1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "x86-isel"
17 #include "X86InstrBuilder.h"
18 #include "X86ISelLowering.h"
19 #include "X86TargetMachine.h"
20 #include "X86TargetObjectFile.h"
21 #include "llvm/CallingConv.h"
22 #include "llvm/Constants.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/GlobalAlias.h"
25 #include "llvm/GlobalVariable.h"
26 #include "llvm/Function.h"
27 #include "llvm/Instructions.h"
28 #include "llvm/Intrinsics.h"
29 #include "llvm/LLVMContext.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineJumpTableInfo.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/MC/MCAsmInfo.h"
38 #include "llvm/MC/MCContext.h"
39 #include "llvm/MC/MCExpr.h"
40 #include "llvm/MC/MCSymbol.h"
41 #include "llvm/ADT/BitVector.h"
42 #include "llvm/ADT/SmallSet.h"
43 #include "llvm/ADT/Statistic.h"
44 #include "llvm/ADT/StringExtras.h"
45 #include "llvm/ADT/VectorExtras.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/ErrorHandling.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/raw_ostream.h"
53 STATISTIC(NumTailCalls, "Number of tail calls");
56 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
58 // Disable16Bit - 16-bit operations typically have a larger encoding than
59 // corresponding 32-bit instructions, and 16-bit code is slow on some
60 // processors. This is an experimental flag to disable 16-bit operations
61 // (which forces them to be Legalized to 32-bit operations).
63 Disable16Bit("disable-16bit", cl::Hidden,
64 cl::desc("Disable use of 16-bit instructions"));
66 // Forward declarations.
67 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
70 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
71 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
72 default: llvm_unreachable("unknown subtarget type");
73 case X86Subtarget::isDarwin:
74 if (TM.getSubtarget<X86Subtarget>().is64Bit())
75 return new X8664_MachoTargetObjectFile();
76 return new X8632_MachoTargetObjectFile();
77 case X86Subtarget::isELF:
78 return new TargetLoweringObjectFileELF();
79 case X86Subtarget::isMingw:
80 case X86Subtarget::isCygwin:
81 case X86Subtarget::isWindows:
82 return new TargetLoweringObjectFileCOFF();
87 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
88 : TargetLowering(TM, createTLOF(TM)) {
89 Subtarget = &TM.getSubtarget<X86Subtarget>();
90 X86ScalarSSEf64 = Subtarget->hasSSE2();
91 X86ScalarSSEf32 = Subtarget->hasSSE1();
92 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
94 RegInfo = TM.getRegisterInfo();
97 // Set up the TargetLowering object.
99 // X86 is weird, it always uses i8 for shift amounts and setcc results.
100 setShiftAmountType(MVT::i8);
101 setBooleanContents(ZeroOrOneBooleanContent);
102 setSchedulingPreference(SchedulingForRegPressure);
103 setStackPointerRegisterToSaveRestore(X86StackPtr);
105 if (Subtarget->isTargetDarwin()) {
106 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
107 setUseUnderscoreSetJmp(false);
108 setUseUnderscoreLongJmp(false);
109 } else if (Subtarget->isTargetMingw()) {
110 // MS runtime is weird: it exports _setjmp, but longjmp!
111 setUseUnderscoreSetJmp(true);
112 setUseUnderscoreLongJmp(false);
114 setUseUnderscoreSetJmp(true);
115 setUseUnderscoreLongJmp(true);
118 // Set up the register classes.
119 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
121 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
122 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
123 if (Subtarget->is64Bit())
124 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
126 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
128 // We don't accept any truncstore of integer registers.
129 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
131 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
132 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
134 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
135 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
136 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
138 // SETOEQ and SETUNE require checking two conditions.
139 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
140 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
141 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
142 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
143 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
144 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
146 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
148 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
149 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
150 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
152 if (Subtarget->is64Bit()) {
153 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
154 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
155 } else if (!UseSoftFloat) {
156 if (X86ScalarSSEf64) {
157 // We have an impenetrably clever algorithm for ui64->double only.
158 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
160 // We have an algorithm for SSE2, and we turn this into a 64-bit
161 // FILD for other targets.
162 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
165 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
167 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
168 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
171 // SSE has no i16 to fp conversion, only i32
172 if (X86ScalarSSEf32) {
173 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
174 // f32 and f64 cases are Legal, f80 case is not
175 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
177 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
178 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
181 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
182 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
185 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
186 // are Legal, f80 is custom lowered.
187 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
188 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
190 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
192 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
193 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
195 if (X86ScalarSSEf32) {
196 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
197 // f32 and f64 cases are Legal, f80 case is not
198 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
200 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
201 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
204 // Handle FP_TO_UINT by promoting the destination to a larger signed
206 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
207 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
208 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
210 if (Subtarget->is64Bit()) {
211 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
212 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
213 } else if (!UseSoftFloat) {
214 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
215 // Expand FP_TO_UINT into a select.
216 // FIXME: We would like to use a Custom expander here eventually to do
217 // the optimal thing for SSE vs. the default expansion in the legalizer.
218 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
220 // With SSE3 we can use fisttpll to convert to a signed i64; without
221 // SSE, we're stuck with a fistpll.
222 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
225 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
226 if (!X86ScalarSSEf64) {
227 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
228 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
231 // Scalar integer divide and remainder are lowered to use operations that
232 // produce two results, to match the available instructions. This exposes
233 // the two-result form to trivial CSE, which is able to combine x/y and x%y
234 // into a single instruction.
236 // Scalar integer multiply-high is also lowered to use two-result
237 // operations, to match the available instructions. However, plain multiply
238 // (low) operations are left as Legal, as there are single-result
239 // instructions for this in x86. Using the two-result multiply instructions
240 // when both high and low results are needed must be arranged by dagcombine.
241 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
242 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
243 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
244 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
245 setOperationAction(ISD::SREM , MVT::i8 , Expand);
246 setOperationAction(ISD::UREM , MVT::i8 , Expand);
247 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
248 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
249 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
250 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
251 setOperationAction(ISD::SREM , MVT::i16 , Expand);
252 setOperationAction(ISD::UREM , MVT::i16 , Expand);
253 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
254 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
255 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
256 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
257 setOperationAction(ISD::SREM , MVT::i32 , Expand);
258 setOperationAction(ISD::UREM , MVT::i32 , Expand);
259 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
260 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
261 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
262 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
263 setOperationAction(ISD::SREM , MVT::i64 , Expand);
264 setOperationAction(ISD::UREM , MVT::i64 , Expand);
266 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
267 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
268 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
269 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
270 if (Subtarget->is64Bit())
271 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
272 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
273 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
274 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
275 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
276 setOperationAction(ISD::FREM , MVT::f32 , Expand);
277 setOperationAction(ISD::FREM , MVT::f64 , Expand);
278 setOperationAction(ISD::FREM , MVT::f80 , Expand);
279 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
281 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
282 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
283 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
286 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
287 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
292 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
293 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
294 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
295 if (Subtarget->is64Bit()) {
296 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
297 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
298 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
301 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
302 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
304 // These should be promoted to a larger select which is supported.
305 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
306 // X86 wants to expand cmov itself.
307 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
309 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
311 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
312 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
313 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
314 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
315 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
316 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
318 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
320 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
321 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
322 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
323 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
324 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
325 if (Subtarget->is64Bit()) {
326 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
327 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
329 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
332 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
333 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
334 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
335 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
336 if (Subtarget->is64Bit())
337 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
338 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
339 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
340 if (Subtarget->is64Bit()) {
341 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
342 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
343 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
344 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
345 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
347 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
348 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
349 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
350 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
351 if (Subtarget->is64Bit()) {
352 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
353 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
354 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
357 if (Subtarget->hasSSE1())
358 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
360 if (!Subtarget->hasSSE2())
361 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
363 // Expand certain atomics
364 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
365 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
366 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
367 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
374 if (!Subtarget->is64Bit()) {
375 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
376 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
377 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
378 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
379 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
380 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
381 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
384 // FIXME - use subtarget debug flags
385 if (!Subtarget->isTargetDarwin() &&
386 !Subtarget->isTargetELF() &&
387 !Subtarget->isTargetCygMing()) {
388 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
391 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
392 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
393 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
394 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
395 if (Subtarget->is64Bit()) {
396 setExceptionPointerRegister(X86::RAX);
397 setExceptionSelectorRegister(X86::RDX);
399 setExceptionPointerRegister(X86::EAX);
400 setExceptionSelectorRegister(X86::EDX);
402 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
403 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
405 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
407 setOperationAction(ISD::TRAP, MVT::Other, Legal);
409 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
410 setOperationAction(ISD::VASTART , MVT::Other, Custom);
411 setOperationAction(ISD::VAEND , MVT::Other, Expand);
412 if (Subtarget->is64Bit()) {
413 setOperationAction(ISD::VAARG , MVT::Other, Custom);
414 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
416 setOperationAction(ISD::VAARG , MVT::Other, Expand);
417 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
420 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
421 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
422 if (Subtarget->is64Bit())
423 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
424 if (Subtarget->isTargetCygMing())
425 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
427 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
429 if (!UseSoftFloat && X86ScalarSSEf64) {
430 // f32 and f64 use SSE.
431 // Set up the FP register classes.
432 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
433 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
435 // Use ANDPD to simulate FABS.
436 setOperationAction(ISD::FABS , MVT::f64, Custom);
437 setOperationAction(ISD::FABS , MVT::f32, Custom);
439 // Use XORP to simulate FNEG.
440 setOperationAction(ISD::FNEG , MVT::f64, Custom);
441 setOperationAction(ISD::FNEG , MVT::f32, Custom);
443 // Use ANDPD and ORPD to simulate FCOPYSIGN.
444 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
445 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
447 // We don't support sin/cos/fmod
448 setOperationAction(ISD::FSIN , MVT::f64, Expand);
449 setOperationAction(ISD::FCOS , MVT::f64, Expand);
450 setOperationAction(ISD::FSIN , MVT::f32, Expand);
451 setOperationAction(ISD::FCOS , MVT::f32, Expand);
453 // Expand FP immediates into loads from the stack, except for the special
455 addLegalFPImmediate(APFloat(+0.0)); // xorpd
456 addLegalFPImmediate(APFloat(+0.0f)); // xorps
457 } else if (!UseSoftFloat && X86ScalarSSEf32) {
458 // Use SSE for f32, x87 for f64.
459 // Set up the FP register classes.
460 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
461 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
463 // Use ANDPS to simulate FABS.
464 setOperationAction(ISD::FABS , MVT::f32, Custom);
466 // Use XORP to simulate FNEG.
467 setOperationAction(ISD::FNEG , MVT::f32, Custom);
469 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
471 // Use ANDPS and ORPS to simulate FCOPYSIGN.
472 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
473 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
475 // We don't support sin/cos/fmod
476 setOperationAction(ISD::FSIN , MVT::f32, Expand);
477 setOperationAction(ISD::FCOS , MVT::f32, Expand);
479 // Special cases we handle for FP constants.
480 addLegalFPImmediate(APFloat(+0.0f)); // xorps
481 addLegalFPImmediate(APFloat(+0.0)); // FLD0
482 addLegalFPImmediate(APFloat(+1.0)); // FLD1
483 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
484 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
487 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
488 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
490 } else if (!UseSoftFloat) {
491 // f32 and f64 in x87.
492 // Set up the FP register classes.
493 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
494 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
496 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
497 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
498 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
499 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
502 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
503 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
505 addLegalFPImmediate(APFloat(+0.0)); // FLD0
506 addLegalFPImmediate(APFloat(+1.0)); // FLD1
507 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
508 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
509 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
510 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
511 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
512 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
515 // Long double always uses X87.
517 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
518 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
519 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
522 APFloat TmpFlt(+0.0);
523 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
525 addLegalFPImmediate(TmpFlt); // FLD0
527 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
528 APFloat TmpFlt2(+1.0);
529 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
531 addLegalFPImmediate(TmpFlt2); // FLD1
532 TmpFlt2.changeSign();
533 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
537 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
538 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
542 // Always use a library call for pow.
543 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
544 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
545 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
547 setOperationAction(ISD::FLOG, MVT::f80, Expand);
548 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
549 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
550 setOperationAction(ISD::FEXP, MVT::f80, Expand);
551 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
553 // First set operation action for all vector types to either promote
554 // (for widening) or expand (for scalarization). Then we will selectively
555 // turn on ones that can be effectively codegen'd.
556 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
557 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
558 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
573 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
574 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
601 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
602 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
603 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
604 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
605 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
606 setOperationAction(ISD::SIGN_EXTEND_INREG, (MVT::SimpleValueType)VT,Expand);
607 setOperationAction(ISD::TRUNCATE, (MVT::SimpleValueType)VT, Expand);
608 setOperationAction(ISD::SIGN_EXTEND, (MVT::SimpleValueType)VT, Expand);
609 setOperationAction(ISD::ZERO_EXTEND, (MVT::SimpleValueType)VT, Expand);
610 setOperationAction(ISD::ANY_EXTEND, (MVT::SimpleValueType)VT, Expand);
611 for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
612 InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
613 setTruncStoreAction((MVT::SimpleValueType)VT,
614 (MVT::SimpleValueType)InnerVT, Expand);
615 setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
616 setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
617 setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
620 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
621 // with -msoft-float, disable use of MMX as well.
622 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
623 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
624 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
625 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
626 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
627 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
629 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
630 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
631 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
632 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
634 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
635 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
636 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
637 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
639 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
640 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
642 setOperationAction(ISD::AND, MVT::v8i8, Promote);
643 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
644 setOperationAction(ISD::AND, MVT::v4i16, Promote);
645 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
646 setOperationAction(ISD::AND, MVT::v2i32, Promote);
647 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
648 setOperationAction(ISD::AND, MVT::v1i64, Legal);
650 setOperationAction(ISD::OR, MVT::v8i8, Promote);
651 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
652 setOperationAction(ISD::OR, MVT::v4i16, Promote);
653 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
654 setOperationAction(ISD::OR, MVT::v2i32, Promote);
655 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
656 setOperationAction(ISD::OR, MVT::v1i64, Legal);
658 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
659 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
660 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
661 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
662 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
663 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
664 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
666 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
667 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
668 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
669 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
670 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
671 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
672 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
673 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
674 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
676 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
677 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
678 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
679 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
680 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
682 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
683 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
684 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
685 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
687 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
688 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
689 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
690 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
692 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
694 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
695 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
696 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
697 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
698 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
699 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
700 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
703 if (!UseSoftFloat && Subtarget->hasSSE1()) {
704 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
706 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
707 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
708 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
709 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
710 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
711 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
712 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
713 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
714 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
715 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
716 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
717 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
720 if (!UseSoftFloat && Subtarget->hasSSE2()) {
721 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
723 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
724 // registers cannot be used even for integer operations.
725 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
726 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
727 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
728 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
730 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
731 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
732 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
733 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
734 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
735 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
736 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
737 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
738 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
739 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
740 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
741 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
742 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
743 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
744 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
745 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
747 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
748 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
749 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
750 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
752 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
753 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
754 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
755 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
756 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
758 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Custom);
759 setOperationAction(ISD::CONCAT_VECTORS, MVT::v2i64, Custom);
760 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
761 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
762 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
764 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
765 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
766 EVT VT = (MVT::SimpleValueType)i;
767 // Do not attempt to custom lower non-power-of-2 vectors
768 if (!isPowerOf2_32(VT.getVectorNumElements()))
770 // Do not attempt to custom lower non-128-bit vectors
771 if (!VT.is128BitVector())
773 setOperationAction(ISD::BUILD_VECTOR,
774 VT.getSimpleVT().SimpleTy, Custom);
775 setOperationAction(ISD::VECTOR_SHUFFLE,
776 VT.getSimpleVT().SimpleTy, Custom);
777 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
778 VT.getSimpleVT().SimpleTy, Custom);
781 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
782 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
783 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
784 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
785 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
786 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
788 if (Subtarget->is64Bit()) {
789 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
790 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
793 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
794 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
795 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
798 // Do not attempt to promote non-128-bit vectors
799 if (!VT.is128BitVector()) {
802 setOperationAction(ISD::AND, SVT, Promote);
803 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
804 setOperationAction(ISD::OR, SVT, Promote);
805 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
806 setOperationAction(ISD::XOR, SVT, Promote);
807 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
808 setOperationAction(ISD::LOAD, SVT, Promote);
809 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
810 setOperationAction(ISD::SELECT, SVT, Promote);
811 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
814 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
816 // Custom lower v2i64 and v2f64 selects.
817 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
818 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
819 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
820 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
822 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
823 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
824 if (!DisableMMX && Subtarget->hasMMX()) {
825 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
826 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
830 if (Subtarget->hasSSE41()) {
831 // FIXME: Do we need to handle scalar-to-vector here?
832 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
834 // i8 and i16 vectors are custom , because the source register and source
835 // source memory operand types are not the same width. f32 vectors are
836 // custom since the immediate controlling the insert encodes additional
838 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
839 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
840 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
841 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
843 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
844 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
845 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
846 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
848 if (Subtarget->is64Bit()) {
849 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
850 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
854 if (Subtarget->hasSSE42()) {
855 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
858 if (!UseSoftFloat && Subtarget->hasAVX()) {
859 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
860 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
861 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
862 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
864 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
865 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
866 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
867 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
868 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
869 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
870 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
871 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
872 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
873 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
874 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
875 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
876 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
877 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
878 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
880 // Operations to consider commented out -v16i16 v32i8
881 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
882 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
883 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
884 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
885 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
886 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
887 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
888 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
889 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
890 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
891 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
892 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
893 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
894 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
896 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
897 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
898 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
899 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
901 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
902 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
903 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
904 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
905 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
907 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
908 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
909 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
910 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
911 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
912 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
915 // Not sure we want to do this since there are no 256-bit integer
918 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
919 // This includes 256-bit vectors
920 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
921 EVT VT = (MVT::SimpleValueType)i;
923 // Do not attempt to custom lower non-power-of-2 vectors
924 if (!isPowerOf2_32(VT.getVectorNumElements()))
927 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
928 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
929 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
932 if (Subtarget->is64Bit()) {
933 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
934 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
939 // Not sure we want to do this since there are no 256-bit integer
942 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
943 // Including 256-bit vectors
944 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
945 EVT VT = (MVT::SimpleValueType)i;
947 if (!VT.is256BitVector()) {
950 setOperationAction(ISD::AND, VT, Promote);
951 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
952 setOperationAction(ISD::OR, VT, Promote);
953 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
954 setOperationAction(ISD::XOR, VT, Promote);
955 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
956 setOperationAction(ISD::LOAD, VT, Promote);
957 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
958 setOperationAction(ISD::SELECT, VT, Promote);
959 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
962 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
966 // We want to custom lower some of our intrinsics.
967 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
969 // Add/Sub/Mul with overflow operations are custom lowered.
970 setOperationAction(ISD::SADDO, MVT::i32, Custom);
971 setOperationAction(ISD::SADDO, MVT::i64, Custom);
972 setOperationAction(ISD::UADDO, MVT::i32, Custom);
973 setOperationAction(ISD::UADDO, MVT::i64, Custom);
974 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
975 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
976 setOperationAction(ISD::USUBO, MVT::i32, Custom);
977 setOperationAction(ISD::USUBO, MVT::i64, Custom);
978 setOperationAction(ISD::SMULO, MVT::i32, Custom);
979 setOperationAction(ISD::SMULO, MVT::i64, Custom);
981 if (!Subtarget->is64Bit()) {
982 // These libcalls are not available in 32-bit.
983 setLibcallName(RTLIB::SHL_I128, 0);
984 setLibcallName(RTLIB::SRL_I128, 0);
985 setLibcallName(RTLIB::SRA_I128, 0);
988 // We have target-specific dag combine patterns for the following nodes:
989 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
990 setTargetDAGCombine(ISD::BUILD_VECTOR);
991 setTargetDAGCombine(ISD::SELECT);
992 setTargetDAGCombine(ISD::SHL);
993 setTargetDAGCombine(ISD::SRA);
994 setTargetDAGCombine(ISD::SRL);
995 setTargetDAGCombine(ISD::OR);
996 setTargetDAGCombine(ISD::STORE);
997 setTargetDAGCombine(ISD::MEMBARRIER);
998 setTargetDAGCombine(ISD::ZERO_EXTEND);
999 if (Subtarget->is64Bit())
1000 setTargetDAGCombine(ISD::MUL);
1002 computeRegisterProperties();
1004 // Divide and reminder operations have no vector equivalent and can
1005 // trap. Do a custom widening for these operations in which we never
1006 // generate more divides/remainder than the original vector width.
1007 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
1008 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
1009 if (!isTypeLegal((MVT::SimpleValueType)VT)) {
1010 setOperationAction(ISD::SDIV, (MVT::SimpleValueType) VT, Custom);
1011 setOperationAction(ISD::UDIV, (MVT::SimpleValueType) VT, Custom);
1012 setOperationAction(ISD::SREM, (MVT::SimpleValueType) VT, Custom);
1013 setOperationAction(ISD::UREM, (MVT::SimpleValueType) VT, Custom);
1017 // FIXME: These should be based on subtarget info. Plus, the values should
1018 // be smaller when we are in optimizing for size mode.
1019 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1020 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1021 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1022 setPrefLoopAlignment(16);
1023 benefitFromCodePlacementOpt = true;
1027 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1032 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1033 /// the desired ByVal argument alignment.
1034 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1037 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1038 if (VTy->getBitWidth() == 128)
1040 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1041 unsigned EltAlign = 0;
1042 getMaxByValAlign(ATy->getElementType(), EltAlign);
1043 if (EltAlign > MaxAlign)
1044 MaxAlign = EltAlign;
1045 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1046 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1047 unsigned EltAlign = 0;
1048 getMaxByValAlign(STy->getElementType(i), EltAlign);
1049 if (EltAlign > MaxAlign)
1050 MaxAlign = EltAlign;
1058 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1059 /// function arguments in the caller parameter area. For X86, aggregates
1060 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1061 /// are at 4-byte boundaries.
1062 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1063 if (Subtarget->is64Bit()) {
1064 // Max of 8 and alignment of type.
1065 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1072 if (Subtarget->hasSSE1())
1073 getMaxByValAlign(Ty, Align);
1077 /// getOptimalMemOpType - Returns the target specific optimal type for load
1078 /// and store operations as a result of memset, memcpy, and memmove
1079 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1082 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1083 bool isSrcConst, bool isSrcStr,
1084 SelectionDAG &DAG) const {
1085 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1086 // linux. This is because the stack realignment code can't handle certain
1087 // cases like PR2962. This should be removed when PR2962 is fixed.
1088 const Function *F = DAG.getMachineFunction().getFunction();
1089 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1090 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1091 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1093 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1096 if (Subtarget->is64Bit() && Size >= 8)
1101 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1102 /// current function. The returned value is a member of the
1103 /// MachineJumpTableInfo::JTEntryKind enum.
1104 unsigned X86TargetLowering::getJumpTableEncoding() const {
1105 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1107 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1108 Subtarget->isPICStyleGOT())
1109 return MachineJumpTableInfo::EK_Custom32;
1111 // Otherwise, use the normal jump table encoding heuristics.
1112 return TargetLowering::getJumpTableEncoding();
1115 /// getPICBaseSymbol - Return the X86-32 PIC base.
