Val == CmpVal)
2393 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2394 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2395 /// the second operand.
2396 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2397 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2398 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2399 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2400 return (Mask[0] < 2 && Mask[1] < 2);
2404 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2405 SmallVector<int, 8> M;
2407 return ::isPSHUFDMask(M, N->getValueType(0));
2410 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2411 /// is suitable for input to PSHUFHW.
2412 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2413 if (VT != MVT::v8i16)
2416 // Lower quadword copied in order or undef.
2417 for (int i = 0; i != 4; ++i)
2418 if (Mask[i] >= 0 && Mask[i] != i)
2421 // Upper quadword shuffled.
2422 for (int i = 4; i != 8; ++i)
2423 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2429 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2430 SmallVector<int, 8> M;
2432 return ::isPSHUFHWMask(M, N->getValueType(0));
2435 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2436 /// is suitable for input to PSHUFLW.
2437 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2438 if (VT != MVT::v8i16)
2441 // Upper quadword copied in order.
2442 for (int i = 4; i != 8; ++i)
2443 if (Mask[i] >= 0 && Mask[i] != i)
2446 // Lower quadword shuffled.
2447 for (int i = 0; i != 4; ++i)
2454 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2455 SmallVector<int, 8> M;
2457 return ::isPSHUFLWMask(M, N->getValueType(0));
2460 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2461 /// is suitable for input to PALIGNR.
2462 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2464 int i, e = VT.getVectorNumElements();
2466 // Do not handle v2i64 / v2f64 shuffles with palignr.
2467 if (e < 4 || !hasSSSE3)
2470 for (i = 0; i != e; ++i)
2474 // All undef, not a palignr.
2478 // Determine if it's ok to perform a palignr with only the LHS, since we
2479 // don't have access to the actual shuffle elements to see if RHS is undef.
2480 bool Unary = Mask[i] < (int)e;
2481 bool NeedsUnary = false;
2483 int s = Mask[i] - i;
2485 // Check the rest of the elements to see if they are consecutive.
2486 for (++i; i != e; ++i) {
2491 Unary = Unary && (m < (int)e);
2492 NeedsUnary = NeedsUnary || (m < s);
2494 if (NeedsUnary && !Unary)
2496 if (Unary && m != ((s+i) & (e-1)))
2498 if (!Unary && m != (s+i))
2504 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2505 SmallVector<int, 8> M;
2507 return ::isPALIGNRMask(M, N->getValueType(0), true);
2510 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2511 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2512 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2513 int NumElems = VT.getVectorNumElements();
2514 if (NumElems != 2 && NumElems != 4)
2517 int Half = NumElems / 2;
2518 for (int i = 0; i < Half; ++i)
2519 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2521 for (int i = Half; i < NumElems; ++i)
2522 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2528 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2529 SmallVector<int, 8> M;
2531 return ::isSHUFPMask(M, N->getValueType(0));
2534 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2535 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2536 /// half elements to come from vector 1 (which would equal the dest.) and
2537 /// the upper half to come from vector 2.
2538 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2539 int NumElems = VT.getVectorNumElements();
2541 if (NumElems != 2 && NumElems != 4)
2544 int Half = NumElems / 2;
2545 for (int i = 0; i < Half; ++i)
2546 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2548 for (int i = Half; i < NumElems; ++i)
2549 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2554 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2555 SmallVector<int, 8> M;
2557 return isCommutedSHUFPMask(M, N->getValueType(0));
2560 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2561 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2562 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2563 if (N->getValueType(0).getVectorNumElements() != 4)
2566 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2567 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2568 isUndefOrEqual(N->getMaskElt(1), 7) &&
2569 isUndefOrEqual(N->getMaskElt(2), 2) &&
2570 isUndefOrEqual(N->getMaskElt(3), 3);
2573 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2574 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2576 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2577 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2582 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2583 isUndefOrEqual(N->getMaskElt(1), 3) &&
2584 isUndefOrEqual(N->getMaskElt(2), 2) &&
2585 isUndefOrEqual(N->getMaskElt(3), 3);
2588 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2589 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2590 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2591 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2593 if (NumElems != 2 && NumElems != 4)
2596 for (unsigned i = 0; i < NumElems/2; ++i)
2597 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2600 for (unsigned i = NumElems/2; i < NumElems; ++i)
2601 if (!isUndefOrEqual(N->getMaskElt(i), i))
2607 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2608 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2609 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2610 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2612 if (NumElems != 2 && NumElems != 4)
2615 for (unsigned i = 0; i < NumElems/2; ++i)
2616 if (!isUndefOrEqual(N->getMaskElt(i), i))
2619 for (unsigned i = 0; i < NumElems/2; ++i)
2620 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2626 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2627 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2628 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2629 bool V2IsSplat = false) {
2630 int NumElts = VT.getVectorNumElements();
2631 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2634 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2636 int BitI1 = Mask[i+1];
2637 if (!isUndefOrEqual(BitI, j))
2640 if (!isUndefOrEqual(BitI1, NumElts))
2643 if (!isUndefOrEqual(BitI1, j + NumElts))
2650 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2651 SmallVector<int, 8> M;
2653 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2656 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2657 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2658 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2659 bool V2IsSplat = false) {
2660 int NumElts = VT.getVectorNumElements();
2661 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2664 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2666 int BitI1 = Mask[i+1];
2667 if (!isUndefOrEqual(BitI, j + NumElts/2))
2670 if (isUndefOrEqual(BitI1, NumElts))
2673 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2680 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2681 SmallVector<int, 8> M;
2683 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2686 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2687 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2689 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2690 int NumElems = VT.getVectorNumElements();
2691 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2694 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2696 int BitI1 = Mask[i+1];
2697 if (!isUndefOrEqual(BitI, j))
2699 if (!isUndefOrEqual(BitI1, j))
2705 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2706 SmallVector<int, 8> M;
2708 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2711 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2712 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2714 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2715 int NumElems = VT.getVectorNumElements();
2716 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2719 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2721 int BitI1 = Mask[i+1];
2722 if (!isUndefOrEqual(BitI, j))
2724 if (!isUndefOrEqual(BitI1, j))
2730 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2731 SmallVector<int, 8> M;
2733 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2736 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2737 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2738 /// MOVSD, and MOVD, i.e. setting the lowest element.
2739 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2740 if (VT.getVectorElementType().getSizeInBits() < 32)
2743 int NumElts = VT.getVectorNumElements();
2745 if (!isUndefOrEqual(Mask[0], NumElts))
2748 for (int i = 1; i < NumElts; ++i)
2749 if (!isUndefOrEqual(Mask[i], i))
2755 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2756 SmallVector<int, 8> M;
2758 return ::isMOVLMask(M, N->getValueType(0));
2761 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2762 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2763 /// element of vector 2 and the other elements to come from vector 1 in order.
2764 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2765 bool V2IsSplat = false, bool V2IsUndef = false) {
2766 int NumOps = VT.getVectorNumElements();
2767 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2770 if (!isUndefOrEqual(Mask[0], 0))
2773 for (int i = 1; i < NumOps; ++i)
2774 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2775 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2776 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2782 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2783 bool V2IsUndef = false) {
2784 SmallVector<int, 8> M;
2786 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2789 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2790 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2791 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2792 if (N->getValueType(0).getVectorNumElements() != 4)
2795 // Expect 1, 1, 3, 3
2796 for (unsigned i = 0; i < 2; ++i) {
2797 int Elt = N->getMaskElt(i);
2798 if (Elt >= 0 && Elt != 1)
2803 for (unsigned i = 2; i < 4; ++i) {
2804 int Elt = N->getMaskElt(i);
2805 if (Elt >= 0 && Elt != 3)
2810 // Don't use movshdup if it can be done with a shufps.
2811 // FIXME: verify that matching u, u, 3, 3 is what we want.
2815 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2816 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2817 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2818 if (N->getValueType(0).getVectorNumElements() != 4)
2821 // Expect 0, 0, 2, 2
2822 for (unsigned i = 0; i < 2; ++i)
2823 if (N->getMaskElt(i) > 0)
2827 for (unsigned i = 2; i < 4; ++i) {
2828 int Elt = N->getMaskElt(i);
2829 if (Elt >= 0 && Elt != 2)
2834 // Don't use movsldup if it can be done with a shufps.
2838 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2839 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2840 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2841 int e = N->getValueType(0).getVectorNumElements() / 2;
2843 for (int i = 0; i < e; ++i)
2844 if (!isUndefOrEqual(N->getMaskElt(i), i))
2846 for (int i = 0; i < e; ++i)
2847 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2852 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2853 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
2854 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2855 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2856 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2858 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2860 for (int i = 0; i < NumOperands; ++i) {
2861 int Val = SVOp->getMaskElt(NumOperands-i-1);
2862 if (Val < 0) Val = 0;
2863 if (Val >= NumOperands) Val -= NumOperands;
2865 if (i != NumOperands - 1)
2871 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2872 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
2873 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2874 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2876 // 8 nodes, but we only care about the last 4.
2877 for (unsigned i = 7; i >= 4; --i) {
2878 int Val = SVOp->getMaskElt(i);
2887 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2888 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
2889 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2890 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2892 // 8 nodes, but we only care about the first 4.
2893 for (int i = 3; i >= 0; --i) {
2894 int Val = SVOp->getMaskElt(i);
2903 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
2904 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
2905 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
2906 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2907 EVT VVT = N->getValueType(0);
2908 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
2912 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
2913 Val = SVOp->getMaskElt(i);
2917 return (Val - i) * EltSize;
2920 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2922 bool X86::isZeroNode(SDValue Elt) {
2923 return ((isa<ConstantSDNode>(Elt) &&
2924 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2925 (isa<ConstantFPSDNode>(Elt) &&
2926 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2929 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2930 /// their permute mask.
2931 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2932 SelectionDAG &DAG) {
2933 EVT VT = SVOp->getValueType(0);
2934 unsigned NumElems = VT.getVectorNumElements();
2935 SmallVector<int, 8> MaskVec;
2937 for (unsigned i = 0; i != NumElems; ++i) {
2938 int idx = SVOp->getMaskElt(i);
2940 MaskVec.push_back(idx);
2941 else if (idx < (int)NumElems)
2942 MaskVec.push_back(idx + NumElems);
2944 MaskVec.push_back(idx - NumElems);
2946 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2947 SVOp->getOperand(0), &MaskVec[0]);
2950 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2951 /// the two vector operands have swapped position.
2952 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
2953 unsigned NumElems = VT.getVectorNumElements();
2954 for (unsigned i = 0; i != NumElems; ++i) {
2958 else if (idx < (int)NumElems)
2959 Mask[i] = idx + NumElems;
2961 Mask[i] = idx - NumElems;
2965 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2966 /// match movhlps. The lower half elements should come from upper half of
2967 /// V1 (and in order), and the upper half elements should come from the upper
2968 /// half of V2 (and in order).
2969 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2970 if (Op->getValueType(0).getVectorNumElements() != 4)
2972 for (unsigned i = 0, e = 2; i != e; ++i)
2973 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2975 for (unsigned i = 2; i != 4; ++i)
2976 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2981 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2982 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2984 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2985 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2987 N = N->getOperand(0).getNode();
2988 if (!ISD::isNON_EXTLoad(N))
2991 *LD = cast<LoadSDNode>(N);
2995 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2996 /// match movlp{s|d}. The lower half elements should come from lower half of
2997 /// V1 (and in order), and the upper half elements should come from the upper
2998 /// half of V2 (and in order). And since V1 will become the source of the
2999 /// MOVLP, it must be either a vector load or a scalar load to vector.
3000 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3001 ShuffleVectorSDNode *Op) {
3002 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3004 // Is V2 is a vector load, don't do this transformation. We will try to use
3005 // load folding shufps op.
3006 if (ISD::isNON_EXTLoad(V2))
3009 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3011 if (NumElems != 2 && NumElems != 4)
3013 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3014 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3016 for (unsigned i = NumElems/2; i != NumElems; ++i)
3017 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3022 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3024 static bool isSplatVector(SDNode *N) {
3025 if (N->getOpcode() != ISD::BUILD_VECTOR)
3028 SDValue SplatValue = N->getOperand(0);
3029 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3030 if (N->getOperand(i) != SplatValue)
3035 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3036 /// to an zero vector.
3037 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3038 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3039 SDValue V1 = N->getOperand(0);
3040 SDValue V2 = N->getOperand(1);
3041 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3042 for (unsigned i = 0; i != NumElems; ++i) {
3043 int Idx = N->getMaskElt(i);
3044 if (Idx >= (int)NumElems) {
3045 unsigned Opc = V2.getOpcode();
3046 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3048 if (Opc != ISD::BUILD_VECTOR ||
3049 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3051 } else if (Idx >= 0) {
3052 unsigned Opc = V1.getOpcode();
3053 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3055 if (Opc != ISD::BUILD_VECTOR ||
3056 !X86::isZeroNode(V1.getOperand(Idx)))
3063 /// getZeroVector - Returns a vector of specified type with all zero elements.
3065 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3067 assert(VT.isVector() && "Expected a vector type");
3069 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3070 // type. This ensures they get CSE'd.
3072 if (VT.getSizeInBits() == 64) { // MMX
3073 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3074 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3075 } else if (HasSSE2) { // SSE2
3076 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3077 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3079 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3080 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3082 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3085 /// getOnesVector - Returns a vector of specified type with all bits set.
3087 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3088 assert(VT.isVector() && "Expected a vector type");
3090 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3091 // type. This ensures they get CSE'd.
3092 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3094 if (VT.getSizeInBits() == 64) // MMX
3095 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3097 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3098 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3102 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3103 /// that point to V2 points to its first element.
3104 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3105 EVT VT = SVOp->getValueType(0);
3106 unsigned NumElems = VT.getVectorNumElements();
3108 bool Changed = false;
3109 SmallVector<int, 8> MaskVec;
3110 SVOp->getMask(MaskVec);
3112 for (unsigned i = 0; i != NumElems; ++i) {
3113 if (MaskVec[i] > (int)NumElems) {
3114 MaskVec[i] = NumElems;
3119 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3120 SVOp->getOperand(1), &MaskVec[0]);
3121 return SDValue(SVOp, 0);
3124 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3125 /// operation of specified width.
3126 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3128 unsigned NumElems = VT.getVectorNumElements();
3129 SmallVector<int, 8> Mask;
3130 Mask.push_back(NumElems);
3131 for (unsigned i = 1; i != NumElems; ++i)
3133 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3136 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3137 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3139 unsigned NumElems = VT.getVectorNumElements();
3140 SmallVector<int, 8> Mask;
3141 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3143 Mask.push_back(i + NumElems);
3145 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3148 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3149 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3151 unsigned NumElems = VT.getVectorNumElements();
3152 unsigned Half = NumElems/2;
3153 SmallVector<int, 8> Mask;
3154 for (unsigned i = 0; i != Half; ++i) {
3155 Mask.push_back(i + Half);
3156 Mask.push_back(i + NumElems + Half);
3158 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3161 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3162 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3164 if (SV->getValueType(0).getVectorNumElements() <= 4)
3165 return SDValue(SV, 0);
3167 EVT PVT = MVT::v4f32;
3168 EVT VT = SV->getValueType(0);
3169 DebugLoc dl = SV->getDebugLoc();
3170 SDValue V1 = SV->getOperand(0);
3171 int NumElems = VT.getVectorNumElements();
3172 int EltNo = SV->getSplatIndex();
3174 // unpack elements to the correct location
3175 while (NumElems > 4) {
3176 if (EltNo < NumElems/2) {
3177 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3179 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3180 EltNo -= NumElems/2;
3185 // Perform the splat.
3186 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3187 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3188 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3189 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3192 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3193 /// vector of zero or undef vector. This produces a shuffle where the low
3194 /// element of V2 is swizzled into the zero/undef vector, landing at element
3195 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3196 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3197 bool isZero, bool HasSSE2,
3198 SelectionDAG &DAG) {
3199 EVT VT = V2.getValueType();
3201 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3202 unsigned NumElems = VT.getVectorNumElements();
3203 SmallVector<int, 16> MaskVec;
3204 for (unsigned i = 0; i != NumElems; ++i)
3205 // If this is the insertion idx, put the low elt of V2 here.
3206 MaskVec.push_back(i == Idx ? NumElems : i);
3207 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3210 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3211 /// a shuffle that is zero.
3213 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3214 bool Low, SelectionDAG &DAG) {
3215 unsigned NumZeros = 0;
3216 for (int i = 0; i < NumElems; ++i) {
3217 unsigned Index = Low ? i : NumElems-i-1;
3218 int Idx = SVOp->getMaskElt(Index);
3223 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3224 if (Elt.getNode() && X86::isZeroNode(Elt))
3232 /// isVectorShift - Returns true if the shuffle can be implemented as a
3233 /// logical left or right shift of a vector.
3234 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3235 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3236 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3237 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3240 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3243 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3247 bool SeenV1 = false;
3248 bool SeenV2 = false;
3249 for (int i = NumZeros; i < NumElems; ++i) {
3250 int Val = isLeft ? (i - NumZeros) : i;
3251 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3263 if (SeenV1 && SeenV2)
3266 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3272 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3274 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3275 unsigned NumNonZero, unsigned NumZero,
3276 SelectionDAG &DAG, TargetLowering &TLI) {
3280 DebugLoc dl = Op.getDebugLoc();
3283 for (unsigned i = 0; i < 16; ++i) {
3284 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3285 if (ThisIsNonZero && First) {
3287 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3289 V = DAG.getUNDEF(MVT::v8i16);
3294 SDValue ThisElt(0, 0), LastElt(0, 0);
3295 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3296 if (LastIsNonZero) {
3297 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3298 MVT::i16, Op.getOperand(i-1));
3300 if (ThisIsNonZero) {
3301 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3302 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3303 ThisElt, DAG.getConstant(8, MVT::i8));
3305 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3309 if (ThisElt.getNode())
3310 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3311 DAG.getIntPtrConstant(i/2));
3315 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3318 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3320 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3321 unsigned NumNonZero, unsigned NumZero,
3322 SelectionDAG &DAG, TargetLowering &TLI) {
3326 DebugLoc dl = Op.getDebugLoc();
3329 for (unsigned i = 0; i < 8; ++i) {
3330 bool isNonZero = (NonZeros & (1 << i)) != 0;
3334 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3336 V = DAG.getUNDEF(MVT::v8i16);
3339 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3340 MVT::v8i16, V, Op.getOperand(i),
3341 DAG.getIntPtrConstant(i));
3348 /// getVShift - Return a vector logical shift node.
3350 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3351 unsigned NumBits, SelectionDAG &DAG,
3352 const TargetLowering &TLI, DebugLoc dl) {
3353 bool isMMX = VT.getSizeInBits() == 64;
3354 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3355 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3356 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3357 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3358 DAG.getNode(Opc, dl, ShVT, SrcOp,
3359 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3363 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3364 SelectionDAG &DAG) {
3366 // Check if the scalar load can be widened into a vector load. And if
3367 // the address is "base + cst" see if the cst can be "absorbed" into
3368 // the shuffle mask.
3369 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3370 SDValue Ptr = LD->getBasePtr();
3371 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3373 EVT PVT = LD->getValueType(0);
3374 if (PVT != MVT::i32 && PVT != MVT::f32)
3379 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3380 FI = FINode->getIndex();
3382 } else if (Ptr.getOpcode() == ISD::ADD &&
3383 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3384 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3385 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3386 Offset = Ptr.getConstantOperandVal(1);
3387 Ptr = Ptr.getOperand(0);
3392 SDValue Chain = LD->getChain();
3393 // Make sure the stack object alignment is at least 16.
3394 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3395 if (DAG.InferPtrAlignment(Ptr) < 16) {
3396 if (MFI->isFixedObjectIndex(FI)) {
3397 // Can't change the alignment. FIXME: It's possible to compute
3398 // the exact stack offset and reference FI + adjust offset instead.
3399 // If someone *really* cares about this. That's the way to implement it.
3402 MFI->setObjectAlignment(FI, 16);
3406 // (Offset % 16) must be multiple of 4. Then address is then
3407 // Ptr + (Offset & ~15).
3410 if ((Offset % 16) & 3)
3412 int64_t StartOffset = Offset & ~15;
3414 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3415 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3417 int EltNo = (Offset - StartOffset) >> 2;
3418 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3419 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3420 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0);
3421 // Canonicalize it to a v4i32 shuffle.
3422 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3423 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3424 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3425 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3432 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3433 DebugLoc dl = Op.getDebugLoc();
3434 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3435 if (ISD::isBuildVectorAllZeros(Op.getNode())
3436 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3437 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3438 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3439 // eliminated on x86-32 hosts.
3440 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3443 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3444 return getOnesVector(Op.getValueType(), DAG, dl);
3445 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3448 EVT VT = Op.getValueType();
3449 EVT ExtVT = VT.getVectorElementType();
3450 unsigned EVTBits = ExtVT.getSizeInBits();
3452 unsigned NumElems = Op.getNumOperands();
3453 unsigned NumZero = 0;
3454 unsigned NumNonZero = 0;
3455 unsigned NonZeros = 0;
3456 bool IsAllConstants = true;
3457 SmallSet<SDValue, 8> Values;
3458 for (unsigned i = 0; i < NumElems; ++i) {
3459 SDValue Elt = Op.getOperand(i);
3460 if (Elt.getOpcode() == ISD::UNDEF)
3463 if (Elt.getOpcode() != ISD::Constant &&
3464 Elt.getOpcode() != ISD::ConstantFP)
3465 IsAllConstants = false;
3466 if (X86::isZeroNode(Elt))
3469 NonZeros |= (1 << i);
3474 if (NumNonZero == 0) {
3475 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3476 return DAG.getUNDEF(VT);
3479 // Special case for single non-zero, non-undef, element.
3480 if (NumNonZero == 1) {
3481 unsigned Idx = CountTrailingZeros_32(NonZeros);
3482 SDValue Item = Op.getOperand(Idx);
3484 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3485 // the value are obviously zero, truncate the value to i32 and do the
3486 // insertion that way. Only do this if the value is non-constant or if the
3487 // value is a constant being inserted into element 0. It is cheaper to do
3488 // a constant pool load than it is to do a movd + shuffle.
3489 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3490 (!IsAllConstants || Idx == 0)) {
3491 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3492 // Handle MMX and SSE both.
3493 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3494 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3496 // Truncate the value (which may itself be a constant) to i32, and
3497 // convert it to a vector with movd (S2V+shuffle to zero extend).
3498 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3499 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3500 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3501 Subtarget->hasSSE2(), DAG);
3503 // Now we have our 32-bit value zero extended in the low element of
3504 // a vector. If Idx != 0, swizzle it into place.
3506 SmallVector<int, 4> Mask;
3507 Mask.push_back(Idx);
3508 for (unsigned i = 1; i != VecElts; ++i)
3510 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3511 DAG.getUNDEF(Item.getValueType()),
3514 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3518 // If we have a constant or non-constant insertion into the low element of
3519 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3520 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3521 // depending on what the source datatype is.
3524 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3525 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3526 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3527 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3528 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3529 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3531 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3532 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3533 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3534 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3535 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3536 Subtarget->hasSSE2(), DAG);
3537 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3541 // Is it a vector logical left shift?