1117 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1118 MCContext &Ctx) const {
1119 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1120 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1121 Twine(MF->getFunctionNumber())+"$pb");
1126 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1127 const MachineBasicBlock *MBB,
1128 unsigned uid,MCContext &Ctx) const{
1129 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1130 Subtarget->isPICStyleGOT());
1131 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1134 // FIXME: @GOTOFF should be a property of MCSymbolRefExpr not in the MCSymbol.
1135 std::string Name = MBB->getSymbol(Ctx)->getName() + "@GOTOFF";
1136 return MCSymbolRefExpr::Create(Ctx.GetOrCreateSymbol(StringRef(Name)), Ctx);
1139 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1141 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1142 SelectionDAG &DAG) const {
1143 if (!Subtarget->is64Bit())
1144 // This doesn't have DebugLoc associated with it, but is not really the
1145 // same as a Register.
1146 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1151 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1152 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1154 const MCExpr *X86TargetLowering::
1155 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1156 MCContext &Ctx) const {
1157 // X86-64 uses RIP relative addressing based on the jump table label.
1158 if (Subtarget->isPICStyleRIPRel())
1159 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1161 // Otherwise, the reference is relative to the PIC base.
1162 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1165 /// getFunctionAlignment - Return the Log2 alignment of this function.
1166 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1167 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1170 //===----------------------------------------------------------------------===//
1171 // Return Value Calling Convention Implementation
1172 //===----------------------------------------------------------------------===//
1174 #include "X86GenCallingConv.inc"
1177 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1178 const SmallVectorImpl<EVT> &OutTys,
1179 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1180 SelectionDAG &DAG) {
1181 SmallVector<CCValAssign, 16> RVLocs;
1182 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1183 RVLocs, *DAG.getContext());
1184 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1188 X86TargetLowering::LowerReturn(SDValue Chain,
1189 CallingConv::ID CallConv, bool isVarArg,
1190 const SmallVectorImpl<ISD::OutputArg> &Outs,
1191 DebugLoc dl, SelectionDAG &DAG) {
1193 SmallVector<CCValAssign, 16> RVLocs;
1194 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1195 RVLocs, *DAG.getContext());
1196 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1198 // If this is the first return lowered for this function, add the regs to the
1199 // liveout set for the function.
1200 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1201 for (unsigned i = 0; i != RVLocs.size(); ++i)
1202 if (RVLocs[i].isRegLoc())
1203 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1208 SmallVector<SDValue, 6> RetOps;
1209 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1210 // Operand #1 = Bytes To Pop
1211 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1213 // Copy the result values into the output registers.
1214 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1215 CCValAssign &VA = RVLocs[i];
1216 assert(VA.isRegLoc() && "Can only return in registers!");
1217 SDValue ValToCopy = Outs[i].Val;
1219 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1220 // the RET instruction and handled by the FP Stackifier.
1221 if (VA.getLocReg() == X86::ST0 ||
1222 VA.getLocReg() == X86::ST1) {
1223 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1224 // change the value to the FP stack register class.
1225 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1226 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1227 RetOps.push_back(ValToCopy);
1228 // Don't emit a copytoreg.
1232 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1233 // which is returned in RAX / RDX.
1234 if (Subtarget->is64Bit()) {
1235 EVT ValVT = ValToCopy.getValueType();
1236 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1237 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1238 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1239 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1243 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1244 Flag = Chain.getValue(1);
1247 // The x86-64 ABI for returning structs by value requires that we copy
1248 // the sret argument into %rax for the return. We saved the argument into
1249 // a virtual register in the entry block, so now we copy the value out
1251 if (Subtarget->is64Bit() &&
1252 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1253 MachineFunction &MF = DAG.getMachineFunction();
1254 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1255 unsigned Reg = FuncInfo->getSRetReturnReg();
1257 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1258 FuncInfo->setSRetReturnReg(Reg);
1260 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1262 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1263 Flag = Chain.getValue(1);
1265 // RAX now acts like a return value.
1266 MF.getRegInfo().addLiveOut(X86::RAX);
1269 RetOps[0] = Chain; // Update chain.
1271 // Add the flag if we have it.
1273 RetOps.push_back(Flag);
1275 return DAG.getNode(X86ISD::RET_FLAG, dl,
1276 MVT::Other, &RetOps[0], RetOps.size());
1279 /// LowerCallResult - Lower the result values of a call into the
1280 /// appropriate copies out of appropriate physical registers.
1283 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1284 CallingConv::ID CallConv, bool isVarArg,
1285 const SmallVectorImpl<ISD::InputArg> &Ins,
1286 DebugLoc dl, SelectionDAG &DAG,
1287 SmallVectorImpl<SDValue> &InVals) {
1289 // Assign locations to each value returned by this call.
1290 SmallVector<CCValAssign, 16> RVLocs;
1291 bool Is64Bit = Subtarget->is64Bit();
1292 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1293 RVLocs, *DAG.getContext());
1294 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1296 // Copy all of the result registers out of their specified physreg.
1297 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1298 CCValAssign &VA = RVLocs[i];
1299 EVT CopyVT = VA.getValVT();
1301 // If this is x86-64, and we disabled SSE, we can't return FP values
1302 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1303 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1304 llvm_report_error("SSE register return with SSE disabled");
1307 // If this is a call to a function that returns an fp value on the floating
1308 // point stack, but where we prefer to use the value in xmm registers, copy
1309 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1310 if ((VA.getLocReg() == X86::ST0 ||
1311 VA.getLocReg() == X86::ST1) &&
1312 isScalarFPTypeInSSEReg(VA.getValVT())) {
1317 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1318 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1319 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1320 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1321 MVT::v2i64, InFlag).getValue(1);
1322 Val = Chain.getValue(0);
1323 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1324 Val, DAG.getConstant(0, MVT::i64));
1326 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1327 MVT::i64, InFlag).getValue(1);
1328 Val = Chain.getValue(0);
1330 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1332 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1333 CopyVT, InFlag).getValue(1);
1334 Val = Chain.getValue(0);
1336 InFlag = Chain.getValue(2);
1338 if (CopyVT != VA.getValVT()) {
1339 // Round the F80 the right size, which also moves to the appropriate xmm
1341 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1342 // This truncation won't change the value.
1343 DAG.getIntPtrConstant(1));
1346 InVals.push_back(Val);
1353 //===----------------------------------------------------------------------===//
1354 // C & StdCall & Fast Calling Convention implementation
1355 //===----------------------------------------------------------------------===//
1356 // StdCall calling convention seems to be standard for many Windows' API
1357 // routines and around. It differs from C calling convention just a little:
1358 // callee should clean up the stack, not caller. Symbols should be also
1359 // decorated in some fancy way :) It doesn't support any vector arguments.
1360 // For info on fast calling convention see Fast Calling Convention (tail call)
1361 // implementation LowerX86_32FastCCCallTo.
1363 /// CallIsStructReturn - Determines whether a call uses struct return
1365 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1369 return Outs[0].Flags.isSRet();
1372 /// ArgsAreStructReturn - Determines whether a function uses struct
1373 /// return semantics.
1375 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1379 return Ins[0].Flags.isSRet();
1382 /// IsCalleePop - Determines whether the callee is required to pop its
1383 /// own arguments. Callee pop is necessary to support tail calls.
1384 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1388 switch (CallingConv) {
1391 case CallingConv::X86_StdCall:
1392 return !Subtarget->is64Bit();
1393 case CallingConv::X86_FastCall:
1394 return !Subtarget->is64Bit();
1395 case CallingConv::Fast:
1396 return PerformTailCallOpt;
1400 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1401 /// given CallingConvention value.
1402 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1403 if (Subtarget->is64Bit()) {
1404 if (Subtarget->isTargetWin64())
1405 return CC_X86_Win64_C;
1410 if (CC == CallingConv::X86_FastCall)
1411 return CC_X86_32_FastCall;
1412 else if (CC == CallingConv::Fast)
1413 return CC_X86_32_FastCC;
1418 /// NameDecorationForCallConv - Selects the appropriate decoration to
1419 /// apply to a MachineFunction containing a given calling convention.
1421 X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
1422 if (CallConv == CallingConv::X86_FastCall)
1424 else if (CallConv == CallingConv::X86_StdCall)
1430 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1431 /// by "Src" to address "Dst" with size and alignment information specified by
1432 /// the specific parameter attribute. The copy will be passed as a byval
1433 /// function parameter.
1435 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1436 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1438 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1439 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1440 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1443 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1444 /// a tailcall target by changing its ABI.
1445 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1446 return PerformTailCallOpt && CC == CallingConv::Fast;
1450 X86TargetLowering::LowerMemArgument(SDValue Chain,
1451 CallingConv::ID CallConv,
1452 const SmallVectorImpl<ISD::InputArg> &Ins,
1453 DebugLoc dl, SelectionDAG &DAG,
1454 const CCValAssign &VA,
1455 MachineFrameInfo *MFI,
1457 // Create the nodes corresponding to a load from this parameter slot.
1458 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1459 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1460 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1463 // If value is passed by pointer we have address passed instead of the value
1465 if (VA.getLocInfo() == CCValAssign::Indirect)
1466 ValVT = VA.getLocVT();
1468 ValVT = VA.getValVT();
1470 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1471 // changed with more analysis.
1472 // In case of tail call optimization mark all arguments mutable. Since they
1473 // could be overwritten by lowering of arguments in case of a tail call.
1474 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1475 VA.getLocMemOffset(), isImmutable, false);
1476 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1477 if (Flags.isByVal())
1479 return DAG.getLoad(ValVT, dl, Chain, FIN,
1480 PseudoSourceValue::getFixedStack(FI), 0);
1484 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1485 CallingConv::ID CallConv,
1487 const SmallVectorImpl<ISD::InputArg> &Ins,
1490 SmallVectorImpl<SDValue> &InVals) {
1492 MachineFunction &MF = DAG.getMachineFunction();
1493 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1495 const Function* Fn = MF.getFunction();
1496 if (Fn->hasExternalLinkage() &&
1497 Subtarget->isTargetCygMing() &&
1498 Fn->getName() == "main")
1499 FuncInfo->setForceFramePointer(true);
1501 // Decorate the function name.
1502 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1504 MachineFrameInfo *MFI = MF.getFrameInfo();
1505 bool Is64Bit = Subtarget->is64Bit();
1506 bool IsWin64 = Subtarget->isTargetWin64();
1508 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1509 "Var args not supported with calling convention fastcc");
1511 // Assign locations to all of the incoming arguments.
1512 SmallVector<CCValAssign, 16> ArgLocs;
1513 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1514 ArgLocs, *DAG.getContext());
1515 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1517 unsigned LastVal = ~0U;
1519 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1520 CCValAssign &VA = ArgLocs[i];
1521 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1523 assert(VA.getValNo() != LastVal &&
1524 "Don't support value assigned to multiple locs yet");
1525 LastVal = VA.getValNo();
1527 if (VA.isRegLoc()) {
1528 EVT RegVT = VA.getLocVT();
1529 TargetRegisterClass *RC = NULL;
1530 if (RegVT == MVT::i32)
1531 RC = X86::GR32RegisterClass;
1532 else if (Is64Bit && RegVT == MVT::i64)
1533 RC = X86::GR64RegisterClass;
1534 else if (RegVT == MVT::f32)
1535 RC = X86::FR32RegisterClass;
1536 else if (RegVT == MVT::f64)
1537 RC = X86::FR64RegisterClass;
1538 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1539 RC = X86::VR128RegisterClass;
1540 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1541 RC = X86::VR64RegisterClass;
1543 llvm_unreachable("Unknown argument type!");
1545 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1546 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1548 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1549 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1551 if (VA.getLocInfo() == CCValAssign::SExt)
1552 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1553 DAG.getValueType(VA.getValVT()));
1554 else if (VA.getLocInfo() == CCValAssign::ZExt)
1555 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1556 DAG.getValueType(VA.getValVT()));
1557 else if (VA.getLocInfo() == CCValAssign::BCvt)
1558 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1560 if (VA.isExtInLoc()) {
1561 // Handle MMX values passed in XMM regs.
1562 if (RegVT.isVector()) {
1563 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1564 ArgValue, DAG.getConstant(0, MVT::i64));
1565 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1567 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1570 assert(VA.isMemLoc());
1571 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1574 // If value is passed via pointer - do a load.
1575 if (VA.getLocInfo() == CCValAssign::Indirect)
1576 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1578 InVals.push_back(ArgValue);
1581 // The x86-64 ABI for returning structs by value requires that we copy
1582 // the sret argument into %rax for the return. Save the argument into
1583 // a virtual register so that we can access it from the return points.
1584 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1585 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1586 unsigned Reg = FuncInfo->getSRetReturnReg();
1588 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1589 FuncInfo->setSRetReturnReg(Reg);
1591 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1592 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1595 unsigned StackSize = CCInfo.getNextStackOffset();
1596 // Align stack specially for tail calls.
1597 if (FuncIsMadeTailCallSafe(CallConv))
1598 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1600 // If the function takes variable number of arguments, make a frame index for
1601 // the start of the first vararg value... for expansion of llvm.va_start.
1603 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1604 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1607 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1609 // FIXME: We should really autogenerate these arrays
1610 static const unsigned GPR64ArgRegsWin64[] = {
1611 X86::RCX, X86::RDX, X86::R8, X86::R9
1613 static const unsigned XMMArgRegsWin64[] = {
1614 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1616 static const unsigned GPR64ArgRegs64Bit[] = {
1617 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1619 static const unsigned XMMArgRegs64Bit[] = {
1620 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1621 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1623 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1626 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1627 GPR64ArgRegs = GPR64ArgRegsWin64;
1628 XMMArgRegs = XMMArgRegsWin64;
1630 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1631 GPR64ArgRegs = GPR64ArgRegs64Bit;
1632 XMMArgRegs = XMMArgRegs64Bit;
1634 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1636 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1639 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1640 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1641 "SSE register cannot be used when SSE is disabled!");
1642 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1643 "SSE register cannot be used when SSE is disabled!");
1644 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1645 // Kernel mode asks for SSE to be disabled, so don't push them
1647 TotalNumXMMRegs = 0;
1649 // For X86-64, if there are vararg parameters that are passed via
1650 // registers, then we must store them to their spots on the stack so they
1651 // may be loaded by deferencing the result of va_next.
1652 VarArgsGPOffset = NumIntRegs * 8;
1653 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1654 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1655 TotalNumXMMRegs * 16, 16,
1658 // Store the integer parameter registers.
1659 SmallVector<SDValue, 8> MemOps;
1660 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1661 unsigned Offset = VarArgsGPOffset;
1662 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1663 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1664 DAG.getIntPtrConstant(Offset));
1665 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1666 X86::GR64RegisterClass);
1667 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1669 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1670 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1672 MemOps.push_back(Store);
1676 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1677 // Now store the XMM (fp + vector) parameter registers.
1678 SmallVector<SDValue, 11> SaveXMMOps;
1679 SaveXMMOps.push_back(Chain);
1681 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1682 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1683 SaveXMMOps.push_back(ALVal);
1685 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1686 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1688 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1689 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1690 X86::VR128RegisterClass);
1691 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1692 SaveXMMOps.push_back(Val);
1694 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1696 &SaveXMMOps[0], SaveXMMOps.size()));
1699 if (!MemOps.empty())
1700 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1701 &MemOps[0], MemOps.size());
1705 // Some CCs need callee pop.
1706 if (IsCalleePop(isVarArg, CallConv)) {
1707 BytesToPopOnReturn = StackSize; // Callee pops everything.
1709 BytesToPopOnReturn = 0; // Callee pops nothing.
1710 // If this is an sret function, the return should pop the hidden pointer.
1711 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1712 BytesToPopOnReturn = 4;
1716 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1717 if (CallConv == CallingConv::X86_FastCall)
1718 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1721 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1727 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1728 SDValue StackPtr, SDValue Arg,
1729 DebugLoc dl, SelectionDAG &DAG,
1730 const CCValAssign &VA,
1731 ISD::ArgFlagsTy Flags) {
1732 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1733 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1734 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1735 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1736 if (Flags.isByVal()) {
1737 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1739 return DAG.getStore(Chain, dl, Arg, PtrOff,
1740 PseudoSourceValue::getStack(), LocMemOffset);
1743 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1744 /// optimization is performed and it is required.
1746 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1747 SDValue &OutRetAddr, SDValue Chain,
1748 bool IsTailCall, bool Is64Bit,
1749 int FPDiff, DebugLoc dl) {
1750 if (!IsTailCall || FPDiff==0) return Chain;
1752 // Adjust the Return address stack slot.
1753 EVT VT = getPointerTy();
1754 OutRetAddr = getReturnAddressFrameIndex(DAG);
1756 // Load the "old" Return address.
1757 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1758 return SDValue(OutRetAddr.getNode(), 1);
1761 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1762 /// optimization is performed and it is required (FPDiff!=0).
1764 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1765 SDValue Chain, SDValue RetAddrFrIdx,
1766 bool Is64Bit, int FPDiff, DebugLoc dl) {
1767 // Store the return address to the appropriate stack slot.
1768 if (!FPDiff) return Chain;
1769 // Calculate the new stack slot for the return address.
1770 int SlotSize = Is64Bit ? 8 : 4;
1771 int NewReturnAddrFI =
1772 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, true,false);
1773 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1774 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1775 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1776 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1781 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1782 CallingConv::ID CallConv, bool isVarArg,
1784 const SmallVectorImpl<ISD::OutputArg> &Outs,
1785 const SmallVectorImpl<ISD::InputArg> &Ins,
1786 DebugLoc dl, SelectionDAG &DAG,
1787 SmallVectorImpl<SDValue> &InVals) {
1788 MachineFunction &MF = DAG.getMachineFunction();
1789 bool Is64Bit = Subtarget->is64Bit();
1790 bool IsStructRet = CallIsStructReturn(Outs);
1793 // Check if it's really possible to do a tail call.
1794 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
1797 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1798 "Var args not supported with calling convention fastcc");
1800 // Analyze operands of the call, assigning locations to each operand.
1801 SmallVector<CCValAssign, 16> ArgLocs;
1802 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1803 ArgLocs, *DAG.getContext());
1804 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1806 // Get a count of how many bytes are to be pushed on the stack.
1807 unsigned NumBytes = CCInfo.getNextStackOffset();
1808 if (FuncIsMadeTailCallSafe(CallConv))
1809 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1815 // Lower arguments at fp - stackoffset + fpdiff.
1816 unsigned NumBytesCallerPushed =
1817 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1818 FPDiff = NumBytesCallerPushed - NumBytes;
1820 // Set the delta of movement of the returnaddr stackslot.
1821 // But only set if delta is greater than previous delta.
1822 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1823 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1826 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1828 SDValue RetAddrFrIdx;
1829 // Load return adress for tail calls.
1830 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
1833 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1834 SmallVector<SDValue, 8> MemOpChains;
1837 // Walk the register/memloc assignments, inserting copies/loads. In the case
1838 // of tail call optimization arguments are handle later.
1839 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1840 CCValAssign &VA = ArgLocs[i];
1841 EVT RegVT = VA.getLocVT();
1842 SDValue Arg = Outs[i].Val;
1843 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1844 bool isByVal = Flags.isByVal();
1846 // Promote the value if needed.
1847 switch (VA.getLocInfo()) {
1848 default: llvm_unreachable("Unknown loc info!");
1849 case CCValAssign::Full: break;
1850 case CCValAssign::SExt:
1851 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1853 case CCValAssign::ZExt:
1854 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1856 case CCValAssign::AExt:
1857 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1858 // Special case: passing MMX values in XMM registers.
1859 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1860 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1861 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1863 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1865 case CCValAssign::BCvt:
1866 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1868 case CCValAssign::Indirect: {
1869 // Store the argument.
1870 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1871 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1872 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1873 PseudoSourceValue::getFixedStack(FI), 0);
1879 if (VA.isRegLoc()) {
1880 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1882 if (!isTailCall || (isTailCall && isByVal)) {
1883 assert(VA.isMemLoc());
1884 if (StackPtr.getNode() == 0)
1885 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1887 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1888 dl, DAG, VA, Flags));
1893 if (!MemOpChains.empty())
1894 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1895 &MemOpChains[0], MemOpChains.size());
1897 // Build a sequence of copy-to-reg nodes chained together with token chain
1898 // and flag operands which copy the outgoing args into registers.
1900 // Tail call byval lowering might overwrite argument registers so in case of
1901 // tail call optimization the copies to registers are lowered later.
1903 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1904 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1905 RegsToPass[i].second, InFlag);
1906 InFlag = Chain.getValue(1);
1910 if (Subtarget->isPICStyleGOT()) {
1911 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1914 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1915 DAG.getNode(X86ISD::GlobalBaseReg,
1916 DebugLoc::getUnknownLoc(),
1919 InFlag = Chain.getValue(1);
1921 // If we are tail calling and generating PIC/GOT style code load the
1922 // address of the callee into ECX. The value in ecx is used as target of
1923 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1924 // for tail calls on PIC/GOT architectures. Normally we would just put the
1925 // address of GOT into ebx and then call target@PLT. But for tail calls
1926 // ebx would be restored (since ebx is callee saved) before jumping to the
1929 // Note: The actual moving to ECX is done further down.
1930 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1931 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1932 !G->getGlobal()->hasProtectedVisibility())
1933 Callee = LowerGlobalAddress(Callee, DAG);
1934 else if (isa<ExternalSymbolSDNode>(Callee))
1935 Callee = LowerExternalSymbol(Callee, DAG);
1939 if (Is64Bit && isVarArg) {
1940 // From AMD64 ABI document:
1941 // For calls that may call functions that use varargs or stdargs
1942 // (prototype-less calls or calls to functions containing ellipsis (...) in
1943 // the declaration) %al is used as hidden argument to specify the number
1944 // of SSE registers used. The contents of %al do not need to match exactly
1945 // the number of registers, but must be an ubound on the number of SSE
1946 // registers used and is in the range 0 - 8 inclusive.
1948 // FIXME: Verify this on Win64
1949 // Count the number of XMM registers allocated.
1950 static const unsigned XMMArgRegs[] = {
1951 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1952 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1954 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1955 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1956 && "SSE registers cannot be used when SSE is disabled");
1958 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1959 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1960 InFlag = Chain.getValue(1);
1964 // For tail calls lower the arguments to the 'real' stack slot.
1966 // Force all the incoming stack arguments to be loaded from the stack
1967 // before any new outgoing arguments are stored to the stack, because the
1968 // outgoing stack slots may alias the incoming argument stack slots, and
1969 // the alias isn't otherwise explicit. This is slightly more conservative
1970 // than necessary, because it means that each store effectively depends
1971 // on every argument instead of just those arguments it would clobber.
1972 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1974 SmallVector<SDValue, 8> MemOpChains2;
1977 // Do not flag preceeding copytoreg stuff together with the following stuff.
1979 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1980 CCValAssign &VA = ArgLocs[i];
1981 if (!VA.isRegLoc()) {
1982 assert(VA.isMemLoc());
1983 SDValue Arg = Outs[i].Val;
1984 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1985 // Create frame index.
1986 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1987 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1988 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1989 FIN = DAG.getFrameIndex(FI, getPointerTy());
1991 if (Flags.isByVal()) {
1992 // Copy relative to framepointer.
1993 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1994 if (StackPtr.getNode() == 0)
1995 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1997 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1999 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2003 // Store relative to framepointer.
2004 MemOpChains2.push_back(
2005 DAG.getStore(ArgChain, dl, Arg, FIN,
2006 PseudoSourceValue::getFixedStack(FI), 0));
2011 if (!MemOpChains2.empty())
2012 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2013 &MemOpChains2[0], MemOpChains2.size());
2015 // Copy arguments to their registers.
2016 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2017 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2018 RegsToPass[i].second, InFlag);
2019 InFlag = Chain.getValue(1);
2023 // Store the return address to the appropriate stack slot.
2024 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2028 bool WasGlobalOrExternal = false;
2029 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2030 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2031 // In the 64-bit large code model, we have to make all calls
2032 // through a register, since the call instruction's 32-bit
2033 // pc-relative offset may not be large enough to hold the whole
2035 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2036 WasGlobalOrExternal = true;
2037 // If the callee is a GlobalAddress node (quite common, every direct call
2038 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2041 // We should use extra load for direct calls to dllimported functions in
2043 GlobalValue *GV = G->getGlobal();
2044 if (!GV->hasDLLImportLinkage()) {
2045 unsigned char OpFlags = 0;
2047 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2048 // external symbols most go through the PLT in PIC mode. If the symbol
2049 // has hidden or protected visibility, or if it is static or local, then
2050 // we don't need to use the PLT - we can directly call it.
2051 if (Subtarget->isTargetELF() &&
2052 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2053 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2054 OpFlags = X86II::MO_PLT;
2055 } else if (Subtarget->isPICStyleStubAny() &&
2056 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2057 Subtarget->getDarwinVers() < 9) {
2058 // PC-relative references to external symbols should go through $stub,
2059 // unless we're building with the leopard linker or later, which
2060 // automatically synthesizes these stubs.
2061 OpFlags = X86II::MO_DARWIN_STUB;
2064 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2065 G->getOffset(), OpFlags);
2067 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2068 WasGlobalOrExternal = true;
2069 unsigned char OpFlags = 0;
2071 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2072 // symbols should go through the PLT.
2073 if (Subtarget->isTargetELF() &&
2074 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2075 OpFlags = X86II::MO_PLT;
2076 } else if (Subtarget->isPICStyleStubAny() &&
2077 Subtarget->getDarwinVers() < 9) {
2078 // PC-relative references to external symbols should go through $stub,
2079 // unless we're building with the leopard linker or later, which
2080 // automatically synthesizes these stubs.
2081 OpFlags = X86II::MO_DARWIN_STUB;
2084 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2088 if (isTailCall && !WasGlobalOrExternal) {
2089 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
2091 Chain = DAG.getCopyToReg(Chain, dl,
2092 DAG.getRegister(Opc, getPointerTy()),
2094 Callee = DAG.getRegister(Opc, getPointerTy());
2095 // Add register as live out.
2096 MF.getRegInfo().addLiveOut(Opc);
2099 // Returns a chain & a flag for retval copy to use.
2100 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2101 SmallVector<SDValue, 8> Ops;
2104 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2105 DAG.getIntPtrConstant(0, true), InFlag);
2106 InFlag = Chain.getValue(1);
2109 Ops.push_back(Chain);
2110 Ops.push_back(Callee);
2113 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2115 // Add argument registers to the end of the list so that they are known live
2117 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2118 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2119 RegsToPass[i].second.getValueType()));
2121 // Add an implicit use GOT pointer in EBX.
2122 if (!isTailCall && Subtarget->isPICStyleGOT())
2123 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2125 // Add an implicit use of AL for x86 vararg functions.
2126 if (Is64Bit && isVarArg)
2127 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2129 if (InFlag.getNode())
2130 Ops.push_back(InFlag);
2133 // If this is the first return lowered for this function, add the regs
2134 // to the liveout set for the function.
2135 if (MF.getRegInfo().liveout_empty()) {
2136 SmallVector<CCValAssign, 16> RVLocs;
2137 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2139 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2140 for (unsigned i = 0; i != RVLocs.size(); ++i)
2141 if (RVLocs[i].isRegLoc())
2142 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2145 assert(((Callee.getOpcode() == ISD::Register &&
2146 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2147 cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
2148 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2149 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2150 "Expecting a global address, external symbol, or scratch register");
2152 return DAG.getNode(X86ISD::TC_RETURN, dl,
2153 NodeTys, &Ops[0], Ops.size());
2156 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2157 InFlag = Chain.getValue(1);
2159 // Create the CALLSEQ_END node.
2160 unsigned NumBytesForCalleeToPush;
2161 if (IsCalleePop(isVarArg, CallConv))
2162 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2163 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2164 // If this is is a call to a struct-return function, the callee
2165 // pops the hidden struct pointer, so we have to push it back.
2166 // This is common for Darwin/X86, Linux & Mingw32 targets.
2167 NumBytesForCalleeToPush = 4;
2169 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2171 // Returns a flag for retval copy to use.
2172 Chain = DAG.getCALLSEQ_END(Chain,
2173 DAG.getIntPtrConstant(NumBytes, true),
2174 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2177 InFlag = Chain.getValue(1);
2179 // Handle result values, copying them out of physregs into vregs that we
2181 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2182 Ins, dl, DAG, InVals);
2186 //===----------------------------------------------------------------------===//
2187 // Fast Calling Convention (tail call) implementation
2188 //===----------------------------------------------------------------------===//
2190 // Like std call, callee cleans arguments, convention except that ECX is
2191 // reserved for storing the tail called function address. Only 2 registers are
2192 // free for argument passing (inreg). Tail call optimization is performed
2194 // * tailcallopt is enabled
2195 // * caller/callee are fastcc
2196 // On X86_64 architecture with GOT-style position independent code only local
2197 // (within module) calls are supported at the moment.
2198 // To keep the stack aligned according to platform abi the function
2199 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2200 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2201 // If a tail called function callee has more arguments than the caller the
2202 // caller needs to make sure that there is room to move the RETADDR to. This is
2203 // achieved by reserving an area the size of the argument delta right after the
2204 // original REtADDR, but before the saved framepointer or the spilled registers
2205 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2217 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2218 /// for a 16 byte align requirement.
2219 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2220 SelectionDAG& DAG) {
2221 MachineFunction &MF = DAG.getMachineFunction();
2222 const TargetMachine &TM = MF.getTarget();
2223 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2224 unsigned StackAlignment = TFI.getStackAlignment();
2225 uint64_t AlignMask = StackAlignment - 1;
2226 int64_t Offset = StackSize;
2227 uint64_t SlotSize = TD->getPointerSize();
2228 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2229 // Number smaller than 12 so just add the difference.
2230 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2232 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2233 Offset = ((~AlignMask) & Offset) + StackAlignment +
2234 (StackAlignment-SlotSize);
2239 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2240 /// for tail call optimization. Targets which want to do tail call
2241 /// optimization should implement this function.
2243 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2244 CallingConv::ID CalleeCC,
2246 const SmallVectorImpl<ISD::OutputArg> &Outs,
2247 const SmallVectorImpl<ISD::InputArg> &Ins,
2248 SelectionDAG& DAG) const {
2249 if (CalleeCC != CallingConv::Fast &&
2250 CalleeCC != CallingConv::C)
2253 // If -tailcallopt is specified, make fastcc functions tail-callable.
2254 const Function *CallerF = DAG.getMachineFunction().getFunction();
2255 if (PerformTailCallOpt &&
2256 CalleeCC == CallingConv::Fast &&
2257 CallerF->getCallingConv() == CalleeCC)
2260 // Look for obvious safe cases to perform tail call optimization.
2261 // For now, only consider callees which take no arguments.
2265 // If the caller does not return a value, then this is obviously safe.
2266 // This is one case where it's safe to perform this optimization even
2267 // if the return types do not match.
2268 const Type *CallerRetTy = CallerF->getReturnType();
2269 if (CallerRetTy->isVoidTy())
2272 // If the return types match, then it's safe.
2273 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2274 if (!G) return false; // FIXME: common external symbols?
2275 Function *CalleeF = cast<Function>(G->getGlobal());
2276 const Type *CalleeRetTy = CalleeF->getReturnType();
2277 return CallerRetTy == CalleeRetTy;
2281 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2283 DenseMap<const Value *, unsigned> &vm,
2284 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2285 DenseMap<const AllocaInst *, int> &am
2287 , SmallSet<Instruction*, 8> &cil
2290 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2298 //===----------------------------------------------------------------------===//
2299 // Other Lowering Hooks
2300 //===----------------------------------------------------------------------===//
2303 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2304 MachineFunction &MF = DAG.getMachineFunction();
2305 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2306 int ReturnAddrIndex = FuncInfo->getRAIndex();
2308 if (ReturnAddrIndex == 0) {
2309 // Set up a frame object for the return address.
2310 uint64_t SlotSize = TD->getPointerSize();
2311 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2313 FuncInfo->setRAIndex(ReturnAddrIndex);
2316 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2320 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2321 bool hasSymbolicDisplacement) {
2322 // Offset should fit into 32 bit immediate field.
2323 if (!isInt32(Offset))
2326 // If we don't have a symbolic displacement - we don't have any extra
2328 if (!hasSymbolicDisplacement)
2331 // FIXME: Some tweaks might be needed for medium code model.
2332 if (M != CodeModel::Small && M != CodeModel::Kernel)
2335 // For small code model we assume that latest object is 16MB before end of 31
2336 // bits boundary. We may also accept pretty large negative constants knowing
2337 // that all objects are in the positive half of address space.
2338 if (M == CodeModel::Small && Offset < 16*1024*1024)
2341 // For kernel code model we know that all object resist in the negative half
2342 // of 32bits address space. We may not accept negative offsets, since they may
2343 // be just off and we may accept pretty large positive ones.
2344 if (M == CodeModel::Kernel && Offset > 0)
2350 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2351 /// specific condition code, returning the condition code and the LHS/RHS of the
2352 /// comparison to make.
2353 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2354 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2356 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2357 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2358 // X > -1 -> X == 0, jump !sign.
2359 RHS = DAG.getConstant(0, RHS.getValueType());
2360 return X86::COND_NS;
2361 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2362 // X < 0 -> X == 0, jump on sign.
2364 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2366 RHS = DAG.getConstant(0, RHS.getValueType());
2367 return X86::COND_LE;
2371 switch (SetCCOpcode) {
2372 default: llvm_unreachable("Invalid integer condition!");
2373 case ISD::SETEQ: return X86::COND_E;
2374 case ISD::SETGT: return X86::COND_G;
2375 case ISD::SETGE: return X86::COND_GE;
2376 case ISD::SETLT: return X86::COND_L;
2377 case ISD::SETLE: return X86::COND_LE;
2378 case ISD::SETNE: return X86::COND_NE;
2379 case ISD::SETULT: return X86::COND_B;
2380 case ISD::SETUGT: return X86::COND_A;
2381 case ISD::SETULE: return X86::COND_BE;
2382 case ISD::SETUGE: return X86::COND_AE;
2386 // First determine if it is required or is profitable to flip the operands.
2388 // If LHS is a foldable load, but RHS is not, flip the condition.
2389 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2390 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2391 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2392 std::swap(LHS, RHS);
2395 switch (SetCCOpcode) {
2401 std::swap(LHS, RHS);
2405 // On a floating point condition, the flags are set as follows:
2407 // 0 | 0 | 0 | X > Y
2408 // 0 | 0 | 1 | X < Y
2409 // 1 | 0 | 0 | X == Y
2410 // 1 | 1 | 1 | unordered
2411 switch (SetCCOpcode) {
2412 default: llvm_unreachable("Condcode should be pre-legalized away");
2414 case ISD::SETEQ: return X86::COND_E;
2415 case ISD::SETOLT: // flipped
2417 case ISD::SETGT: return X86::COND_A;
2418 case ISD::SETOLE: // flipped
2420 case ISD::SETGE: return X86::COND_AE;
2421 case ISD::SETUGT: // flipped
2423 case ISD::SETLT: return X86::COND_B;
2424 case ISD::SETUGE: // flipped
2426 case ISD::SETLE: return X86::COND_BE;
2428 case ISD::SETNE: return X86::COND_NE;
2429 case ISD::SETUO: return X86::COND_P;
2430 case ISD::SETO: return X86::COND_NP;
2432 case ISD::SETUNE: return X86::COND_INVALID;
2436 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2437 /// code. Current x86 isa includes the following FP cmov instructions:
2438 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2439 static bool hasFPCMov(unsigned X86CC) {
2455 /// isFPImmLegal - Returns true if the target can instruction select the
2456 /// specified FP immediate natively. If false, the legalizer will
2457 /// materialize the FP immediate as a load from a constant pool.
2458 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2459 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2460 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2466 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2467 /// the specified range (L, H].
2468 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2469 return (Val < 0) || (Val >= Low && Val < Hi);
2472 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2473 /// specified value.
2474 static bool isUndefOrEqual(int Val, int CmpVal) {
2475 if (Val < 0 || Val == CmpVal)
2480 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2481 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2482 /// the second operand.
2483 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2484 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2485 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2486 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2487 return (Mask[0] < 2 && Mask[1] < 2);
2491 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2492 SmallVector<int, 8> M;
2494 return ::isPSHUFDMask(M, N->getValueType(0));
2497 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2498 /// is suitable for input to PSHUFHW.
2499 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2500 if (VT != MVT::v8i16)
2503 // Lower quadword copied in order or undef.
2504 for (int i = 0; i != 4; ++i)
2505 if (Mask[i] >= 0 && Mask[i] != i)
2508 // Upper quadword shuffled.
2509 for (int i = 4; i != 8; ++i)
2510 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2516 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2517 SmallVector<int, 8> M;
2519 return ::isPSHUFHWMask(M, N->getValueType(0));
2522 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2523 /// is suitable for input to PSHUFLW.
2524 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2525 if (VT != MVT::v8i16)
2528 // Upper quadword copied in order.
2529 for (int i = 4; i != 8; ++i)
2530 if (Mask[i] >= 0 && Mask[i] != i)
2533 // Lower quadword shuffled.
2534 for (int i = 0; i != 4; ++i)
2541 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2542 SmallVector<int, 8> M;
2544 return ::isPSHUFLWMask(M, N->getValueType(0));
2547 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2548 /// is suitable for input to PALIGNR.
2549 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2551 int i, e = VT.getVectorNumElements();
2553 // Do not handle v2i64 / v2f64 shuffles with palignr.
2554 if (e < 4 || !hasSSSE3)
2557 for (i = 0; i != e; ++i)
2561 // All undef, not a palignr.
2565 // Determine if it's ok to perform a palignr with only the LHS, since we
2566 // don't have access to the actual shuffle elements to see if RHS is undef.
2567 bool Unary = Mask[i] < (int)e;
2568 bool NeedsUnary = false;
2570 int s = Mask[i] - i;
2572 // Check the rest of the elements to see if they are consecutive.
2573 for (++i; i != e; ++i) {
2578 Unary = Unary && (m < (int)e);
2579 NeedsUnary = NeedsUnary || (m < s);
2581 if (NeedsUnary && !Unary)
2583 if (Unary && m != ((s+i) & (e-1)))
2585 if (!Unary && m != (s+i))
2591 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2592 SmallVector<int, 8> M;
2594 return ::isPALIGNRMask(M, N->getValueType(0), true);
2597 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2598 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2599 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2600 int NumElems = VT.getVectorNumElements();
2601 if (NumElems != 2 && NumElems != 4)
2604 int Half = NumElems / 2;
2605 for (int i = 0; i < Half; ++i)
2606 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2608 for (int i = Half; i < NumElems; ++i)
2609 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2615 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2616 SmallVector<int, 8> M;
2618 return ::isSHUFPMask(M, N->getValueType(0));
2621 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2622 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2623 /// half elements to come from vector 1 (which would equal the dest.) and
2624 /// the upper half to come from vector 2.
2625 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2626 int NumElems = VT.getVectorNumElements();
2628 if (NumElems != 2 && NumElems != 4)
2631 int Half = NumElems / 2;
2632 for (int i = 0; i < Half; ++i)
2633 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2635 for (int i = Half; i < NumElems; ++i)
2636 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2641 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2642 SmallVector<int, 8> M;
2644 return isCommutedSHUFPMask(M, N->getValueType(0));
2647 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2648 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2649 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2650 if (N->getValueType(0).getVectorNumElements() != 4)
2653 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2654 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2655 isUndefOrEqual(N->getMaskElt(1), 7) &&
2656 isUndefOrEqual(N->getMaskElt(2), 2) &&
2657 isUndefOrEqual(N->getMaskElt(3), 3);
2660 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2661 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2663 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2664 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2669 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2670 isUndefOrEqual(N->getMaskElt(1), 3) &&
2671 isUndefOrEqual(N->getMaskElt(2), 2) &&
2672 isUndefOrEqual(N->getMaskElt(3), 3);
2675 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2676 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2677 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2678 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2680 if (NumElems != 2 && NumElems != 4)
2683 for (unsigned i = 0; i < NumElems/2; ++i)
2684 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2687 for (unsigned i = NumElems/2; i < NumElems; ++i)
2688 if (!isUndefOrEqual(N->getMaskElt(i), i))
2694 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2695 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2696 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2697 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2699 if (NumElems != 2 && NumElems != 4)
2702 for (unsigned i = 0; i < NumElems/2; ++i)
2703 if (!isUndefOrEqual(N->getMaskElt(i), i))
2706 for (unsigned i = 0; i < NumElems/2; ++i)
2707 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2713 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2714 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2715 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2716 bool V2IsSplat = false) {
2717 int NumElts = VT.getVectorNumElements();
2718 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2721 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2723 int BitI1 = Mask[i+1];
2724 if (!isUndefOrEqual(BitI, j))
2727 if (!isUndefOrEqual(BitI1, NumElts))
2730 if (!isUndefOrEqual(BitI1, j + NumElts))
2737 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2738 SmallVector<int, 8> M;
2740 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2743 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2744 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2745 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2746 bool V2IsSplat = false) {
2747 int NumElts = VT.getVectorNumElements();
2748 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2751 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2753 int BitI1 = Mask[i+1];
2754 if (!isUndefOrEqual(BitI, j + NumElts/2))
2757 if (isUndefOrEqual(BitI1, NumElts))
2760 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2767 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2768 SmallVector<int, 8> M;
2770 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2773 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2774 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2776 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2777 int NumElems = VT.getVectorNumElements();
2778 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2781 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2783 int BitI1 = Mask[i+1];
2784 if (!isUndefOrEqual(BitI, j))
2786 if (!isUndefOrEqual(BitI1, j))
2792 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2793 SmallVector<int, 8> M;
2795 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2798 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2799 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2801 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2802 int NumElems = VT.getVectorNumElements();
2803 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2806 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2808 int BitI1 = Mask[i+1];
2809 if (!isUndefOrEqual(BitI, j))
2811 if (!isUndefOrEqual(BitI1, j))
2817 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2818 SmallVector<int, 8> M;
2820 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2823 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2824 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2825 /// MOVSD, and MOVD, i.e. setting the lowest element.
2826 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2827 if (VT.getVectorElementType().getSizeInBits() < 32)
2830 int NumElts = VT.getVectorNumElements();
2832 if (!isUndefOrEqual(Mask[0], NumElts))
2835 for (int i = 1; i < NumElts; ++i)
2836 if (!isUndefOrEqual(Mask[i], i))
2842 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2843 SmallVector<int, 8> M;
2845 return ::isMOVLMask(M, N->getValueType(0));
2848 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2849 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2850 /// element of vector 2 and the other elements to come from vector 1 in order.
2851 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2852 bool V2IsSplat = false, bool V2IsUndef = false) {
2853 int NumOps = VT.getVectorNumElements();
2854 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2857 if (!isUndefOrEqual(Mask[0], 0))
2860 for (int i = 1; i < NumOps; ++i)
2861 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2862 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2863 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2869 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2870 bool V2IsUndef = false) {
2871 SmallVector<int, 8> M;
2873 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2876 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2877 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2878 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2879 if (N->getValueType(0).getVectorNumElements() != 4)
2882 // Expect 1, 1, 3, 3
2883 for (unsigned i = 0; i < 2; ++i) {
2884 int Elt = N->getMaskElt(i);
2885 if (Elt >= 0 && Elt != 1)
2890 for (unsigned i = 2; i < 4; ++i) {
2891 int Elt = N->getMaskElt(i);
2892 if (Elt >= 0 && Elt != 3)
2897 // Don't use movshdup if it can be done with a shufps.
2898 // FIXME: verify that matching u, u, 3, 3 is what we want.
2902 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2903 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2904 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2905 if (N->getValueType(0).getVectorNumElements() != 4)
2908 // Expect 0, 0, 2, 2
2909 for (unsigned i = 0; i < 2; ++i)
2910 if (N->getMaskElt(i) > 0)
2914 for (unsigned i = 2; i < 4; ++i) {
2915 int Elt = N->getMaskElt(i);
2916 if (Elt >= 0 && Elt != 2)
2921 // Don't use movsldup if it can be done with a shufps.
2925 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2926 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2927 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2928 int e = N->getValueType(0).getVectorNumElements() / 2;
2930 for (int i = 0; i < e; ++i)
2931 if (!isUndefOrEqual(N->getMaskElt(i), i))
2933 for (int i = 0; i < e; ++i)
2934 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2939 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2940 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
2941 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2942 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2943 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2945 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2947 for (int i = 0; i < NumOperands; ++i) {
2948 int Val = SVOp->getMaskElt(NumOperands-i-1);
2949 if (Val < 0) Val = 0;
2950 if (Val >= NumOperands) Val -= NumOperands;
2952 if (i != NumOperands - 1)
2958 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2959 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
2960 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2961 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2963 // 8 nodes, but we only care about the last 4.
2964 for (unsigned i = 7; i >= 4; --i) {
2965 int Val = SVOp->getMaskElt(i);
2974 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2975 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
2976 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2977 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2979 // 8 nodes, but we only care about the first 4.
2980 for (int i = 3; i >= 0; --i) {
2981 int Val = SVOp->getMaskElt(i);
2990 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
2991 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
2992 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
2993 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2994 EVT VVT = N->getValueType(0);
2995 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
2999 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3000 Val = SVOp->getMaskElt(i);
3004 return (Val - i) * EltSize;
3007 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3009 bool X86::isZeroNode(SDValue Elt) {
3010 return ((isa<ConstantSDNode>(Elt) &&
3011 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3012 (isa<ConstantFPSDNode>(Elt) &&
3013 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3016 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3017 /// their permute mask.
3018 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3019 SelectionDAG &DAG) {
3020 EVT VT = SVOp->getValueType(0);
3021 unsigned NumElems = VT.getVectorNumElements();
3022 SmallVector<int, 8> MaskVec;
3024 for (unsigned i = 0; i != NumElems; ++i) {
3025 int idx = SVOp->getMaskElt(i);
3027 MaskVec.push_back(idx);
3028 else if (idx < (int)NumElems)
3029 MaskVec.push_back(idx + NumElems);
3031 MaskVec.push_back(idx - NumElems);
3033 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3034 SVOp->getOperand(0), &MaskVec[0]);
3037 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3038 /// the two vector operands have swapped position.
3039 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3040 unsigned NumElems = VT.getVectorNumElements();
3041 for (unsigned i = 0; i != NumElems; ++i) {
3045 else if (idx < (int)NumElems)
3046 Mask[i] = idx + NumElems;
3048 Mask[i] = idx - NumElems;
3052 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3053 /// match movhlps. The lower half elements should come from upper half of
3054 /// V1 (and in order), and the upper half elements should come from the upper
3055 /// half of V2 (and in order).