3542 if (NumElems == 2 && Idx == 1 &&
3543 X86::isZeroNode(Op.getOperand(0)) &&
3544 !X86::isZeroNode(Op.getOperand(1))) {
3545 unsigned NumBits = VT.getSizeInBits();
3546 return getVShift(true, VT,
3547 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3548 VT, Op.getOperand(1)),
3549 NumBits/2, DAG, *this, dl);
3552 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3555 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3556 // is a non-constant being inserted into an element other than the low one,
3557 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3558 // movd/movss) to move this into the low element, then shuffle it into
3560 if (EVTBits == 32) {
3561 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3563 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3564 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3565 Subtarget->hasSSE2(), DAG);
3566 SmallVector<int, 8> MaskVec;
3567 for (unsigned i = 0; i < NumElems; i++)
3568 MaskVec.push_back(i == Idx ? 0 : 1);
3569 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3573 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3574 if (Values.size() == 1) {
3575 if (EVTBits == 32) {
3576 // Instead of a shuffle like this:
3577 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3578 // Check if it's possible to issue this instead.
3579 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3580 unsigned Idx = CountTrailingZeros_32(NonZeros);
3581 SDValue Item = Op.getOperand(Idx);
3582 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3583 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3588 // A vector full of immediates; various special cases are already
3589 // handled, so this is best done with a single constant-pool load.
3593 // Let legalizer expand 2-wide build_vectors.
3594 if (EVTBits == 64) {
3595 if (NumNonZero == 1) {
3596 // One half is zero or undef.
3597 unsigned Idx = CountTrailingZeros_32(NonZeros);
3598 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3599 Op.getOperand(Idx));
3600 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3601 Subtarget->hasSSE2(), DAG);
3606 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3607 if (EVTBits == 8 && NumElems == 16) {
3608 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3610 if (V.getNode()) return V;
3613 if (EVTBits == 16 && NumElems == 8) {
3614 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3616 if (V.getNode()) return V;
3619 // If element VT is == 32 bits, turn it into a number of shuffles.
3620 SmallVector<SDValue, 8> V;
3622 if (NumElems == 4 && NumZero > 0) {
3623 for (unsigned i = 0; i < 4; ++i) {
3624 bool isZero = !(NonZeros & (1 << i));
3626 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3628 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3631 for (unsigned i = 0; i < 2; ++i) {
3632 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3635 V[i] = V[i*2]; // Must be a zero vector.
3638 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3641 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3644 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3649 SmallVector<int, 8> MaskVec;
3650 bool Reverse = (NonZeros & 0x3) == 2;
3651 for (unsigned i = 0; i < 2; ++i)
3652 MaskVec.push_back(Reverse ? 1-i : i);
3653 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3654 for (unsigned i = 0; i < 2; ++i)
3655 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3656 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3659 if (Values.size() > 2) {
3660 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3661 // values to be inserted is equal to the number of elements, in which case
3662 // use the unpack code below in the hopes of matching the consecutive elts
3663 // load merge pattern for shuffles.
3664 // FIXME: We could probably just check that here directly.
3665 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3666 getSubtarget()->hasSSE41()) {
3667 V[0] = DAG.getUNDEF(VT);
3668 for (unsigned i = 0; i < NumElems; ++i)
3669 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3670 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3671 Op.getOperand(i), DAG.getIntPtrConstant(i));
3674 // Expand into a number of unpckl*.
3676 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3677 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3678 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3679 for (unsigned i = 0; i < NumElems; ++i)
3680 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3682 while (NumElems != 0) {
3683 for (unsigned i = 0; i < NumElems; ++i)
3684 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3694 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3695 // We support concatenate two MMX registers and place them in a MMX
3696 // register. This is better than doing a stack convert.
3697 DebugLoc dl = Op.getDebugLoc();
3698 EVT ResVT = Op.getValueType();
3699 assert(Op.getNumOperands() == 2);
3700 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3701 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3703 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3704 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3705 InVec = Op.getOperand(1);
3706 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3707 unsigned NumElts = ResVT.getVectorNumElements();
3708 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3709 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3710 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3712 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3713 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3714 Mask[0] = 0; Mask[1] = 2;
3715 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3717 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3720 // v8i16 shuffles - Prefer shuffles in the following order:
3721 // 1. [all] pshuflw, pshufhw, optional move
3722 // 2. [ssse3] 1 x pshufb
3723 // 3. [ssse3] 2 x pshufb + 1 x por
3724 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3726 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3727 SelectionDAG &DAG, X86TargetLowering &TLI) {
3728 SDValue V1 = SVOp->getOperand(0);
3729 SDValue V2 = SVOp->getOperand(1);
3730 DebugLoc dl = SVOp->getDebugLoc();
3731 SmallVector<int, 8> MaskVals;
3733 // Determine if more than 1 of the words in each of the low and high quadwords
3734 // of the result come from the same quadword of one of the two inputs. Undef
3735 // mask values count as coming from any quadword, for better codegen.
3736 SmallVector<unsigned, 4> LoQuad(4);
3737 SmallVector<unsigned, 4> HiQuad(4);
3738 BitVector InputQuads(4);
3739 for (unsigned i = 0; i < 8; ++i) {
3740 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3741 int EltIdx = SVOp->getMaskElt(i);
3742 MaskVals.push_back(EltIdx);
3751 InputQuads.set(EltIdx / 4);
3754 int BestLoQuad = -1;
3755 unsigned MaxQuad = 1;
3756 for (unsigned i = 0; i < 4; ++i) {
3757 if (LoQuad[i] > MaxQuad) {
3759 MaxQuad = LoQuad[i];
3763 int BestHiQuad = -1;
3765 for (unsigned i = 0; i < 4; ++i) {
3766 if (HiQuad[i] > MaxQuad) {
3768 MaxQuad = HiQuad[i];
3772 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3773 // of the two input vectors, shuffle them into one input vector so only a
3774 // single pshufb instruction is necessary. If There are more than 2 input
3775 // quads, disable the next transformation since it does not help SSSE3.
3776 bool V1Used = InputQuads[0] || InputQuads[1];
3777 bool V2Used = InputQuads[2] || InputQuads[3];
3778 if (TLI.getSubtarget()->hasSSSE3()) {
3779 if (InputQuads.count() == 2 && V1Used && V2Used) {
3780 BestLoQuad = InputQuads.find_first();
3781 BestHiQuad = InputQuads.find_next(BestLoQuad);
3783 if (InputQuads.count() > 2) {
3789 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3790 // the shuffle mask. If a quad is scored as -1, that means that it contains
3791 // words from all 4 input quadwords.
3793 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3794 SmallVector<int, 8> MaskV;
3795 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3796 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3797 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3798 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3799 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3800 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3802 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3803 // source words for the shuffle, to aid later transformations.
3804 bool AllWordsInNewV = true;
3805 bool InOrder[2] = { true, true };
3806 for (unsigned i = 0; i != 8; ++i) {
3807 int idx = MaskVals[i];
3809 InOrder[i/4] = false;
3810 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3812 AllWordsInNewV = false;
3816 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3817 if (AllWordsInNewV) {
3818 for (int i = 0; i != 8; ++i) {
3819 int idx = MaskVals[i];
3822 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3823 if ((idx != i) && idx < 4)
3825 if ((idx != i) && idx > 3)
3834 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3835 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3836 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3837 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3838 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3842 // If we have SSSE3, and all words of the result are from 1 input vector,
3843 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3844 // is present, fall back to case 4.
3845 if (TLI.getSubtarget()->hasSSSE3()) {
3846 SmallVector<SDValue,16> pshufbMask;
3848 // If we have elements from both input vectors, set the high bit of the
3849 // shuffle mask element to zero out elements that come from V2 in the V1
3850 // mask, and elements that come from V1 in the V2 mask, so that the two
3851 // results can be OR'd together.
3852 bool TwoInputs = V1Used && V2Used;
3853 for (unsigned i = 0; i != 8; ++i) {
3854 int EltIdx = MaskVals[i] * 2;
3855 if (TwoInputs && (EltIdx >= 16)) {
3856 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3857 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3860 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3861 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3863 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3864 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3865 DAG.getNode(ISD::BUILD_VECTOR, dl,
3866 MVT::v16i8, &pshufbMask[0], 16));
3868 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3870 // Calculate the shuffle mask for the second input, shuffle it, and
3871 // OR it with the first shuffled input.
3873 for (unsigned i = 0; i != 8; ++i) {
3874 int EltIdx = MaskVals[i] * 2;
3876 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3877 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3880 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3881 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3883 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3884 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3885 DAG.getNode(ISD::BUILD_VECTOR, dl,
3886 MVT::v16i8, &pshufbMask[0], 16));
3887 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3888 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3891 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3892 // and update MaskVals with new element order.
3893 BitVector InOrder(8);
3894 if (BestLoQuad >= 0) {
3895 SmallVector<int, 8> MaskV;
3896 for (int i = 0; i != 4; ++i) {
3897 int idx = MaskVals[i];
3899 MaskV.push_back(-1);
3901 } else if ((idx / 4) == BestLoQuad) {
3902 MaskV.push_back(idx & 3);
3905 MaskV.push_back(-1);
3908 for (unsigned i = 4; i != 8; ++i)
3910 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3914 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3915 // and update MaskVals with the new element order.
3916 if (BestHiQuad >= 0) {
3917 SmallVector<int, 8> MaskV;
3918 for (unsigned i = 0; i != 4; ++i)
3920 for (unsigned i = 4; i != 8; ++i) {
3921 int idx = MaskVals[i];
3923 MaskV.push_back(-1);
3925 } else if ((idx / 4) == BestHiQuad) {
3926 MaskV.push_back((idx & 3) + 4);
3929 MaskV.push_back(-1);
3932 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3936 // In case BestHi & BestLo were both -1, which means each quadword has a word
3937 // from each of the four input quadwords, calculate the InOrder bitvector now
3938 // before falling through to the insert/extract cleanup.
3939 if (BestLoQuad == -1 && BestHiQuad == -1) {
3941 for (int i = 0; i != 8; ++i)
3942 if (MaskVals[i] < 0 || MaskVals[i] == i)
3946 // The other elements are put in the right place using pextrw and pinsrw.
3947 for (unsigned i = 0; i != 8; ++i) {
3950 int EltIdx = MaskVals[i];
3953 SDValue ExtOp = (EltIdx < 8)
3954 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3955 DAG.getIntPtrConstant(EltIdx))
3956 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3957 DAG.getIntPtrConstant(EltIdx - 8));
3958 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3959 DAG.getIntPtrConstant(i));
3964 // v16i8 shuffles - Prefer shuffles in the following order:
3965 // 1. [ssse3] 1 x pshufb
3966 // 2. [ssse3] 2 x pshufb + 1 x por
3967 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3969 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3970 SelectionDAG &DAG, X86TargetLowering &TLI) {
3971 SDValue V1 = SVOp->getOperand(0);
3972 SDValue V2 = SVOp->getOperand(1);
3973 DebugLoc dl = SVOp->getDebugLoc();
3974 SmallVector<int, 16> MaskVals;
3975 SVOp->getMask(MaskVals);
3977 // If we have SSSE3, case 1 is generated when all result bytes come from
3978 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3979 // present, fall back to case 3.
3980 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3983 for (unsigned i = 0; i < 16; ++i) {
3984 int EltIdx = MaskVals[i];
3993 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3994 if (TLI.getSubtarget()->hasSSSE3()) {
3995 SmallVector<SDValue,16> pshufbMask;
3997 // If all result elements are from one input vector, then only translate
3998 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4000 // Otherwise, we have elements from both input vectors, and must zero out
4001 // elements that come from V2 in the first mask, and V1 in the second mask
4002 // so that we can OR them together.
4003 bool TwoInputs = !(V1Only || V2Only);
4004 for (unsigned i = 0; i != 16; ++i) {
4005 int EltIdx = MaskVals[i];
4006 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4007 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4010 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4012 // If all the elements are from V2, assign it to V1 and return after
4013 // building the first pshufb.
4016 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4017 DAG.getNode(ISD::BUILD_VECTOR, dl,
4018 MVT::v16i8, &pshufbMask[0], 16));
4022 // Calculate the shuffle mask for the second input, shuffle it, and
4023 // OR it with the first shuffled input.
4025 for (unsigned i = 0; i != 16; ++i) {
4026 int EltIdx = MaskVals[i];
4028 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4031 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4033 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4034 DAG.getNode(ISD::BUILD_VECTOR, dl,
4035 MVT::v16i8, &pshufbMask[0], 16));
4036 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4039 // No SSSE3 - Calculate in place words and then fix all out of place words
4040 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4041 // the 16 different words that comprise the two doublequadword input vectors.
4042 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4043 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4044 SDValue NewV = V2Only ? V2 : V1;
4045 for (int i = 0; i != 8; ++i) {
4046 int Elt0 = MaskVals[i*2];
4047 int Elt1 = MaskVals[i*2+1];
4049 // This word of the result is all undef, skip it.
4050 if (Elt0 < 0 && Elt1 < 0)
4053 // This word of the result is already in the correct place, skip it.
4054 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4056 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4059 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4060 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4063 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4064 // using a single extract together, load it and store it.
4065 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4066 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4067 DAG.getIntPtrConstant(Elt1 / 2));
4068 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4069 DAG.getIntPtrConstant(i));
4073 // If Elt1 is defined, extract it from the appropriate source. If the
4074 // source byte is not also odd, shift the extracted word left 8 bits
4075 // otherwise clear the bottom 8 bits if we need to do an or.
4077 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4078 DAG.getIntPtrConstant(Elt1 / 2));
4079 if ((Elt1 & 1) == 0)
4080 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4081 DAG.getConstant(8, TLI.getShiftAmountTy()));
4083 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4084 DAG.getConstant(0xFF00, MVT::i16));
4086 // If Elt0 is defined, extract it from the appropriate source. If the
4087 // source byte is not also even, shift the extracted word right 8 bits. If
4088 // Elt1 was also defined, OR the extracted values together before
4089 // inserting them in the result.
4091 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4092 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4093 if ((Elt0 & 1) != 0)
4094 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4095 DAG.getConstant(8, TLI.getShiftAmountTy()));
4097 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4098 DAG.getConstant(0x00FF, MVT::i16));
4099 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4102 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4103 DAG.getIntPtrConstant(i));
4105 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4108 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4109 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4110 /// done when every pair / quad of shuffle mask elements point to elements in
4111 /// the right sequence. e.g.
4112 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4114 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4116 TargetLowering &TLI, DebugLoc dl) {
4117 EVT VT = SVOp->getValueType(0);
4118 SDValue V1 = SVOp->getOperand(0);
4119 SDValue V2 = SVOp->getOperand(1);
4120 unsigned NumElems = VT.getVectorNumElements();
4121 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4122 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4123 EVT MaskEltVT = MaskVT.getVectorElementType();
4125 switch (VT.getSimpleVT().SimpleTy) {
4126 default: assert(false && "Unexpected!");
4127 case MVT::v4f32: NewVT = MVT::v2f64; break;
4128 case MVT::v4i32: NewVT = MVT::v2i64; break;
4129 case MVT::v8i16: NewVT = MVT::v4i32; break;
4130 case MVT::v16i8: NewVT = MVT::v4i32; break;
4133 if (NewWidth == 2) {
4139 int Scale = NumElems / NewWidth;
4140 SmallVector<int, 8> MaskVec;
4141 for (unsigned i = 0; i < NumElems; i += Scale) {
4143 for (int j = 0; j < Scale; ++j) {
4144 int EltIdx = SVOp->getMaskElt(i+j);
4148 StartIdx = EltIdx - (EltIdx % Scale);
4149 if (EltIdx != StartIdx + j)
4153 MaskVec.push_back(-1);
4155 MaskVec.push_back(StartIdx / Scale);
4158 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4159 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4160 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4163 /// getVZextMovL - Return a zero-extending vector move low node.
4165 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4166 SDValue SrcOp, SelectionDAG &DAG,
4167 const X86Subtarget *Subtarget, DebugLoc dl) {
4168 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4169 LoadSDNode *LD = NULL;
4170 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4171 LD = dyn_cast<LoadSDNode>(SrcOp);
4173 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4175 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4176 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4177 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4178 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4179 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4181 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4182 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4183 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4184 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4192 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4193 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4194 DAG.getNode(ISD::BIT_CONVERT, dl,
4198 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4201 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4202 SDValue V1 = SVOp->getOperand(0);
4203 SDValue V2 = SVOp->getOperand(1);
4204 DebugLoc dl = SVOp->getDebugLoc();
4205 EVT VT = SVOp->getValueType(0);
4207 SmallVector<std::pair<int, int>, 8> Locs;
4209 SmallVector<int, 8> Mask1(4U, -1);
4210 SmallVector<int, 8> PermMask;
4211 SVOp->getMask(PermMask);
4215 for (unsigned i = 0; i != 4; ++i) {
4216 int Idx = PermMask[i];
4218 Locs[i] = std::make_pair(-1, -1);
4220 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4222 Locs[i] = std::make_pair(0, NumLo);
4226 Locs[i] = std::make_pair(1, NumHi);
4228 Mask1[2+NumHi] = Idx;
4234 if (NumLo <= 2 && NumHi <= 2) {
4235 // If no more than two elements come from either vector. This can be
4236 // implemented with two shuffles. First shuffle gather the elements.
4237 // The second shuffle, which takes the first shuffle as both of its
4238 // vector operands, put the elements into the right order.
4239 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4241 SmallVector<int, 8> Mask2(4U, -1);
4243 for (unsigned i = 0; i != 4; ++i) {
4244 if (Locs[i].first == -1)
4247 unsigned Idx = (i < 2) ? 0 : 4;
4248 Idx += Locs[i].first * 2 + Locs[i].second;
4253 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4254 } else if (NumLo == 3 || NumHi == 3) {
4255 // Otherwise, we must have three elements from one vector, call it X, and
4256 // one element from the other, call it Y. First, use a shufps to build an
4257 // intermediate vector with the one element from Y and the element from X
4258 // that will be in the same half in the final destination (the indexes don't
4259 // matter). Then, use a shufps to build the final vector, taking the half
4260 // containing the element from Y from the intermediate, and the other half
4263 // Normalize it so the 3 elements come from V1.
4264 CommuteVectorShuffleMask(PermMask, VT);
4268 // Find the element from V2.
4270 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4271 int Val = PermMask[HiIndex];
4278 Mask1[0] = PermMask[HiIndex];
4280 Mask1[2] = PermMask[HiIndex^1];
4282 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4285 Mask1[0] = PermMask[0];
4286 Mask1[1] = PermMask[1];
4287 Mask1[2] = HiIndex & 1 ? 6 : 4;
4288 Mask1[3] = HiIndex & 1 ? 4 : 6;
4289 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4291 Mask1[0] = HiIndex & 1 ? 2 : 0;
4292 Mask1[1] = HiIndex & 1 ? 0 : 2;
4293 Mask1[2] = PermMask[2];
4294 Mask1[3] = PermMask[3];
4299 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4303 // Break it into (shuffle shuffle_hi, shuffle_lo).
4305 SmallVector<int,8> LoMask(4U, -1);
4306 SmallVector<int,8> HiMask(4U, -1);
4308 SmallVector<int,8> *MaskPtr = &LoMask;
4309 unsigned MaskIdx = 0;
4312 for (unsigned i = 0; i != 4; ++i) {
4319 int Idx = PermMask[i];
4321 Locs[i] = std::make_pair(-1, -1);
4322 } else if (Idx < 4) {
4323 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4324 (*MaskPtr)[LoIdx] = Idx;
4327 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4328 (*MaskPtr)[HiIdx] = Idx;
4333 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4334 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4335 SmallVector<int, 8> MaskOps;
4336 for (unsigned i = 0; i != 4; ++i) {
4337 if (Locs[i].first == -1) {
4338 MaskOps.push_back(-1);
4340 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4341 MaskOps.push_back(Idx);
4344 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4348 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4349 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4350 SDValue V1 = Op.getOperand(0);
4351 SDValue V2 = Op.getOperand(1);
4352 EVT VT = Op.getValueType();
4353 DebugLoc dl = Op.getDebugLoc();
4354 unsigned NumElems = VT.getVectorNumElements();
4355 bool isMMX = VT.getSizeInBits() == 64;
4356 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4357 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4358 bool V1IsSplat = false;
4359 bool V2IsSplat = false;
4361 if (isZeroShuffle(SVOp))
4362 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4364 // Promote splats to v4f32.
4365 if (SVOp->isSplat()) {
4366 if (isMMX || NumElems < 4)
4368 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4371 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4373 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4374 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4375 if (NewOp.getNode())
4376 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4377 LowerVECTOR_SHUFFLE(NewOp, DAG));
4378 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4379 // FIXME: Figure out a cleaner way to do this.
4380 // Try to make use of movq to zero out the top part.
4381 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4382 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4383 if (NewOp.getNode()) {
4384 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4385 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4386 DAG, Subtarget, dl);
4388 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4389 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4390 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4391 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4392 DAG, Subtarget, dl);
4396 if (X86::isPSHUFDMask(SVOp))
4399 // Check if this can be converted into a logical shift.
4400 bool isLeft = false;
4403 bool isShift = getSubtarget()->hasSSE2() &&
4404 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4405 if (isShift && ShVal.hasOneUse()) {
4406 // If the shifted value has multiple uses, it may be cheaper to use
4407 // v_set0 + movlhps or movhlps, etc.
4408 EVT EltVT = VT.getVectorElementType();
4409 ShAmt *= EltVT.getSizeInBits();
4410 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4413 if (X86::isMOVLMask(SVOp)) {
4416 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4417 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4422 // FIXME: fold these into legal mask.
4423 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4424 X86::isMOVSLDUPMask(SVOp) ||
4425 X86::isMOVHLPSMask(SVOp) ||
4426 X86::isMOVLHPSMask(SVOp) ||
4427 X86::isMOVLPMask(SVOp)))
4430 if (ShouldXformToMOVHLPS(SVOp) ||
4431 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4432 return CommuteVectorShuffle(SVOp, DAG);
4435 // No better options. Use a vshl / vsrl.
4436 EVT EltVT = VT.getVectorElementType();
4437 ShAmt *= EltVT.getSizeInBits();
4438 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4441 bool Commuted = false;
4442 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4443 // 1,1,1,1 -> v8i16 though.
4444 V1IsSplat = isSplatVector(V1.getNode());
4445 V2IsSplat = isSplatVector(V2.getNode());
4447 // Canonicalize the splat or undef, if present, to be on the RHS.
4448 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4449 Op = CommuteVectorShuffle(SVOp, DAG);
4450 SVOp = cast<ShuffleVectorSDNode>(Op);
4451 V1 = SVOp->getOperand(0);
4452 V2 = SVOp->getOperand(1);
4453 std::swap(V1IsSplat, V2IsSplat);
4454 std::swap(V1IsUndef, V2IsUndef);
4458 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4459 // Shuffling low element of v1 into undef, just return v1.
4462 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4463 // the instruction selector will not match, so get a canonical MOVL with
4464 // swapped operands to undo the commute.
4465 return getMOVL(DAG, dl, VT, V2, V1);
4468 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4469 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4470 X86::isUNPCKLMask(SVOp) ||
4471 X86::isUNPCKHMask(SVOp))
4475 // Normalize mask so all entries that point to V2 points to its first
4476 // element then try to match unpck{h|l} again. If match, return a
4477 // new vector_shuffle with the corrected mask.