3056 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3057 if (Op->getValueType(0).getVectorNumElements() != 4)
3059 for (unsigned i = 0, e = 2; i != e; ++i)
3060 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3062 for (unsigned i = 2; i != 4; ++i)
3063 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3068 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3069 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3071 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3072 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3074 N = N->getOperand(0).getNode();
3075 if (!ISD::isNON_EXTLoad(N))
3078 *LD = cast<LoadSDNode>(N);
3082 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3083 /// match movlp{s|d}. The lower half elements should come from lower half of
3084 /// V1 (and in order), and the upper half elements should come from the upper
3085 /// half of V2 (and in order). And since V1 will become the source of the
3086 /// MOVLP, it must be either a vector load or a scalar load to vector.
3087 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3088 ShuffleVectorSDNode *Op) {
3089 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3091 // Is V2 is a vector load, don't do this transformation. We will try to use
3092 // load folding shufps op.
3093 if (ISD::isNON_EXTLoad(V2))
3096 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3098 if (NumElems != 2 && NumElems != 4)
3100 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3101 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3103 for (unsigned i = NumElems/2; i != NumElems; ++i)
3104 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3109 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3111 static bool isSplatVector(SDNode *N) {
3112 if (N->getOpcode() != ISD::BUILD_VECTOR)
3115 SDValue SplatValue = N->getOperand(0);
3116 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3117 if (N->getOperand(i) != SplatValue)
3122 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3123 /// to an zero vector.
3124 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3125 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3126 SDValue V1 = N->getOperand(0);
3127 SDValue V2 = N->getOperand(1);
3128 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3129 for (unsigned i = 0; i != NumElems; ++i) {
3130 int Idx = N->getMaskElt(i);
3131 if (Idx >= (int)NumElems) {
3132 unsigned Opc = V2.getOpcode();
3133 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3135 if (Opc != ISD::BUILD_VECTOR ||
3136 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3138 } else if (Idx >= 0) {
3139 unsigned Opc = V1.getOpcode();
3140 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3142 if (Opc != ISD::BUILD_VECTOR ||
3143 !X86::isZeroNode(V1.getOperand(Idx)))
3150 /// getZeroVector - Returns a vector of specified type with all zero elements.
3152 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3154 assert(VT.isVector() && "Expected a vector type");
3156 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3157 // type. This ensures they get CSE'd.
3159 if (VT.getSizeInBits() == 64) { // MMX
3160 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3161 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3162 } else if (HasSSE2) { // SSE2
3163 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3164 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3166 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3167 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3169 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3172 /// getOnesVector - Returns a vector of specified type with all bits set.
3174 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3175 assert(VT.isVector() && "Expected a vector type");
3177 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3178 // type. This ensures they get CSE'd.
3179 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3181 if (VT.getSizeInBits() == 64) // MMX
3182 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3184 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3185 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3189 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3190 /// that point to V2 points to its first element.
3191 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3192 EVT VT = SVOp->getValueType(0);
3193 unsigned NumElems = VT.getVectorNumElements();
3195 bool Changed = false;
3196 SmallVector<int, 8> MaskVec;
3197 SVOp->getMask(MaskVec);
3199 for (unsigned i = 0; i != NumElems; ++i) {
3200 if (MaskVec[i] > (int)NumElems) {
3201 MaskVec[i] = NumElems;
3206 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3207 SVOp->getOperand(1), &MaskVec[0]);
3208 return SDValue(SVOp, 0);
3211 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3212 /// operation of specified width.
3213 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3215 unsigned NumElems = VT.getVectorNumElements();
3216 SmallVector<int, 8> Mask;
3217 Mask.push_back(NumElems);
3218 for (unsigned i = 1; i != NumElems; ++i)
3220 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3223 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3224 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3226 unsigned NumElems = VT.getVectorNumElements();
3227 SmallVector<int, 8> Mask;
3228 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3230 Mask.push_back(i + NumElems);
3232 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3235 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3236 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3238 unsigned NumElems = VT.getVectorNumElements();
3239 unsigned Half = NumElems/2;
3240 SmallVector<int, 8> Mask;
3241 for (unsigned i = 0; i != Half; ++i) {
3242 Mask.push_back(i + Half);
3243 Mask.push_back(i + NumElems + Half);
3245 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3248 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3249 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3251 if (SV->getValueType(0).getVectorNumElements() <= 4)
3252 return SDValue(SV, 0);
3254 EVT PVT = MVT::v4f32;
3255 EVT VT = SV->getValueType(0);
3256 DebugLoc dl = SV->getDebugLoc();
3257 SDValue V1 = SV->getOperand(0);
3258 int NumElems = VT.getVectorNumElements();
3259 int EltNo = SV->getSplatIndex();
3261 // unpack elements to the correct location
3262 while (NumElems > 4) {
3263 if (EltNo < NumElems/2) {
3264 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3266 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3267 EltNo -= NumElems/2;
3272 // Perform the splat.
3273 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3274 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3275 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3276 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3279 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3280 /// vector of zero or undef vector. This produces a shuffle where the low
3281 /// element of V2 is swizzled into the zero/undef vector, landing at element
3282 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3283 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3284 bool isZero, bool HasSSE2,
3285 SelectionDAG &DAG) {
3286 EVT VT = V2.getValueType();
3288 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3289 unsigned NumElems = VT.getVectorNumElements();
3290 SmallVector<int, 16> MaskVec;
3291 for (unsigned i = 0; i != NumElems; ++i)
3292 // If this is the insertion idx, put the low elt of V2 here.
3293 MaskVec.push_back(i == Idx ? NumElems : i);
3294 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3297 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3298 /// a shuffle that is zero.
3300 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3301 bool Low, SelectionDAG &DAG) {
3302 unsigned NumZeros = 0;
3303 for (int i = 0; i < NumElems; ++i) {
3304 unsigned Index = Low ? i : NumElems-i-1;
3305 int Idx = SVOp->getMaskElt(Index);
3310 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3311 if (Elt.getNode() && X86::isZeroNode(Elt))
3319 /// isVectorShift - Returns true if the shuffle can be implemented as a
3320 /// logical left or right shift of a vector.
3321 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3322 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3323 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3324 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3327 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3330 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3334 bool SeenV1 = false;
3335 bool SeenV2 = false;
3336 for (int i = NumZeros; i < NumElems; ++i) {
3337 int Val = isLeft ? (i - NumZeros) : i;
3338 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3350 if (SeenV1 && SeenV2)
3353 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3359 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3361 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3362 unsigned NumNonZero, unsigned NumZero,
3363 SelectionDAG &DAG, TargetLowering &TLI) {
3367 DebugLoc dl = Op.getDebugLoc();
3370 for (unsigned i = 0; i < 16; ++i) {
3371 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3372 if (ThisIsNonZero && First) {
3374 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3376 V = DAG.getUNDEF(MVT::v8i16);
3381 SDValue ThisElt(0, 0), LastElt(0, 0);
3382 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3383 if (LastIsNonZero) {
3384 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3385 MVT::i16, Op.getOperand(i-1));
3387 if (ThisIsNonZero) {
3388 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3389 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3390 ThisElt, DAG.getConstant(8, MVT::i8));
3392 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3396 if (ThisElt.getNode())
3397 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3398 DAG.getIntPtrConstant(i/2));
3402 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3405 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3407 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3408 unsigned NumNonZero, unsigned NumZero,
3409 SelectionDAG &DAG, TargetLowering &TLI) {
3413 DebugLoc dl = Op.getDebugLoc();
3416 for (unsigned i = 0; i < 8; ++i) {
3417 bool isNonZero = (NonZeros & (1 << i)) != 0;
3421 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3423 V = DAG.getUNDEF(MVT::v8i16);
3426 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3427 MVT::v8i16, V, Op.getOperand(i),
3428 DAG.getIntPtrConstant(i));
3435 /// getVShift - Return a vector logical shift node.
3437 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3438 unsigned NumBits, SelectionDAG &DAG,
3439 const TargetLowering &TLI, DebugLoc dl) {
3440 bool isMMX = VT.getSizeInBits() == 64;
3441 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3442 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3443 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3444 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3445 DAG.getNode(Opc, dl, ShVT, SrcOp,
3446 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3450 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3451 SelectionDAG &DAG) {
3453 // Check if the scalar load can be widened into a vector load. And if
3454 // the address is "base + cst" see if the cst can be "absorbed" into
3455 // the shuffle mask.
3456 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3457 SDValue Ptr = LD->getBasePtr();
3458 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3460 EVT PVT = LD->getValueType(0);
3461 if (PVT != MVT::i32 && PVT != MVT::f32)
3466 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3467 FI = FINode->getIndex();
3469 } else if (Ptr.getOpcode() == ISD::ADD &&
3470 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3471 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3472 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3473 Offset = Ptr.getConstantOperandVal(1);
3474 Ptr = Ptr.getOperand(0);
3479 SDValue Chain = LD->getChain();
3480 // Make sure the stack object alignment is at least 16.
3481 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3482 if (DAG.InferPtrAlignment(Ptr) < 16) {
3483 if (MFI->isFixedObjectIndex(FI)) {
3484 // Can't change the alignment. FIXME: It's possible to compute
3485 // the exact stack offset and reference FI + adjust offset instead.
3486 // If someone *really* cares about this. That's the way to implement it.
3489 MFI->setObjectAlignment(FI, 16);
3493 // (Offset % 16) must be multiple of 4. Then address is then
3494 // Ptr + (Offset & ~15).
3497 if ((Offset % 16) & 3)
3499 int64_t StartOffset = Offset & ~15;
3501 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3502 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3504 int EltNo = (Offset - StartOffset) >> 2;
3505 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3506 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3507 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0);
3508 // Canonicalize it to a v4i32 shuffle.
3509 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3510 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3511 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3512 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3519 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3520 DebugLoc dl = Op.getDebugLoc();
3521 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3522 if (ISD::isBuildVectorAllZeros(Op.getNode())
3523 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3524 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3525 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3526 // eliminated on x86-32 hosts.
3527 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3530 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3531 return getOnesVector(Op.getValueType(), DAG, dl);
3532 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3535 EVT VT = Op.getValueType();
3536 EVT ExtVT = VT.getVectorElementType();
3537 unsigned EVTBits = ExtVT.getSizeInBits();
3539 unsigned NumElems = Op.getNumOperands();
3540 unsigned NumZero = 0;
3541 unsigned NumNonZero = 0;
3542 unsigned NonZeros = 0;
3543 bool IsAllConstants = true;
3544 SmallSet<SDValue, 8> Values;
3545 for (unsigned i = 0; i < NumElems; ++i) {
3546 SDValue Elt = Op.getOperand(i);
3547 if (Elt.getOpcode() == ISD::UNDEF)
3550 if (Elt.getOpcode() != ISD::Constant &&
3551 Elt.getOpcode() != ISD::ConstantFP)
3552 IsAllConstants = false;
3553 if (X86::isZeroNode(Elt))
3556 NonZeros |= (1 << i);
3561 if (NumNonZero == 0) {
3562 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3563 return DAG.getUNDEF(VT);
3566 // Special case for single non-zero, non-undef, element.
3567 if (NumNonZero == 1) {
3568 unsigned Idx = CountTrailingZeros_32(NonZeros);
3569 SDValue Item = Op.getOperand(Idx);
3571 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3572 // the value are obviously zero, truncate the value to i32 and do the
3573 // insertion that way. Only do this if the value is non-constant or if the
3574 // value is a constant being inserted into element 0. It is cheaper to do
3575 // a constant pool load than it is to do a movd + shuffle.
3576 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3577 (!IsAllConstants || Idx == 0)) {
3578 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3579 // Handle MMX and SSE both.
3580 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3581 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3583 // Truncate the value (which may itself be a constant) to i32, and
3584 // convert it to a vector with movd (S2V+shuffle to zero extend).
3585 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3586 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3587 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3588 Subtarget->hasSSE2(), DAG);
3590 // Now we have our 32-bit value zero extended in the low element of
3591 // a vector. If Idx != 0, swizzle it into place.
3593 SmallVector<int, 4> Mask;
3594 Mask.push_back(Idx);
3595 for (unsigned i = 1; i != VecElts; ++i)
3597 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3598 DAG.getUNDEF(Item.getValueType()),
3601 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3605 // If we have a constant or non-constant insertion into the low element of
3606 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3607 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3608 // depending on what the source datatype is.
3611 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3612 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3613 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3614 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3615 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3616 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3618 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3619 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3620 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3621 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3622 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3623 Subtarget->hasSSE2(), DAG);
3624 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3628 // Is it a vector logical left shift?
3629 if (NumElems == 2 && Idx == 1 &&
3630 X86::isZeroNode(Op.getOperand(0)) &&
3631 !X86::isZeroNode(Op.getOperand(1))) {
3632 unsigned NumBits = VT.getSizeInBits();
3633 return getVShift(true, VT,
3634 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3635 VT, Op.getOperand(1)),
3636 NumBits/2, DAG, *this, dl);
3639 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3642 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3643 // is a non-constant being inserted into an element other than the low one,
3644 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3645 // movd/movss) to move this into the low element, then shuffle it into
3647 if (EVTBits == 32) {
3648 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3650 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3651 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3652 Subtarget->hasSSE2(), DAG);
3653 SmallVector<int, 8> MaskVec;
3654 for (unsigned i = 0; i < NumElems; i++)
3655 MaskVec.push_back(i == Idx ? 0 : 1);
3656 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3660 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3661 if (Values.size() == 1) {
3662 if (EVTBits == 32) {
3663 // Instead of a shuffle like this:
3664 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3665 // Check if it's possible to issue this instead.
3666 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3667 unsigned Idx = CountTrailingZeros_32(NonZeros);
3668 SDValue Item = Op.getOperand(Idx);
3669 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3670 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3675 // A vector full of immediates; various special cases are already
3676 // handled, so this is best done with a single constant-pool load.
3680 // Let legalizer expand 2-wide build_vectors.
3681 if (EVTBits == 64) {
3682 if (NumNonZero == 1) {
3683 // One half is zero or undef.
3684 unsigned Idx = CountTrailingZeros_32(NonZeros);
3685 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3686 Op.getOperand(Idx));
3687 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3688 Subtarget->hasSSE2(), DAG);
3693 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3694 if (EVTBits == 8 && NumElems == 16) {
3695 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3697 if (V.getNode()) return V;
3700 if (EVTBits == 16 && NumElems == 8) {
3701 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3703 if (V.getNode()) return V;
3706 // If element VT is == 32 bits, turn it into a number of shuffles.
3707 SmallVector<SDValue, 8> V;
3709 if (NumElems == 4 && NumZero > 0) {
3710 for (unsigned i = 0; i < 4; ++i) {
3711 bool isZero = !(NonZeros & (1 << i));
3713 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3715 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3718 for (unsigned i = 0; i < 2; ++i) {
3719 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3722 V[i] = V[i*2]; // Must be a zero vector.
3725 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3728 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3731 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3736 SmallVector<int, 8> MaskVec;
3737 bool Reverse = (NonZeros & 0x3) == 2;
3738 for (unsigned i = 0; i < 2; ++i)
3739 MaskVec.push_back(Reverse ? 1-i : i);
3740 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3741 for (unsigned i = 0; i < 2; ++i)
3742 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3743 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3746 if (Values.size() > 2) {
3747 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3748 // values to be inserted is equal to the number of elements, in which case
3749 // use the unpack code below in the hopes of matching the consecutive elts
3750 // load merge pattern for shuffles.
3751 // FIXME: We could probably just check that here directly.
3752 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3753 getSubtarget()->hasSSE41()) {
3754 V[0] = DAG.getUNDEF(VT);
3755 for (unsigned i = 0; i < NumElems; ++i)
3756 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3757 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3758 Op.getOperand(i), DAG.getIntPtrConstant(i));
3761 // Expand into a number of unpckl*.
3763 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3764 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3765 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3766 for (unsigned i = 0; i < NumElems; ++i)
3767 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3769 while (NumElems != 0) {
3770 for (unsigned i = 0; i < NumElems; ++i)
3771 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3781 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3782 // We support concatenate two MMX registers and place them in a MMX
3783 // register. This is better than doing a stack convert.
3784 DebugLoc dl = Op.getDebugLoc();
3785 EVT ResVT = Op.getValueType();
3786 assert(Op.getNumOperands() == 2);
3787 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3788 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3790 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3791 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3792 InVec = Op.getOperand(1);
3793 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3794 unsigned NumElts = ResVT.getVectorNumElements();
3795 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3796 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3797 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3799 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3800 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3801 Mask[0] = 0; Mask[1] = 2;
3802 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3804 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3807 // v8i16 shuffles - Prefer shuffles in the following order:
3808 // 1. [all] pshuflw, pshufhw, optional move
3809 // 2. [ssse3] 1 x pshufb
3810 // 3. [ssse3] 2 x pshufb + 1 x por
3811 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3813 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3814 SelectionDAG &DAG, X86TargetLowering &TLI) {
3815 SDValue V1 = SVOp->getOperand(0);
3816 SDValue V2 = SVOp->getOperand(1);
3817 DebugLoc dl = SVOp->getDebugLoc();
3818 SmallVector<int, 8> MaskVals;
3820 // Determine if more than 1 of the words in each of the low and high quadwords
3821 // of the result come from the same quadword of one of the two inputs. Undef
3822 // mask values count as coming from any quadword, for better codegen.
3823 SmallVector<unsigned, 4> LoQuad(4);
3824 SmallVector<unsigned, 4> HiQuad(4);
3825 BitVector InputQuads(4);
3826 for (unsigned i = 0; i < 8; ++i) {
3827 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3828 int EltIdx = SVOp->getMaskElt(i);
3829 MaskVals.push_back(EltIdx);
3838 InputQuads.set(EltIdx / 4);
3841 int BestLoQuad = -1;
3842 unsigned MaxQuad = 1;
3843 for (unsigned i = 0; i < 4; ++i) {
3844 if (LoQuad[i] > MaxQuad) {
3846 MaxQuad = LoQuad[i];
3850 int BestHiQuad = -1;
3852 for (unsigned i = 0; i < 4; ++i) {
3853 if (HiQuad[i] > MaxQuad) {
3855 MaxQuad = HiQuad[i];
3859 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3860 // of the two input vectors, shuffle them into one input vector so only a
3861 // single pshufb instruction is necessary. If There are more than 2 input
3862 // quads, disable the next transformation since it does not help SSSE3.
3863 bool V1Used = InputQuads[0] || InputQuads[1];
3864 bool V2Used = InputQuads[2] || InputQuads[3];
3865 if (TLI.getSubtarget()->hasSSSE3()) {
3866 if (InputQuads.count() == 2 && V1Used && V2Used) {
3867 BestLoQuad = InputQuads.find_first();
3868 BestHiQuad = InputQuads.find_next(BestLoQuad);
3870 if (InputQuads.count() > 2) {
3876 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3877 // the shuffle mask. If a quad is scored as -1, that means that it contains
3878 // words from all 4 input quadwords.
3880 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3881 SmallVector<int, 8> MaskV;
3882 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3883 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3884 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3885 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3886 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3887 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3889 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3890 // source words for the shuffle, to aid later transformations.
3891 bool AllWordsInNewV = true;
3892 bool InOrder[2] = { true, true };
3893 for (unsigned i = 0; i != 8; ++i) {
3894 int idx = MaskVals[i];
3896 InOrder[i/4] = false;
3897 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3899 AllWordsInNewV = false;
3903 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3904 if (AllWordsInNewV) {
3905 for (int i = 0; i != 8; ++i) {
3906 int idx = MaskVals[i];
3909 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3910 if ((idx != i) && idx < 4)
3912 if ((idx != i) && idx > 3)
3921 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3922 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3923 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3924 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3925 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3929 // If we have SSSE3, and all words of the result are from 1 input vector,
3930 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3931 // is present, fall back to case 4.
3932 if (TLI.getSubtarget()->hasSSSE3()) {
3933 SmallVector<SDValue,16> pshufbMask;
3935 // If we have elements from both input vectors, set the high bit of the
3936 // shuffle mask element to zero out elements that come from V2 in the V1
3937 // mask, and elements that come from V1 in the V2 mask, so that the two
3938 // results can be OR'd together.
3939 bool TwoInputs = V1Used && V2Used;
3940 for (unsigned i = 0; i != 8; ++i) {
3941 int EltIdx = MaskVals[i] * 2;
3942 if (TwoInputs && (EltIdx >= 16)) {
3943 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3944 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3947 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3948 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3950 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3951 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3952 DAG.getNode(ISD::BUILD_VECTOR, dl,
3953 MVT::v16i8, &pshufbMask[0], 16));
3955 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3957 // Calculate the shuffle mask for the second input, shuffle it, and
3958 // OR it with the first shuffled input.
3960 for (unsigned i = 0; i != 8; ++i) {
3961 int EltIdx = MaskVals[i] * 2;
3963 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3964 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3967 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3968 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3970 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3971 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3972 DAG.getNode(ISD::BUILD_VECTOR, dl,
3973 MVT::v16i8, &pshufbMask[0], 16));
3974 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3975 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3978 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3979 // and update MaskVals with new element order.
3980 BitVector InOrder(8);
3981 if (BestLoQuad >= 0) {
3982 SmallVector<int, 8> MaskV;
3983 for (int i = 0; i != 4; ++i) {
3984 int idx = MaskVals[i];
3986 MaskV.push_back(-1);
3988 } else if ((idx / 4) == BestLoQuad) {
3989 MaskV.push_back(idx & 3);
3992 MaskV.push_back(-1);
3995 for (unsigned i = 4; i != 8; ++i)
3997 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4001 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4002 // and update MaskVals with the new element order.
4003 if (BestHiQuad >= 0) {
4004 SmallVector<int, 8> MaskV;
4005 for (unsigned i = 0; i != 4; ++i)
4007 for (unsigned i = 4; i != 8; ++i) {
4008 int idx = MaskVals[i];
4010 MaskV.push_back(-1);
4012 } else if ((idx / 4) == BestHiQuad) {
4013 MaskV.push_back((idx & 3) + 4);
4016 MaskV.push_back(-1);
4019 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4023 // In case BestHi & BestLo were both -1, which means each quadword has a word
4024 // from each of the four input quadwords, calculate the InOrder bitvector now
4025 // before falling through to the insert/extract cleanup.
4026 if (BestLoQuad == -1 && BestHiQuad == -1) {
4028 for (int i = 0; i != 8; ++i)
4029 if (MaskVals[i] < 0 || MaskVals[i] == i)
4033 // The other elements are put in the right place using pextrw and pinsrw.
4034 for (unsigned i = 0; i != 8; ++i) {
4037 int EltIdx = MaskVals[i];
4040 SDValue ExtOp = (EltIdx < 8)
4041 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4042 DAG.getIntPtrConstant(EltIdx))
4043 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4044 DAG.getIntPtrConstant(EltIdx - 8));
4045 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4046 DAG.getIntPtrConstant(i));
4051 // v16i8 shuffles - Prefer shuffles in the following order:
4052 // 1. [ssse3] 1 x pshufb
4053 // 2. [ssse3] 2 x pshufb + 1 x por
4054 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4056 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4057 SelectionDAG &DAG, X86TargetLowering &TLI) {
4058 SDValue V1 = SVOp->getOperand(0);
4059 SDValue V2 = SVOp->getOperand(1);
4060 DebugLoc dl = SVOp->getDebugLoc();
4061 SmallVector<int, 16> MaskVals;
4062 SVOp->getMask(MaskVals);
4064 // If we have SSSE3, case 1 is generated when all result bytes come from
4065 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4066 // present, fall back to case 3.
4067 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4070 for (unsigned i = 0; i < 16; ++i) {
4071 int EltIdx = MaskVals[i];
4080 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4081 if (TLI.getSubtarget()->hasSSSE3()) {
4082 SmallVector<SDValue,16> pshufbMask;
4084 // If all result elements are from one input vector, then only translate
4085 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4087 // Otherwise, we have elements from both input vectors, and must zero out
4088 // elements that come from V2 in the first mask, and V1 in the second mask
4089 // so that we can OR them together.
4090 bool TwoInputs = !(V1Only || V2Only);
4091 for (unsigned i = 0; i != 16; ++i) {
4092 int EltIdx = MaskVals[i];
4093 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4094 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4097 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4099 // If all the elements are from V2, assign it to V1 and return after
4100 // building the first pshufb.
4103 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4104 DAG.getNode(ISD::BUILD_VECTOR, dl,
4105 MVT::v16i8, &pshufbMask[0], 16));
4109 // Calculate the shuffle mask for the second input, shuffle it, and
4110 // OR it with the first shuffled input.
4112 for (unsigned i = 0; i != 16; ++i) {
4113 int EltIdx = MaskVals[i];
4115 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4118 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4120 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4121 DAG.getNode(ISD::BUILD_VECTOR, dl,
4122 MVT::v16i8, &pshufbMask[0], 16));
4123 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4126 // No SSSE3 - Calculate in place words and then fix all out of place words
4127 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4128 // the 16 different words that comprise the two doublequadword input vectors.
4129 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4130 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4131 SDValue NewV = V2Only ? V2 : V1;
4132 for (int i = 0; i != 8; ++i) {
4133 int Elt0 = MaskVals[i*2];
4134 int Elt1 = MaskVals[i*2+1];
4136 // This word of the result is all undef, skip it.
4137 if (Elt0 < 0 && Elt1 < 0)
4140 // This word of the result is already in the correct place, skip it.
4141 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4143 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4146 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4147 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4150 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4151 // using a single extract together, load it and store it.
4152 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4153 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4154 DAG.getIntPtrConstant(Elt1 / 2));
4155 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4156 DAG.getIntPtrConstant(i));
4160 // If Elt1 is defined, extract it from the appropriate source. If the
4161 // source byte is not also odd, shift the extracted word left 8 bits
4162 // otherwise clear the bottom 8 bits if we need to do an or.
4164 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4165 DAG.getIntPtrConstant(Elt1 / 2));
4166 if ((Elt1 & 1) == 0)
4167 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4168 DAG.getConstant(8, TLI.getShiftAmountTy()));
4170 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4171 DAG.getConstant(0xFF00, MVT::i16));
4173 // If Elt0 is defined, extract it from the appropriate source. If the
4174 // source byte is not also even, shift the extracted word right 8 bits. If
4175 // Elt1 was also defined, OR the extracted values together before
4176 // inserting them in the result.
4178 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4179 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4180 if ((Elt0 & 1) != 0)
4181 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4182 DAG.getConstant(8, TLI.getShiftAmountTy()));
4184 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4185 DAG.getConstant(0x00FF, MVT::i16));
4186 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4189 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4190 DAG.getIntPtrConstant(i));
4192 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4195 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4196 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4197 /// done when every pair / quad of shuffle mask elements point to elements in
4198 /// the right sequence. e.g.
4199 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4201 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4203 TargetLowering &TLI, DebugLoc dl) {
4204 EVT VT = SVOp->getValueType(0);
4205 SDValue V1 = SVOp->getOperand(0);
4206 SDValue V2 = SVOp->getOperand(1);
4207 unsigned NumElems = VT.getVectorNumElements();
4208 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4209 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4210 EVT MaskEltVT = MaskVT.getVectorElementType();
4212 switch (VT.getSimpleVT().SimpleTy) {
4213 default: assert(false && "Unexpected!");
4214 case MVT::v4f32: NewVT = MVT::v2f64; break;
4215 case MVT::v4i32: NewVT = MVT::v2i64; break;
4216 case MVT::v8i16: NewVT = MVT::v4i32; break;
4217 case MVT::v16i8: NewVT = MVT::v4i32; break;
4220 if (NewWidth == 2) {
4226 int Scale = NumElems / NewWidth;
4227 SmallVector<int, 8> MaskVec;
4228 for (unsigned i = 0; i < NumElems; i += Scale) {
4230 for (int j = 0; j < Scale; ++j) {
4231 int EltIdx = SVOp->getMaskElt(i+j);
4235 StartIdx = EltIdx - (EltIdx % Scale);
4236 if (EltIdx != StartIdx + j)
4240 MaskVec.push_back(-1);
4242 MaskVec.push_back(StartIdx / Scale);
4245 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4246 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4247 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4250 /// getVZextMovL - Return a zero-extending vector move low node.
4252 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4253 SDValue SrcOp, SelectionDAG &DAG,
4254 const X86Subtarget *Subtarget, DebugLoc dl) {
4255 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4256 LoadSDNode *LD = NULL;
4257 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4258 LD = dyn_cast<LoadSDNode>(SrcOp);
4260 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4262 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4263 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4264 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4265 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4266 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4268 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4269 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4270 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4271 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4279 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4280 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4281 DAG.getNode(ISD::BIT_CONVERT, dl,
4285 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4288 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4289 SDValue V1 = SVOp->getOperand(0);
4290 SDValue V2 = SVOp->getOperand(1);
4291 DebugLoc dl = SVOp->getDebugLoc();
4292 EVT VT = SVOp->getValueType(0);
4294 SmallVector<std::pair<int, int>, 8> Locs;
4296 SmallVector<int, 8> Mask1(4U, -1);
4297 SmallVector<int, 8> PermMask;
4298 SVOp->getMask(PermMask);
4302 for (unsigned i = 0; i != 4; ++i) {
4303 int Idx = PermMask[i];
4305 Locs[i] = std::make_pair(-1, -1);
4307 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4309 Locs[i] = std::make_pair(0, NumLo);
4313 Locs[i] = std::make_pair(1, NumHi);
4315 Mask1[2+NumHi] = Idx;
4321 if (NumLo <= 2 && NumHi <= 2) {
4322 // If no more than two elements come from either vector. This can be
4323 // implemented with two shuffles. First shuffle gather the elements.
4324 // The second shuffle, which takes the first shuffle as both of its
4325 // vector operands, put the elements into the right order.
4326 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4328 SmallVector<int, 8> Mask2(4U, -1);
4330 for (unsigned i = 0; i != 4; ++i) {
4331 if (Locs[i].first == -1)
4334 unsigned Idx = (i < 2) ? 0 : 4;
4335 Idx += Locs[i].first * 2 + Locs[i].second;
4340 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4341 } else if (NumLo == 3 || NumHi == 3) {
4342 // Otherwise, we must have three elements from one vector, call it X, and
4343 // one element from the other, call it Y. First, use a shufps to build an
4344 // intermediate vector with the one element from Y and the element from X
4345 // that will be in the same half in the final destination (the indexes don't
4346 // matter). Then, use a shufps to build the final vector, taking the half
4347 // containing the element from Y from the intermediate, and the other half
4350 // Normalize it so the 3 elements come from V1.
4351 CommuteVectorShuffleMask(PermMask, VT);
4355 // Find the element from V2.
4357 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4358 int Val = PermMask[HiIndex];
4365 Mask1[0] = PermMask[HiIndex];
4367 Mask1[2] = PermMask[HiIndex^1];
4369 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4372 Mask1[0] = PermMask[0];
4373 Mask1[1] = PermMask[1];
4374 Mask1[2] = HiIndex & 1 ? 6 : 4;
4375 Mask1[3] = HiIndex & 1 ? 4 : 6;
4376 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4378 Mask1[0] = HiIndex & 1 ? 2 : 0;
4379 Mask1[1] = HiIndex & 1 ? 0 : 2;
4380 Mask1[2] = PermMask[2];
4381 Mask1[3] = PermMask[3];
4386 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4390 // Break it into (shuffle shuffle_hi, shuffle_lo).
4392 SmallVector<int,8> LoMask(4U, -1);
4393 SmallVector<int,8> HiMask(4U, -1);
4395 SmallVector<int,8> *MaskPtr = &LoMask;
4396 unsigned MaskIdx = 0;
4399 for (unsigned i = 0; i != 4; ++i) {
4406 int Idx = PermMask[i];
4408 Locs[i] = std::make_pair(-1, -1);
4409 } else if (Idx < 4) {
4410 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4411 (*MaskPtr)[LoIdx] = Idx;
4414 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4415 (*MaskPtr)[HiIdx] = Idx;
4420 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4421 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4422 SmallVector<int, 8> MaskOps;
4423 for (unsigned i = 0; i != 4; ++i) {
4424 if (Locs[i].first == -1) {
4425 MaskOps.push_back(-1);
4427 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4428 MaskOps.push_back(Idx);
4431 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4435 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4436 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4437 SDValue V1 = Op.getOperand(0);
4438 SDValue V2 = Op.getOperand(1);
4439 EVT VT = Op.getValueType();
4440 DebugLoc dl = Op.getDebugLoc();
4441 unsigned NumElems = VT.getVectorNumElements();
4442 bool isMMX = VT.getSizeInBits() == 64;
4443 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4444 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4445 bool V1IsSplat = false;
4446 bool V2IsSplat = false;
4448 if (isZeroShuffle(SVOp))
4449 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4451 // Promote splats to v4f32.
4452 if (SVOp->isSplat()) {
4453 if (isMMX || NumElems < 4)
4455 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4458 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4460 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4461 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4462 if (NewOp.getNode())
4463 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4464 LowerVECTOR_SHUFFLE(NewOp, DAG));
4465 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4466 // FIXME: Figure out a cleaner way to do this.
4467 // Try to make use of movq to zero out the top part.
4468 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4469 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4470 if (NewOp.getNode()) {
4471 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4472 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4473 DAG, Subtarget, dl);
4475 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4476 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4477 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4478 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4479 DAG, Subtarget, dl);
4483 if (X86::isPSHUFDMask(SVOp))
4486 // Check if this can be converted into a logical shift.
4487 bool isLeft = false;
4490 bool isShift = getSubtarget()->hasSSE2() &&
4491 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4492 if (isShift && ShVal.hasOneUse()) {
4493 // If the shifted value has multiple uses, it may be cheaper to use
4494 // v_set0 + movlhps or movhlps, etc.
4495 EVT EltVT = VT.getVectorElementType();
4496 ShAmt *= EltVT.getSizeInBits();
4497 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4500 if (X86::isMOVLMask(SVOp)) {
4503 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4504 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4509 // FIXME: fold these into legal mask.
4510 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4511 X86::isMOVSLDUPMask(SVOp) ||
4512 X86::isMOVHLPSMask(SVOp) ||
4513 X86::isMOVLHPSMask(SVOp) ||
4514 X86::isMOVLPMask(SVOp)))
4517 if (ShouldXformToMOVHLPS(SVOp) ||
4518 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4519 return CommuteVectorShuffle(SVOp, DAG);
4522 // No better options. Use a vshl / vsrl.
4523 EVT EltVT = VT.getVectorElementType();
4524 ShAmt *= EltVT.getSizeInBits();
4525 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4528 bool Commuted = false;
4529 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4530 // 1,1,1,1 -> v8i16 though.
4531 V1IsSplat = isSplatVector(V1.getNode());
4532 V2IsSplat = isSplatVector(V2.getNode());
4534 // Canonicalize the splat or undef, if present, to be on the RHS.
4535 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4536 Op = CommuteVectorShuffle(SVOp, DAG);
4537 SVOp = cast<ShuffleVectorSDNode>(Op);
4538 V1 = SVOp->getOperand(0);
4539 V2 = SVOp->getOperand(1);
4540 std::swap(V1IsSplat, V2IsSplat);
4541 std::swap(V1IsUndef, V2IsUndef);
4545 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4546 // Shuffling low element of v1 into undef, just return v1.
4549 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4550 // the instruction selector will not match, so get a canonical MOVL with
4551 // swapped operands to undo the commute.
4552 return getMOVL(DAG, dl, VT, V2, V1);
4555 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4556 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4557 X86::isUNPCKLMask(SVOp) ||
4558 X86::isUNPCKHMask(SVOp))
4562 // Normalize mask so all entries that point to V2 points to its first
4563 // element then try to match unpck{h|l} again. If match, return a
4564 // new vector_shuffle with the corrected mask.
4565 SDValue NewMask = NormalizeMask(SVOp, DAG);
4566 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4567 if (NSVOp != SVOp) {
4568 if (X86::isUNPCKLMask(NSVOp, true)) {
4570 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4577 // Commute is back and try unpck* again.
4578 // FIXME: this seems wrong.
4579 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4580 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4581 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4582 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4583 X86::isUNPCKLMask(NewSVOp) ||
4584 X86::isUNPCKHMask(NewSVOp))
4588 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4590 // Normalize the node to match x86 shuffle ops if needed
4591 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4592 return CommuteVectorShuffle(SVOp, DAG);
4594 // Check for legal shuffle and return?
4595 SmallVector<int, 16> PermMask;
4596 SVOp->getMask(PermMask);
4597 if (isShuffleMaskLegal(PermMask, VT))
4600 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4601 if (VT == MVT::v8i16) {
4602 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4603 if (NewOp.getNode())
4607 if (VT == MVT::v16i8) {
4608 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4609 if (NewOp.getNode())
4613 // Handle all 4 wide cases with a number of shuffles except for MMX.
4614 if (NumElems == 4 && !isMMX)
4615 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4621 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4622 SelectionDAG &DAG) {
4623 EVT VT = Op.getValueType();
4624 DebugLoc dl = Op.getDebugLoc();
4625 if (VT.getSizeInBits() == 8) {
4626 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4627 Op.getOperand(0), Op.getOperand(1));
4628 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4629 DAG.getValueType(VT));
4630 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4631 } else if (VT.getSizeInBits() == 16) {
4632 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4633 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4635 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4636 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4637 DAG.getNode(ISD::BIT_CONVERT, dl,
4641 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4642 Op.getOperand(0), Op.getOperand(1));
4643 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4644 DAG.getValueType(VT));
4645 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4646 } else if (VT == MVT::f32) {
4647 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4648 // the result back to FR32 register. It's only worth matching if the
4649 // result has a single use which is a store or a bitcast to i32. And in
4650 // the case of a store, it's not worth it if the index is a constant 0,
4651 // because a MOVSSmr can be used instead, which is smaller and faster.
4652 if (!Op.hasOneUse())
4654 SDNode *User = *Op.getNode()->use_begin();
4655 if ((User->getOpcode() != ISD::STORE ||
4656 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4657 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4658 (User->getOpcode() != ISD::BIT_CONVERT ||
4659 User->getValueType(0) != MVT::i32))
4661 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4662 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4665 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4666 } else if (VT == MVT::i32) {
4667 // ExtractPS works with constant index.
4668 if (isa<ConstantSDNode>(Op.getOperand(1)))
4676 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4677 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4680 if (Subtarget->hasSSE41()) {
4681 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4686 EVT VT = Op.getValueType();
4687 DebugLoc dl = Op.getDebugLoc();
4688 // TODO: handle v16i8.
4689 if (VT.getSizeInBits() == 16) {
4690 SDValue Vec = Op.getOperand(0);
4691 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4693 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4694 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4695 DAG.getNode(ISD::BIT_CONVERT, dl,
4698 // Transform it so it match pextrw which produces a 32-bit result.
4699 EVT EltVT = MVT::i32;
4700 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4701 Op.getOperand(0), Op.getOperand(1));
4702 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4703 DAG.getValueType(VT));
4704 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4705 } else if (VT.getSizeInBits() == 32) {
4706 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4710 // SHUFPS the element to the lowest double word, then movss.
4711 int Mask[4] = { Idx, -1, -1, -1 };
4712 EVT VVT = Op.getOperand(0).getValueType();
4713 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4714 DAG.getUNDEF(VVT), Mask);
4715 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4716 DAG.getIntPtrConstant(0));
4717 } else if (VT.getSizeInBits() == 64) {
4718 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4719 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4720 // to match extract_elt for f64.
4721 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4725 // UNPCKHPD the element to the lowest double word, then movsd.
4726 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4727 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4728 int Mask[2] = { 1, -1 };
4729 EVT VVT = Op.getOperand(0).getValueType();
4730 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4731 DAG.getUNDEF(VVT), Mask);
4732 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4733 DAG.getIntPtrConstant(0));
4740 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4741 EVT VT = Op.getValueType();
4742 EVT EltVT = VT.getVectorElementType();
4743 DebugLoc dl = Op.getDebugLoc();
4745 SDValue N0 = Op.getOperand(0);
4746 SDValue N1 = Op.getOperand(1);
4747 SDValue N2 = Op.getOperand(2);
4749 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4750 isa<ConstantSDNode>(N2)) {
4751 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4753 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4755 if (N1.getValueType() != MVT::i32)
4756 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4757 if (N2.getValueType() != MVT::i32)
4758 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4759 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4760 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4761 // Bits [7:6] of the constant are the source select. This will always be
4762 // zero here. The DAG Combiner may combine an extract_elt index into these
4763 // bits. For example (insert (extract, 3), 2) could be matched by putting
4764 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4765 // Bits [5:4] of the constant are the destination select. This is the
4766 // value of the incoming immediate.
4767 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4768 // combine either bitwise AND or insert of float 0.0 to set these bits.
4769 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4770 // Create this as a scalar to vector..
4771 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4772 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4773 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4774 // PINSR* works with constant index.
4781 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4782 EVT VT = Op.getValueType();
4783 EVT EltVT = VT.getVectorElementType();
4785 if (Subtarget->hasSSE41())
4786 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4788 if (EltVT == MVT::i8)
4791 DebugLoc dl = Op.getDebugLoc();
4792 SDValue N0 = Op.getOperand(0);
4793 SDValue N1 = Op.getOperand(1);
4794 SDValue N2 = Op.getOperand(2);
4796 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4797 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4798 // as its second argument.
4799 if (N1.getValueType() != MVT::i32)
4800 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4801 if (N2.getValueType() != MVT::i32)
4802 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4803 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4809 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4810 DebugLoc dl = Op.getDebugLoc();
4811 if (Op.getValueType() == MVT::v2f32)
4812 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4813 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4814 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4815 Op.getOperand(0))));
4817 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4818 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4820 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4821 EVT VT = MVT::v2i32;
4822 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4829 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4830 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4833 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4834 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4835 // one of the above mentioned nodes. It has to be wrapped because otherwise
4836 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4837 // be used to form addressing mode. These wrapped nodes will be selected
4840 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4841 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4843 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4845 unsigned char OpFlag = 0;
4846 unsigned WrapperKind = X86ISD::Wrapper;
4847 CodeModel::Model M = getTargetMachine().getCodeModel();
4849 if (Subtarget->isPICStyleRIPRel() &&
4850 (M == CodeModel::Small || M == CodeModel::Kernel))
4851 WrapperKind = X86ISD::WrapperRIP;
4852 else if (Subtarget->isPICStyleGOT())
4853 OpFlag = X86II::MO_GOTOFF;
4854 else if (Subtarget->isPICStyleStubPIC())
4855 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4857 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4859 CP->getOffset(), OpFlag);
4860 DebugLoc DL = CP->getDebugLoc();
4861 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4862 // With PIC, the address is actually $g + Offset.
4864 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4865 DAG.getNode(X86ISD::GlobalBaseReg,
4866 DebugLoc::getUnknownLoc(), getPointerTy()),
4873 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4874 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4876 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4878 unsigned char OpFlag = 0;
4879 unsigned WrapperKind = X86ISD::Wrapper;
4880 CodeModel::Model M = getTargetMachine().getCodeModel();
4882 if (Subtarget->isPICStyleRIPRel() &&
4883 (M == CodeModel::Small || M == CodeModel::Kernel))
4884 WrapperKind = X86ISD::WrapperRIP;
4885 else if (Subtarget->isPICStyleGOT())
4886 OpFlag = X86II::MO_GOTOFF;
4887 else if (Subtarget->isPICStyleStubPIC())
4888 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4890 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4892 DebugLoc DL = JT->getDebugLoc();
4893 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4895 // With PIC, the address is actually $g + Offset.
4897 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4898 DAG.getNode(X86ISD::GlobalBaseReg,
4899 DebugLoc::getUnknownLoc(), getPointerTy()),
4907 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4908 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4910 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4912 unsigned char OpFlag = 0;
4913 unsigned WrapperKind = X86ISD::Wrapper;
4914 CodeModel::Model M = getTargetMachine().getCodeModel();
4916 if (Subtarget->isPICStyleRIPRel() &&
4917 (M == CodeModel::Small || M == CodeModel::Kernel))
4918 WrapperKind = X86ISD::WrapperRIP;
4919 else if (Subtarget->isPICStyleGOT())
4920 OpFlag = X86II::MO_GOTOFF;
4921 else if (Subtarget->isPICStyleStubPIC())
4922 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4924 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4926 DebugLoc DL = Op.getDebugLoc();
4927 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4930 // With PIC, the address is actually $g + Offset.
4931 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4932 !Subtarget->is64Bit()) {
4933 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4934 DAG.getNode(X86ISD::GlobalBaseReg,
4935 DebugLoc::getUnknownLoc(),
4944 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
4945 // Create the TargetBlockAddressAddress node.
4946 unsigned char OpFlags =
4947 Subtarget->ClassifyBlockAddressReference();
4948 CodeModel::Model M = getTargetMachine().getCodeModel();
4949 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4950 DebugLoc dl = Op.getDebugLoc();
4951 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
4952 /*isTarget=*/true, OpFlags);
4954 if (Subtarget->isPICStyleRIPRel() &&
4955 (M == CodeModel::Small || M == CodeModel::Kernel))
4956 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4958 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4960 // With PIC, the address is actually $g + Offset.
4961 if (isGlobalRelativeToPICBase(OpFlags)) {
4962 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4963 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4971 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4973 SelectionDAG &DAG) const {
4974 // Create the TargetGlobalAddress node, folding in the constant
4975 // offset if it is legal.
4976 unsigned char OpFlags =
4977 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4978 CodeModel::Model M = getTargetMachine().getCodeModel();
4980 if (OpFlags == X86II::MO_NO_FLAG &&
4981 X86::isOffsetSuitableForCodeModel(Offset, M)) {
4982 // A direct static reference to a global.
4983 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4986 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4989 if (Subtarget->isPICStyleRIPRel() &&
4990 (M == CodeModel::Small || M == CodeModel::Kernel))
4991 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4993 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4995 // With PIC, the address is actually $g + Offset.
4996 if (isGlobalRelativeToPICBase(OpFlags)) {
4997 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4998 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5002 // For globals that require a load from a stub to get the address, emit the
5004 if (isGlobalStubReference(OpFlags))
5005 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5006 PseudoSourceValue::getGOT(), 0);
5008 // If there was a non-zero offset that we didn't fold, create an explicit
5011 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5012 DAG.getConstant(Offset, getPointerTy()));
5018 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5019 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5020 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5021 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5025 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5026 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5027 unsigned char OperandFlags) {
5028 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5029 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5030 DebugLoc dl = GA->getDebugLoc();
5031 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5032 GA->getValueType(0),
5036 SDValue Ops[] = { Chain, TGA, *InFlag };
5037 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5039 SDValue Ops[] = { Chain, TGA };
5040 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5043 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5044 MFI->setHasCalls(true);
5046 SDValue Flag = Chain.getValue(1);
5047 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5050 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5052 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5055 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5056 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5057 DAG.getNode(X86ISD::GlobalBaseReg,
5058 DebugLoc::getUnknownLoc(),
5060 InFlag = Chain.getValue(1);
5062 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5065 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5067 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5069 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5070 X86::RAX, X86II::MO_TLSGD);
5073 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5074 // "local exec" model.