4478 SDValue NewMask = NormalizeMask(SVOp, DAG);
4479 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4480 if (NSVOp != SVOp) {
4481 if (X86::isUNPCKLMask(NSVOp, true)) {
4483 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4490 // Commute is back and try unpck* again.
4491 // FIXME: this seems wrong.
4492 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4493 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4494 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4495 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4496 X86::isUNPCKLMask(NewSVOp) ||
4497 X86::isUNPCKHMask(NewSVOp))
4501 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4503 // Normalize the node to match x86 shuffle ops if needed
4504 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4505 return CommuteVectorShuffle(SVOp, DAG);
4507 // Check for legal shuffle and return?
4508 SmallVector<int, 16> PermMask;
4509 SVOp->getMask(PermMask);
4510 if (isShuffleMaskLegal(PermMask, VT))
4513 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4514 if (VT == MVT::v8i16) {
4515 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4516 if (NewOp.getNode())
4520 if (VT == MVT::v16i8) {
4521 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4522 if (NewOp.getNode())
4526 // Handle all 4 wide cases with a number of shuffles except for MMX.
4527 if (NumElems == 4 && !isMMX)
4528 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4534 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4535 SelectionDAG &DAG) {
4536 EVT VT = Op.getValueType();
4537 DebugLoc dl = Op.getDebugLoc();
4538 if (VT.getSizeInBits() == 8) {
4539 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4540 Op.getOperand(0), Op.getOperand(1));
4541 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4542 DAG.getValueType(VT));
4543 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4544 } else if (VT.getSizeInBits() == 16) {
4545 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4546 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4548 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4549 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4550 DAG.getNode(ISD::BIT_CONVERT, dl,
4554 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4555 Op.getOperand(0), Op.getOperand(1));
4556 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4557 DAG.getValueType(VT));
4558 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4559 } else if (VT == MVT::f32) {
4560 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4561 // the result back to FR32 register. It's only worth matching if the
4562 // result has a single use which is a store or a bitcast to i32. And in
4563 // the case of a store, it's not worth it if the index is a constant 0,
4564 // because a MOVSSmr can be used instead, which is smaller and faster.
4565 if (!Op.hasOneUse())
4567 SDNode *User = *Op.getNode()->use_begin();
4568 if ((User->getOpcode() != ISD::STORE ||
4569 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4570 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4571 (User->getOpcode() != ISD::BIT_CONVERT ||
4572 User->getValueType(0) != MVT::i32))
4574 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4575 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4578 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4579 } else if (VT == MVT::i32) {
4580 // ExtractPS works with constant index.
4581 if (isa<ConstantSDNode>(Op.getOperand(1)))
4589 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4590 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4593 if (Subtarget->hasSSE41()) {
4594 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4599 EVT VT = Op.getValueType();
4600 DebugLoc dl = Op.getDebugLoc();
4601 // TODO: handle v16i8.
4602 if (VT.getSizeInBits() == 16) {
4603 SDValue Vec = Op.getOperand(0);
4604 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4606 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4607 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4608 DAG.getNode(ISD::BIT_CONVERT, dl,
4611 // Transform it so it match pextrw which produces a 32-bit result.
4612 EVT EltVT = MVT::i32;
4613 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4614 Op.getOperand(0), Op.getOperand(1));
4615 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4616 DAG.getValueType(VT));
4617 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4618 } else if (VT.getSizeInBits() == 32) {
4619 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4623 // SHUFPS the element to the lowest double word, then movss.
4624 int Mask[4] = { Idx, -1, -1, -1 };
4625 EVT VVT = Op.getOperand(0).getValueType();
4626 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4627 DAG.getUNDEF(VVT), Mask);
4628 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4629 DAG.getIntPtrConstant(0));
4630 } else if (VT.getSizeInBits() == 64) {
4631 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4632 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4633 // to match extract_elt for f64.
4634 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4638 // UNPCKHPD the element to the lowest double word, then movsd.
4639 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4640 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4641 int Mask[2] = { 1, -1 };
4642 EVT VVT = Op.getOperand(0).getValueType();
4643 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4644 DAG.getUNDEF(VVT), Mask);
4645 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4646 DAG.getIntPtrConstant(0));
4653 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4654 EVT VT = Op.getValueType();
4655 EVT EltVT = VT.getVectorElementType();
4656 DebugLoc dl = Op.getDebugLoc();
4658 SDValue N0 = Op.getOperand(0);
4659 SDValue N1 = Op.getOperand(1);
4660 SDValue N2 = Op.getOperand(2);
4662 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4663 isa<ConstantSDNode>(N2)) {
4664 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4666 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4668 if (N1.getValueType() != MVT::i32)
4669 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4670 if (N2.getValueType() != MVT::i32)
4671 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4672 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4673 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4674 // Bits [7:6] of the constant are the source select. This will always be
4675 // zero here. The DAG Combiner may combine an extract_elt index into these
4676 // bits. For example (insert (extract, 3), 2) could be matched by putting
4677 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4678 // Bits [5:4] of the constant are the destination select. This is the
4679 // value of the incoming immediate.
4680 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4681 // combine either bitwise AND or insert of float 0.0 to set these bits.
4682 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4683 // Create this as a scalar to vector..
4684 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4685 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4686 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4687 // PINSR* works with constant index.
4694 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4695 EVT VT = Op.getValueType();
4696 EVT EltVT = VT.getVectorElementType();
4698 if (Subtarget->hasSSE41())
4699 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4701 if (EltVT == MVT::i8)
4704 DebugLoc dl = Op.getDebugLoc();
4705 SDValue N0 = Op.getOperand(0);
4706 SDValue N1 = Op.getOperand(1);
4707 SDValue N2 = Op.getOperand(2);
4709 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4710 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4711 // as its second argument.
4712 if (N1.getValueType() != MVT::i32)
4713 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4714 if (N2.getValueType() != MVT::i32)
4715 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4716 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4722 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4723 DebugLoc dl = Op.getDebugLoc();
4724 if (Op.getValueType() == MVT::v2f32)
4725 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4726 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4727 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4728 Op.getOperand(0))));
4730 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4731 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4733 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4734 EVT VT = MVT::v2i32;
4735 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4742 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4743 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4746 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4747 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4748 // one of the above mentioned nodes. It has to be wrapped because otherwise
4749 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4750 // be used to form addressing mode. These wrapped nodes will be selected
4753 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4754 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4756 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4758 unsigned char OpFlag = 0;
4759 unsigned WrapperKind = X86ISD::Wrapper;
4760 CodeModel::Model M = getTargetMachine().getCodeModel();
4762 if (Subtarget->isPICStyleRIPRel() &&
4763 (M == CodeModel::Small || M == CodeModel::Kernel))
4764 WrapperKind = X86ISD::WrapperRIP;
4765 else if (Subtarget->isPICStyleGOT())
4766 OpFlag = X86II::MO_GOTOFF;
4767 else if (Subtarget->isPICStyleStubPIC())
4768 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4770 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4772 CP->getOffset(), OpFlag);
4773 DebugLoc DL = CP->getDebugLoc();
4774 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4775 // With PIC, the address is actually $g + Offset.
4777 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4778 DAG.getNode(X86ISD::GlobalBaseReg,
4779 DebugLoc::getUnknownLoc(), getPointerTy()),
4786 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4787 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4789 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4791 unsigned char OpFlag = 0;
4792 unsigned WrapperKind = X86ISD::Wrapper;
4793 CodeModel::Model M = getTargetMachine().getCodeModel();
4795 if (Subtarget->isPICStyleRIPRel() &&
4796 (M == CodeModel::Small || M == CodeModel::Kernel))
4797 WrapperKind = X86ISD::WrapperRIP;
4798 else if (Subtarget->isPICStyleGOT())
4799 OpFlag = X86II::MO_GOTOFF;
4800 else if (Subtarget->isPICStyleStubPIC())
4801 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4803 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4805 DebugLoc DL = JT->getDebugLoc();
4806 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4808 // With PIC, the address is actually $g + Offset.
4810 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4811 DAG.getNode(X86ISD::GlobalBaseReg,
4812 DebugLoc::getUnknownLoc(), getPointerTy()),
4820 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4821 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4823 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4825 unsigned char OpFlag = 0;
4826 unsigned WrapperKind = X86ISD::Wrapper;
4827 CodeModel::Model M = getTargetMachine().getCodeModel();
4829 if (Subtarget->isPICStyleRIPRel() &&
4830 (M == CodeModel::Small || M == CodeModel::Kernel))
4831 WrapperKind = X86ISD::WrapperRIP;
4832 else if (Subtarget->isPICStyleGOT())
4833 OpFlag = X86II::MO_GOTOFF;
4834 else if (Subtarget->isPICStyleStubPIC())
4835 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4837 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4839 DebugLoc DL = Op.getDebugLoc();
4840 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4843 // With PIC, the address is actually $g + Offset.
4844 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4845 !Subtarget->is64Bit()) {
4846 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4847 DAG.getNode(X86ISD::GlobalBaseReg,
4848 DebugLoc::getUnknownLoc(),
4857 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
4858 // Create the TargetBlockAddressAddress node.
4859 unsigned char OpFlags =
4860 Subtarget->ClassifyBlockAddressReference();
4861 CodeModel::Model M = getTargetMachine().getCodeModel();
4862 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4863 DebugLoc dl = Op.getDebugLoc();
4864 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
4865 /*isTarget=*/true, OpFlags);
4867 if (Subtarget->isPICStyleRIPRel() &&
4868 (M == CodeModel::Small || M == CodeModel::Kernel))
4869 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4871 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4873 // With PIC, the address is actually $g + Offset.
4874 if (isGlobalRelativeToPICBase(OpFlags)) {
4875 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4876 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4884 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4886 SelectionDAG &DAG) const {
4887 // Create the TargetGlobalAddress node, folding in the constant
4888 // offset if it is legal.
4889 unsigned char OpFlags =
4890 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4891 CodeModel::Model M = getTargetMachine().getCodeModel();
4893 if (OpFlags == X86II::MO_NO_FLAG &&
4894 X86::isOffsetSuitableForCodeModel(Offset, M)) {
4895 // A direct static reference to a global.
4896 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4899 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4902 if (Subtarget->isPICStyleRIPRel() &&
4903 (M == CodeModel::Small || M == CodeModel::Kernel))
4904 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4906 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4908 // With PIC, the address is actually $g + Offset.
4909 if (isGlobalRelativeToPICBase(OpFlags)) {
4910 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4911 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4915 // For globals that require a load from a stub to get the address, emit the
4917 if (isGlobalStubReference(OpFlags))
4918 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4919 PseudoSourceValue::getGOT(), 0);
4921 // If there was a non-zero offset that we didn't fold, create an explicit
4924 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4925 DAG.getConstant(Offset, getPointerTy()));
4931 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4932 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4933 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4934 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4938 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4939 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
4940 unsigned char OperandFlags) {
4941 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4942 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4943 DebugLoc dl = GA->getDebugLoc();
4944 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4945 GA->getValueType(0),
4949 SDValue Ops[] = { Chain, TGA, *InFlag };
4950 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4952 SDValue Ops[] = { Chain, TGA };
4953 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4956 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
4957 MFI->setHasCalls(true);
4959 SDValue Flag = Chain.getValue(1);
4960 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4963 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4965 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4968 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4969 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4970 DAG.getNode(X86ISD::GlobalBaseReg,
4971 DebugLoc::getUnknownLoc(),
4973 InFlag = Chain.getValue(1);
4975 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4978 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4980 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4982 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4983 X86::RAX, X86II::MO_TLSGD);
4986 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4987 // "local exec" model.
4988 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4989 const EVT PtrVT, TLSModel::Model model,
4991 DebugLoc dl = GA->getDebugLoc();
4992 // Get the Thread Pointer
4993 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4994 DebugLoc::getUnknownLoc(), PtrVT,
4995 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4998 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5001 unsigned char OperandFlags = 0;
5002 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5004 unsigned WrapperKind = X86ISD::Wrapper;
5005 if (model == TLSModel::LocalExec) {
5006 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5007 } else if (is64Bit) {
5008 assert(model == TLSModel::InitialExec);
5009 OperandFlags = X86II::MO_GOTTPOFF;
5010 WrapperKind = X86ISD::WrapperRIP;
5012 assert(model == TLSModel::InitialExec);
5013 OperandFlags = X86II::MO_INDNTPOFF;
5016 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5018 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5019 GA->getOffset(), OperandFlags);
5020 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5022 if (model == TLSModel::InitialExec)
5023 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5024 PseudoSourceValue::getGOT(), 0);
5026 // The address of the thread local variable is the add of the thread
5027 // pointer with the offset of the variable.
5028 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5032 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5033 // TODO: implement the "local dynamic" model
5034 // TODO: implement the "initial exec"model for pic executables
5035 assert(Subtarget->isTargetELF() &&
5036 "TLS not implemented for non-ELF targets");
5037 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5038 const GlobalValue *GV = GA->getGlobal();
5040 // If GV is an alias then use the aliasee for determining
5041 // thread-localness.
5042 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5043 GV = GA->resolveAliasedGlobal(false);
5045 TLSModel::Model model = getTLSModel(GV,
5046 getTargetMachine().getRelocationModel());
5049 case TLSModel::GeneralDynamic:
5050 case TLSModel::LocalDynamic: // not implemented
5051 if (Subtarget->is64Bit())
5052 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5053 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5055 case TLSModel::InitialExec:
5056 case TLSModel::LocalExec:
5057 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5058 Subtarget->is64Bit());
5061 llvm_unreachable("Unreachable");
5066 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5067 /// take a 2 x i32 value to shift plus a shift amount.
5068 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5069 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5070 EVT VT = Op.getValueType();
5071 unsigned VTBits = VT.getSizeInBits();
5072 DebugLoc dl = Op.getDebugLoc();
5073 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5074 SDValue ShOpLo = Op.getOperand(0);
5075 SDValue ShOpHi = Op.getOperand(1);
5076 SDValue ShAmt = Op.getOperand(2);
5077 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5078 DAG.getConstant(VTBits - 1, MVT::i8))
5079 : DAG.getConstant(0, VT);
5082 if (Op.getOpcode() == ISD::SHL_PARTS) {
5083 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5084 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5086 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5087 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5090 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5091 DAG.getConstant(VTBits, MVT::i8));
5092 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
5093 AndNode, DAG.getConstant(0, MVT::i8));
5096 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5097 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5098 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5100 if (Op.getOpcode() == ISD::SHL_PARTS) {
5101 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5102 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5104 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5105 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5108 SDValue Ops[2] = { Lo, Hi };
5109 return DAG.getMergeValues(Ops, 2, dl);
5112 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5113 EVT SrcVT = Op.getOperand(0).getValueType();
5115 if (SrcVT.isVector()) {
5116 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5122 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5123 "Unknown SINT_TO_FP to lower!");
5125 // These are really Legal; return the operand so the caller accepts it as
5127 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5129 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5130 Subtarget->is64Bit()) {
5134 DebugLoc dl = Op.getDebugLoc();
5135 unsigned Size = SrcVT.getSizeInBits()/8;
5136 MachineFunction &MF = DAG.getMachineFunction();
5137 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5138 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5139 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5141 PseudoSourceValue::getFixedStack(SSFI), 0);
5142 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5145 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5147 SelectionDAG &DAG) {
5149 DebugLoc dl = Op.getDebugLoc();
5151 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5153 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5155 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5156 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5157 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5158 Tys, Ops, array_lengthof(Ops));
5161 Chain = Result.getValue(1);
5162 SDValue InFlag = Result.getValue(2);
5164 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5165 // shouldn't be necessary except that RFP cannot be live across
5166 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5167 MachineFunction &MF = DAG.getMachineFunction();
5168 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5169 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5170 Tys = DAG.getVTList(MVT::Other);
5172 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5174 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5175 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5176 PseudoSourceValue::getFixedStack(SSFI), 0);
5182 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5183 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5184 // This algorithm is not obvious. Here it is in C code, more or less:
5186 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5187 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5188 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5190 // Copy ints to xmm registers.
5191 __m128i xh = _mm_cvtsi32_si128( hi );
5192 __m128i xl = _mm_cvtsi32_si128( lo );
5194 // Combine into low half of a single xmm register.
5195 __m128i x = _mm_unpacklo_epi32( xh, xl );
5199 // Merge in appropriate exponents to give the integer bits the right
5201 x = _mm_unpacklo_epi32( x, exp );
5203 // Subtract away the biases to deal with the IEEE-754 double precision
5205 d = _mm_sub_pd( (__m128d) x, bias );
5207 // All conversions up to here are exact. The correctly rounded result is
5208 // calculated using the current rounding mode using the following
5210 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5211 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5212 // store doesn't really need to be here (except
5213 // maybe to zero the other double)
5218 DebugLoc dl = Op.getDebugLoc();
5219 LLVMContext *Context = DAG.getContext();
5221 // Build some magic constants.
5222 std::vector<Constant*> CV0;
5223 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5224 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5225 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5226 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5227 Constant *C0 = ConstantVector::get(CV0);
5228 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5230 std::vector<Constant*> CV1;
5232 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5234 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5235 Constant *C1 = ConstantVector::get(CV1);
5236 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5238 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5239 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5241 DAG.getIntPtrConstant(1)));
5242 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5243 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5245 DAG.getIntPtrConstant(0)));
5246 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5247 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5248 PseudoSourceValue::getConstantPool(), 0,
5250 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5251 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5252 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5253 PseudoSourceValue::getConstantPool(), 0,
5255 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5257 // Add the halves; easiest way is to swap them into another reg first.
5258 int ShufMask[2] = { 1, -1 };
5259 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5260 DAG.getUNDEF(MVT::v2f64), ShufMask);
5261 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5262 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5263 DAG.getIntPtrConstant(0));
5266 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5267 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5268 DebugLoc dl = Op.getDebugLoc();
5269 // FP constant to bias correct the final result.
5270 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5273 // Load the 32-bit value into an XMM register.
5274 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5275 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5277 DAG.getIntPtrConstant(0)));
5279 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5280 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5281 DAG.getIntPtrConstant(0));
5283 // Or the load with the bias.
5284 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5285 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5286 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5288 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5289 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5290 MVT::v2f64, Bias)));
5291 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5292 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5293 DAG.getIntPtrConstant(0));
5295 // Subtract the bias.
5296 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5298 // Handle final rounding.
5299 EVT DestVT = Op.getValueType();
5301 if (DestVT.bitsLT(MVT::f64)) {
5302 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5303 DAG.getIntPtrConstant(0));
5304 } else if (DestVT.bitsGT(MVT::f64)) {
5305 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5308 // Handle final rounding.
5312 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5313 SDValue N0 = Op.getOperand(0);
5314 DebugLoc dl = Op.getDebugLoc();
5316 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5317 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5318 // the optimization here.
5319 if (DAG.SignBitIsZero(N0))
5320 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5322 EVT SrcVT = N0.getValueType();
5323 if (SrcVT == MVT::i64) {
5324 // We only handle SSE2 f64 target here; caller can expand the rest.
5325 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5328 return LowerUINT_TO_FP_i64(Op, DAG);
5329 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5330 return LowerUINT_TO_FP_i32(Op, DAG);
5333 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5335 // Make a 64-bit buffer, and use it to build an FILD.
5336 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5337 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5338 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5339 getPointerTy(), StackSlot, WordOff);
5340 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5341 StackSlot, NULL, 0);
5342 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5343 OffsetSlot, NULL, 0);
5344 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5347 std::pair<SDValue,SDValue> X86TargetLowering::
5348 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5349 DebugLoc dl = Op.getDebugLoc();
5351 EVT DstTy = Op.getValueType();
5354 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5358 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5359 DstTy.getSimpleVT() >= MVT::i16 &&
5360 "Unknown FP_TO_SINT to lower!");
5362 // These are really Legal.
5363 if (DstTy == MVT::i32 &&
5364 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5365 return std::make_pair(SDValue(), SDValue());
5366 if (Subtarget->is64Bit() &&
5367 DstTy == MVT::i64 &&
5368 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5369 return std::make_pair(SDValue(), SDValue());
5371 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5373 MachineFunction &MF = DAG.getMachineFunction();
5374 unsigned MemSize = DstTy.getSizeInBits()/8;
5375 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5376 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5379 switch (DstTy.getSimpleVT().SimpleTy) {
5380 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5381 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5382 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5383 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5386 SDValue Chain = DAG.getEntryNode();
5387 SDValue Value = Op.getOperand(0);
5388 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5389 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5390 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5391 PseudoSourceValue::getFixedStack(SSFI), 0);
5392 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5394 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5396 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5397 Chain = Value.getValue(1);
5398 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5399 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5402 // Build the FP_TO_INT*_IN_MEM
5403 SDValue Ops[] = { Chain, Value, StackSlot };
5404 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5406 return std::make_pair(FIST, StackSlot);
5409 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5410 if (Op.getValueType().isVector()) {
5411 if (Op.getValueType() == MVT::v2i32 &&
5412 Op.getOperand(0).getValueType() == MVT::v2f64) {
5418 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5419 SDValue FIST = Vals.first, StackSlot = Vals.second;
5420 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5421 if (FIST.getNode() == 0) return Op;
5424 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5425 FIST, StackSlot, NULL, 0);
5428 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5429 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5430 SDValue FIST = Vals.first, StackSlot = Vals.second;
5431 assert(FIST.getNode() && "Unexpected failure");
5434 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5435 FIST, StackSlot, NULL, 0);
5438 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5439 LLVMContext *Context = DAG.getContext();
5440 DebugLoc dl = Op.getDebugLoc();
5441 EVT VT = Op.getValueType();
5444 EltVT = VT.getVectorElementType();
5445 std::vector<Constant*> CV;
5446 if (EltVT == MVT::f64) {
5447 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5451 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5457 Constant *C = ConstantVector::get(CV);
5458 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5459 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5460 PseudoSourceValue::getConstantPool(), 0,
5462 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5465 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5466 LLVMContext *Context = DAG.getContext();
5467 DebugLoc dl = Op.getDebugLoc();
5468 EVT VT = Op.getValueType();
5471 EltVT = VT.getVectorElementType();
5472 std::vector<Constant*> CV;
5473 if (EltVT == MVT::f64) {
5474 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5478 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5484 Constant *C = ConstantVector::get(CV);
5485 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5486 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5487 PseudoSourceValue::getConstantPool(), 0,
5489 if (VT.isVector()) {
5490 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5491 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5492 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5494 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5496 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5500 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5501 LLVMContext *Context = DAG.getContext();
5502 SDValue Op0 = Op.getOperand(0);
5503 SDValue Op1 = Op.getOperand(1);
5504 DebugLoc dl = Op.getDebugLoc();
5505 EVT VT = Op.getValueType();
5506 EVT SrcVT = Op1.getValueType();
5508 // If second operand is smaller, extend it first.
5509 if (SrcVT.bitsLT(VT)) {
5510 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5513 // And if it is bigger, shrink it first.
5514 if (SrcVT.bitsGT(VT)) {
5515 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5519 // At this point the operands and the result should have the same
5520 // type, and that won't be f80 since that is not custom lowered.
5522 // First get the sign bit of second operand.
5523 std::vector<Constant*> CV;
5524 if (SrcVT == MVT::f64) {
5525 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5526 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5528 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5529 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5530 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5531 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5533 Constant *C = ConstantVector::get(CV);
5534 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5535 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5536 PseudoSourceValue::getConstantPool(), 0,
5538 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5540 // Shift sign bit right or left if the two operands have different types.
5541 if (SrcVT.bitsGT(VT)) {
5542 // Op0 is MVT::f32, Op1 is MVT::f64.
5543 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5544 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5545 DAG.getConstant(32, MVT::i32));
5546 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5547 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5548 DAG.getIntPtrConstant(0));
5551 // Clear first operand sign bit.
5553 if (VT == MVT::f64) {
5554 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5555 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5557 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5558 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5559 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5560 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5562 C = ConstantVector::get(CV);
5563 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5564 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5565 PseudoSourceValue::getConstantPool(), 0,
5567 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5569 // Or the value with the sign bit.
5570 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5573 /// Emit nodes that will be selected as "test Op0,Op0", or something
5575 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5576 SelectionDAG &DAG) {
5577 DebugLoc dl = Op.getDebugLoc();
5579 // CF and OF aren't always set the way we want. Determine which
5580 // of these we need.
5581 bool NeedCF = false;
5582 bool NeedOF = false;
5584 case X86::COND_A: case X86::COND_AE:
5585 case X86::COND_B: case X86::COND_BE:
5588 case X86::COND_G: case X86::COND_GE:
5589 case X86::COND_L: case X86::COND_LE:
5590 case X86::COND_O: case X86::COND_NO:
5596 // See if we can use the EFLAGS value from the operand instead of
5597 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5598 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5599 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5600 unsigned Opcode = 0;
5601 unsigned NumOperands = 0;
5602 switch (Op.getNode()->getOpcode()) {
5604 // Due to an isel shortcoming, be conservative if this add is likely to
5605 // be selected as part of a load-modify-store instruction. When the root
5606 // node in a match is a store, isel doesn't know how to remap non-chain
5607 // non-flag uses of other nodes in the match, such as the ADD in this
5608 // case. This leads to the ADD being left around and reselected, with
5609 // the result being two adds in the output.
5610 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5611 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5612 if (UI->getOpcode() == ISD::STORE)
5614 if (ConstantSDNode *C =
5615 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5616 // An add of one will be selected as an INC.
5617 if (C->getAPIntValue() == 1) {
5618 Opcode = X86ISD::INC;
5622 // An add of negative one (subtract of one) will be selected as a DEC.
5623 if (C->getAPIntValue().isAllOnesValue()) {
5624 Opcode = X86ISD::DEC;
5629 // Otherwise use a regular EFLAGS-setting add.
5630 Opcode = X86ISD::ADD;
5634 // If the primary and result isn't used, don't bother using X86ISD::AND,
5635 // because a TEST instruction will be better.
5636 bool NonFlagUse = false;
5637 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5638 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5640 unsigned UOpNo = UI.getOperandNo();
5641 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5642 // Look pass truncate.
5643 UOpNo = User->use_begin().getOperandNo();
5644 User = *User->use_begin();
5646 if (User->getOpcode() != ISD::BRCOND &&
5647 User->getOpcode() != ISD::SETCC &&
5648 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5660 // Due to the ISEL shortcoming noted above, be conservative if this op is
5661 // likely to be selected as part of a load-modify-store instruction.
5662 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5663 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5664 if (UI->getOpcode() == ISD::STORE)
5666 // Otherwise use a regular EFLAGS-setting instruction.
5667 switch (Op.getNode()->getOpcode()) {
5668 case ISD::SUB: Opcode = X86ISD::SUB; break;
5669 case ISD::OR: Opcode = X86ISD::OR; break;
5670 case ISD::XOR: Opcode = X86ISD::XOR; break;
5671 case ISD::AND: Opcode = X86ISD::AND; break;
5672 default: llvm_unreachable("unexpected operator!");
5683 return SDValue(Op.getNode(), 1);
5689 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5690 SmallVector<SDValue, 4> Ops;
5691 for (unsigned i = 0; i != NumOperands; ++i)
5692 Ops.push_back(Op.getOperand(i));
5693 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5694 DAG.ReplaceAllUsesWith(Op, New);
5695 return SDValue(New.getNode(), 1);
5699 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5700 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5701 DAG.getConstant(0, Op.getValueType()));
5704 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5706 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5707 SelectionDAG &DAG) {
5708 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5709 if (C->getAPIntValue() == 0)
5710 return EmitTest(Op0, X86CC, DAG);
5712 DebugLoc dl = Op0.getDebugLoc();
5713 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5716 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5717 /// if it's possible.
5718 static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
5719 DebugLoc dl, SelectionDAG &DAG) {
5721 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5722 if (ConstantSDNode *Op010C =
5723 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5724 if (Op010C->getZExtValue() == 1) {
5725 LHS = Op0.getOperand(0);
5726 RHS = Op0.getOperand(1).getOperand(1);
5728 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5729 if (ConstantSDNode *Op000C =
5730 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5731 if (Op000C->getZExtValue() == 1) {
5732 LHS = Op0.getOperand(1);
5733 RHS = Op0.getOperand(0).getOperand(1);
5735 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5736 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5737 SDValue AndLHS = Op0.getOperand(0);
5738 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5739 LHS = AndLHS.getOperand(0);
5740 RHS = AndLHS.getOperand(1);
5744 if (LHS.getNode()) {
5745 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5746 // instruction. Since the shift amount is in-range-or-undefined, we know
5747 // that doing a bittest on the i16 value is ok. We extend to i32 because
5748 // the encoding for the i16 version is larger than the i32 version.
5749 if (LHS.getValueType() == MVT::i8)
5750 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5752 // If the operand types disagree, extend the shift amount to match. Since
5753 // BT ignores high bits (like shifts) we can use anyextend.
5754 if (LHS.getValueType() != RHS.getValueType())
5755 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5757 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5758 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5759 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5760 DAG.getConstant(Cond, MVT::i8), BT);
5766 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5767 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5768 SDValue Op0 = Op.getOperand(0);
5769 SDValue Op1 = Op.getOperand(1);
5770 DebugLoc dl = Op.getDebugLoc();
5771 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5773 // Optimize to BT if possible.
5774 // Lower (X & (1 << N)) == 0 to BT(X, N).
5775 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5776 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5777 if (Op0.getOpcode() == ISD::AND &&
5779 Op1.getOpcode() == ISD::Constant &&
5780 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5781 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5782 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5783 if (NewSetCC.getNode())
5787 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5788 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5789 if (X86CC == X86::COND_INVALID)
5792 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5794 // Use sbb x, x to materialize carry bit into a GPR.
5795 if (X86CC == X86::COND_B)
5796 return DAG.getNode(ISD::AND, dl, MVT::i8,
5797 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5798 DAG.getConstant(X86CC, MVT::i8), Cond),
5799 DAG.getConstant(1, MVT::i8));
5801 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5802 DAG.getConstant(X86CC, MVT::i8), Cond);
5805 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5807 SDValue Op0 = Op.getOperand(0);
5808 SDValue Op1 = Op.getOperand(1);
5809 SDValue CC = Op.getOperand(2);
5810 EVT VT = Op.getValueType();
5811 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5812 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5813 DebugLoc dl = Op.getDebugLoc();
5817 EVT VT0 = Op0.getValueType();
5818 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5819 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5822 switch (SetCCOpcode) {
5825 case ISD::SETEQ: SSECC = 0; break;
5827 case ISD::SETGT: Swap = true; // Fallthrough
5829 case ISD::SETOLT: SSECC = 1; break;
5831 case ISD::SETGE: Swap = true; // Fallthrough
5833 case ISD::SETOLE: SSECC = 2; break;
5834 case ISD::SETUO: SSECC = 3; break;
5836 case ISD::SETNE: SSECC = 4; break;
5837 case ISD::SETULE: Swap = true;
5838 case ISD::SETUGE: SSECC = 5; break;
5839 case ISD::SETULT: Swap = true;
5840 case ISD::SETUGT: SSECC = 6; break;
5841 case ISD::SETO: SSECC = 7; break;
5844 std::swap(Op0, Op1);
5846 // In the two special cases we can't handle, emit two comparisons.
5848 if (SetCCOpcode == ISD::SETUEQ) {
5850 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5851 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5852 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5854 else if (SetCCOpcode == ISD::SETONE) {
5856 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5857 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5858 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5860 llvm_unreachable("Illegal FP comparison");
5862 // Handle all other FP comparisons here.
5863 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5866 // We are handling one of the integer comparisons here. Since SSE only has
5867 // GT and EQ comparisons for integer, swapping operands and multiple
5868 // operations may be required for some comparisons.
5869 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5870 bool Swap = false, Invert = false, FlipSigns = false;
5872 switch (VT.getSimpleVT().SimpleTy) {
5875 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5877 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5879 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5880 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5883 switch (SetCCOpcode) {
5885 case ISD::SETNE: Invert = true;
5886 case ISD::SETEQ: Opc = EQOpc; break;
5887 case ISD::SETLT: Swap = true;
5888 case ISD::SETGT: Opc = GTOpc; break;
5889 case ISD::SETGE: Swap = true;
5890 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5891 case ISD::SETULT: Swap = true;
5892 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5893 case ISD::SETUGE: Swap = true;
5894 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5897 std::swap(Op0, Op1);
5899 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5900 // bits of the inputs before performing those operations.
5902 EVT EltVT = VT.getVectorElementType();
5903 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5905 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5906 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5908 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5909 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5912 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5914 // If the logical-not of the result is required, perform that now.
5916 Result = DAG.getNOT(dl, Result, VT);
5921 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5922 static bool isX86LogicalCmp(SDValue Op) {
5923 unsigned Opc = Op.getNode()->getOpcode();
5924 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5926 if (Op.getResNo() == 1 &&
5927 (Opc == X86ISD::ADD ||
5928 Opc == X86ISD::SUB ||
5929 Opc == X86ISD::SMUL ||
5930 Opc == X86ISD::UMUL ||
5931 Opc == X86ISD::INC ||
5932 Opc == X86ISD::DEC ||
5933 Opc == X86ISD::OR ||
5934 Opc == X86ISD::XOR ||
5935 Opc == X86ISD::AND))
5941 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5942 bool addTest = true;
5943 SDValue Cond = Op.getOperand(0);
5944 DebugLoc dl = Op.getDebugLoc();
5947 if (Cond.getOpcode() == ISD::SETCC) {
5948 SDValue NewCond = LowerSETCC(Cond, DAG);
5949 if (NewCond.getNode())
5953 // Look pass (and (setcc_carry (cmp ...)), 1).
5954 if (Cond.getOpcode() == ISD::AND &&
5955 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
5956 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
5957 if (C && C->getAPIntValue() == 1)
5958 Cond = Cond.getOperand(0);
5961 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5962 // setting operand in place of the X86ISD::SETCC.
5963 if (Cond.getOpcode() == X86ISD::SETCC ||
5964 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
5965 CC = Cond.getOperand(0);
5967 SDValue Cmp = Cond.getOperand(1);
5968 unsigned Opc = Cmp.getOpcode();
5969 EVT VT = Op.getValueType();
5971 bool IllegalFPCMov = false;
5972 if (VT.isFloatingPoint() && !VT.isVector() &&
5973 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5974 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5976 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5977 Opc == X86ISD::BT) { // FIXME
5984 // Look pass the truncate.
5985 if (Cond.getOpcode() == ISD::TRUNCATE)
5986 Cond = Cond.getOperand(0);
5988 // We know the result of AND is compared against zero. Try to match
5990 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
5991 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
5992 if (NewSetCC.getNode()) {
5993 CC = NewSetCC.getOperand(0);
5994 Cond = NewSetCC.getOperand(1);
6001 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6002 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6005 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6006 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6007 // condition is true.
6008 SDValue Ops[] = { Op.getOperand(2), Op.getOperand(1), CC, Cond };
6009 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6012 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6013 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6014 // from the AND / OR.
6015 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6016 Opc = Op.getOpcode();
6017 if (Opc != ISD::OR && Opc != ISD::AND)
6019 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6020 Op.getOperand(0).hasOneUse() &&
6021 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6022 Op.getOperand(1).hasOneUse());
6025 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6026 // 1 and that the SETCC node has a single use.
6027 static bool isXor1OfSetCC(SDValue Op) {
6028 if (Op.getOpcode() != ISD::XOR)
6030 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6031 if (N1C && N1C->getAPIntValue() == 1) {
6032 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6033 Op.getOperand(0).hasOneUse();
6038 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6039 bool addTest = true;
6040 SDValue Chain = Op.getOperand(0);
6041 SDValue Cond = Op.getOperand(1);
6042 SDValue Dest = Op.getOperand(2);
6043 DebugLoc dl = Op.getDebugLoc();
6046 if (Cond.getOpcode() == ISD::SETCC) {
6047 SDValue NewCond = LowerSETCC(Cond, DAG);
6048 if (NewCond.getNode())
6052 // FIXME: LowerXALUO doesn't handle these!!
6053 else if (Cond.getOpcode() == X86ISD::ADD ||
6054 Cond.getOpcode() == X86ISD::SUB ||
6055 Cond.getOpcode() == X86ISD::SMUL ||
6056 Cond.getOpcode() == X86ISD::UMUL)
6057 Cond = LowerXALUO(Cond, DAG);
6060 // Look pass (and (setcc_carry (cmp ...)), 1).
6061 if (Cond.getOpcode() == ISD::AND &&
6062 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6063 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6064 if (C && C->getAPIntValue() == 1)
6065 Cond = Cond.getOperand(0);
6068 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6069 // setting operand in place of the X86ISD::SETCC.
6070 if (Cond.getOpcode() == X86ISD::SETCC ||
6071 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6072 CC = Cond.getOperand(0);
6074 SDValue Cmp = Cond.getOperand(1);
6075 unsigned Opc = Cmp.getOpcode();
6076 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6077 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6081 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6085 // These can only come from an arithmetic instruction with overflow,
6086 // e.g. SADDO, UADDO.
6087 Cond = Cond.getNode()->getOperand(1);
6094 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6095 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6096 if (CondOpc == ISD::OR) {
6097 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6098 // two branches instead of an explicit OR instruction with a
6100 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6101 isX86LogicalCmp(Cmp)) {
6102 CC = Cond.getOperand(0).getOperand(0);
6103 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6104 Chain, Dest, CC, Cmp);
6105 CC = Cond.getOperand(1).getOperand(0);
6109 } else { // ISD::AND
6110 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6111 // two branches instead of an explicit AND instruction with a
6112 // separate test. However, we only do this if this block doesn't
6113 // have a fall-through edge, because this requires an explicit
6114 // jmp when the condition is false.
6115 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6116 isX86LogicalCmp(Cmp) &&
6117 Op.getNode()->hasOneUse()) {
6118 X86::CondCode CCode =
6119 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6120 CCode = X86::GetOppositeBranchCondition(CCode);
6121 CC = DAG.getConstant(CCode, MVT::i8);
6122 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6123 // Look for an unconditional branch following this conditional branch.
6124 // We need this because we need to reverse the successors in order
6125 // to implement FCMP_OEQ.
6126 if (User.getOpcode() == ISD::BR) {
6127 SDValue FalseBB = User.getOperand(1);
6129 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6130 assert(NewBR == User);
6133 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6134 Chain, Dest, CC, Cmp);
6135 X86::CondCode CCode =
6136 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6137 CCode = X86::GetOppositeBranchCondition(CCode);
6138 CC = DAG.getConstant(CCode, MVT::i8);
6144 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6145 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6146 // It should be transformed during dag combiner except when the condition
6147 // is set by a arithmetics with overflow node.
6148 X86::CondCode CCode =
6149 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6150 CCode = X86::GetOppositeBranchCondition(CCode);
6151 CC = DAG.getConstant(CCode, MVT::i8);
6152 Cond = Cond.getOperand(0).getOperand(1);
6158 // Look pass the truncate.
6159 if (Cond.getOpcode() == ISD::TRUNCATE)
6160 Cond = Cond.getOperand(0);
6162 // We know the result of AND is compared against zero. Try to match
6164 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6165 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6166 if (NewSetCC.getNode()) {
6167 CC = NewSetCC.getOperand(0);
6168 Cond = NewSetCC.getOperand(1);
6175 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6176 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6178 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6179 Chain, Dest, CC, Cond);
6183 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6184 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6185 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6186 // that the guard pages used by the OS virtual memory manager are allocated in
6187 // correct sequence.
6189 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6190 SelectionDAG &DAG) {
6191 assert(Subtarget->isTargetCygMing() &&
6192 "This should be used only on Cygwin/Mingw targets");
6193 DebugLoc dl = Op.getDebugLoc();
6196 SDValue Chain = Op.getOperand(0);
6197 SDValue Size = Op.getOperand(1);
6198 // FIXME: Ensure alignment here
6202 EVT IntPtr = getPointerTy();
6203 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6205 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6207 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6208 Flag = Chain.getValue(1);
6210 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6211 SDValue Ops[] = { Chain,
6212 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6213 DAG.getRegister(X86::EAX, IntPtr),
6214 DAG.getRegister(X86StackPtr, SPTy),
6216 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6217 Flag = Chain.getValue(1);
6219 Chain = DAG.getCALLSEQ_END(Chain,
6220 DAG.getIntPtrConstant(0, true),
6221 DAG.getIntPtrConstant(0, true),
6224 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6226 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6227 return DAG.getMergeValues(Ops1, 2, dl);
6231 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6233 SDValue Dst, SDValue Src,
6234 SDValue Size, unsigned Align,
6236 uint64_t DstSVOff) {
6237 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6239 // If not DWORD aligned or size is more than the threshold, call the library.
6240 // The libc version is likely to be faster for these cases. It can use the
6241 // address value and run time information about the CPU.
6242 if ((Align & 3) != 0 ||
6244 ConstantSize->getZExtValue() >
6245 getSubtarget()->getMaxInlineSizeThreshold()) {
6246 SDValue InFlag(0, 0);
6248 // Check to see if there is a specialized entry-point for memory zeroing.
6249 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6251 if (const char *bzeroEntry = V &&
6252 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6253 EVT IntPtr = getPointerTy();
6254 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6255 TargetLowering::ArgListTy Args;
6256 TargetLowering::ArgListEntry Entry;
6258 Entry.Ty = IntPtrTy;
6259 Args.push_back(Entry);
6261 Args.push_back(Entry);
6262 std::pair<SDValue,SDValue> CallResult =
6263 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6264 false, false, false, false,
6265 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6266 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6267 DAG.GetOrdering(Chain.getNode()));
6268 return CallResult.second;
6271 // Otherwise have the target-independent code call memset.
6275 uint64_t SizeVal = ConstantSize->getZExtValue();
6276 SDValue InFlag(0, 0);
6279 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6280 unsigned BytesLeft = 0;
6281 bool TwoRepStos = false;
6284 uint64_t Val = ValC->getZExtValue() & 255;
6286 // If the value is a constant, then we can potentially use larger sets.
6287 switch (Align & 3) {
6288 case 2: // WORD aligned
6291 Val = (Val << 8) | Val;
6293 case 0: // DWORD aligned
6296 Val = (Val << 8) | Val;
6297 Val = (Val << 16) | Val;
6298 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6301 Val = (Val << 32) | Val;
6304 default: // Byte aligned
6307 Count = DAG.getIntPtrConstant(SizeVal);
6311 if (AVT.bitsGT(MVT::i8)) {
6312 unsigned UBytes = AVT.getSizeInBits() / 8;
6313 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6314 BytesLeft = SizeVal % UBytes;
6317 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6319 InFlag = Chain.getValue(1);
6322 Count = DAG.getIntPtrConstant(SizeVal);
6323 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6324 InFlag = Chain.getValue(1);
6327 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6330 InFlag = Chain.getValue(1);
6331 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6334 InFlag = Chain.getValue(1);
6336 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6337 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6338 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6341 InFlag = Chain.getValue(1);
6343 EVT CVT = Count.getValueType();
6344 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6345 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6346 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6349 InFlag = Chain.getValue(1);
6350 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6351 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6352 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6353 } else if (BytesLeft) {
6354 // Handle the last 1 - 7 bytes.
6355 unsigned Offset = SizeVal - BytesLeft;
6356 EVT AddrVT = Dst.getValueType();
6357 EVT SizeVT = Size.getValueType();
6359 Chain = DAG.getMemset(Chain, dl,
6360 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6361 DAG.getConstant(Offset, AddrVT)),
6363 DAG.getConstant(BytesLeft, SizeVT),
6364 Align, DstSV, DstSVOff + Offset);
6367 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6372 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6373 SDValue Chain, SDValue Dst, SDValue Src,
6374 SDValue Size, unsigned Align,
6376 const Value *DstSV, uint64_t DstSVOff,
6377 const Value *SrcSV, uint64_t SrcSVOff) {
6378 // This requires the copy size to be a constant, preferrably
6379 // within a subtarget-specific limit.
6380 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6383 uint64_t SizeVal = ConstantSize->getZExtValue();
6384 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6387 /// If not DWORD aligned, call the library.
6388 if ((Align & 3) != 0)
6393 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6396 unsigned UBytes = AVT.getSizeInBits() / 8;
6397 unsigned CountVal = SizeVal / UBytes;
6398 SDValue Count = DAG.getIntPtrConstant(CountVal);
6399 unsigned BytesLeft = SizeVal % UBytes;
6401 SDValue InFlag(0, 0);
6402 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6405 InFlag = Chain.getValue(1);
6406 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6409 InFlag = Chain.getValue(1);
6410 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6413 InFlag = Chain.getValue(1);
6415 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6416 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6417 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6418 array_lengthof(Ops));
6420 SmallVector<SDValue, 4> Results;
6421 Results.push_back(RepMovs);
6423 // Handle the last 1 - 7 bytes.
6424 unsigned Offset = SizeVal - BytesLeft;
6425 EVT DstVT = Dst.getValueType();
6426 EVT SrcVT = Src.getValueType();
6427 EVT SizeVT = Size.getValueType();
6428 Results.push_back(DAG.getMemcpy(Chain, dl,
6429 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6430 DAG.getConstant(Offset, DstVT)),
6431 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6432 DAG.getConstant(Offset, SrcVT)),
6433 DAG.getConstant(BytesLeft, SizeVT),
6434 Align, AlwaysInline,
6435 DstSV, DstSVOff + Offset,
6436 SrcSV, SrcSVOff + Offset));
6439 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6440 &Results[0], Results.size());
6443 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6444 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6445 DebugLoc dl = Op.getDebugLoc();
6447 if (!Subtarget->is64Bit()) {
6448 // vastart just stores the address of the VarArgsFrameIndex slot into the
6449 // memory location argument.
6450 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6451 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6455 // gp_offset (0 - 6 * 8)
6456 // fp_offset (48 - 48 + 8 * 16)
6457 // overflow_arg_area (point to parameters coming in memory).
6459 SmallVector<SDValue, 8> MemOps;
6460 SDValue FIN = Op.getOperand(1);
6462 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6463 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6465 MemOps.push_back(Store);
6468 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6469 FIN, DAG.getIntPtrConstant(4));
6470 Store = DAG.getStore(Op.getOperand(0), dl,
6471 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6473 MemOps.push_back(Store);
6475 // Store ptr to overflow_arg_area
6476 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6477 FIN, DAG.getIntPtrConstant(4));
6478 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6479 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6480 MemOps.push_back(Store);
6482 // Store ptr to reg_save_area.