5075 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5076 const EVT PtrVT, TLSModel::Model model,
5078 DebugLoc dl = GA->getDebugLoc();
5079 // Get the Thread Pointer
5080 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5081 DebugLoc::getUnknownLoc(), PtrVT,
5082 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5085 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5088 unsigned char OperandFlags = 0;
5089 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5091 unsigned WrapperKind = X86ISD::Wrapper;
5092 if (model == TLSModel::LocalExec) {
5093 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5094 } else if (is64Bit) {
5095 assert(model == TLSModel::InitialExec);
5096 OperandFlags = X86II::MO_GOTTPOFF;
5097 WrapperKind = X86ISD::WrapperRIP;
5099 assert(model == TLSModel::InitialExec);
5100 OperandFlags = X86II::MO_INDNTPOFF;
5103 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5105 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5106 GA->getOffset(), OperandFlags);
5107 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5109 if (model == TLSModel::InitialExec)
5110 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5111 PseudoSourceValue::getGOT(), 0);
5113 // The address of the thread local variable is the add of the thread
5114 // pointer with the offset of the variable.
5115 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5119 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5120 // TODO: implement the "local dynamic" model
5121 // TODO: implement the "initial exec"model for pic executables
5122 assert(Subtarget->isTargetELF() &&
5123 "TLS not implemented for non-ELF targets");
5124 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5125 const GlobalValue *GV = GA->getGlobal();
5127 // If GV is an alias then use the aliasee for determining
5128 // thread-localness.
5129 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5130 GV = GA->resolveAliasedGlobal(false);
5132 TLSModel::Model model = getTLSModel(GV,
5133 getTargetMachine().getRelocationModel());
5136 case TLSModel::GeneralDynamic:
5137 case TLSModel::LocalDynamic: // not implemented
5138 if (Subtarget->is64Bit())
5139 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5140 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5142 case TLSModel::InitialExec:
5143 case TLSModel::LocalExec:
5144 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5145 Subtarget->is64Bit());
5148 llvm_unreachable("Unreachable");
5153 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5154 /// take a 2 x i32 value to shift plus a shift amount.
5155 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5156 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5157 EVT VT = Op.getValueType();
5158 unsigned VTBits = VT.getSizeInBits();
5159 DebugLoc dl = Op.getDebugLoc();
5160 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5161 SDValue ShOpLo = Op.getOperand(0);
5162 SDValue ShOpHi = Op.getOperand(1);
5163 SDValue ShAmt = Op.getOperand(2);
5164 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5165 DAG.getConstant(VTBits - 1, MVT::i8))
5166 : DAG.getConstant(0, VT);
5169 if (Op.getOpcode() == ISD::SHL_PARTS) {
5170 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5171 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5173 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5174 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5177 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5178 DAG.getConstant(VTBits, MVT::i8));
5179 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
5180 AndNode, DAG.getConstant(0, MVT::i8));
5183 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5184 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5185 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5187 if (Op.getOpcode() == ISD::SHL_PARTS) {
5188 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5189 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5191 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5192 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5195 SDValue Ops[2] = { Lo, Hi };
5196 return DAG.getMergeValues(Ops, 2, dl);
5199 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5200 EVT SrcVT = Op.getOperand(0).getValueType();
5202 if (SrcVT.isVector()) {
5203 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5209 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5210 "Unknown SINT_TO_FP to lower!");
5212 // These are really Legal; return the operand so the caller accepts it as
5214 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5216 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5217 Subtarget->is64Bit()) {
5221 DebugLoc dl = Op.getDebugLoc();
5222 unsigned Size = SrcVT.getSizeInBits()/8;
5223 MachineFunction &MF = DAG.getMachineFunction();
5224 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5225 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5226 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5228 PseudoSourceValue::getFixedStack(SSFI), 0);
5229 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5232 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5234 SelectionDAG &DAG) {
5236 DebugLoc dl = Op.getDebugLoc();
5238 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5240 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5242 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5243 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5244 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5245 Tys, Ops, array_lengthof(Ops));
5248 Chain = Result.getValue(1);
5249 SDValue InFlag = Result.getValue(2);
5251 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5252 // shouldn't be necessary except that RFP cannot be live across
5253 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5254 MachineFunction &MF = DAG.getMachineFunction();
5255 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5256 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5257 Tys = DAG.getVTList(MVT::Other);
5259 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5261 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5262 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5263 PseudoSourceValue::getFixedStack(SSFI), 0);
5269 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5270 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5271 // This algorithm is not obvious. Here it is in C code, more or less:
5273 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5274 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5275 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5277 // Copy ints to xmm registers.
5278 __m128i xh = _mm_cvtsi32_si128( hi );
5279 __m128i xl = _mm_cvtsi32_si128( lo );
5281 // Combine into low half of a single xmm register.
5282 __m128i x = _mm_unpacklo_epi32( xh, xl );
5286 // Merge in appropriate exponents to give the integer bits the right
5288 x = _mm_unpacklo_epi32( x, exp );
5290 // Subtract away the biases to deal with the IEEE-754 double precision
5292 d = _mm_sub_pd( (__m128d) x, bias );
5294 // All conversions up to here are exact. The correctly rounded result is
5295 // calculated using the current rounding mode using the following
5297 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5298 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5299 // store doesn't really need to be here (except
5300 // maybe to zero the other double)
5305 DebugLoc dl = Op.getDebugLoc();
5306 LLVMContext *Context = DAG.getContext();
5308 // Build some magic constants.
5309 std::vector<Constant*> CV0;
5310 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5311 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5312 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5313 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5314 Constant *C0 = ConstantVector::get(CV0);
5315 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5317 std::vector<Constant*> CV1;
5319 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5321 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5322 Constant *C1 = ConstantVector::get(CV1);
5323 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5325 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5326 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5328 DAG.getIntPtrConstant(1)));
5329 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5330 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5332 DAG.getIntPtrConstant(0)));
5333 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5334 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5335 PseudoSourceValue::getConstantPool(), 0,
5337 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5338 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5339 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5340 PseudoSourceValue::getConstantPool(), 0,
5342 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5344 // Add the halves; easiest way is to swap them into another reg first.
5345 int ShufMask[2] = { 1, -1 };
5346 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5347 DAG.getUNDEF(MVT::v2f64), ShufMask);
5348 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5349 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5350 DAG.getIntPtrConstant(0));
5353 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5354 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5355 DebugLoc dl = Op.getDebugLoc();
5356 // FP constant to bias correct the final result.
5357 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5360 // Load the 32-bit value into an XMM register.
5361 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5362 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5364 DAG.getIntPtrConstant(0)));
5366 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5367 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5368 DAG.getIntPtrConstant(0));
5370 // Or the load with the bias.
5371 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5372 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5373 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5375 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5376 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5377 MVT::v2f64, Bias)));
5378 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5379 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5380 DAG.getIntPtrConstant(0));
5382 // Subtract the bias.
5383 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5385 // Handle final rounding.
5386 EVT DestVT = Op.getValueType();
5388 if (DestVT.bitsLT(MVT::f64)) {
5389 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5390 DAG.getIntPtrConstant(0));
5391 } else if (DestVT.bitsGT(MVT::f64)) {
5392 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5395 // Handle final rounding.
5399 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5400 SDValue N0 = Op.getOperand(0);
5401 DebugLoc dl = Op.getDebugLoc();
5403 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5404 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5405 // the optimization here.
5406 if (DAG.SignBitIsZero(N0))
5407 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5409 EVT SrcVT = N0.getValueType();
5410 if (SrcVT == MVT::i64) {
5411 // We only handle SSE2 f64 target here; caller can expand the rest.
5412 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5415 return LowerUINT_TO_FP_i64(Op, DAG);
5416 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5417 return LowerUINT_TO_FP_i32(Op, DAG);
5420 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5422 // Make a 64-bit buffer, and use it to build an FILD.
5423 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5424 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5425 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5426 getPointerTy(), StackSlot, WordOff);
5427 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5428 StackSlot, NULL, 0);
5429 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5430 OffsetSlot, NULL, 0);
5431 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5434 std::pair<SDValue,SDValue> X86TargetLowering::
5435 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5436 DebugLoc dl = Op.getDebugLoc();
5438 EVT DstTy = Op.getValueType();
5441 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5445 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5446 DstTy.getSimpleVT() >= MVT::i16 &&
5447 "Unknown FP_TO_SINT to lower!");
5449 // These are really Legal.
5450 if (DstTy == MVT::i32 &&
5451 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5452 return std::make_pair(SDValue(), SDValue());
5453 if (Subtarget->is64Bit() &&
5454 DstTy == MVT::i64 &&
5455 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5456 return std::make_pair(SDValue(), SDValue());
5458 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5460 MachineFunction &MF = DAG.getMachineFunction();
5461 unsigned MemSize = DstTy.getSizeInBits()/8;
5462 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5463 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5466 switch (DstTy.getSimpleVT().SimpleTy) {
5467 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5468 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5469 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5470 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5473 SDValue Chain = DAG.getEntryNode();
5474 SDValue Value = Op.getOperand(0);
5475 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5476 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5477 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5478 PseudoSourceValue::getFixedStack(SSFI), 0);
5479 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5481 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5483 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5484 Chain = Value.getValue(1);
5485 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5486 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5489 // Build the FP_TO_INT*_IN_MEM
5490 SDValue Ops[] = { Chain, Value, StackSlot };
5491 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5493 return std::make_pair(FIST, StackSlot);
5496 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5497 if (Op.getValueType().isVector()) {
5498 if (Op.getValueType() == MVT::v2i32 &&
5499 Op.getOperand(0).getValueType() == MVT::v2f64) {
5505 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5506 SDValue FIST = Vals.first, StackSlot = Vals.second;
5507 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5508 if (FIST.getNode() == 0) return Op;
5511 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5512 FIST, StackSlot, NULL, 0);
5515 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5516 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5517 SDValue FIST = Vals.first, StackSlot = Vals.second;
5518 assert(FIST.getNode() && "Unexpected failure");
5521 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5522 FIST, StackSlot, NULL, 0);
5525 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5526 LLVMContext *Context = DAG.getContext();
5527 DebugLoc dl = Op.getDebugLoc();
5528 EVT VT = Op.getValueType();
5531 EltVT = VT.getVectorElementType();
5532 std::vector<Constant*> CV;
5533 if (EltVT == MVT::f64) {
5534 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5538 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5544 Constant *C = ConstantVector::get(CV);
5545 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5546 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5547 PseudoSourceValue::getConstantPool(), 0,
5549 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5552 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5553 LLVMContext *Context = DAG.getContext();
5554 DebugLoc dl = Op.getDebugLoc();
5555 EVT VT = Op.getValueType();
5558 EltVT = VT.getVectorElementType();
5559 std::vector<Constant*> CV;
5560 if (EltVT == MVT::f64) {
5561 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5565 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5571 Constant *C = ConstantVector::get(CV);
5572 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5573 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5574 PseudoSourceValue::getConstantPool(), 0,
5576 if (VT.isVector()) {
5577 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5578 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5579 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5581 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5583 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5587 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5588 LLVMContext *Context = DAG.getContext();
5589 SDValue Op0 = Op.getOperand(0);
5590 SDValue Op1 = Op.getOperand(1);
5591 DebugLoc dl = Op.getDebugLoc();
5592 EVT VT = Op.getValueType();
5593 EVT SrcVT = Op1.getValueType();
5595 // If second operand is smaller, extend it first.
5596 if (SrcVT.bitsLT(VT)) {
5597 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5600 // And if it is bigger, shrink it first.
5601 if (SrcVT.bitsGT(VT)) {
5602 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5606 // At this point the operands and the result should have the same
5607 // type, and that won't be f80 since that is not custom lowered.
5609 // First get the sign bit of second operand.
5610 std::vector<Constant*> CV;
5611 if (SrcVT == MVT::f64) {
5612 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5613 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5615 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5616 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5617 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5618 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5620 Constant *C = ConstantVector::get(CV);
5621 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5622 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5623 PseudoSourceValue::getConstantPool(), 0,
5625 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5627 // Shift sign bit right or left if the two operands have different types.
5628 if (SrcVT.bitsGT(VT)) {
5629 // Op0 is MVT::f32, Op1 is MVT::f64.
5630 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5631 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5632 DAG.getConstant(32, MVT::i32));
5633 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5634 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5635 DAG.getIntPtrConstant(0));
5638 // Clear first operand sign bit.
5640 if (VT == MVT::f64) {
5641 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5642 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5644 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5645 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5646 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5647 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5649 C = ConstantVector::get(CV);
5650 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5651 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5652 PseudoSourceValue::getConstantPool(), 0,
5654 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5656 // Or the value with the sign bit.
5657 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5660 /// Emit nodes that will be selected as "test Op0,Op0", or something
5662 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5663 SelectionDAG &DAG) {
5664 DebugLoc dl = Op.getDebugLoc();
5666 // CF and OF aren't always set the way we want. Determine which
5667 // of these we need.
5668 bool NeedCF = false;
5669 bool NeedOF = false;
5671 case X86::COND_A: case X86::COND_AE:
5672 case X86::COND_B: case X86::COND_BE:
5675 case X86::COND_G: case X86::COND_GE:
5676 case X86::COND_L: case X86::COND_LE:
5677 case X86::COND_O: case X86::COND_NO:
5683 // See if we can use the EFLAGS value from the operand instead of
5684 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5685 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5686 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5687 unsigned Opcode = 0;
5688 unsigned NumOperands = 0;
5689 switch (Op.getNode()->getOpcode()) {
5691 // Due to an isel shortcoming, be conservative if this add is likely to
5692 // be selected as part of a load-modify-store instruction. When the root
5693 // node in a match is a store, isel doesn't know how to remap non-chain
5694 // non-flag uses of other nodes in the match, such as the ADD in this
5695 // case. This leads to the ADD being left around and reselected, with
5696 // the result being two adds in the output.
5697 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5698 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5699 if (UI->getOpcode() == ISD::STORE)
5701 if (ConstantSDNode *C =
5702 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5703 // An add of one will be selected as an INC.
5704 if (C->getAPIntValue() == 1) {
5705 Opcode = X86ISD::INC;
5709 // An add of negative one (subtract of one) will be selected as a DEC.
5710 if (C->getAPIntValue().isAllOnesValue()) {
5711 Opcode = X86ISD::DEC;
5716 // Otherwise use a regular EFLAGS-setting add.
5717 Opcode = X86ISD::ADD;
5721 // If the primary and result isn't used, don't bother using X86ISD::AND,
5722 // because a TEST instruction will be better.
5723 bool NonFlagUse = false;
5724 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5725 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5727 unsigned UOpNo = UI.getOperandNo();
5728 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5729 // Look pass truncate.
5730 UOpNo = User->use_begin().getOperandNo();
5731 User = *User->use_begin();
5733 if (User->getOpcode() != ISD::BRCOND &&
5734 User->getOpcode() != ISD::SETCC &&
5735 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5747 // Due to the ISEL shortcoming noted above, be conservative if this op is
5748 // likely to be selected as part of a load-modify-store instruction.
5749 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5750 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5751 if (UI->getOpcode() == ISD::STORE)
5753 // Otherwise use a regular EFLAGS-setting instruction.
5754 switch (Op.getNode()->getOpcode()) {
5755 case ISD::SUB: Opcode = X86ISD::SUB; break;
5756 case ISD::OR: Opcode = X86ISD::OR; break;
5757 case ISD::XOR: Opcode = X86ISD::XOR; break;
5758 case ISD::AND: Opcode = X86ISD::AND; break;
5759 default: llvm_unreachable("unexpected operator!");
5770 return SDValue(Op.getNode(), 1);
5776 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5777 SmallVector<SDValue, 4> Ops;
5778 for (unsigned i = 0; i != NumOperands; ++i)
5779 Ops.push_back(Op.getOperand(i));
5780 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5781 DAG.ReplaceAllUsesWith(Op, New);
5782 return SDValue(New.getNode(), 1);
5786 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5787 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5788 DAG.getConstant(0, Op.getValueType()));
5791 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5793 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5794 SelectionDAG &DAG) {
5795 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5796 if (C->getAPIntValue() == 0)
5797 return EmitTest(Op0, X86CC, DAG);
5799 DebugLoc dl = Op0.getDebugLoc();
5800 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5803 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5804 /// if it's possible.
5805 static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
5806 DebugLoc dl, SelectionDAG &DAG) {
5808 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5809 if (ConstantSDNode *Op010C =
5810 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5811 if (Op010C->getZExtValue() == 1) {
5812 LHS = Op0.getOperand(0);
5813 RHS = Op0.getOperand(1).getOperand(1);
5815 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5816 if (ConstantSDNode *Op000C =
5817 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5818 if (Op000C->getZExtValue() == 1) {
5819 LHS = Op0.getOperand(1);
5820 RHS = Op0.getOperand(0).getOperand(1);
5822 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5823 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5824 SDValue AndLHS = Op0.getOperand(0);
5825 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5826 LHS = AndLHS.getOperand(0);
5827 RHS = AndLHS.getOperand(1);
5831 if (LHS.getNode()) {
5832 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5833 // instruction. Since the shift amount is in-range-or-undefined, we know
5834 // that doing a bittest on the i16 value is ok. We extend to i32 because
5835 // the encoding for the i16 version is larger than the i32 version.
5836 if (LHS.getValueType() == MVT::i8)
5837 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5839 // If the operand types disagree, extend the shift amount to match. Since
5840 // BT ignores high bits (like shifts) we can use anyextend.
5841 if (LHS.getValueType() != RHS.getValueType())
5842 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5844 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5845 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5846 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5847 DAG.getConstant(Cond, MVT::i8), BT);
5853 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5854 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5855 SDValue Op0 = Op.getOperand(0);
5856 SDValue Op1 = Op.getOperand(1);
5857 DebugLoc dl = Op.getDebugLoc();
5858 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5860 // Optimize to BT if possible.
5861 // Lower (X & (1 << N)) == 0 to BT(X, N).
5862 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5863 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5864 if (Op0.getOpcode() == ISD::AND &&
5866 Op1.getOpcode() == ISD::Constant &&
5867 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5868 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5869 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5870 if (NewSetCC.getNode())
5874 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5875 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5876 if (X86CC == X86::COND_INVALID)
5879 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5881 // Use sbb x, x to materialize carry bit into a GPR.
5882 if (X86CC == X86::COND_B)
5883 return DAG.getNode(ISD::AND, dl, MVT::i8,
5884 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5885 DAG.getConstant(X86CC, MVT::i8), Cond),
5886 DAG.getConstant(1, MVT::i8));
5888 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5889 DAG.getConstant(X86CC, MVT::i8), Cond);
5892 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5894 SDValue Op0 = Op.getOperand(0);
5895 SDValue Op1 = Op.getOperand(1);
5896 SDValue CC = Op.getOperand(2);
5897 EVT VT = Op.getValueType();
5898 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5899 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5900 DebugLoc dl = Op.getDebugLoc();
5904 EVT VT0 = Op0.getValueType();
5905 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5906 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5909 switch (SetCCOpcode) {
5912 case ISD::SETEQ: SSECC = 0; break;
5914 case ISD::SETGT: Swap = true; // Fallthrough
5916 case ISD::SETOLT: SSECC = 1; break;
5918 case ISD::SETGE: Swap = true; // Fallthrough
5920 case ISD::SETOLE: SSECC = 2; break;
5921 case ISD::SETUO: SSECC = 3; break;
5923 case ISD::SETNE: SSECC = 4; break;
5924 case ISD::SETULE: Swap = true;
5925 case ISD::SETUGE: SSECC = 5; break;
5926 case ISD::SETULT: Swap = true;
5927 case ISD::SETUGT: SSECC = 6; break;
5928 case ISD::SETO: SSECC = 7; break;
5931 std::swap(Op0, Op1);
5933 // In the two special cases we can't handle, emit two comparisons.
5935 if (SetCCOpcode == ISD::SETUEQ) {
5937 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5938 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5939 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5941 else if (SetCCOpcode == ISD::SETONE) {
5943 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5944 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5945 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5947 llvm_unreachable("Illegal FP comparison");
5949 // Handle all other FP comparisons here.
5950 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5953 // We are handling one of the integer comparisons here. Since SSE only has
5954 // GT and EQ comparisons for integer, swapping operands and multiple
5955 // operations may be required for some comparisons.
5956 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5957 bool Swap = false, Invert = false, FlipSigns = false;
5959 switch (VT.getSimpleVT().SimpleTy) {
5962 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5964 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5966 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5967 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5970 switch (SetCCOpcode) {
5972 case ISD::SETNE: Invert = true;
5973 case ISD::SETEQ: Opc = EQOpc; break;
5974 case ISD::SETLT: Swap = true;
5975 case ISD::SETGT: Opc = GTOpc; break;
5976 case ISD::SETGE: Swap = true;
5977 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5978 case ISD::SETULT: Swap = true;
5979 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5980 case ISD::SETUGE: Swap = true;
5981 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5984 std::swap(Op0, Op1);
5986 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5987 // bits of the inputs before performing those operations.
5989 EVT EltVT = VT.getVectorElementType();
5990 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5992 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5993 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5995 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5996 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5999 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6001 // If the logical-not of the result is required, perform that now.
6003 Result = DAG.getNOT(dl, Result, VT);
6008 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6009 static bool isX86LogicalCmp(SDValue Op) {
6010 unsigned Opc = Op.getNode()->getOpcode();
6011 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6013 if (Op.getResNo() == 1 &&
6014 (Opc == X86ISD::ADD ||
6015 Opc == X86ISD::SUB ||
6016 Opc == X86ISD::SMUL ||
6017 Opc == X86ISD::UMUL ||
6018 Opc == X86ISD::INC ||
6019 Opc == X86ISD::DEC ||
6020 Opc == X86ISD::OR ||
6021 Opc == X86ISD::XOR ||
6022 Opc == X86ISD::AND))
6028 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6029 bool addTest = true;
6030 SDValue Cond = Op.getOperand(0);
6031 DebugLoc dl = Op.getDebugLoc();
6034 if (Cond.getOpcode() == ISD::SETCC) {
6035 SDValue NewCond = LowerSETCC(Cond, DAG);
6036 if (NewCond.getNode())
6040 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6041 SDValue Op1 = Op.getOperand(1);
6042 SDValue Op2 = Op.getOperand(2);
6043 if (Cond.getOpcode() == X86ISD::SETCC &&
6044 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6045 SDValue Cmp = Cond.getOperand(1);
6046 if (Cmp.getOpcode() == X86ISD::CMP) {
6047 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6048 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6049 ConstantSDNode *RHSC =
6050 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6051 if (N1C && N1C->isAllOnesValue() &&
6052 N2C && N2C->isNullValue() &&
6053 RHSC && RHSC->isNullValue()) {
6054 SDValue CmpOp0 = Cmp.getOperand(0);
6055 Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
6056 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6057 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6058 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6063 // Look pass (and (setcc_carry (cmp ...)), 1).