6483 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6484 FIN, DAG.getIntPtrConstant(8));
6485 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6486 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6487 MemOps.push_back(Store);
6488 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6489 &MemOps[0], MemOps.size());
6492 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6493 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6494 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6495 SDValue Chain = Op.getOperand(0);
6496 SDValue SrcPtr = Op.getOperand(1);
6497 SDValue SrcSV = Op.getOperand(2);
6499 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6503 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6504 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6505 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6506 SDValue Chain = Op.getOperand(0);
6507 SDValue DstPtr = Op.getOperand(1);
6508 SDValue SrcPtr = Op.getOperand(2);
6509 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6510 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6511 DebugLoc dl = Op.getDebugLoc();
6513 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6514 DAG.getIntPtrConstant(24), 8, false,
6515 DstSV, 0, SrcSV, 0);
6519 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6520 DebugLoc dl = Op.getDebugLoc();
6521 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6523 default: return SDValue(); // Don't custom lower most intrinsics.
6524 // Comparison intrinsics.
6525 case Intrinsic::x86_sse_comieq_ss:
6526 case Intrinsic::x86_sse_comilt_ss:
6527 case Intrinsic::x86_sse_comile_ss:
6528 case Intrinsic::x86_sse_comigt_ss:
6529 case Intrinsic::x86_sse_comige_ss:
6530 case Intrinsic::x86_sse_comineq_ss:
6531 case Intrinsic::x86_sse_ucomieq_ss:
6532 case Intrinsic::x86_sse_ucomilt_ss:
6533 case Intrinsic::x86_sse_ucomile_ss:
6534 case Intrinsic::x86_sse_ucomigt_ss:
6535 case Intrinsic::x86_sse_ucomige_ss:
6536 case Intrinsic::x86_sse_ucomineq_ss:
6537 case Intrinsic::x86_sse2_comieq_sd:
6538 case Intrinsic::x86_sse2_comilt_sd:
6539 case Intrinsic::x86_sse2_comile_sd:
6540 case Intrinsic::x86_sse2_comigt_sd:
6541 case Intrinsic::x86_sse2_comige_sd:
6542 case Intrinsic::x86_sse2_comineq_sd:
6543 case Intrinsic::x86_sse2_ucomieq_sd:
6544 case Intrinsic::x86_sse2_ucomilt_sd:
6545 case Intrinsic::x86_sse2_ucomile_sd:
6546 case Intrinsic::x86_sse2_ucomigt_sd:
6547 case Intrinsic::x86_sse2_ucomige_sd:
6548 case Intrinsic::x86_sse2_ucomineq_sd: {
6550 ISD::CondCode CC = ISD::SETCC_INVALID;
6553 case Intrinsic::x86_sse_comieq_ss:
6554 case Intrinsic::x86_sse2_comieq_sd:
6558 case Intrinsic::x86_sse_comilt_ss:
6559 case Intrinsic::x86_sse2_comilt_sd:
6563 case Intrinsic::x86_sse_comile_ss:
6564 case Intrinsic::x86_sse2_comile_sd:
6568 case Intrinsic::x86_sse_comigt_ss:
6569 case Intrinsic::x86_sse2_comigt_sd:
6573 case Intrinsic::x86_sse_comige_ss:
6574 case Intrinsic::x86_sse2_comige_sd:
6578 case Intrinsic::x86_sse_comineq_ss:
6579 case Intrinsic::x86_sse2_comineq_sd:
6583 case Intrinsic::x86_sse_ucomieq_ss:
6584 case Intrinsic::x86_sse2_ucomieq_sd:
6585 Opc = X86ISD::UCOMI;
6588 case Intrinsic::x86_sse_ucomilt_ss:
6589 case Intrinsic::x86_sse2_ucomilt_sd:
6590 Opc = X86ISD::UCOMI;
6593 case Intrinsic::x86_sse_ucomile_ss:
6594 case Intrinsic::x86_sse2_ucomile_sd:
6595 Opc = X86ISD::UCOMI;
6598 case Intrinsic::x86_sse_ucomigt_ss:
6599 case Intrinsic::x86_sse2_ucomigt_sd:
6600 Opc = X86ISD::UCOMI;
6603 case Intrinsic::x86_sse_ucomige_ss:
6604 case Intrinsic::x86_sse2_ucomige_sd:
6605 Opc = X86ISD::UCOMI;
6608 case Intrinsic::x86_sse_ucomineq_ss:
6609 case Intrinsic::x86_sse2_ucomineq_sd:
6610 Opc = X86ISD::UCOMI;
6615 SDValue LHS = Op.getOperand(1);
6616 SDValue RHS = Op.getOperand(2);
6617 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6618 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6619 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6620 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6621 DAG.getConstant(X86CC, MVT::i8), Cond);
6622 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6624 // ptest intrinsics. The intrinsic these come from are designed to return
6625 // an integer value, not just an instruction so lower it to the ptest
6626 // pattern and a setcc for the result.
6627 case Intrinsic::x86_sse41_ptestz:
6628 case Intrinsic::x86_sse41_ptestc:
6629 case Intrinsic::x86_sse41_ptestnzc:{
6632 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6633 case Intrinsic::x86_sse41_ptestz:
6635 X86CC = X86::COND_E;
6637 case Intrinsic::x86_sse41_ptestc:
6639 X86CC = X86::COND_B;
6641 case Intrinsic::x86_sse41_ptestnzc:
6643 X86CC = X86::COND_A;
6647 SDValue LHS = Op.getOperand(1);
6648 SDValue RHS = Op.getOperand(2);
6649 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6650 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6651 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6652 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6655 // Fix vector shift instructions where the last operand is a non-immediate
6657 case Intrinsic::x86_sse2_pslli_w:
6658 case Intrinsic::x86_sse2_pslli_d:
6659 case Intrinsic::x86_sse2_pslli_q:
6660 case Intrinsic::x86_sse2_psrli_w:
6661 case Intrinsic::x86_sse2_psrli_d:
6662 case Intrinsic::x86_sse2_psrli_q:
6663 case Intrinsic::x86_sse2_psrai_w:
6664 case Intrinsic::x86_sse2_psrai_d:
6665 case Intrinsic::x86_mmx_pslli_w:
6666 case Intrinsic::x86_mmx_pslli_d:
6667 case Intrinsic::x86_mmx_pslli_q:
6668 case Intrinsic::x86_mmx_psrli_w:
6669 case Intrinsic::x86_mmx_psrli_d:
6670 case Intrinsic::x86_mmx_psrli_q:
6671 case Intrinsic::x86_mmx_psrai_w:
6672 case Intrinsic::x86_mmx_psrai_d: {
6673 SDValue ShAmt = Op.getOperand(2);
6674 if (isa<ConstantSDNode>(ShAmt))
6677 unsigned NewIntNo = 0;
6678 EVT ShAmtVT = MVT::v4i32;
6680 case Intrinsic::x86_sse2_pslli_w:
6681 NewIntNo = Intrinsic::x86_sse2_psll_w;
6683 case Intrinsic::x86_sse2_pslli_d:
6684 NewIntNo = Intrinsic::x86_sse2_psll_d;
6686 case Intrinsic::x86_sse2_pslli_q:
6687 NewIntNo = Intrinsic::x86_sse2_psll_q;
6689 case Intrinsic::x86_sse2_psrli_w:
6690 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6692 case Intrinsic::x86_sse2_psrli_d:
6693 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6695 case Intrinsic::x86_sse2_psrli_q:
6696 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6698 case Intrinsic::x86_sse2_psrai_w:
6699 NewIntNo = Intrinsic::x86_sse2_psra_w;
6701 case Intrinsic::x86_sse2_psrai_d:
6702 NewIntNo = Intrinsic::x86_sse2_psra_d;
6705 ShAmtVT = MVT::v2i32;
6707 case Intrinsic::x86_mmx_pslli_w:
6708 NewIntNo = Intrinsic::x86_mmx_psll_w;
6710 case Intrinsic::x86_mmx_pslli_d:
6711 NewIntNo = Intrinsic::x86_mmx_psll_d;
6713 case Intrinsic::x86_mmx_pslli_q:
6714 NewIntNo = Intrinsic::x86_mmx_psll_q;
6716 case Intrinsic::x86_mmx_psrli_w:
6717 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6719 case Intrinsic::x86_mmx_psrli_d:
6720 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6722 case Intrinsic::x86_mmx_psrli_q:
6723 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6725 case Intrinsic::x86_mmx_psrai_w:
6726 NewIntNo = Intrinsic::x86_mmx_psra_w;
6728 case Intrinsic::x86_mmx_psrai_d:
6729 NewIntNo = Intrinsic::x86_mmx_psra_d;
6731 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6737 // The vector shift intrinsics with scalars uses 32b shift amounts but
6738 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6742 ShOps[1] = DAG.getConstant(0, MVT::i32);
6743 if (ShAmtVT == MVT::v4i32) {
6744 ShOps[2] = DAG.getUNDEF(MVT::i32);
6745 ShOps[3] = DAG.getUNDEF(MVT::i32);
6746 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6748 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6751 EVT VT = Op.getValueType();
6752 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6753 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6754 DAG.getConstant(NewIntNo, MVT::i32),
6755 Op.getOperand(1), ShAmt);
6760 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6761 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6762 DebugLoc dl = Op.getDebugLoc();
6765 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6767 DAG.getConstant(TD->getPointerSize(),
6768 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6769 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6770 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6775 // Just load the return address.
6776 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6777 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6778 RetAddrFI, NULL, 0);
6781 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6782 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6783 MFI->setFrameAddressIsTaken(true);
6784 EVT VT = Op.getValueType();
6785 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6786 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6787 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6788 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6790 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6794 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6795 SelectionDAG &DAG) {
6796 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6799 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6801 MachineFunction &MF = DAG.getMachineFunction();
6802 SDValue Chain = Op.getOperand(0);
6803 SDValue Offset = Op.getOperand(1);
6804 SDValue Handler = Op.getOperand(2);
6805 DebugLoc dl = Op.getDebugLoc();
6807 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6809 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6811 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6812 DAG.getIntPtrConstant(-TD->getPointerSize()));
6813 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6814 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6815 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6816 MF.getRegInfo().addLiveOut(StoreAddrReg);
6818 return DAG.getNode(X86ISD::EH_RETURN, dl,
6820 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6823 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6824 SelectionDAG &DAG) {
6825 SDValue Root = Op.getOperand(0);
6826 SDValue Trmp = Op.getOperand(1); // trampoline
6827 SDValue FPtr = Op.getOperand(2); // nested function
6828 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6829 DebugLoc dl = Op.getDebugLoc();
6831 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6833 const X86InstrInfo *TII =
6834 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6836 if (Subtarget->is64Bit()) {
6837 SDValue OutChains[6];
6839 // Large code-model.
6841 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6842 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6844 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6845 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6847 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6849 // Load the pointer to the nested function into R11.
6850 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6851 SDValue Addr = Trmp;
6852 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6855 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6856 DAG.getConstant(2, MVT::i64));
6857 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6859 // Load the 'nest' parameter value into R10.
6860 // R10 is specified in X86CallingConv.td
6861 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6862 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6863 DAG.getConstant(10, MVT::i64));
6864 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6865 Addr, TrmpAddr, 10);
6867 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6868 DAG.getConstant(12, MVT::i64));
6869 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6871 // Jump to the nested function.
6872 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6873 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6874 DAG.getConstant(20, MVT::i64));
6875 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6876 Addr, TrmpAddr, 20);
6878 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6879 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6880 DAG.getConstant(22, MVT::i64));
6881 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6885 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6886 return DAG.getMergeValues(Ops, 2, dl);
6888 const Function *Func =
6889 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6890 CallingConv::ID CC = Func->getCallingConv();
6895 llvm_unreachable("Unsupported calling convention");
6896 case CallingConv::C:
6897 case CallingConv::X86_StdCall: {
6898 // Pass 'nest' parameter in ECX.
6899 // Must be kept in sync with X86CallingConv.td
6902 // Check that ECX wasn't needed by an 'inreg' parameter.
6903 const FunctionType *FTy = Func->getFunctionType();
6904 const AttrListPtr &Attrs = Func->getAttributes();
6906 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6907 unsigned InRegCount = 0;
6910 for (FunctionType::param_iterator I = FTy->param_begin(),
6911 E = FTy->param_end(); I != E; ++I, ++Idx)
6912 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6913 // FIXME: should only count parameters that are lowered to integers.
6914 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6916 if (InRegCount > 2) {
6917 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6922 case CallingConv::X86_FastCall:
6923 case CallingConv::Fast:
6924 // Pass 'nest' parameter in EAX.
6925 // Must be kept in sync with X86CallingConv.td
6930 SDValue OutChains[4];
6933 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6934 DAG.getConstant(10, MVT::i32));
6935 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6937 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6938 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6939 OutChains[0] = DAG.getStore(Root, dl,
6940 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6943 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6944 DAG.getConstant(1, MVT::i32));
6945 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6947 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6948 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6949 DAG.getConstant(5, MVT::i32));
6950 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6951 TrmpAddr, 5, false, 1);
6953 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6954 DAG.getConstant(6, MVT::i32));
6955 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6958 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6959 return DAG.getMergeValues(Ops, 2, dl);
6963 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6965 The rounding mode is in bits 11:10 of FPSR, and has the following
6972 FLT_ROUNDS, on the other hand, expects the following:
6979 To perform the conversion, we do:
6980 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6983 MachineFunction &MF = DAG.getMachineFunction();
6984 const TargetMachine &TM = MF.getTarget();
6985 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6986 unsigned StackAlignment = TFI.getStackAlignment();
6987 EVT VT = Op.getValueType();
6988 DebugLoc dl = Op.getDebugLoc();
6990 // Save FP Control Word to stack slot
6991 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
6992 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6994 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6995 DAG.getEntryNode(), StackSlot);
6997 // Load FP Control Word from stack slot
6998 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
7000 // Transform as necessary
7002 DAG.getNode(ISD::SRL, dl, MVT::i16,
7003 DAG.getNode(ISD::AND, dl, MVT::i16,
7004 CWD, DAG.getConstant(0x800, MVT::i16)),
7005 DAG.getConstant(11, MVT::i8));
7007 DAG.getNode(ISD::SRL, dl, MVT::i16,
7008 DAG.getNode(ISD::AND, dl, MVT::i16,
7009 CWD, DAG.getConstant(0x400, MVT::i16)),
7010 DAG.getConstant(9, MVT::i8));
7013 DAG.getNode(ISD::AND, dl, MVT::i16,
7014 DAG.getNode(ISD::ADD, dl, MVT::i16,
7015 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7016 DAG.getConstant(1, MVT::i16)),
7017 DAG.getConstant(3, MVT::i16));
7020 return DAG.getNode((VT.getSizeInBits() < 16 ?
7021 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7024 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7025 EVT VT = Op.getValueType();
7027 unsigned NumBits = VT.getSizeInBits();
7028 DebugLoc dl = Op.getDebugLoc();
7030 Op = Op.getOperand(0);
7031 if (VT == MVT::i8) {
7032 // Zero extend to i32 since there is not an i8 bsr.
7034 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7037 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7038 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7039 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7041 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7044 DAG.getConstant(NumBits+NumBits-1, OpVT),
7045 DAG.getConstant(X86::COND_E, MVT::i8),
7048 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7050 // Finally xor with NumBits-1.
7051 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7054 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7058 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7059 EVT VT = Op.getValueType();
7061 unsigned NumBits = VT.getSizeInBits();
7062 DebugLoc dl = Op.getDebugLoc();
7064 Op = Op.getOperand(0);
7065 if (VT == MVT::i8) {
7067 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7070 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7071 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7072 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7074 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7077 DAG.getConstant(NumBits, OpVT),
7078 DAG.getConstant(X86::COND_E, MVT::i8),
7081 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7084 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7088 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7089 EVT VT = Op.getValueType();
7090 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7091 DebugLoc dl = Op.getDebugLoc();
7093 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7094 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7095 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7096 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7097 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7099 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7100 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7101 // return AloBlo + AloBhi + AhiBlo;
7103 SDValue A = Op.getOperand(0);
7104 SDValue B = Op.getOperand(1);
7106 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7107 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7108 A, DAG.getConstant(32, MVT::i32));
7109 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7110 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7111 B, DAG.getConstant(32, MVT::i32));
7112 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7113 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7115 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7116 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7118 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7119 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7121 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7122 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7123 AloBhi, DAG.getConstant(32, MVT::i32));
7124 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7125 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7126 AhiBlo, DAG.getConstant(32, MVT::i32));
7127 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7128 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7133 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7134 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7135 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7136 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7137 // has only one use.
7138 SDNode *N = Op.getNode();
7139 SDValue LHS = N->getOperand(0);
7140 SDValue RHS = N->getOperand(1);
7141 unsigned BaseOp = 0;
7143 DebugLoc dl = Op.getDebugLoc();
7145 switch (Op.getOpcode()) {
7146 default: llvm_unreachable("Unknown ovf instruction!");
7148 // A subtract of one will be selected as a INC. Note that INC doesn't
7149 // set CF, so we can't do this for UADDO.
7150 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7151 if (C->getAPIntValue() == 1) {
7152 BaseOp = X86ISD::INC;
7156 BaseOp = X86ISD::ADD;
7160 BaseOp = X86ISD::ADD;
7164 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7165 // set CF, so we can't do this for USUBO.
7166 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7167 if (C->getAPIntValue() == 1) {
7168 BaseOp = X86ISD::DEC;
7172 BaseOp = X86ISD::SUB;
7176 BaseOp = X86ISD::SUB;
7180 BaseOp = X86ISD::SMUL;
7184 BaseOp = X86ISD::UMUL;
7189 // Also sets EFLAGS.
7190 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7191 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7194 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7195 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7197 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7201 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7202 EVT T = Op.getValueType();
7203 DebugLoc dl = Op.getDebugLoc();
7206 switch(T.getSimpleVT().SimpleTy) {
7208 assert(false && "Invalid value type!");
7209 case MVT::i8: Reg = X86::AL; size = 1; break;
7210 case MVT::i16: Reg = X86::AX; size = 2; break;
7211 case MVT::i32: Reg = X86::EAX; size = 4; break;
7213 assert(Subtarget->is64Bit() && "Node not type legal!");
7214 Reg = X86::RAX; size = 8;
7217 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7218 Op.getOperand(2), SDValue());
7219 SDValue Ops[] = { cpIn.getValue(0),
7222 DAG.getTargetConstant(size, MVT::i8),
7224 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7225 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7227 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7231 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7232 SelectionDAG &DAG) {
7233 assert(Subtarget->is64Bit() && "Result not type legalized?");
7234 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7235 SDValue TheChain = Op.getOperand(0);
7236 DebugLoc dl = Op.getDebugLoc();
7237 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7238 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7239 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7241 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7242 DAG.getConstant(32, MVT::i8));
7244 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7247 return DAG.getMergeValues(Ops, 2, dl);
7250 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7251 SDNode *Node = Op.getNode();
7252 DebugLoc dl = Node->getDebugLoc();
7253 EVT T = Node->getValueType(0);
7254 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7255 DAG.getConstant(0, T), Node->getOperand(2));
7256 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7257 cast<AtomicSDNode>(Node)->getMemoryVT(),
7258 Node->getOperand(0),
7259 Node->getOperand(1), negOp,
7260 cast<AtomicSDNode>(Node)->getSrcValue(),
7261 cast<AtomicSDNode>(Node)->getAlignment());
7264 /// LowerOperation - Provide custom lowering hooks for some operations.
7266 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7267 switch (Op.getOpcode()) {
7268 default: llvm_unreachable("Should not custom lower this!");
7269 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7270 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7271 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7272 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7273 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7274 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7275 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7276 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7277 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7278 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7279 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7280 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7281 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7282 case ISD::SHL_PARTS:
7283 case ISD::SRA_PARTS:
7284 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7285 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7286 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7287 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7288 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7289 case ISD::FABS: return LowerFABS(Op, DAG);
7290 case ISD::FNEG: return LowerFNEG(Op, DAG);
7291 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7292 case ISD::SETCC: return LowerSETCC(Op, DAG);
7293 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7294 case ISD::SELECT: return LowerSELECT(Op, DAG);
7295 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7296 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7297 case ISD::VASTART: return LowerVASTART(Op, DAG);
7298 case ISD::VAARG: return LowerVAARG(Op, DAG);
7299 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7300 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7301 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7302 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7303 case ISD::FRAME_TO_ARGS_OFFSET:
7304 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7305 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7306 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7307 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7308 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7309 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7310 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7311 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7317 case ISD::UMULO: return LowerXALUO(Op, DAG);
7318 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7322 void X86TargetLowering::
7323 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7324 SelectionDAG &DAG, unsigned NewOp) {
7325 EVT T = Node->getValueType(0);
7326 DebugLoc dl = Node->getDebugLoc();
7327 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7329 SDValue Chain = Node->getOperand(0);
7330 SDValue In1 = Node->getOperand(1);
7331 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7332 Node->getOperand(2), DAG.getIntPtrConstant(0));
7333 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7334 Node->getOperand(2), DAG.getIntPtrConstant(1));
7335 SDValue Ops[] = { Chain, In1, In2L, In2H };
7336 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7338 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7339 cast<MemSDNode>(Node)->getMemOperand());
7340 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7341 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7342 Results.push_back(Result.getValue(2));
7345 /// ReplaceNodeResults - Replace a node with an illegal result type
7346 /// with a new node built out of custom code.
7347 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7348 SmallVectorImpl<SDValue>&Results,
7349 SelectionDAG &DAG) {
7350 DebugLoc dl = N->getDebugLoc();
7351 switch (N->getOpcode()) {
7353 assert(false && "Do not know how to custom type legalize this operation!");
7355 case ISD::FP_TO_SINT: {
7356 std::pair<SDValue,SDValue> Vals =
7357 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7358 SDValue FIST = Vals.first, StackSlot = Vals.second;
7359 if (FIST.getNode() != 0) {
7360 EVT VT = N->getValueType(0);
7361 // Return a load from the stack slot.