6064 if (Cond.getOpcode() == ISD::AND &&
6065 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6066 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6067 if (C && C->getAPIntValue() == 1)
6068 Cond = Cond.getOperand(0);
6071 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6072 // setting operand in place of the X86ISD::SETCC.
6073 if (Cond.getOpcode() == X86ISD::SETCC ||
6074 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6075 CC = Cond.getOperand(0);
6077 SDValue Cmp = Cond.getOperand(1);
6078 unsigned Opc = Cmp.getOpcode();
6079 EVT VT = Op.getValueType();
6081 bool IllegalFPCMov = false;
6082 if (VT.isFloatingPoint() && !VT.isVector() &&
6083 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6084 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6086 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6087 Opc == X86ISD::BT) { // FIXME
6094 // Look pass the truncate.
6095 if (Cond.getOpcode() == ISD::TRUNCATE)
6096 Cond = Cond.getOperand(0);
6098 // We know the result of AND is compared against zero. Try to match
6100 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6101 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6102 if (NewSetCC.getNode()) {
6103 CC = NewSetCC.getOperand(0);
6104 Cond = NewSetCC.getOperand(1);
6111 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6112 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6115 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6116 // condition is true.
6117 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6118 SDValue Ops[] = { Op2, Op1, CC, Cond };
6119 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6122 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6123 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6124 // from the AND / OR.
6125 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6126 Opc = Op.getOpcode();
6127 if (Opc != ISD::OR && Opc != ISD::AND)
6129 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6130 Op.getOperand(0).hasOneUse() &&
6131 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6132 Op.getOperand(1).hasOneUse());
6135 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6136 // 1 and that the SETCC node has a single use.
6137 static bool isXor1OfSetCC(SDValue Op) {
6138 if (Op.getOpcode() != ISD::XOR)
6140 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6141 if (N1C && N1C->getAPIntValue() == 1) {
6142 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6143 Op.getOperand(0).hasOneUse();
6148 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6149 bool addTest = true;
6150 SDValue Chain = Op.getOperand(0);
6151 SDValue Cond = Op.getOperand(1);
6152 SDValue Dest = Op.getOperand(2);
6153 DebugLoc dl = Op.getDebugLoc();
6156 if (Cond.getOpcode() == ISD::SETCC) {
6157 SDValue NewCond = LowerSETCC(Cond, DAG);
6158 if (NewCond.getNode())
6162 // FIXME: LowerXALUO doesn't handle these!!
6163 else if (Cond.getOpcode() == X86ISD::ADD ||
6164 Cond.getOpcode() == X86ISD::SUB ||
6165 Cond.getOpcode() == X86ISD::SMUL ||
6166 Cond.getOpcode() == X86ISD::UMUL)
6167 Cond = LowerXALUO(Cond, DAG);
6170 // Look pass (and (setcc_carry (cmp ...)), 1).
6171 if (Cond.getOpcode() == ISD::AND &&
6172 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6173 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6174 if (C && C->getAPIntValue() == 1)
6175 Cond = Cond.getOperand(0);
6178 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6179 // setting operand in place of the X86ISD::SETCC.
6180 if (Cond.getOpcode() == X86ISD::SETCC ||
6181 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6182 CC = Cond.getOperand(0);
6184 SDValue Cmp = Cond.getOperand(1);
6185 unsigned Opc = Cmp.getOpcode();
6186 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6187 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6191 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6195 // These can only come from an arithmetic instruction with overflow,
6196 // e.g. SADDO, UADDO.
6197 Cond = Cond.getNode()->getOperand(1);
6204 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6205 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6206 if (CondOpc == ISD::OR) {
6207 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6208 // two branches instead of an explicit OR instruction with a
6210 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6211 isX86LogicalCmp(Cmp)) {
6212 CC = Cond.getOperand(0).getOperand(0);
6213 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6214 Chain, Dest, CC, Cmp);
6215 CC = Cond.getOperand(1).getOperand(0);
6219 } else { // ISD::AND
6220 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6221 // two branches instead of an explicit AND instruction with a
6222 // separate test. However, we only do this if this block doesn't
6223 // have a fall-through edge, because this requires an explicit
6224 // jmp when the condition is false.
6225 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6226 isX86LogicalCmp(Cmp) &&
6227 Op.getNode()->hasOneUse()) {
6228 X86::CondCode CCode =
6229 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6230 CCode = X86::GetOppositeBranchCondition(CCode);
6231 CC = DAG.getConstant(CCode, MVT::i8);
6232 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6233 // Look for an unconditional branch following this conditional branch.
6234 // We need this because we need to reverse the successors in order
6235 // to implement FCMP_OEQ.
6236 if (User.getOpcode() == ISD::BR) {
6237 SDValue FalseBB = User.getOperand(1);
6239 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6240 assert(NewBR == User);
6243 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6244 Chain, Dest, CC, Cmp);
6245 X86::CondCode CCode =
6246 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6247 CCode = X86::GetOppositeBranchCondition(CCode);
6248 CC = DAG.getConstant(CCode, MVT::i8);
6254 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6255 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6256 // It should be transformed during dag combiner except when the condition
6257 // is set by a arithmetics with overflow node.
6258 X86::CondCode CCode =
6259 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6260 CCode = X86::GetOppositeBranchCondition(CCode);
6261 CC = DAG.getConstant(CCode, MVT::i8);
6262 Cond = Cond.getOperand(0).getOperand(1);
6268 // Look pass the truncate.
6269 if (Cond.getOpcode() == ISD::TRUNCATE)
6270 Cond = Cond.getOperand(0);
6272 // We know the result of AND is compared against zero. Try to match
6274 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6275 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6276 if (NewSetCC.getNode()) {
6277 CC = NewSetCC.getOperand(0);
6278 Cond = NewSetCC.getOperand(1);
6285 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6286 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6288 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6289 Chain, Dest, CC, Cond);
6293 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6294 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6295 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6296 // that the guard pages used by the OS virtual memory manager are allocated in
6297 // correct sequence.
6299 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6300 SelectionDAG &DAG) {
6301 assert(Subtarget->isTargetCygMing() &&
6302 "This should be used only on Cygwin/Mingw targets");
6303 DebugLoc dl = Op.getDebugLoc();
6306 SDValue Chain = Op.getOperand(0);
6307 SDValue Size = Op.getOperand(1);
6308 // FIXME: Ensure alignment here
6312 EVT IntPtr = getPointerTy();
6313 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6315 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6317 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6318 Flag = Chain.getValue(1);
6320 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6321 SDValue Ops[] = { Chain,
6322 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6323 DAG.getRegister(X86::EAX, IntPtr),
6324 DAG.getRegister(X86StackPtr, SPTy),
6326 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6327 Flag = Chain.getValue(1);
6329 Chain = DAG.getCALLSEQ_END(Chain,
6330 DAG.getIntPtrConstant(0, true),
6331 DAG.getIntPtrConstant(0, true),
6334 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6336 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6337 return DAG.getMergeValues(Ops1, 2, dl);
6341 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6343 SDValue Dst, SDValue Src,
6344 SDValue Size, unsigned Align,
6346 uint64_t DstSVOff) {
6347 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6349 // If not DWORD aligned or size is more than the threshold, call the library.
6350 // The libc version is likely to be faster for these cases. It can use the
6351 // address value and run time information about the CPU.
6352 if ((Align & 3) != 0 ||
6354 ConstantSize->getZExtValue() >
6355 getSubtarget()->getMaxInlineSizeThreshold()) {
6356 SDValue InFlag(0, 0);
6358 // Check to see if there is a specialized entry-point for memory zeroing.
6359 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6361 if (const char *bzeroEntry = V &&
6362 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6363 EVT IntPtr = getPointerTy();
6364 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6365 TargetLowering::ArgListTy Args;
6366 TargetLowering::ArgListEntry Entry;
6368 Entry.Ty = IntPtrTy;
6369 Args.push_back(Entry);
6371 Args.push_back(Entry);
6372 std::pair<SDValue,SDValue> CallResult =
6373 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6374 false, false, false, false,
6375 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6376 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6377 DAG.GetOrdering(Chain.getNode()));
6378 return CallResult.second;
6381 // Otherwise have the target-independent code call memset.
6385 uint64_t SizeVal = ConstantSize->getZExtValue();
6386 SDValue InFlag(0, 0);
6389 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6390 unsigned BytesLeft = 0;
6391 bool TwoRepStos = false;
6394 uint64_t Val = ValC->getZExtValue() & 255;
6396 // If the value is a constant, then we can potentially use larger sets.
6397 switch (Align & 3) {
6398 case 2: // WORD aligned
6401 Val = (Val << 8) | Val;
6403 case 0: // DWORD aligned
6406 Val = (Val << 8) | Val;
6407 Val = (Val << 16) | Val;
6408 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6411 Val = (Val << 32) | Val;
6414 default: // Byte aligned
6417 Count = DAG.getIntPtrConstant(SizeVal);
6421 if (AVT.bitsGT(MVT::i8)) {
6422 unsigned UBytes = AVT.getSizeInBits() / 8;
6423 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6424 BytesLeft = SizeVal % UBytes;
6427 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6429 InFlag = Chain.getValue(1);
6432 Count = DAG.getIntPtrConstant(SizeVal);
6433 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6434 InFlag = Chain.getValue(1);
6437 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6440 InFlag = Chain.getValue(1);
6441 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6444 InFlag = Chain.getValue(1);
6446 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6447 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6448 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6451 InFlag = Chain.getValue(1);
6453 EVT CVT = Count.getValueType();
6454 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6455 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6456 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6459 InFlag = Chain.getValue(1);
6460 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6461 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6462 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6463 } else if (BytesLeft) {
6464 // Handle the last 1 - 7 bytes.
6465 unsigned Offset = SizeVal - BytesLeft;
6466 EVT AddrVT = Dst.getValueType();
6467 EVT SizeVT = Size.getValueType();
6469 Chain = DAG.getMemset(Chain, dl,
6470 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6471 DAG.getConstant(Offset, AddrVT)),
6473 DAG.getConstant(BytesLeft, SizeVT),
6474 Align, DstSV, DstSVOff + Offset);
6477 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6482 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6483 SDValue Chain, SDValue Dst, SDValue Src,
6484 SDValue Size, unsigned Align,
6486 const Value *DstSV, uint64_t DstSVOff,
6487 const Value *SrcSV, uint64_t SrcSVOff) {
6488 // This requires the copy size to be a constant, preferrably
6489 // within a subtarget-specific limit.
6490 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6493 uint64_t SizeVal = ConstantSize->getZExtValue();
6494 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6497 /// If not DWORD aligned, call the library.
6498 if ((Align & 3) != 0)
6503 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6506 unsigned UBytes = AVT.getSizeInBits() / 8;
6507 unsigned CountVal = SizeVal / UBytes;
6508 SDValue Count = DAG.getIntPtrConstant(CountVal);
6509 unsigned BytesLeft = SizeVal % UBytes;
6511 SDValue InFlag(0, 0);
6512 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6515 InFlag = Chain.getValue(1);
6516 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6519 InFlag = Chain.getValue(1);
6520 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6523 InFlag = Chain.getValue(1);
6525 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6526 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6527 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6528 array_lengthof(Ops));
6530 SmallVector<SDValue, 4> Results;
6531 Results.push_back(RepMovs);
6533 // Handle the last 1 - 7 bytes.
6534 unsigned Offset = SizeVal - BytesLeft;
6535 EVT DstVT = Dst.getValueType();
6536 EVT SrcVT = Src.getValueType();
6537 EVT SizeVT = Size.getValueType();
6538 Results.push_back(DAG.getMemcpy(Chain, dl,
6539 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6540 DAG.getConstant(Offset, DstVT)),
6541 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6542 DAG.getConstant(Offset, SrcVT)),
6543 DAG.getConstant(BytesLeft, SizeVT),
6544 Align, AlwaysInline,
6545 DstSV, DstSVOff + Offset,
6546 SrcSV, SrcSVOff + Offset));
6549 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6550 &Results[0], Results.size());
6553 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6554 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6555 DebugLoc dl = Op.getDebugLoc();
6557 if (!Subtarget->is64Bit()) {
6558 // vastart just stores the address of the VarArgsFrameIndex slot into the
6559 // memory location argument.
6560 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6561 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6565 // gp_offset (0 - 6 * 8)
6566 // fp_offset (48 - 48 + 8 * 16)
6567 // overflow_arg_area (point to parameters coming in memory).
6569 SmallVector<SDValue, 8> MemOps;
6570 SDValue FIN = Op.getOperand(1);
6572 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6573 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6575 MemOps.push_back(Store);
6578 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6579 FIN, DAG.getIntPtrConstant(4));
6580 Store = DAG.getStore(Op.getOperand(0), dl,
6581 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6583 MemOps.push_back(Store);
6585 // Store ptr to overflow_arg_area
6586 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6587 FIN, DAG.getIntPtrConstant(4));
6588 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6589 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6590 MemOps.push_back(Store);
6592 // Store ptr to reg_save_area.
6593 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6594 FIN, DAG.getIntPtrConstant(8));
6595 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6596 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6597 MemOps.push_back(Store);
6598 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6599 &MemOps[0], MemOps.size());
6602 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6603 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6604 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6605 SDValue Chain = Op.getOperand(0);
6606 SDValue SrcPtr = Op.getOperand(1);
6607 SDValue SrcSV = Op.getOperand(2);
6609 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6613 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6614 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6615 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6616 SDValue Chain = Op.getOperand(0);
6617 SDValue DstPtr = Op.getOperand(1);
6618 SDValue SrcPtr = Op.getOperand(2);
6619 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6620 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6621 DebugLoc dl = Op.getDebugLoc();
6623 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6624 DAG.getIntPtrConstant(24), 8, false,
6625 DstSV, 0, SrcSV, 0);
6629 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6630 DebugLoc dl = Op.getDebugLoc();
6631 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6633 default: return SDValue(); // Don't custom lower most intrinsics.
6634 // Comparison intrinsics.
6635 case Intrinsic::x86_sse_comieq_ss:
6636 case Intrinsic::x86_sse_comilt_ss:
6637 case Intrinsic::x86_sse_comile_ss:
6638 case Intrinsic::x86_sse_comigt_ss:
6639 case Intrinsic::x86_sse_comige_ss:
6640 case Intrinsic::x86_sse_comineq_ss:
6641 case Intrinsic::x86_sse_ucomieq_ss:
6642 case Intrinsic::x86_sse_ucomilt_ss:
6643 case Intrinsic::x86_sse_ucomile_ss:
6644 case Intrinsic::x86_sse_ucomigt_ss:
6645 case Intrinsic::x86_sse_ucomige_ss:
6646 case Intrinsic::x86_sse_ucomineq_ss:
6647 case Intrinsic::x86_sse2_comieq_sd:
6648 case Intrinsic::x86_sse2_comilt_sd:
6649 case Intrinsic::x86_sse2_comile_sd:
6650 case Intrinsic::x86_sse2_comigt_sd:
6651 case Intrinsic::x86_sse2_comige_sd:
6652 case Intrinsic::x86_sse2_comineq_sd:
6653 case Intrinsic::x86_sse2_ucomieq_sd:
6654 case Intrinsic::x86_sse2_ucomilt_sd:
6655 case Intrinsic::x86_sse2_ucomile_sd:
6656 case Intrinsic::x86_sse2_ucomigt_sd:
6657 case Intrinsic::x86_sse2_ucomige_sd:
6658 case Intrinsic::x86_sse2_ucomineq_sd: {
6660 ISD::CondCode CC = ISD::SETCC_INVALID;
6663 case Intrinsic::x86_sse_comieq_ss:
6664 case Intrinsic::x86_sse2_comieq_sd:
6668 case Intrinsic::x86_sse_comilt_ss:
6669 case Intrinsic::x86_sse2_comilt_sd:
6673 case Intrinsic::x86_sse_comile_ss:
6674 case Intrinsic::x86_sse2_comile_sd:
6678 case Intrinsic::x86_sse_comigt_ss:
6679 case Intrinsic::x86_sse2_comigt_sd:
6683 case Intrinsic::x86_sse_comige_ss:
6684 case Intrinsic::x86_sse2_comige_sd:
6688 case Intrinsic::x86_sse_comineq_ss:
6689 case Intrinsic::x86_sse2_comineq_sd:
6693 case Intrinsic::x86_sse_ucomieq_ss:
6694 case Intrinsic::x86_sse2_ucomieq_sd:
6695 Opc = X86ISD::UCOMI;
6698 case Intrinsic::x86_sse_ucomilt_ss:
6699 case Intrinsic::x86_sse2_ucomilt_sd:
6700 Opc = X86ISD::UCOMI;
6703 case Intrinsic::x86_sse_ucomile_ss:
6704 case Intrinsic::x86_sse2_ucomile_sd:
6705 Opc = X86ISD::UCOMI;
6708 case Intrinsic::x86_sse_ucomigt_ss:
6709 case Intrinsic::x86_sse2_ucomigt_sd:
6710 Opc = X86ISD::UCOMI;
6713 case Intrinsic::x86_sse_ucomige_ss:
6714 case Intrinsic::x86_sse2_ucomige_sd:
6715 Opc = X86ISD::UCOMI;
6718 case Intrinsic::x86_sse_ucomineq_ss:
6719 case Intrinsic::x86_sse2_ucomineq_sd:
6720 Opc = X86ISD::UCOMI;
6725 SDValue LHS = Op.getOperand(1);
6726 SDValue RHS = Op.getOperand(2);
6727 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6728 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6729 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6730 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6731 DAG.getConstant(X86CC, MVT::i8), Cond);
6732 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6734 // ptest intrinsics. The intrinsic these come from are designed to return
6735 // an integer value, not just an instruction so lower it to the ptest
6736 // pattern and a setcc for the result.
6737 case Intrinsic::x86_sse41_ptestz:
6738 case Intrinsic::x86_sse41_ptestc:
6739 case Intrinsic::x86_sse41_ptestnzc:{
6742 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6743 case Intrinsic::x86_sse41_ptestz:
6745 X86CC = X86::COND_E;
6747 case Intrinsic::x86_sse41_ptestc:
6749 X86CC = X86::COND_B;
6751 case Intrinsic::x86_sse41_ptestnzc:
6753 X86CC = X86::COND_A;
6757 SDValue LHS = Op.getOperand(1);
6758 SDValue RHS = Op.getOperand(2);
6759 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6760 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6761 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6762 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6765 // Fix vector shift instructions where the last operand is a non-immediate
6767 case Intrinsic::x86_sse2_pslli_w:
6768 case Intrinsic::x86_sse2_pslli_d:
6769 case Intrinsic::x86_sse2_pslli_q:
6770 case Intrinsic::x86_sse2_psrli_w:
6771 case Intrinsic::x86_sse2_psrli_d:
6772 case Intrinsic::x86_sse2_psrli_q:
6773 case Intrinsic::x86_sse2_psrai_w:
6774 case Intrinsic::x86_sse2_psrai_d:
6775 case Intrinsic::x86_mmx_pslli_w:
6776 case Intrinsic::x86_mmx_pslli_d:
6777 case Intrinsic::x86_mmx_pslli_q:
6778 case Intrinsic::x86_mmx_psrli_w:
6779 case Intrinsic::x86_mmx_psrli_d:
6780 case Intrinsic::x86_mmx_psrli_q:
6781 case Intrinsic::x86_mmx_psrai_w:
6782 case Intrinsic::x86_mmx_psrai_d: {
6783 SDValue ShAmt = Op.getOperand(2);
6784 if (isa<ConstantSDNode>(ShAmt))
6787 unsigned NewIntNo = 0;
6788 EVT ShAmtVT = MVT::v4i32;
6790 case Intrinsic::x86_sse2_pslli_w:
6791 NewIntNo = Intrinsic::x86_sse2_psll_w;
6793 case Intrinsic::x86_sse2_pslli_d:
6794 NewIntNo = Intrinsic::x86_sse2_psll_d;
6796 case Intrinsic::x86_sse2_pslli_q:
6797 NewIntNo = Intrinsic::x86_sse2_psll_q;
6799 case Intrinsic::x86_sse2_psrli_w:
6800 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6802 case Intrinsic::x86_sse2_psrli_d:
6803 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6805 case Intrinsic::x86_sse2_psrli_q:
6806 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6808 case Intrinsic::x86_sse2_psrai_w:
6809 NewIntNo = Intrinsic::x86_sse2_psra_w;
6811 case Intrinsic::x86_sse2_psrai_d:
6812 NewIntNo = Intrinsic::x86_sse2_psra_d;
6815 ShAmtVT = MVT::v2i32;
6817 case Intrinsic::x86_mmx_pslli_w:
6818 NewIntNo = Intrinsic::x86_mmx_psll_w;
6820 case Intrinsic::x86_mmx_pslli_d:
6821 NewIntNo = Intrinsic::x86_mmx_psll_d;
6823 case Intrinsic::x86_mmx_pslli_q:
6824 NewIntNo = Intrinsic::x86_mmx_psll_q;
6826 case Intrinsic::x86_mmx_psrli_w:
6827 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6829 case Intrinsic::x86_mmx_psrli_d:
6830 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6832 case Intrinsic::x86_mmx_psrli_q:
6833 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6835 case Intrinsic::x86_mmx_psrai_w:
6836 NewIntNo = Intrinsic::x86_mmx_psra_w;
6838 case Intrinsic::x86_mmx_psrai_d:
6839 NewIntNo = Intrinsic::x86_mmx_psra_d;
6841 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6847 // The vector shift intrinsics with scalars uses 32b shift amounts but
6848 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6852 ShOps[1] = DAG.getConstant(0, MVT::i32);
6853 if (ShAmtVT == MVT::v4i32) {
6854 ShOps[2] = DAG.getUNDEF(MVT::i32);
6855 ShOps[3] = DAG.getUNDEF(MVT::i32);
6856 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6858 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6861 EVT VT = Op.getValueType();
6862 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6863 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6864 DAG.getConstant(NewIntNo, MVT::i32),
6865 Op.getOperand(1), ShAmt);
6870 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6871 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6872 DebugLoc dl = Op.getDebugLoc();
6875 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6877 DAG.getConstant(TD->getPointerSize(),
6878 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6879 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6880 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6885 // Just load the return address.
6886 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6887 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6888 RetAddrFI, NULL, 0);
6891 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6892 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6893 MFI->setFrameAddressIsTaken(true);
6894 EVT VT = Op.getValueType();
6895 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6896 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6897 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6898 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6900 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6904 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6905 SelectionDAG &DAG) {
6906 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6909 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6911 MachineFunction &MF = DAG.getMachineFunction();
6912 SDValue Chain = Op.getOperand(0);
6913 SDValue Offset = Op.getOperand(1);
6914 SDValue Handler = Op.getOperand(2);
6915 DebugLoc dl = Op.getDebugLoc();
6917 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6919 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6921 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6922 DAG.getIntPtrConstant(-TD->getPointerSize()));
6923 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6924 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6925 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6926 MF.getRegInfo().addLiveOut(StoreAddrReg);
6928 return DAG.getNode(X86ISD::EH_RETURN, dl,
6930 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6933 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6934 SelectionDAG &DAG) {
6935 SDValue Root = Op.getOperand(0);
6936 SDValue Trmp = Op.getOperand(1); // trampoline
6937 SDValue FPtr = Op.getOperand(2); // nested function
6938 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6939 DebugLoc dl = Op.getDebugLoc();
6941 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6943 const X86InstrInfo *TII =
6944 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6946 if (Subtarget->is64Bit()) {
6947 SDValue OutChains[6];
6949 // Large code-model.