7362 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7366 case ISD::READCYCLECOUNTER: {
7367 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7368 SDValue TheChain = N->getOperand(0);
7369 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7370 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7372 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7374 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7375 SDValue Ops[] = { eax, edx };
7376 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7377 Results.push_back(edx.getValue(1));
7384 EVT WidenVT = getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
7385 Results.push_back(DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()));
7388 case ISD::ATOMIC_CMP_SWAP: {
7389 EVT T = N->getValueType(0);
7390 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7391 SDValue cpInL, cpInH;
7392 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7393 DAG.getConstant(0, MVT::i32));
7394 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7395 DAG.getConstant(1, MVT::i32));
7396 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7397 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7399 SDValue swapInL, swapInH;
7400 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7401 DAG.getConstant(0, MVT::i32));
7402 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7403 DAG.getConstant(1, MVT::i32));
7404 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7406 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7407 swapInL.getValue(1));
7408 SDValue Ops[] = { swapInH.getValue(0),
7410 swapInH.getValue(1) };
7411 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7412 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7413 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7414 MVT::i32, Result.getValue(1));
7415 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7416 MVT::i32, cpOutL.getValue(2));
7417 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7418 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7419 Results.push_back(cpOutH.getValue(1));
7422 case ISD::ATOMIC_LOAD_ADD:
7423 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7425 case ISD::ATOMIC_LOAD_AND:
7426 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7428 case ISD::ATOMIC_LOAD_NAND:
7429 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7431 case ISD::ATOMIC_LOAD_OR:
7432 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7434 case ISD::ATOMIC_LOAD_SUB:
7435 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7437 case ISD::ATOMIC_LOAD_XOR:
7438 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7440 case ISD::ATOMIC_SWAP:
7441 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7446 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7448 default: return NULL;
7449 case X86ISD::BSF: return "X86ISD::BSF";
7450 case X86ISD::BSR: return "X86ISD::BSR";
7451 case X86ISD::SHLD: return "X86ISD::SHLD";
7452 case X86ISD::SHRD: return "X86ISD::SHRD";
7453 case X86ISD::FAND: return "X86ISD::FAND";
7454 case X86ISD::FOR: return "X86ISD::FOR";
7455 case X86ISD::FXOR: return "X86ISD::FXOR";
7456 case X86ISD::FSRL: return "X86ISD::FSRL";
7457 case X86ISD::FILD: return "X86ISD::FILD";
7458 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7459 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7460 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7461 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7462 case X86ISD::FLD: return "X86ISD::FLD";
7463 case X86ISD::FST: return "X86ISD::FST";
7464 case X86ISD::CALL: return "X86ISD::CALL";
7465 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7466 case X86ISD::BT: return "X86ISD::BT";
7467 case X86ISD::CMP: return "X86ISD::CMP";
7468 case X86ISD::COMI: return "X86ISD::COMI";
7469 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7470 case X86ISD::SETCC: return "X86ISD::SETCC";
7471 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7472 case X86ISD::CMOV: return "X86ISD::CMOV";
7473 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7474 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7475 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7476 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7477 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7478 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7479 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7480 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7481 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7482 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7483 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7484 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7485 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7486 case X86ISD::FMAX: return "X86ISD::FMAX";
7487 case X86ISD::FMIN: return "X86ISD::FMIN";
7488 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7489 case X86ISD::FRCP: return "X86ISD::FRCP";
7490 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7491 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7492 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7493 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7494 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7495 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7496 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7497 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7498 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7499 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7500 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7501 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7502 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7503 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7504 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7505 case X86ISD::VSHL: return "X86ISD::VSHL";
7506 case X86ISD::VSRL: return "X86ISD::VSRL";
7507 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7508 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7509 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7510 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7511 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7512 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7513 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7514 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7515 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7516 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7517 case X86ISD::ADD: return "X86ISD::ADD";
7518 case X86ISD::SUB: return "X86ISD::SUB";
7519 case X86ISD::SMUL: return "X86ISD::SMUL";
7520 case X86ISD::UMUL: return "X86ISD::UMUL";
7521 case X86ISD::INC: return "X86ISD::INC";
7522 case X86ISD::DEC: return "X86ISD::DEC";
7523 case X86ISD::OR: return "X86ISD::OR";
7524 case X86ISD::XOR: return "X86ISD::XOR";
7525 case X86ISD::AND: return "X86ISD::AND";
7526 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7527 case X86ISD::PTEST: return "X86ISD::PTEST";
7528 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7532 // isLegalAddressingMode - Return true if the addressing mode represented
7533 // by AM is legal for this target, for a load/store of the specified type.
7534 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7535 const Type *Ty) const {
7536 // X86 supports extremely general addressing modes.
7537 CodeModel::Model M = getTargetMachine().getCodeModel();
7539 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7540 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7545 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7547 // If a reference to this global requires an extra load, we can't fold it.
7548 if (isGlobalStubReference(GVFlags))
7551 // If BaseGV requires a register for the PIC base, we cannot also have a
7552 // BaseReg specified.
7553 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7556 // If lower 4G is not available, then we must use rip-relative addressing.
7557 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7567 // These scales always work.
7572 // These scales are formed with basereg+scalereg. Only accept if there is
7577 default: // Other stuff never works.
7585 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7586 if (!Ty1->isInteger() || !Ty2->isInteger())
7588 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7589 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7590 if (NumBits1 <= NumBits2)
7592 return Subtarget->is64Bit() || NumBits1 < 64;
7595 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7596 if (!VT1.isInteger() || !VT2.isInteger())
7598 unsigned NumBits1 = VT1.getSizeInBits();
7599 unsigned NumBits2 = VT2.getSizeInBits();
7600 if (NumBits1 <= NumBits2)
7602 return Subtarget->is64Bit() || NumBits1 < 64;
7605 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7606 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7607 return Ty1->isInteger(32) && Ty2->isInteger(64) && Subtarget->is64Bit();
7610 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7611 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7612 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7615 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7616 // i16 instructions are longer (0x66 prefix) and potentially slower.
7617 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7620 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7621 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7622 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7623 /// are assumed to be legal.
7625 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7627 // Only do shuffles on 128-bit vector types for now.
7628 if (VT.getSizeInBits() == 64)
7631 // FIXME: pshufb, blends, shifts.
7632 return (VT.getVectorNumElements() == 2 ||
7633 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7634 isMOVLMask(M, VT) ||
7635 isSHUFPMask(M, VT) ||
7636 isPSHUFDMask(M, VT) ||
7637 isPSHUFHWMask(M, VT) ||
7638 isPSHUFLWMask(M, VT) ||
7639 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7640 isUNPCKLMask(M, VT) ||
7641 isUNPCKHMask(M, VT) ||
7642 isUNPCKL_v_undef_Mask(M, VT) ||
7643 isUNPCKH_v_undef_Mask(M, VT));
7647 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7649 unsigned NumElts = VT.getVectorNumElements();
7650 // FIXME: This collection of masks seems suspect.
7653 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7654 return (isMOVLMask(Mask, VT) ||
7655 isCommutedMOVLMask(Mask, VT, true) ||
7656 isSHUFPMask(Mask, VT) ||
7657 isCommutedSHUFPMask(Mask, VT));
7662 //===----------------------------------------------------------------------===//
7663 // X86 Scheduler Hooks
7664 //===----------------------------------------------------------------------===//
7666 // private utility function
7668 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7669 MachineBasicBlock *MBB,
7677 TargetRegisterClass *RC,
7678 bool invSrc) const {
7679 // For the atomic bitwise operator, we generate
7682 // ld t1 = [bitinstr.addr]
7683 // op t2 = t1, [bitinstr.val]
7685 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7687 // fallthrough -->nextMBB
7688 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7689 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7690 MachineFunction::iterator MBBIter = MBB;
7693 /// First build the CFG
7694 MachineFunction *F = MBB->getParent();
7695 MachineBasicBlock *thisMBB = MBB;
7696 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7697 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7698 F->insert(MBBIter, newMBB);
7699 F->insert(MBBIter, nextMBB);
7701 // Move all successors to thisMBB to nextMBB
7702 nextMBB->transferSuccessors(thisMBB);
7704 // Update thisMBB to fall through to newMBB
7705 thisMBB->addSuccessor(newMBB);
7707 // newMBB jumps to itself and fall through to nextMBB
7708 newMBB->addSuccessor(nextMBB);
7709 newMBB->addSuccessor(newMBB);
7711 // Insert instructions into newMBB based on incoming instruction
7712 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7713 "unexpected number of operands");
7714 DebugLoc dl = bInstr->getDebugLoc();
7715 MachineOperand& destOper = bInstr->getOperand(0);
7716 MachineOperand* argOpers[2 + X86AddrNumOperands];
7717 int numArgs = bInstr->getNumOperands() - 1;
7718 for (int i=0; i < numArgs; ++i)
7719 argOpers[i] = &bInstr->getOperand(i+1);
7721 // x86 address has 4 operands: base, index, scale, and displacement
7722 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7723 int valArgIndx = lastAddrIndx + 1;
7725 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7726 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7727 for (int i=0; i <= lastAddrIndx; ++i)
7728 (*MIB).addOperand(*argOpers[i]);
7730 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7732 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7737 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7738 assert((argOpers[valArgIndx]->isReg() ||
7739 argOpers[valArgIndx]->isImm()) &&
7741 if (argOpers[valArgIndx]->isReg())
7742 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7744 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7746 (*MIB).addOperand(*argOpers[valArgIndx]);
7748 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7751 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7752 for (int i=0; i <= lastAddrIndx; ++i)
7753 (*MIB).addOperand(*argOpers[i]);
7755 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7756 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7757 bInstr->memoperands_end());
7759 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7763 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7765 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7769 // private utility function: 64 bit atomics on 32 bit host.
7771 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7772 MachineBasicBlock *MBB,
7777 bool invSrc) const {
7778 // For the atomic bitwise operator, we generate
7779 // thisMBB (instructions are in pairs, except cmpxchg8b)
7780 // ld t1,t2 = [bitinstr.addr]
7782 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7783 // op t5, t6 <- out1, out2, [bitinstr.val]
7784 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7785 // mov ECX, EBX <- t5, t6
7786 // mov EAX, EDX <- t1, t2
7787 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7788 // mov t3, t4 <- EAX, EDX
7790 // result in out1, out2
7791 // fallthrough -->nextMBB
7793 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7794 const unsigned LoadOpc = X86::MOV32rm;
7795 const unsigned copyOpc = X86::MOV32rr;
7796 const unsigned NotOpc = X86::NOT32r;
7797 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7798 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7799 MachineFunction::iterator MBBIter = MBB;
7802 /// First build the CFG
7803 MachineFunction *F = MBB->getParent();
7804 MachineBasicBlock *thisMBB = MBB;
7805 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7806 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7807 F->insert(MBBIter, newMBB);
7808 F->insert(MBBIter, nextMBB);
7810 // Move all successors to thisMBB to nextMBB
7811 nextMBB->transferSuccessors(thisMBB);
7813 // Update thisMBB to fall through to newMBB
7814 thisMBB->addSuccessor(newMBB);
7816 // newMBB jumps to itself and fall through to nextMBB
7817 newMBB->addSuccessor(nextMBB);
7818 newMBB->addSuccessor(newMBB);
7820 DebugLoc dl = bInstr->getDebugLoc();
7821 // Insert instructions into newMBB based on incoming instruction
7822 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7823 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7824 "unexpected number of operands");
7825 MachineOperand& dest1Oper = bInstr->getOperand(0);
7826 MachineOperand& dest2Oper = bInstr->getOperand(1);
7827 MachineOperand* argOpers[2 + X86AddrNumOperands];
7828 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7829 argOpers[i] = &bInstr->getOperand(i+2);
7831 // x86 address has 5 operands: base, index, scale, displacement, and segment.
7832 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7834 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7835 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7836 for (int i=0; i <= lastAddrIndx; ++i)
7837 (*MIB).addOperand(*argOpers[i]);
7838 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7839 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7840 // add 4 to displacement.
7841 for (int i=0; i <= lastAddrIndx-2; ++i)
7842 (*MIB).addOperand(*argOpers[i]);
7843 MachineOperand newOp3 = *(argOpers[3]);
7845 newOp3.setImm(newOp3.getImm()+4);
7847 newOp3.setOffset(newOp3.getOffset()+4);
7848 (*MIB).addOperand(newOp3);
7849 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7851 // t3/4 are defined later, at the bottom of the loop
7852 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7853 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7854 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7855 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7856 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7857 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7859 // The subsequent operations should be using the destination registers of
7860 //the PHI instructions.
7862 t1 = F->getRegInfo().createVirtualRegister(RC);
7863 t2 = F->getRegInfo().createVirtualRegister(RC);
7864 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
7865 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
7867 t1 = dest1Oper.getReg();
7868 t2 = dest2Oper.getReg();
7871 int valArgIndx = lastAddrIndx + 1;
7872 assert((argOpers[valArgIndx]->isReg() ||
7873 argOpers[valArgIndx]->isImm()) &&
7875 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7876 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7877 if (argOpers[valArgIndx]->isReg())
7878 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7880 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7881 if (regOpcL != X86::MOV32rr)
7883 (*MIB).addOperand(*argOpers[valArgIndx]);
7884 assert(argOpers[valArgIndx + 1]->isReg() ==
7885 argOpers[valArgIndx]->isReg());
7886 assert(argOpers[valArgIndx + 1]->isImm() ==
7887 argOpers[valArgIndx]->isImm());
7888 if (argOpers[valArgIndx + 1]->isReg())
7889 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7891 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7892 if (regOpcH != X86::MOV32rr)
7894 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7896 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7898 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7901 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7903 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7906 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7907 for (int i=0; i <= lastAddrIndx; ++i)
7908 (*MIB).addOperand(*argOpers[i]);
7910 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7911 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7912 bInstr->memoperands_end());
7914 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7915 MIB.addReg(X86::EAX);
7916 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7917 MIB.addReg(X86::EDX);
7920 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7922 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7926 // private utility function
7928 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7929 MachineBasicBlock *MBB,
7930 unsigned cmovOpc) const {
7931 // For the atomic min/max operator, we generate
7934 // ld t1 = [min/max.addr]
7935 // mov t2 = [min/max.val]
7937 // cmov[cond] t2 = t1
7939 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7941 // fallthrough -->nextMBB
7943 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7944 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7945 MachineFunction::iterator MBBIter = MBB;
7948 /// First build the CFG
7949 MachineFunction *F = MBB->getParent();
7950 MachineBasicBlock *thisMBB = MBB;
7951 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7952 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7953 F->insert(MBBIter, newMBB);
7954 F->insert(MBBIter, nextMBB);
7956 // Move all successors of thisMBB to nextMBB
7957 nextMBB->transferSuccessors(thisMBB);
7959 // Update thisMBB to fall through to newMBB
7960 thisMBB->addSuccessor(newMBB);
7962 // newMBB jumps to newMBB and fall through to nextMBB
7963 newMBB->addSuccessor(nextMBB);
7964 newMBB->addSuccessor(newMBB);
7966 DebugLoc dl = mInstr->getDebugLoc();
7967 // Insert instructions into newMBB based on incoming instruction
7968 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7969 "unexpected number of operands");
7970 MachineOperand& destOper = mInstr->getOperand(0);
7971 MachineOperand* argOpers[2 + X86AddrNumOperands];
7972 int numArgs = mInstr->getNumOperands() - 1;
7973 for (int i=0; i < numArgs; ++i)
7974 argOpers[i] = &mInstr->getOperand(i+1);
7976 // x86 address has 4 operands: base, index, scale, and displacement
7977 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7978 int valArgIndx = lastAddrIndx + 1;
7980 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7981 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7982 for (int i=0; i <= lastAddrIndx; ++i)
7983 (*MIB).addOperand(*argOpers[i]);
7985 // We only support register and immediate values
7986 assert((argOpers[valArgIndx]->isReg() ||
7987 argOpers[valArgIndx]->isImm()) &&
7990 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7991 if (argOpers[valArgIndx]->isReg())
7992 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7994 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7995 (*MIB).addOperand(*argOpers[valArgIndx]);
7997 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8000 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8005 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8006 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8010 // Cmp and exchange if none has modified the memory location
8011 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8012 for (int i=0; i <= lastAddrIndx; ++i)
8013 (*MIB).addOperand(*argOpers[i]);
8015 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8016 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8017 mInstr->memoperands_end());
8019 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8020 MIB.addReg(X86::EAX);
8023 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8025 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8029 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8030 // all of this code can be replaced with that in the .td file.
8032 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8033 unsigned numArgs, bool memArg) const {
8035 MachineFunction *F = BB->getParent();
8036 DebugLoc dl = MI->getDebugLoc();
8037 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8041 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8043 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8045 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8047 for (unsigned i = 0; i < numArgs; ++i) {
8048 MachineOperand &Op = MI->getOperand(i+1);
8050 if (!(Op.isReg() && Op.isImplicit()))
8054 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8057 F->DeleteMachineInstr(MI);
8063 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8065 MachineBasicBlock *MBB) const {
8066 // Emit code to save XMM registers to the stack. The ABI says that the
8067 // number of registers to save is given in %al, so it's theoretically
8068 // possible to do an indirect jump trick to avoid saving all of them,
8069 // however this code takes a simpler approach and just executes all
8070 // of the stores if %al is non-zero. It's less code, and it's probably
8071 // easier on the hardware branch predictor, and stores aren't all that
8072 // expensive anyway.
8074 // Create the new basic blocks. One block contains all the XMM stores,
8075 // and one block is the final destination regardless of whether any
8076 // stores were performed.
8077 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8078 MachineFunction *F = MBB->getParent();
8079 MachineFunction::iterator MBBIter = MBB;
8081 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8082 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8083 F->insert(MBBIter, XMMSaveMBB);
8084 F->insert(MBBIter, EndMBB);
8087 // Move any original successors of MBB to the end block.
8088 EndMBB->transferSuccessors(MBB);
8089 // The original block will now fall through to the XMM save block.
8090 MBB->addSuccessor(XMMSaveMBB);
8091 // The XMMSaveMBB will fall through to the end block.
8092 XMMSaveMBB->addSuccessor(EndMBB);
8094 // Now add the instructions.
8095 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8096 DebugLoc DL = MI->getDebugLoc();
8098 unsigned CountReg = MI->getOperand(0).getReg();
8099 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8100 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8102 if (!Subtarget->isTargetWin64()) {
8103 // If %al is 0, branch around the XMM save block.
8104 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8105 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
8106 MBB->addSuccessor(EndMBB);
8109 // In the XMM save block, save all the XMM argument registers.
8110 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8111 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8112 MachineMemOperand *MMO =
8113 F->getMachineMemOperand(
8114 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8115 MachineMemOperand::MOStore, Offset,
8116 /*Size=*/16, /*Align=*/16);
8117 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8118 .addFrameIndex(RegSaveFrameIndex)
8119 .addImm(/*Scale=*/1)
8120 .addReg(/*IndexReg=*/0)
8121 .addImm(/*Disp=*/Offset)
8122 .addReg(/*Segment=*/0)
8123 .addReg(MI->getOperand(i).getReg())
8124 .addMemOperand(MMO);
8127 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8133 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8134 MachineBasicBlock *BB,
8135 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8136 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8137 DebugLoc DL = MI->getDebugLoc();
8139 // To "insert" a SELECT_CC instruction, we actually have to insert the
8140 // diamond control-flow pattern. The incoming instruction knows the
8141 // destination vreg to set, the condition code register to branch on, the
8142 // true/false values to select between, and a branch opcode to use.
8143 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8144 MachineFunction::iterator It = BB;
8150 // cmpTY ccX, r1, r2
8152 // fallthrough --> copy0MBB
8153 MachineBasicBlock *thisMBB = BB;
8154 MachineFunction *F = BB->getParent();
8155 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8156 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8158 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8159 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8160 F->insert(It, copy0MBB);
8161 F->insert(It, sinkMBB);
8162 // Update machine-CFG edges by first adding all successors of the current
8163 // block to the new block which will contain the Phi node for the select.
8164 // Also inform sdisel of the edge changes.
8165 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8166 E = BB->succ_end(); I != E; ++I) {
8167 EM->insert(std::make_pair(*I, sinkMBB));
8168 sinkMBB->addSuccessor(*I);
8170 // Next, remove all successors of the current block, and add the true
8171 // and fallthrough blocks as its successors.
8172 while (!BB->succ_empty())
8173 BB->removeSuccessor(BB->succ_begin());
8174 // Add the true and fallthrough blocks as its successors.
8175 BB->addSuccessor(copy0MBB);
8176 BB->addSuccessor(sinkMBB);
8179 // %FalseValue = ...
8180 // # fallthrough to sinkMBB
8183 // Update machine-CFG edges
8184 BB->addSuccessor(sinkMBB);
8187 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8190 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8191 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8192 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8194 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8200 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8201 MachineBasicBlock *BB,
8202 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8203 switch (MI->getOpcode()) {
8204 default: assert(false && "Unexpected instr type to insert");
8206 case X86::CMOV_V1I64:
8207 case X86::CMOV_FR32:
8208 case X86::CMOV_FR64:
8209 case X86::CMOV_V4F32:
8210 case X86::CMOV_V2F64:
8211 case X86::CMOV_V2I64:
8212 return EmitLoweredSelect(MI, BB, EM);
8214 case X86::FP32_TO_INT16_IN_MEM:
8215 case X86::FP32_TO_INT32_IN_MEM:
8216 case X86::FP32_TO_INT64_IN_MEM:
8217 case X86::FP64_TO_INT16_IN_MEM:
8218 case X86::FP64_TO_INT32_IN_MEM:
8219 case X86::FP64_TO_INT64_IN_MEM:
8220 case X86::FP80_TO_INT16_IN_MEM:
8221 case X86::FP80_TO_INT32_IN_MEM:
8222 case X86::FP80_TO_INT64_IN_MEM: {
8223 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8224 DebugLoc DL = MI->getDebugLoc();
8226 // Change the floating point control register to use "round towards zero"
8227 // mode when truncating to an integer value.
8228 MachineFunction *F = BB->getParent();
8229 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8230 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8232 // Load the old value of the high byte of the control word...
8234 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8235 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8238 // Set the high part to be round to zero...
8239 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8242 // Reload the modified control word now...
8243 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8245 // Restore the memory image of control word to original value
8246 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8249 // Get the X86 opcode to use.
8251 switch (MI->getOpcode()) {
8252 default: llvm_unreachable("illegal opcode!");
8253 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8254 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8255 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8256 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8257 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8258 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8259 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8260 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8261 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8265 MachineOperand &Op = MI->getOperand(0);
8267 AM.BaseType = X86AddressMode::RegBase;
8268 AM.Base.Reg = Op.getReg();
8270 AM.BaseType = X86AddressMode::FrameIndexBase;
8271 AM.Base.FrameIndex = Op.getIndex();
8273 Op = MI->getOperand(1);
8275 AM.Scale = Op.getImm();
8276 Op = MI->getOperand(2);
8278 AM.IndexReg = Op.getImm();
8279 Op = MI->getOperand(3);
8280 if (Op.isGlobal()) {
8281 AM.GV = Op.getGlobal();
8283 AM.Disp = Op.getImm();
8285 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8286 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8288 // Reload the original control word now.
8289 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8291 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8294 // String/text processing lowering.
8295 case X86::PCMPISTRM128REG:
8296 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8297 case X86::PCMPISTRM128MEM:
8298 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8299 case X86::PCMPESTRM128REG:
8300 return EmitPCMP(MI, BB, 5, false /* in mem */);
8301 case X86::PCMPESTRM128MEM:
8302 return EmitPCMP(MI, BB, 5, true /* in mem */);
8305 case X86::ATOMAND32:
8306 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8307 X86::AND32ri, X86::MOV32rm,
8308 X86::LCMPXCHG32, X86::MOV32rr,
8309 X86::NOT32r, X86::EAX,
8310 X86::GR32RegisterClass);
8312 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8313 X86::OR32ri, X86::MOV32rm,
8314 X86::LCMPXCHG32, X86::MOV32rr,
8315 X86::NOT32r, X86::EAX,
8316 X86::GR32RegisterClass);
8317 case X86::ATOMXOR32:
8318 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8319 X86::XOR32ri, X86::MOV32rm,
8320 X86::LCMPXCHG32, X86::MOV32rr,
8321 X86::NOT32r, X86::EAX,
8322 X86::GR32RegisterClass);
8323 case X86::ATOMNAND32:
8324 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8325 X86::AND32ri, X86::MOV32rm,
8326 X86::LCMPXCHG32, X86::MOV32rr,
8327 X86::NOT32r, X86::EAX,
8328 X86::GR32RegisterClass, true);
8329 case X86::ATOMMIN32:
8330 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8331 case X86::ATOMMAX32:
8332 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8333 case X86::ATOMUMIN32:
8334 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8335 case X86::ATOMUMAX32:
8336 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8338 case X86::ATOMAND16:
8339 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8340 X86::AND16ri, X86::MOV16rm,
8341 X86::LCMPXCHG16, X86::MOV16rr,
8342 X86::NOT16r, X86::AX,
8343 X86::GR16RegisterClass);
8345 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8346 X86::OR16ri, X86::MOV16rm,
8347 X86::LCMPXCHG16, X86::MOV16rr,
8348 X86::NOT16r, X86::AX,
8349 X86::GR16RegisterClass);
8350 case X86::ATOMXOR16:
8351 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8352 X86::XOR16ri, X86::MOV16rm,
8353 X86::LCMPXCHG16, X86::MOV16rr,
8354 X86::NOT16r, X86::AX,
8355 X86::GR16RegisterClass);
8356 case X86::ATOMNAND16:
8357 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8358 X86::AND16ri, X86::MOV16rm,
8359 X86::LCMPXCHG16, X86::MOV16rr,
8360 X86::NOT16r, X86::AX,
8361 X86::GR16RegisterClass, true);
8362 case X86::ATOMMIN16:
8363 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8364 case X86::ATOMMAX16:
8365 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8366 case X86::ATOMUMIN16:
8367 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8368 case X86::ATOMUMAX16:
8369 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8372 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8373 X86::AND8ri, X86::MOV8rm,
8374 X86::LCMPXCHG8, X86::MOV8rr,
8375 X86::NOT8r, X86::AL,
8376 X86::GR8RegisterClass);
8378 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8379 X86::OR8ri, X86::MOV8rm,
8380 X86::LCMPXCHG8, X86::MOV8rr,
8381 X86::NOT8r, X86::AL,
8382 X86::GR8RegisterClass);
8384 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8385 X86::XOR8ri, X86::MOV8rm,
8386 X86::LCMPXCHG8, X86::MOV8rr,
8387 X86::NOT8r, X86::AL,
8388 X86::GR8RegisterClass);
8389 case X86::ATOMNAND8:
8390 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8391 X86::AND8ri, X86::MOV8rm,
8392 X86::LCMPXCHG8, X86::MOV8rr,
8393 X86::NOT8r, X86::AL,
8394 X86::GR8RegisterClass, true);
8395 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8396 // This group is for 64-bit host.
8397 case X86::ATOMAND64:
8398 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8399 X86::AND64ri32, X86::MOV64rm,
8400 X86::LCMPXCHG64, X86::MOV64rr,
8401 X86::NOT64r, X86::RAX,
8402 X86::GR64RegisterClass);
8404 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8405 X86::OR64ri32, X86::MOV64rm,
8406 X86::LCMPXCHG64, X86::MOV64rr,
8407 X86::NOT64r, X86::RAX,
8408 X86::GR64RegisterClass);
8409 case X86::ATOMXOR64:
8410 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8411 X86::XOR64ri32, X86::MOV64rm,
8412 X86::LCMPXCHG64, X86::MOV64rr,
8413 X86::NOT64r, X86::RAX,
8414 X86::GR64RegisterClass);
8415 case X86::ATOMNAND64:
8416 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8417 X86::AND64ri32, X86::MOV64rm,
8418 X86::LCMPXCHG64, X86::MOV64rr,
8419 X86::NOT64r, X86::RAX,
8420 X86::GR64RegisterClass, true);
8421 case X86::ATOMMIN64:
8422 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8423 case X86::ATOMMAX64:
8424 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8425 case X86::ATOMUMIN64:
8426 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8427 case X86::ATOMUMAX64:
8428 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8430 // This group does 64-bit operations on a 32-bit host.
8431 case X86::ATOMAND6432:
8432 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8433 X86::AND32rr, X86::AND32rr,
8434 X86::AND32ri, X86::AND32ri,
8436 case X86::ATOMOR6432:
8437 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8438 X86::OR32rr, X86::OR32rr,
8439 X86::OR32ri, X86::OR32ri,
8441 case X86::ATOMXOR6432:
8442 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8443 X86::XOR32rr, X86::XOR32rr,
8444 X86::XOR32ri, X86::XOR32ri,
8446 case X86::ATOMNAND6432:
8447 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8448 X86::AND32rr, X86::AND32rr,
8449 X86::AND32ri, X86::AND32ri,
8451 case X86::ATOMADD6432:
8452 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8453 X86::ADD32rr, X86::ADC32rr,
8454 X86::ADD32ri, X86::ADC32ri,
8456 case X86::ATOMSUB6432:
8457 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8458 X86::SUB32rr, X86::SBB32rr,
8459 X86::SUB32ri, X86::SBB32ri,
8461 case X86::ATOMSWAP6432:
8462 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8463 X86::MOV32rr, X86::MOV32rr,
8464 X86::MOV32ri, X86::MOV32ri,
8466 case X86::VASTART_SAVE_XMM_REGS:
8467 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8471 //===----------------------------------------------------------------------===//
8472 // X86 Optimization Hooks
8473 //===----------------------------------------------------------------------===//
8475 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8479 const SelectionDAG &DAG,
8480 unsigned Depth) const {
8481 unsigned Opc = Op.getOpcode();
8482 assert((Opc >= ISD::BUILTIN_OP_END ||
8483 Opc == ISD::INTRINSIC_WO_CHAIN ||
8484 Opc == ISD::INTRINSIC_W_CHAIN ||
8485 Opc == ISD::INTRINSIC_VOID) &&
8486 "Should use MaskedValueIsZero if you don't know whether Op"
8487 " is a target node!");
8489 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8501 // These nodes' second result is a boolean.
8502 if (Op.getResNo() == 0)
8506 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8507 Mask.getBitWidth() - 1);
8512 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8513 /// node is a GlobalAddress + offset.
8514 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8515 GlobalValue* &GA, int64_t &Offset) const{
8516 if (N->getOpcode() == X86ISD::Wrapper) {
8517 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8518 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8519 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8523 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8526 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8527 EVT EltVT, LoadSDNode *&LDBase,
8528 unsigned &LastLoadedElt,
8529 SelectionDAG &DAG, MachineFrameInfo *MFI,
8530 const TargetLowering &TLI) {
8532 LastLoadedElt = -1U;
8533 for (unsigned i = 0; i < NumElems; ++i) {
8534 if (N->getMaskElt(i) < 0) {
8540 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8541 if (!Elt.getNode() ||
8542 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8545 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8547 LDBase = cast<LoadSDNode>(Elt.getNode());
8551 if (Elt.getOpcode() == ISD::UNDEF)
8554 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8555 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8562 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8563 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8564 /// if the load addresses are consecutive, non-overlapping, and in the right
8565 /// order. In the case of v2i64, it will see if it can rewrite the
8566 /// shuffle to be an appropriate build vector so it can take advantage of
8567 // performBuildVectorCombine.
8568 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8569 const TargetLowering &TLI) {
8570 DebugLoc dl = N->getDebugLoc();
8571 EVT VT = N->getValueType(0);
8572 EVT EltVT = VT.getVectorElementType();
8573 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8574 unsigned NumElems = VT.getVectorNumElements();
8576 if (VT.getSizeInBits() != 128)
8579 // Try to combine a vector_shuffle into a 128-bit load.
8580 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8581 LoadSDNode *LD = NULL;
8582 unsigned LastLoadedElt;
8583 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8587 if (LastLoadedElt == NumElems - 1) {
8588 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8589 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8590 LD->getSrcValue(), LD->getSrcValueOffset(),
8592 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8593 LD->getSrcValue(), LD->getSrcValueOffset(),
8594 LD->isVolatile(), LD->getAlignment());
8595 } else if (NumElems == 4 && LastLoadedElt == 1) {
8596 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8597 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8598 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8599 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8604 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8605 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8606 const X86Subtarget *Subtarget) {
8607 DebugLoc DL = N->getDebugLoc();
8608 SDValue Cond = N->getOperand(0);
8609 // Get the LHS/RHS of the select.
8610 SDValue LHS = N->getOperand(1);
8611 SDValue RHS = N->getOperand(2);
8613 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8614 // instructions have the peculiarity that if either operand is a NaN,
8615 // they chose what we call the RHS operand (and as such are not symmetric).
8616 // It happens that this matches the semantics of the common C idiom
8617 // x<y?x:y and related forms, so we can recognize these cases.
8618 if (Subtarget->hasSSE2() &&
8619 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8620 Cond.getOpcode() == ISD::SETCC) {
8621 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8623 unsigned Opcode = 0;
8624 // Check for x CC y ? x : y.
8625 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8629 // This can be a min if we can prove that at least one of the operands
8631 if (!FiniteOnlyFPMath()) {
8632 if (DAG.isKnownNeverNaN(RHS)) {
8633 // Put the potential NaN in the RHS so that SSE will preserve it.
8634 std::swap(LHS, RHS);
8635 } else if (!DAG.isKnownNeverNaN(LHS))
8638 Opcode = X86ISD::FMIN;
8641 // This can be a min if we can prove that at least one of the operands
8643 if (!FiniteOnlyFPMath()) {
8644 if (DAG.isKnownNeverNaN(LHS)) {
8645 // Put the potential NaN in the RHS so that SSE will preserve it.
8646 std::swap(LHS, RHS);
8647 } else if (!DAG.isKnownNeverNaN(RHS))
8650 Opcode = X86ISD::FMIN;
8653 // This can be a min, but if either operand is a NaN we need it to
8654 // preserve the original LHS.
8655 std::swap(LHS, RHS);
8659 Opcode = X86ISD::FMIN;
8663 // This can be a max if we can prove that at least one of the operands
8665 if (!FiniteOnlyFPMath()) {
8666 if (DAG.isKnownNeverNaN(LHS)) {
8667 // Put the potential NaN in the RHS so that SSE will preserve it.
8668 std::swap(LHS, RHS);
8669 } else if (!DAG.isKnownNeverNaN(RHS))
8672 Opcode = X86ISD::FMAX;
8675 // This can be a max if we can prove that at least one of the operands
8677 if (!FiniteOnlyFPMath()) {
8678 if (DAG.isKnownNeverNaN(RHS)) {
8679 // Put the potential NaN in the RHS so that SSE will preserve it.
8680 std::swap(LHS, RHS);
8681 } else if (!DAG.isKnownNeverNaN(LHS))
8684 Opcode = X86ISD::FMAX;
8687 // This can be a max, but if either operand is a NaN we need it to
8688 // preserve the original LHS.
8689 std::swap(LHS, RHS);
8693 Opcode = X86ISD::FMAX;
8696 // Check for x CC y ? y : x -- a min/max with reversed arms.
8697 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8701 // This can be a min if we can prove that at least one of the operands
8703 if (!FiniteOnlyFPMath()) {
8704 if (DAG.isKnownNeverNaN(RHS)) {
8705 // Put the potential NaN in the RHS so that SSE will preserve it.
8706 std::swap(LHS, RHS);
8707 } else if (!DAG.isKnownNeverNaN(LHS))
8710 Opcode = X86ISD::FMIN;
8713 // This can be a min if we can prove that at least one of the operands
8715 if (!FiniteOnlyFPMath()) {
8716 if (DAG.isKnownNeverNaN(LHS)) {
8717 // Put the potential NaN in the RHS so that SSE will preserve it.
8718 std::swap(LHS, RHS);
8719 } else if (!DAG.isKnownNeverNaN(RHS))
8722 Opcode = X86ISD::FMIN;
8725 // This can be a min, but if either operand is a NaN we need it to
8726 // preserve the original LHS.
8727 std::swap(LHS, RHS);
8731 Opcode = X86ISD::FMIN;
8735 // This can be a max if we can prove that at least one of the operands
8737 if (!FiniteOnlyFPMath()) {
8738 if (DAG.isKnownNeverNaN(LHS)) {
8739 // Put the potential NaN in the RHS so that SSE will preserve it.
8740 std::swap(LHS, RHS);
8741 } else if (!DAG.isKnownNeverNaN(RHS))
8744 Opcode = X86ISD::FMAX;
8747 // This can be a max if we can prove that at least one of the operands
8749 if (!FiniteOnlyFPMath()) {
8750 if (DAG.isKnownNeverNaN(RHS)) {
8751 // Put the potential NaN in the RHS so that SSE will preserve it.
8752 std::swap(LHS, RHS);
8753 } else if (!DAG.isKnownNeverNaN(LHS))
8756 Opcode = X86ISD::FMAX;
8759 // This can be a max, but if either operand is a NaN we need it to
8760 // preserve the original LHS.
8761 std::swap(LHS, RHS);
8765 Opcode = X86ISD::FMAX;
8771 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8774 // If this is a select between two integer constants, try to do some
8776 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8777 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8778 // Don't do this for crazy integer types.
8779 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8780 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8781 // so that TrueC (the true value) is larger than FalseC.
8782 bool NeedsCondInvert = false;
8784 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8785 // Efficiently invertible.
8786 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8787 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8788 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8789 NeedsCondInvert = true;
8790 std::swap(TrueC, FalseC);
8793 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8794 if (FalseC->getAPIntValue() == 0 &&
8795 TrueC->getAPIntValue().isPowerOf2()) {
8796 if (NeedsCondInvert) // Invert the condition if needed.
8797 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8798 DAG.getConstant(1, Cond.getValueType()));
8800 // Zero extend the condition if needed.
8801 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8803 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8804 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8805 DAG.getConstant(ShAmt, MVT::i8));
8808 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8809 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8810 if (NeedsCondInvert) // Invert the condition if needed.
8811 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8812 DAG.getConstant(1, Cond.getValueType()));
8814 // Zero extend the condition if needed.
8815 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8816 FalseC->getValueType(0), Cond);
8817 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8818 SDValue(FalseC, 0));
8821 // Optimize cases that will turn into an LEA instruction. This requires
8822 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8823 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8824 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8825 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8827 bool isFastMultiplier = false;
8829 switch ((unsigned char)Diff) {
8831 case 1: // result = add base, cond
8832 case 2: // result = lea base( , cond*2)
8833 case 3: // result = lea base(cond, cond*2)
8834 case 4: // result = lea base( , cond*4)
8835 case 5: // result = lea base(cond, cond*4)
8836 case 8: // result = lea base( , cond*8)
8837 case 9: // result = lea base(cond, cond*8)
8838 isFastMultiplier = true;
8843 if (isFastMultiplier) {
8844 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8845 if (NeedsCondInvert) // Invert the condition if needed.
8846 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8847 DAG.getConstant(1, Cond.getValueType()));
8849 // Zero extend the condition if needed.
8850 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8852 // Scale the condition by the difference.
8854 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8855 DAG.getConstant(Diff, Cond.getValueType()));
8857 // Add the base if non-zero.
8858 if (FalseC->getAPIntValue() != 0)
8859 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8860 SDValue(FalseC, 0));
8870 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8871 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8872 TargetLowering::DAGCombinerInfo &DCI) {
8873 DebugLoc DL = N->getDebugLoc();
8875 // If the flag operand isn't dead, don't touch this CMOV.
8876 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8879 // If this is a select between two integer constants, try to do some
8880 // optimizations. Note that the operands are ordered the opposite of SELECT
8882 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8883 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8884 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8885 // larger than FalseC (the false value).
8886 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8888 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8889 CC = X86::GetOppositeBranchCondition(CC);
8890 std::swap(TrueC, FalseC);
8893 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8894 // This is efficient for any integer data type (including i8/i16) and
8896 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8897 SDValue Cond = N->getOperand(3);
8898 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8899 DAG.getConstant(CC, MVT::i8), Cond);
8901 // Zero extend the condition if needed.
8902 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8904 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8905 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8906 DAG.getConstant(ShAmt, MVT::i8));
8907 if (N->getNumValues() == 2) // Dead flag value?
8908 return DCI.CombineTo(N, Cond, SDValue());
8912 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8913 // for any integer data type, including i8/i16.
8914 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8915 SDValue Cond = N->getOperand(3);
8916 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8917 DAG.getConstant(CC, MVT::i8), Cond);
8919 // Zero extend the condition if needed.
8920 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8921 FalseC->getValueType(0), Cond);
8922 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8923 SDValue(FalseC, 0));
8925 if (N->getNumValues() == 2) // Dead flag value?
8926 return DCI.CombineTo(N, Cond, SDValue());
8930 // Optimize cases that will turn into an LEA instruction. This requires
8931 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8932 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8933 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8934 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8936 bool isFastMultiplier = false;
8938 switch ((unsigned char)Diff) {
8940 case 1: // result = add base, cond
8941 case 2: // result = lea base( , cond*2)
8942 case 3: // result = lea base(cond, cond*2)
8943 case 4: // result = lea base( , cond*4)
8944 case 5: // result = lea base(cond, cond*4)
8945 case 8: // result = lea base( , cond*8)
8946 case 9: // result = lea base(cond, cond*8)
8947 isFastMultiplier = true;
8952 if (isFastMultiplier) {
8953 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8954 SDValue Cond = N->getOperand(3);
8955 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8956 DAG.getConstant(CC, MVT::i8), Cond);
8957 // Zero extend the condition if needed.
8958 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8960 // Scale the condition by the difference.
8962 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8963 DAG.getConstant(Diff, Cond.getValueType()));
8965 // Add the base if non-zero.
8966 if (FalseC->getAPIntValue() != 0)
8967 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8968 SDValue(FalseC, 0));
8969 if (N->getNumValues() == 2) // Dead flag value?
8970 return DCI.CombineTo(N, Cond, SDValue());
8980 /// PerformMulCombine - Optimize a single multiply with constant into two
8981 /// in order to implement it with two cheaper instructions, e.g.
8982 /// LEA + SHL, LEA + LEA.
8983 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8984 TargetLowering::DAGCombinerInfo &DCI) {
8985 if (DAG.getMachineFunction().
8986 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8989 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8992 EVT VT = N->getValueType(0);
8996 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8999 uint64_t MulAmt = C->getZExtValue();
9000 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9003 uint64_t MulAmt1 = 0;
9004 uint64_t MulAmt2 = 0;
9005 if ((MulAmt % 9) == 0) {
9007 MulAmt2 = MulAmt / 9;
9008 } else if ((MulAmt % 5) == 0) {
9010 MulAmt2 = MulAmt / 5;
9011 } else if ((MulAmt % 3) == 0) {
9013 MulAmt2 = MulAmt / 3;
9016 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9017 DebugLoc DL = N->getDebugLoc();
9019 if (isPowerOf2_64(MulAmt2) &&
9020 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9021 // If second multiplifer is pow2, issue it first. We want the multiply by
9022 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9024 std::swap(MulAmt1, MulAmt2);
9027 if (isPowerOf2_64(MulAmt1))
9028 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9029 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9031 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9032 DAG.getConstant(MulAmt1, VT));
9034 if (isPowerOf2_64(MulAmt2))
9035 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9036 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9038 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9039 DAG.getConstant(MulAmt2, VT));
9041 // Do not add new nodes to DAG combiner worklist.
9042 DCI.CombineTo(N, NewMul, false);
9047 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9048 SDValue N0 = N->getOperand(0);
9049 SDValue N1 = N->getOperand(1);
9050 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9051 EVT VT = N0.getValueType();
9053 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9054 // since the result of setcc_c is all zero's or all ones.
9055 if (N1C && N0.getOpcode() == ISD::AND &&
9056 N0.getOperand(1).getOpcode() == ISD::Constant) {
9057 SDValue N00 = N0.getOperand(0);
9058 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9059 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9060 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9061 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9062 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9063 APInt ShAmt = N1C->getAPIntValue();
9064 Mask = Mask.shl(ShAmt);
9066 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9067 N00, DAG.getConstant(Mask, VT));
9074 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9076 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9077 const X86Subtarget *Subtarget) {
9078 EVT VT = N->getValueType(0);
9079 if (!VT.isVector() && VT.isInteger() &&
9080 N->getOpcode() == ISD::SHL)
9081 return PerformSHLCombine(N, DAG);
9083 // On X86 with SSE2 support, we can transform this to a vector shift if
9084 // all elements are shifted by the same amount. We can't do this in legalize
9085 // because the a constant vector is typically transformed to a constant pool
9086 // so we have no knowledge of the shift amount.
9087 if (!Subtarget->hasSSE2())
9090 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9093 SDValue ShAmtOp = N->getOperand(1);
9094 EVT EltVT = VT.getVectorElementType();
9095 DebugLoc DL = N->getDebugLoc();
9096 SDValue BaseShAmt = SDValue();
9097 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9098 unsigned NumElts = VT.getVectorNumElements();
9100 for (; i != NumElts; ++i) {
9101 SDValue Arg = ShAmtOp.getOperand(i);
9102 if (Arg.getOpcode() == ISD::UNDEF) continue;
9106 for (; i != NumElts; ++i) {
9107 SDValue Arg = ShAmtOp.getOperand(i);
9108 if (Arg.getOpcode() == ISD::UNDEF) continue;
9109 if (Arg != BaseShAmt) {
9113 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9114 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9115 SDValue InVec = ShAmtOp.getOperand(0);
9116 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9117 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9119 for (; i != NumElts; ++i) {
9120 SDValue Arg = InVec.getOperand(i);
9121 if (Arg.getOpcode() == ISD::UNDEF) continue;
9125 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9126 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9127 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9128 if (C->getZExtValue() == SplatIdx)
9129 BaseShAmt = InVec.getOperand(1);
9132 if (BaseShAmt.getNode() == 0)
9133 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9134 DAG.getIntPtrConstant(0));
9138 // The shift amount is an i32.
9139 if (EltVT.bitsGT(MVT::i32))
9140 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9141 else if (EltVT.bitsLT(MVT::i32))
9142 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9144 // The shift amount is identical so we can do a vector shift.
9145 SDValue ValOp = N->getOperand(0);
9146 switch (N->getOpcode()) {
9148 llvm_unreachable("Unknown shift opcode!");
9151 if (VT == MVT::v2i64)
9152 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9153 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9155 if (VT == MVT::v4i32)
9156 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9157 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9159 if (VT == MVT::v8i16)
9160 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9161 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9165 if (VT == MVT::v4i32)
9166 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9167 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9169 if (VT == MVT::v8i16)
9170 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9171 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9175 if (VT == MVT::v2i64)
9176 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9177 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9179 if (VT == MVT::v4i32)
9180 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9181 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9183 if (VT == MVT::v8i16)
9184 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9185 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9192 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9193 const X86Subtarget *Subtarget) {
9194 EVT VT = N->getValueType(0);
9195 if (VT != MVT::i64 || !Subtarget->is64Bit())
9198 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9199 SDValue N0 = N->getOperand(0);
9200 SDValue N1 = N->getOperand(1);
9201 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9203 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9206 SDValue ShAmt0 = N0.getOperand(1);
9207 if (ShAmt0.getValueType() != MVT::i8)
9209 SDValue ShAmt1 = N1.getOperand(1);
9210 if (ShAmt1.getValueType() != MVT::i8)
9212 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9213 ShAmt0 = ShAmt0.getOperand(0);
9214 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9215 ShAmt1 = ShAmt1.getOperand(0);
9217 DebugLoc DL = N->getDebugLoc();
9218 unsigned Opc = X86ISD::SHLD;
9219 SDValue Op0 = N0.getOperand(0);
9220 SDValue Op1 = N1.getOperand(0);
9221 if (ShAmt0.getOpcode() == ISD::SUB) {
9223 std::swap(Op0, Op1);
9224 std::swap(ShAmt0, ShAmt1);
9227 if (ShAmt1.getOpcode() == ISD::SUB) {
9228 SDValue Sum = ShAmt1.getOperand(0);
9229 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9230 if (SumC->getSExtValue() == 64 &&
9231 ShAmt1.getOperand(1) == ShAmt0)
9232 return DAG.getNode(Opc, DL, VT,
9234 DAG.getNode(ISD::TRUNCATE, DL,
9237 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9238 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9240 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9241 return DAG.getNode(Opc, DL, VT,
9242 N0.getOperand(0), N1.getOperand(0),
9243 DAG.getNode(ISD::TRUNCATE, DL,
9250 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9251 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9252 const X86Subtarget *Subtarget) {
9253 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9254 // the FP state in cases where an emms may be missing.
9255 // A preferable solution to the general problem is to figure out the right
9256 // places to insert EMMS. This qualifies as a quick hack.
9258 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9259 StoreSDNode *St = cast<StoreSDNode>(N);
9260 EVT VT = St->getValue().getValueType();
9261 if (VT.getSizeInBits() != 64)
9264 const Function *F = DAG.getMachineFunction().getFunction();
9265 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9266 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9267 && Subtarget->hasSSE2();
9268 if ((VT.isVector() ||
9269 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9270 isa<LoadSDNode>(St->getValue()) &&
9271 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9272 St->getChain().hasOneUse() && !St->isVolatile()) {
9273 SDNode* LdVal = St->getValue().getNode();
9275 int TokenFactorIndex = -1;
9276 SmallVector<SDValue, 8> Ops;
9277 SDNode* ChainVal = St->getChain().getNode();
9278 // Must be a store of a load. We currently handle two cases: the load
9279 // is a direct child, and it's under an intervening TokenFactor. It is
9280 // possible to dig deeper under nested TokenFactors.
9281 if (ChainVal == LdVal)
9282 Ld = cast<LoadSDNode>(St->getChain());
9283 else if (St->getValue().hasOneUse() &&
9284 ChainVal->getOpcode() == ISD::TokenFactor) {
9285 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9286 if (ChainVal->getOperand(i).getNode() == LdVal) {
9287 TokenFactorIndex = i;
9288 Ld = cast<LoadSDNode>(St->getValue());
9290 Ops.push_back(ChainVal->getOperand(i));
9294 if (!Ld || !ISD::isNormalLoad(Ld))
9297 // If this is not the MMX case, i.e. we are just turning i64 load/store
9298 // into f64 load/store, avoid the transformation if there are multiple
9299 // uses of the loaded value.
9300 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9303 DebugLoc LdDL = Ld->getDebugLoc();
9304 DebugLoc StDL = N->getDebugLoc();
9305 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9306 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9308 if (Subtarget->is64Bit() || F64IsLegal) {
9309 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9310 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9311 Ld->getBasePtr(), Ld->getSrcValue(),
9312 Ld->getSrcValueOffset(), Ld->isVolatile(),
9313 Ld->getAlignment());
9314 SDValue NewChain = NewLd.getValue(1);
9315 if (TokenFactorIndex != -1) {
9316 Ops.push_back(NewChain);
9317 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9320 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9321 St->getSrcValue(), St->getSrcValueOffset(),
9322 St->isVolatile(), St->getAlignment());
9325 // Otherwise, lower to two pairs of 32-bit loads / stores.
9326 SDValue LoAddr = Ld->getBasePtr();
9327 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9328 DAG.getConstant(4, MVT::i32));
9330 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9331 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9332 Ld->isVolatile(), Ld->getAlignment());
9333 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9334 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9336 MinAlign(Ld->getAlignment(), 4));
9338 SDValue NewChain = LoLd.getValue(1);
9339 if (TokenFactorIndex != -1) {
9340 Ops.push_back(LoLd);
9341 Ops.push_back(HiLd);
9342 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9346 LoAddr = St->getBasePtr();
9347 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9348 DAG.getConstant(4, MVT::i32));
9350 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9351 St->getSrcValue(), St->getSrcValueOffset(),
9352 St->isVolatile(), St->getAlignment());
9353 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9355 St->getSrcValueOffset() + 4,
9357 MinAlign(St->getAlignment(), 4));
9358 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9363 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9364 /// X86ISD::FXOR nodes.
9365 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9366 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9367 // F[X]OR(0.0, x) -> x
9368 // F[X]OR(x, 0.0) -> x
9369 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9370 if (C->getValueAPF().isPosZero())
9371 return N->getOperand(1);
9372 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9373 if (C->getValueAPF().isPosZero())
9374 return N->getOperand(0);
9378 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9379 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9380 // FAND(0.0, x) -> 0.0
9381 // FAND(x, 0.0) -> 0.0
9382 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9383 if (C->getValueAPF().isPosZero())
9384 return N->getOperand(0);
9385 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9386 if (C->getValueAPF().isPosZero())
9387 return N->getOperand(1);
9391 static SDValue PerformBTCombine(SDNode *N,
9393 TargetLowering::DAGCombinerInfo &DCI) {
9394 // BT ignores high bits in the bit index operand.
9395 SDValue Op1 = N->getOperand(1);
9396 if (Op1.hasOneUse()) {
9397 unsigned BitWidth = Op1.getValueSizeInBits();
9398 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9399 APInt KnownZero, KnownOne;
9400 TargetLowering::TargetLoweringOpt TLO(DAG);
9401 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9402 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9403 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9404 DCI.CommitTargetLoweringOpt(TLO);
9409 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9410 SDValue Op = N->getOperand(0);
9411 if (Op.getOpcode() == ISD::BIT_CONVERT)
9412 Op = Op.getOperand(0);
9413 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9414 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9415 VT.getVectorElementType().getSizeInBits() ==
9416 OpVT.getVectorElementType().getSizeInBits()) {
9417 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9422 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9423 // Locked instructions, in turn, have implicit fence semantics (all memory
9424 // operations are flushed before issuing the locked instruction, and the
9425 // are not buffered), so we can fold away the common pattern of
9426 // fence-atomic-fence.
9427 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9428 SDValue atomic = N->getOperand(0);
9429 switch (atomic.getOpcode()) {
9430 case ISD::ATOMIC_CMP_SWAP:
9431 case ISD::ATOMIC_SWAP:
9432 case ISD::ATOMIC_LOAD_ADD:
9433 case ISD::ATOMIC_LOAD_SUB:
9434 case ISD::ATOMIC_LOAD_AND:
9435 case ISD::ATOMIC_LOAD_OR:
9436 case ISD::ATOMIC_LOAD_XOR:
9437 case ISD::ATOMIC_LOAD_NAND:
9438 case ISD::ATOMIC_LOAD_MIN:
9439 case ISD::ATOMIC_LOAD_MAX:
9440 case ISD::ATOMIC_LOAD_UMIN:
9441 case ISD::ATOMIC_LOAD_UMAX:
9447 SDValue fence = atomic.getOperand(0);
9448 if (fence.getOpcode() != ISD::MEMBARRIER)
9451 switch (atomic.getOpcode()) {
9452 case ISD::ATOMIC_CMP_SWAP:
9453 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9454 atomic.getOperand(1), atomic.getOperand(2),
9455 atomic.getOperand(3));
9456 case ISD::ATOMIC_SWAP:
9457 case ISD::ATOMIC_LOAD_ADD:
9458 case ISD::ATOMIC_LOAD_SUB:
9459 case ISD::ATOMIC_LOAD_AND:
9460 case ISD::ATOMIC_LOAD_OR:
9461 case ISD::ATOMIC_LOAD_XOR:
9462 case ISD::ATOMIC_LOAD_NAND:
9463 case ISD::ATOMIC_LOAD_MIN:
9464 case ISD::ATOMIC_LOAD_MAX:
9465 case ISD::ATOMIC_LOAD_UMIN:
9466 case ISD::ATOMIC_LOAD_UMAX:
9467 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9468 atomic.getOperand(1), atomic.getOperand(2));
9474 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9475 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9476 // (and (i32 x86isd::setcc_carry), 1)
9477 // This eliminates the zext. This transformation is necessary because
9478 // ISD::SETCC is always legalized to i8.
9479 DebugLoc dl = N->getDebugLoc();
9480 SDValue N0 = N->getOperand(0);
9481 EVT VT = N->getValueType(0);
9482 if (N0.getOpcode() == ISD::AND &&
9484 N0.getOperand(0).hasOneUse()) {
9485 SDValue N00 = N0.getOperand(0);
9486 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9488 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9489 if (!C || C->getZExtValue() != 1)
9491 return DAG.getNode(ISD::AND, dl, VT,
9492 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9493 N00.getOperand(0), N00.getOperand(1)),
9494 DAG.getConstant(1, VT));
9500 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9501 DAGCombinerInfo &DCI) const {
9502 SelectionDAG &DAG = DCI.DAG;
9503 switch (N->getOpcode()) {
9505 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9506 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9507 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9508 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9511 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9512 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9513 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9515 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9516 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9517 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9518 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9519 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9520 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9526 //===----------------------------------------------------------------------===//
9527 // X86 Inline Assembly Support
9528 //===----------------------------------------------------------------------===//
9530 static bool LowerToBSwap(CallInst *CI) {
9531 // FIXME: this should verify that we are targetting a 486 or better. If not,
9532 // we will turn this bswap into something that will be lowered to logical ops
9533 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9534 // so don't worry about this.
9536 // Verify this is a simple bswap.
9537 if (CI->getNumOperands() != 2 ||
9538 CI->getType() != CI->getOperand(1)->getType() ||
9539 !CI->getType()->isInteger())
9542 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9543 if (!Ty || Ty->getBitWidth() % 16 != 0)
9546 // Okay, we can do this xform, do so now.
9547 const Type *Tys[] = { Ty };
9548 Module *M = CI->getParent()->getParent()->getParent();
9549 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9551 Value *Op = CI->getOperand(1);
9552 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9554 CI->replaceAllUsesWith(Op);
9555 CI->eraseFromParent();
9559 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9560 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9561 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9563 std::string AsmStr = IA->getAsmString();
9565 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9566 SmallVector<StringRef, 4> AsmPieces;
9567 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9569 switch (AsmPieces.size()) {
9570 default: return false;
9572 AsmStr = AsmPieces[0];
9574 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9577 if (AsmPieces.size() == 2 &&
9578 (AsmPieces[0] == "bswap" ||
9579 AsmPieces[0] == "bswapq" ||
9580 AsmPieces[0] == "bswapl") &&
9581 (AsmPieces[1] == "$0" ||
9582 AsmPieces[1] == "${0:q}")) {
9583 // No need to check constraints, nothing other than the equivalent of
9584 // "=r,0" would be valid here.
9585 return LowerToBSwap(CI);
9587 // rorw $$8, ${0:w} --> llvm.bswap.i16
9588 if (CI->getType()->isInteger(16) &&
9589 AsmPieces.size() == 3 &&
9590 AsmPieces[0] == "rorw" &&
9591 AsmPieces[1] == "$$8," &&
9592 AsmPieces[2] == "${0:w}" &&
9593 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9594 return LowerToBSwap(CI);
9598 if (CI->getType()->isInteger(64) &&
9599 Constraints.size() >= 2 &&
9600 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9601 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9602 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9603 SmallVector<StringRef, 4> Words;
9604 SplitString(AsmPieces[0], Words, " \t");
9605 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9607 SplitString(AsmPieces[1], Words, " \t");
9608 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9610 SplitString(AsmPieces[2], Words, " \t,");
9611 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9612 Words[2] == "%edx") {
9613 return LowerToBSwap(CI);
9625 /// getConstraintType - Given a constraint letter, return the type of
9626 /// constraint it is for this target.
9627 X86TargetLowering::ConstraintType
9628 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9629 if (Constraint.size() == 1) {
9630 switch (Constraint[0]) {
9642 return C_RegisterClass;
9650 return TargetLowering::getConstraintType(Constraint);
9653 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9654 /// with another that has more specific requirements based on the type of the
9655 /// corresponding operand.
9656 const char *X86TargetLowering::
9657 LowerXConstraint(EVT ConstraintVT) const {
9658 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9659 // 'f' like normal targets.
9660 if (ConstraintVT.isFloatingPoint()) {
9661 if (Subtarget->hasSSE2())
9663 if (Subtarget->hasSSE1())
9667 return TargetLowering::LowerXConstraint(ConstraintVT);
9670 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9671 /// vector. If it is invalid, don't add anything to Ops.
9672 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9675 std::vector<SDValue>&Ops,
9676 SelectionDAG &DAG) const {
9677 SDValue Result(0, 0);
9679 switch (Constraint) {
9682 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9683 if (C->getZExtValue() <= 31) {
9684 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9690 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9691 if (C->getZExtValue() <= 63) {
9692 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9698 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9699 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9700 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9706 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9707 if (C->getZExtValue() <= 255) {
9708 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9714 // 32-bit signed value
9715 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9716 const ConstantInt *CI = C->getConstantIntValue();
9717 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9718 C->getSExtValue())) {
9719 // Widen to 64 bits here to get it sign extended.
9720 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9723 // FIXME gcc accepts some relocatable values here too, but only in certain
9724 // memory models; it's complicated.
9729 // 32-bit unsigned value
9730 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9731 const ConstantInt *CI = C->getConstantIntValue();
9732 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9733 C->getZExtValue())) {
9734 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9738 // FIXME gcc accepts some relocatable values here too, but only in certain
9739 // memory models; it's complicated.
9743 // Literal immediates are always ok.
9744 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9745 // Widen to 64 bits here to get it sign extended.
9746 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9750 // If we are in non-pic codegen mode, we allow the address of a global (with
9751 // an optional displacement) to be used with 'i'.
9752 GlobalAddressSDNode *GA = 0;
9755 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9757 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9758 Offset += GA->getOffset();
9760 } else if (Op.getOpcode() == ISD::ADD) {
9761 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9762 Offset += C->getZExtValue();
9763 Op = Op.getOperand(0);
9766 } else if (Op.getOpcode() == ISD::SUB) {
9767 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9768 Offset += -C->getZExtValue();
9769 Op = Op.getOperand(0);
9774 // Otherwise, this isn't something we can handle, reject it.
9778 GlobalValue *GV = GA->getGlobal();
9779 // If we require an extra load to get this address, as in PIC mode, we
9781 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9782 getTargetMachine())))
9786 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9788 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9794 if (Result.getNode()) {
9795 Ops.push_back(Result);
9798 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9802 std::vector<unsigned> X86TargetLowering::
9803 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9805 if (Constraint.size() == 1) {
9806 // FIXME: not handling fp-stack yet!
9807 switch (Constraint[0]) { // GCC X86 Constraint Letters
9808 default: break; // Unknown constraint letter
9809 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9810 if (Subtarget->is64Bit()) {
9812 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9813 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9814 X86::R10D,X86::R11D,X86::R12D,
9815 X86::R13D,X86::R14D,X86::R15D,
9816 X86::EBP, X86::ESP, 0);
9817 else if (VT == MVT::i16)
9818 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9819 X86::SI, X86::DI, X86::R8W,X86::R9W,
9820 X86::R10W,X86::R11W,X86::R12W,
9821 X86::R13W,X86::R14W,X86::R15W,
9822 X86::BP, X86::SP, 0);
9823 else if (VT == MVT::i8)
9824 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9825 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9826 X86::R10B,X86::R11B,X86::R12B,
9827 X86::R13B,X86::R14B,X86::R15B,
9828 X86::BPL, X86::SPL, 0);
9830 else if (VT == MVT::i64)
9831 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9832 X86::RSI, X86::RDI, X86::R8, X86::R9,
9833 X86::R10, X86::R11, X86::R12,
9834 X86::R13, X86::R14, X86::R15,
9835 X86::RBP, X86::RSP, 0);
9839 // 32-bit fallthrough
9842 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9843 else if (VT == MVT::i16)
9844 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9845 else if (VT == MVT::i8)
9846 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9847 else if (VT == MVT::i64)
9848 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9853 return std::vector<unsigned>();
9856 std::pair<unsigned, const TargetRegisterClass*>
9857 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9859 // First, see if this is a constraint that directly corresponds to an LLVM
9861 if (Constraint.size() == 1) {
9862 // GCC Constraint Letters
9863 switch (Constraint[0]) {
9865 case 'r': // GENERAL_REGS
9866 case 'l': // INDEX_REGS
9868 return std::make_pair(0U, X86::GR8RegisterClass);
9870 return std::make_pair(0U, X86::GR16RegisterClass);
9871 if (VT == MVT::i32 || !Subtarget->is64Bit())
9872 return std::make_pair(0U, X86::GR32RegisterClass);
9873 return std::make_pair(0U, X86::GR64RegisterClass);
9874 case 'R': // LEGACY_REGS
9876 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
9878 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
9879 if (VT == MVT::i32 || !Subtarget->is64Bit())
9880 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
9881 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
9882 case 'f': // FP Stack registers.
9883 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
9884 // value to the correct fpstack register class.
9885 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9886 return std::make_pair(0U, X86::RFP32RegisterClass);
9887 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9888 return std::make_pair(0U, X86::RFP64RegisterClass);
9889 return std::make_pair(0U, X86::RFP80RegisterClass);
9890 case 'y': // MMX_REGS if MMX allowed.
9891 if (!Subtarget->hasMMX()) break;
9892 return std::make_pair(0U, X86::VR64RegisterClass);
9893 case 'Y': // SSE_REGS if SSE2 allowed
9894 if (!Subtarget->hasSSE2()) break;
9896 case 'x': // SSE_REGS if SSE1 allowed
9897 if (!Subtarget->hasSSE1()) break;
9899 switch (VT.getSimpleVT().SimpleTy) {
9901 // Scalar SSE types.
9904 return std::make_pair(0U, X86::FR32RegisterClass);
9907 return std::make_pair(0U, X86::FR64RegisterClass);
9915 return std::make_pair(0U, X86::VR128RegisterClass);
9921 // Use the default implementation in TargetLowering to convert the register
9922 // constraint into a member of a register class.
9923 std::pair<unsigned, const TargetRegisterClass*> Res;
9924 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9926 // Not found as a standard register?
9927 if (Res.second == 0) {
9928 // Map st(0) -> st(7) -> ST0
9929 if (Constraint.size() == 7 && Constraint[0] == '{' &&
9930 tolower(Constraint[1]) == 's' &&
9931 tolower(Constraint[2]) == 't' &&
9932 Constraint[3] == '(' &&
9933 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
9934 Constraint[5] == ')' &&
9935 Constraint[6] == '}') {
9937 Res.first = X86::ST0+Constraint[4]-'0';
9938 Res.second = X86::RFP80RegisterClass;
9942 // GCC allows "st(0)" to be called just plain "st".
9943 if (StringRef("{st}").equals_lower(Constraint)) {
9944 Res.first = X86::ST0;
9945 Res.second = X86::RFP80RegisterClass;
9950 if (StringRef("{flags}").equals_lower(Constraint)) {
9951 Res.first = X86::EFLAGS;
9952 Res.second = X86::CCRRegisterClass;
9956 // 'A' means EAX + EDX.
9957 if (Constraint == "A") {
9958 Res.first = X86::EAX;
9959 Res.second = X86::GR32_ADRegisterClass;
9965 // Otherwise, check to see if this is a register class of the wrong value
9966 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
9967 // turn into {ax},{dx}.
9968 if (Res.second->hasType(VT))
9969 return Res; // Correct type already, nothing to do.
9971 // All of the single-register GCC register classes map their values onto
9972 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
9973 // really want an 8-bit or 32-bit register, map to the appropriate register
9974 // class and return the appropriate register.
9975 if (Res.second == X86::GR16RegisterClass) {
9976 if (VT == MVT::i8) {
9977 unsigned DestReg = 0;
9978 switch (Res.first) {
9980 case X86::AX: DestReg = X86::AL; break;
9981 case X86::DX: DestReg = X86::DL; break;
9982 case X86::CX: DestReg = X86::CL; break;
9983 case X86::BX: DestReg = X86::BL; break;
9986 Res.first = DestReg;
9987 Res.second = X86::GR8RegisterClass;
9989 } else if (VT == MVT::i32) {
9990 unsigned DestReg = 0;
9991 switch (Res.first) {
9993 case X86::AX: DestReg = X86::EAX; break;
9994 case X86::DX: DestReg = X86::EDX; break;
9995 case X86::CX: DestReg = X86::ECX; break;
9996 case X86::BX: DestReg = X86::EBX; break;
9997 case X86::SI: DestReg = X86::ESI; break;
9998 case X86::DI: DestReg = X86::EDI; break;
9999 case X86::BP: DestReg = X86::EBP; break;
10000 case X86::SP: DestReg = X86::ESP; break;
10003 Res.first = DestReg;
10004 Res.second = X86::GR32RegisterClass;
10006 } else if (VT == MVT::i64) {
10007 unsigned DestReg = 0;
10008 switch (Res.first) {
10010 case X86::AX: DestReg = X86::RAX; break;
10011 case X86::DX: DestReg = X86::RDX; break;
10012 case X86::CX: DestReg = X86::RCX; break;
10013 case X86::BX: DestReg = X86::RBX; break;
10014 case X86::SI: DestReg = X86::RSI; break;
10015 case X86::DI: DestReg = X86::RDI; break;
10016 case X86::BP: DestReg = X86::RBP; break;
10017 case X86::SP: DestReg = X86::RSP; break;
10020 Res.first = DestReg;
10021 Res.second = X86::GR64RegisterClass;
10024 } else if (Res.second == X86::FR32RegisterClass ||
10025 Res.second == X86::FR64RegisterClass ||
10026 Res.second == X86::VR128RegisterClass) {
10027 // Handle references to XMM physical registers that got mapped into the
10028 // wrong class. This can happen with constraints like {xmm0} where the
10029 // target independent register mapper will just pick the first match it can
10030 // find, ignoring the required type.
10031 if (VT == MVT::f32)
10032 Res.second = X86::FR32RegisterClass;
10033 else if (VT == MVT::f64)
10034 Res.second = X86::FR64RegisterClass;
10035 else if (X86::VR128RegisterClass->hasType(VT))
10036 Res.second = X86::VR128RegisterClass;
10042 //===----------------------------------------------------------------------===//
10043 // X86 Widen vector type
10044 //===----------------------------------------------------------------------===//
10046 /// getWidenVectorType: given a vector type, returns the type to widen
10047 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10048 /// If there is no vector type that we want to widen to, returns MVT::Other
10049 /// When and where to widen is target dependent based on the cost of
10050 /// scalarizing vs using the wider vector type.
10052 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10053 assert(VT.isVector());
10054 if (isTypeLegal(VT))
10057 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10058 // type based on element type. This would speed up our search (though
10059 // it may not be worth it since the size of the list is relatively
10061 EVT EltVT = VT.getVectorElementType();
10062 unsigned NElts = VT.getVectorNumElements();
10064 // On X86, it make sense to widen any vector wider than 1
10068 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10069 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10070 EVT SVT = (MVT::SimpleValueType)nVT;
10072 if (isTypeLegal(SVT) &&
10073 SVT.getVectorElementType() == EltVT &&
10074 SVT.getVectorNumElements() > NElts)