6951 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6952 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6954 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6955 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6957 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6959 // Load the pointer to the nested function into R11.
6960 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6961 SDValue Addr = Trmp;
6962 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6965 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6966 DAG.getConstant(2, MVT::i64));
6967 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6969 // Load the 'nest' parameter value into R10.
6970 // R10 is specified in X86CallingConv.td
6971 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6972 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6973 DAG.getConstant(10, MVT::i64));
6974 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6975 Addr, TrmpAddr, 10);
6977 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6978 DAG.getConstant(12, MVT::i64));
6979 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6981 // Jump to the nested function.
6982 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6983 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6984 DAG.getConstant(20, MVT::i64));
6985 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6986 Addr, TrmpAddr, 20);
6988 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6989 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6990 DAG.getConstant(22, MVT::i64));
6991 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6995 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6996 return DAG.getMergeValues(Ops, 2, dl);
6998 const Function *Func =
6999 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7000 CallingConv::ID CC = Func->getCallingConv();
7005 llvm_unreachable("Unsupported calling convention");
7006 case CallingConv::C:
7007 case CallingConv::X86_StdCall: {
7008 // Pass 'nest' parameter in ECX.
7009 // Must be kept in sync with X86CallingConv.td
7012 // Check that ECX wasn't needed by an 'inreg' parameter.
7013 const FunctionType *FTy = Func->getFunctionType();
7014 const AttrListPtr &Attrs = Func->getAttributes();
7016 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7017 unsigned InRegCount = 0;
7020 for (FunctionType::param_iterator I = FTy->param_begin(),
7021 E = FTy->param_end(); I != E; ++I, ++Idx)
7022 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7023 // FIXME: should only count parameters that are lowered to integers.
7024 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7026 if (InRegCount > 2) {
7027 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7032 case CallingConv::X86_FastCall:
7033 case CallingConv::Fast:
7034 // Pass 'nest' parameter in EAX.
7035 // Must be kept in sync with X86CallingConv.td
7040 SDValue OutChains[4];
7043 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7044 DAG.getConstant(10, MVT::i32));
7045 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7047 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
7048 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7049 OutChains[0] = DAG.getStore(Root, dl,
7050 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7053 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7054 DAG.getConstant(1, MVT::i32));
7055 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
7057 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
7058 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7059 DAG.getConstant(5, MVT::i32));
7060 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7061 TrmpAddr, 5, false, 1);
7063 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7064 DAG.getConstant(6, MVT::i32));
7065 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
7068 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7069 return DAG.getMergeValues(Ops, 2, dl);
7073 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7075 The rounding mode is in bits 11:10 of FPSR, and has the following
7082 FLT_ROUNDS, on the other hand, expects the following:
7089 To perform the conversion, we do:
7090 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7093 MachineFunction &MF = DAG.getMachineFunction();
7094 const TargetMachine &TM = MF.getTarget();
7095 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7096 unsigned StackAlignment = TFI.getStackAlignment();
7097 EVT VT = Op.getValueType();
7098 DebugLoc dl = Op.getDebugLoc();
7100 // Save FP Control Word to stack slot
7101 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7102 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7104 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7105 DAG.getEntryNode(), StackSlot);
7107 // Load FP Control Word from stack slot
7108 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
7110 // Transform as necessary
7112 DAG.getNode(ISD::SRL, dl, MVT::i16,
7113 DAG.getNode(ISD::AND, dl, MVT::i16,
7114 CWD, DAG.getConstant(0x800, MVT::i16)),
7115 DAG.getConstant(11, MVT::i8));
7117 DAG.getNode(ISD::SRL, dl, MVT::i16,
7118 DAG.getNode(ISD::AND, dl, MVT::i16,
7119 CWD, DAG.getConstant(0x400, MVT::i16)),
7120 DAG.getConstant(9, MVT::i8));
7123 DAG.getNode(ISD::AND, dl, MVT::i16,
7124 DAG.getNode(ISD::ADD, dl, MVT::i16,
7125 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7126 DAG.getConstant(1, MVT::i16)),
7127 DAG.getConstant(3, MVT::i16));
7130 return DAG.getNode((VT.getSizeInBits() < 16 ?
7131 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7134 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7135 EVT VT = Op.getValueType();
7137 unsigned NumBits = VT.getSizeInBits();
7138 DebugLoc dl = Op.getDebugLoc();
7140 Op = Op.getOperand(0);
7141 if (VT == MVT::i8) {
7142 // Zero extend to i32 since there is not an i8 bsr.
7144 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7147 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7148 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7149 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7151 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7154 DAG.getConstant(NumBits+NumBits-1, OpVT),
7155 DAG.getConstant(X86::COND_E, MVT::i8),
7158 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7160 // Finally xor with NumBits-1.
7161 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7164 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7168 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7169 EVT VT = Op.getValueType();
7171 unsigned NumBits = VT.getSizeInBits();
7172 DebugLoc dl = Op.getDebugLoc();
7174 Op = Op.getOperand(0);
7175 if (VT == MVT::i8) {
7177 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7180 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7181 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7182 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7184 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7187 DAG.getConstant(NumBits, OpVT),
7188 DAG.getConstant(X86::COND_E, MVT::i8),
7191 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7194 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7198 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7199 EVT VT = Op.getValueType();
7200 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7201 DebugLoc dl = Op.getDebugLoc();
7203 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7204 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7205 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7206 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7207 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7209 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7210 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7211 // return AloBlo + AloBhi + AhiBlo;
7213 SDValue A = Op.getOperand(0);
7214 SDValue B = Op.getOperand(1);
7216 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7217 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7218 A, DAG.getConstant(32, MVT::i32));
7219 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7220 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7221 B, DAG.getConstant(32, MVT::i32));
7222 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7223 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7225 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7226 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7228 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7229 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7231 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7232 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7233 AloBhi, DAG.getConstant(32, MVT::i32));
7234 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7235 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7236 AhiBlo, DAG.getConstant(32, MVT::i32));
7237 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7238 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7243 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7244 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7245 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7246 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7247 // has only one use.
7248 SDNode *N = Op.getNode();
7249 SDValue LHS = N->getOperand(0);
7250 SDValue RHS = N->getOperand(1);
7251 unsigned BaseOp = 0;
7253 DebugLoc dl = Op.getDebugLoc();
7255 switch (Op.getOpcode()) {
7256 default: llvm_unreachable("Unknown ovf instruction!");
7258 // A subtract of one will be selected as a INC. Note that INC doesn't
7259 // set CF, so we can't do this for UADDO.
7260 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7261 if (C->getAPIntValue() == 1) {
7262 BaseOp = X86ISD::INC;
7266 BaseOp = X86ISD::ADD;
7270 BaseOp = X86ISD::ADD;
7274 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7275 // set CF, so we can't do this for USUBO.
7276 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7277 if (C->getAPIntValue() == 1) {
7278 BaseOp = X86ISD::DEC;
7282 BaseOp = X86ISD::SUB;
7286 BaseOp = X86ISD::SUB;
7290 BaseOp = X86ISD::SMUL;
7294 BaseOp = X86ISD::UMUL;
7299 // Also sets EFLAGS.
7300 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7301 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7304 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7305 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7307 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7311 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7312 EVT T = Op.getValueType();
7313 DebugLoc dl = Op.getDebugLoc();
7316 switch(T.getSimpleVT().SimpleTy) {
7318 assert(false && "Invalid value type!");
7319 case MVT::i8: Reg = X86::AL; size = 1; break;
7320 case MVT::i16: Reg = X86::AX; size = 2; break;
7321 case MVT::i32: Reg = X86::EAX; size = 4; break;
7323 assert(Subtarget->is64Bit() && "Node not type legal!");
7324 Reg = X86::RAX; size = 8;
7327 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7328 Op.getOperand(2), SDValue());
7329 SDValue Ops[] = { cpIn.getValue(0),
7332 DAG.getTargetConstant(size, MVT::i8),
7334 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7335 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7337 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7341 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7342 SelectionDAG &DAG) {
7343 assert(Subtarget->is64Bit() && "Result not type legalized?");
7344 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7345 SDValue TheChain = Op.getOperand(0);
7346 DebugLoc dl = Op.getDebugLoc();
7347 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7348 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7349 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7351 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7352 DAG.getConstant(32, MVT::i8));
7354 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7357 return DAG.getMergeValues(Ops, 2, dl);
7360 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7361 SDNode *Node = Op.getNode();
7362 DebugLoc dl = Node->getDebugLoc();
7363 EVT T = Node->getValueType(0);
7364 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7365 DAG.getConstant(0, T), Node->getOperand(2));
7366 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7367 cast<AtomicSDNode>(Node)->getMemoryVT(),
7368 Node->getOperand(0),
7369 Node->getOperand(1), negOp,
7370 cast<AtomicSDNode>(Node)->getSrcValue(),
7371 cast<AtomicSDNode>(Node)->getAlignment());
7374 /// LowerOperation - Provide custom lowering hooks for some operations.
7376 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7377 switch (Op.getOpcode()) {
7378 default: llvm_unreachable("Should not custom lower this!");
7379 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7380 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7381 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7382 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7383 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7384 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7385 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7386 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7387 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7388 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7389 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7390 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7391 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7392 case ISD::SHL_PARTS:
7393 case ISD::SRA_PARTS:
7394 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7395 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7396 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7397 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7398 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7399 case ISD::FABS: return LowerFABS(Op, DAG);
7400 case ISD::FNEG: return LowerFNEG(Op, DAG);
7401 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7402 case ISD::SETCC: return LowerSETCC(Op, DAG);
7403 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7404 case ISD::SELECT: return LowerSELECT(Op, DAG);
7405 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7406 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7407 case ISD::VASTART: return LowerVASTART(Op, DAG);
7408 case ISD::VAARG: return LowerVAARG(Op, DAG);
7409 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7410 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7411 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7412 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7413 case ISD::FRAME_TO_ARGS_OFFSET:
7414 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7415 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7416 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7417 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7418 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7419 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7420 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7421 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7427 case ISD::UMULO: return LowerXALUO(Op, DAG);
7428 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7432 void X86TargetLowering::
7433 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7434 SelectionDAG &DAG, unsigned NewOp) {
7435 EVT T = Node->getValueType(0);
7436 DebugLoc dl = Node->getDebugLoc();
7437 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7439 SDValue Chain = Node->getOperand(0);
7440 SDValue In1 = Node->getOperand(1);
7441 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7442 Node->getOperand(2), DAG.getIntPtrConstant(0));
7443 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7444 Node->getOperand(2), DAG.getIntPtrConstant(1));
7445 SDValue Ops[] = { Chain, In1, In2L, In2H };
7446 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7448 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7449 cast<MemSDNode>(Node)->getMemOperand());
7450 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7451 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7452 Results.push_back(Result.getValue(2));
7455 /// ReplaceNodeResults - Replace a node with an illegal result type
7456 /// with a new node built out of custom code.
7457 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7458 SmallVectorImpl<SDValue>&Results,
7459 SelectionDAG &DAG) {
7460 DebugLoc dl = N->getDebugLoc();
7461 switch (N->getOpcode()) {
7463 assert(false && "Do not know how to custom type legalize this operation!");
7465 case ISD::FP_TO_SINT: {
7466 std::pair<SDValue,SDValue> Vals =
7467 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7468 SDValue FIST = Vals.first, StackSlot = Vals.second;
7469 if (FIST.getNode() != 0) {
7470 EVT VT = N->getValueType(0);
7471 // Return a load from the stack slot.
7472 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7476 case ISD::READCYCLECOUNTER: {
7477 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7478 SDValue TheChain = N->getOperand(0);
7479 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7480 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7482 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7484 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7485 SDValue Ops[] = { eax, edx };
7486 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7487 Results.push_back(edx.getValue(1));
7494 EVT WidenVT = getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
7495 Results.push_back(DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()));
7498 case ISD::ATOMIC_CMP_SWAP: {
7499 EVT T = N->getValueType(0);
7500 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7501 SDValue cpInL, cpInH;
7502 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7503 DAG.getConstant(0, MVT::i32));
7504 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7505 DAG.getConstant(1, MVT::i32));
7506 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7507 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7509 SDValue swapInL, swapInH;
7510 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7511 DAG.getConstant(0, MVT::i32));
7512 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7513 DAG.getConstant(1, MVT::i32));
7514 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7516 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7517 swapInL.getValue(1));
7518 SDValue Ops[] = { swapInH.getValue(0),
7520 swapInH.getValue(1) };
7521 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7522 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7523 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7524 MVT::i32, Result.getValue(1));
7525 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7526 MVT::i32, cpOutL.getValue(2));
7527 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7528 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7529 Results.push_back(cpOutH.getValue(1));
7532 case ISD::ATOMIC_LOAD_ADD:
7533 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7535 case ISD::ATOMIC_LOAD_AND:
7536 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7538 case ISD::ATOMIC_LOAD_NAND:
7539 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7541 case ISD::ATOMIC_LOAD_OR:
7542 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7544 case ISD::ATOMIC_LOAD_SUB:
7545 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7547 case ISD::ATOMIC_LOAD_XOR:
7548 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7550 case ISD::ATOMIC_SWAP:
7551 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7556 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7558 default: return NULL;
7559 case X86ISD::BSF: return "X86ISD::BSF";
7560 case X86ISD::BSR: return "X86ISD::BSR";
7561 case X86ISD::SHLD: return "X86ISD::SHLD";
7562 case X86ISD::SHRD: return "X86ISD::SHRD";
7563 case X86ISD::FAND: return "X86ISD::FAND";
7564 case X86ISD::FOR: return "X86ISD::FOR";
7565 case X86ISD::FXOR: return "X86ISD::FXOR";
7566 case X86ISD::FSRL: return "X86ISD::FSRL";
7567 case X86ISD::FILD: return "X86ISD::FILD";
7568 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7569 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7570 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7571 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7572 case X86ISD::FLD: return "X86ISD::FLD";
7573 case X86ISD::FST: return "X86ISD::FST";
7574 case X86ISD::CALL: return "X86ISD::CALL";
7575 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7576 case X86ISD::BT: return "X86ISD::BT";
7577 case X86ISD::CMP: return "X86ISD::CMP";
7578 case X86ISD::COMI: return "X86ISD::COMI";
7579 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7580 case X86ISD::SETCC: return "X86ISD::SETCC";
7581 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7582 case X86ISD::CMOV: return "X86ISD::CMOV";
7583 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7584 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7585 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7586 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7587 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7588 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7589 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7590 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7591 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7592 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7593 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7594 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7595 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7596 case X86ISD::FMAX: return "X86ISD::FMAX";
7597 case X86ISD::FMIN: return "X86ISD::FMIN";
7598 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7599 case X86ISD::FRCP: return "X86ISD::FRCP";
7600 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7601 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7602 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7603 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7604 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7605 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7606 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7607 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7608 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7609 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7610 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7611 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7612 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7613 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7614 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7615 case X86ISD::VSHL: return "X86ISD::VSHL";
7616 case X86ISD::VSRL: return "X86ISD::VSRL";
7617 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7618 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7619 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7620 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7621 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7622 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7623 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7624 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7625 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7626 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7627 case X86ISD::ADD: return "X86ISD::ADD";
7628 case X86ISD::SUB: return "X86ISD::SUB";
7629 case X86ISD::SMUL: return "X86ISD::SMUL";
7630 case X86ISD::UMUL: return "X86ISD::UMUL";
7631 case X86ISD::INC: return "X86ISD::INC";
7632 case X86ISD::DEC: return "X86ISD::DEC";
7633 case X86ISD::OR: return "X86ISD::OR";
7634 case X86ISD::XOR: return "X86ISD::XOR";
7635 case X86ISD::AND: return "X86ISD::AND";
7636 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7637 case X86ISD::PTEST: return "X86ISD::PTEST";
7638 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7642 // isLegalAddressingMode - Return true if the addressing mode represented
7643 // by AM is legal for this target, for a load/store of the specified type.
7644 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7645 const Type *Ty) const {
7646 // X86 supports extremely general addressing modes.
7647 CodeModel::Model M = getTargetMachine().getCodeModel();
7649 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7650 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7655 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7657 // If a reference to this global requires an extra load, we can't fold it.
7658 if (isGlobalStubReference(GVFlags))
7661 // If BaseGV requires a register for the PIC base, we cannot also have a
7662 // BaseReg specified.
7663 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7666 // If lower 4G is not available, then we must use rip-relative addressing.
7667 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7677 // These scales always work.
7682 // These scales are formed with basereg+scalereg. Only accept if there is
7687 default: // Other stuff never works.
7695 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7696 if (!Ty1->isInteger() || !Ty2->isInteger())
7698 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7699 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7700 if (NumBits1 <= NumBits2)
7702 return Subtarget->is64Bit() || NumBits1 < 64;
7705 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7706 if (!VT1.isInteger() || !VT2.isInteger())
7708 unsigned NumBits1 = VT1.getSizeInBits();
7709 unsigned NumBits2 = VT2.getSizeInBits();
7710 if (NumBits1 <= NumBits2)
7712 return Subtarget->is64Bit() || NumBits1 < 64;
7715 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7716 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7717 return Ty1->isInteger(32) && Ty2->isInteger(64) && Subtarget->is64Bit();
7720 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7721 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7722 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7725 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7726 // i16 instructions are longer (0x66 prefix) and potentially slower.
7727 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7730 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7731 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7732 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7733 /// are assumed to be legal.
7735 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7737 // Only do shuffles on 128-bit vector types for now.
7738 if (VT.getSizeInBits() == 64)
7741 // FIXME: pshufb, blends, shifts.
7742 return (VT.getVectorNumElements() == 2 ||
7743 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7744 isMOVLMask(M, VT) ||
7745 isSHUFPMask(M, VT) ||
7746 isPSHUFDMask(M, VT) ||
7747 isPSHUFHWMask(M, VT) ||
7748 isPSHUFLWMask(M, VT) ||
7749 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7750 isUNPCKLMask(M, VT) ||
7751 isUNPCKHMask(M, VT) ||
7752 isUNPCKL_v_undef_Mask(M, VT) ||
7753 isUNPCKH_v_undef_Mask(M, VT));
7757 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7759 unsigned NumElts = VT.getVectorNumElements();
7760 // FIXME: This collection of masks seems suspect.
7763 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7764 return (isMOVLMask(Mask, VT) ||
7765 isCommutedMOVLMask(Mask, VT, true) ||
7766 isSHUFPMask(Mask, VT) ||
7767 isCommutedSHUFPMask(Mask, VT));
7772 //===----------------------------------------------------------------------===//
7773 // X86 Scheduler Hooks
7774 //===----------------------------------------------------------------------===//
7776 // private utility function
7778 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7779 MachineBasicBlock *MBB,
7787 TargetRegisterClass *RC,
7788 bool invSrc) const {
7789 // For the atomic bitwise operator, we generate
7792 // ld t1 = [bitinstr.addr]
7793 // op t2 = t1, [bitinstr.val]
7795 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7797 // fallthrough -->nextMBB
7798 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7799 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7800 MachineFunction::iterator MBBIter = MBB;
7803 /// First build the CFG
7804 MachineFunction *F = MBB->getParent();
7805 MachineBasicBlock *thisMBB = MBB;
7806 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7807 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7808 F->insert(MBBIter, newMBB);
7809 F->insert(MBBIter, nextMBB);
7811 // Move all successors to thisMBB to nextMBB
7812 nextMBB->transferSuccessors(thisMBB);
7814 // Update thisMBB to fall through to newMBB
7815 thisMBB->addSuccessor(newMBB);
7817 // newMBB jumps to itself and fall through to nextMBB
7818 newMBB->addSuccessor(nextMBB);
7819 newMBB->addSuccessor(newMBB);
7821 // Insert instructions into newMBB based on incoming instruction
7822 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7823 "unexpected number of operands");
7824 DebugLoc dl = bInstr->getDebugLoc();
7825 MachineOperand& destOper = bInstr->getOperand(0);
7826 MachineOperand* argOpers[2 + X86AddrNumOperands];
7827 int numArgs = bInstr->getNumOperands() - 1;
7828 for (int i=0; i < numArgs; ++i)
7829 argOpers[i] = &bInstr->getOperand(i+1);
7831 // x86 address has 4 operands: base, index, scale, and displacement
7832 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7833 int valArgIndx = lastAddrIndx + 1;
7835 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7836 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7837 for (int i=0; i <= lastAddrIndx; ++i)
7838 (*MIB).addOperand(*argOpers[i]);
7840 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7842 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7847 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7848 assert((argOpers[valArgIndx]->isReg() ||
7849 argOpers[valArgIndx]->isImm()) &&
7851 if (argOpers[valArgIndx]->isReg())
7852 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7854 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7856 (*MIB).addOperand(*argOpers[valArgIndx]);
7858 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7861 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7862 for (int i=0; i <= lastAddrIndx; ++i)
7863 (*MIB).addOperand(*argOpers[i]);
7865 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7866 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7867 bInstr->memoperands_end());
7869 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7873 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7875 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7879 // private utility function: 64 bit atomics on 32 bit host.
7881 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7882 MachineBasicBlock *MBB,
7887 bool invSrc) const {
7888 // For the atomic bitwise operator, we generate
7889 // thisMBB (instructions are in pairs, except cmpxchg8b)
7890 // ld t1,t2 = [bitinstr.addr]
7892 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7893 // op t5, t6 <- out1, out2, [bitinstr.val]
7894 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7895 // mov ECX, EBX <- t5, t6
7896 // mov EAX, EDX <- t1, t2
7897 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7898 // mov t3, t4 <- EAX, EDX
7900 // result in out1, out2
7901 // fallthrough -->nextMBB
7903 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7904 const unsigned LoadOpc = X86::MOV32rm;
7905 const unsigned copyOpc = X86::MOV32rr;
7906 const unsigned NotOpc = X86::NOT32r;
7907 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7908 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7909 MachineFunction::iterator MBBIter = MBB;
7912 /// First build the CFG
7913 MachineFunction *F = MBB->getParent();
7914 MachineBasicBlock *thisMBB = MBB;
7915 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7916 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7917 F->insert(MBBIter, newMBB);
7918 F->insert(MBBIter, nextMBB);
7920 // Move all successors to thisMBB to nextMBB
7921 nextMBB->transferSuccessors(thisMBB);
7923 // Update thisMBB to fall through to newMBB
7924 thisMBB->addSuccessor(newMBB);
7926 // newMBB jumps to itself and fall through to nextMBB
7927 newMBB->addSuccessor(nextMBB);
7928 newMBB->addSuccessor(newMBB);
7930 DebugLoc dl = bInstr->getDebugLoc();
7931 // Insert instructions into newMBB based on incoming instruction
7932 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7933 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7934 "unexpected number of operands");
7935 MachineOperand& dest1Oper = bInstr->getOperand(0);
7936 MachineOperand& dest2Oper = bInstr->getOperand(1);
7937 MachineOperand* argOpers[2 + X86AddrNumOperands];
7938 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7939 argOpers[i] = &bInstr->getOperand(i+2);
7941 // x86 address has 5 operands: base, index, scale, displacement, and segment.
7942 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7944 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7945 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7946 for (int i=0; i <= lastAddrIndx; ++i)
7947 (*MIB).addOperand(*argOpers[i]);
7948 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7949 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7950 // add 4 to displacement.
7951 for (int i=0; i <= lastAddrIndx-2; ++i)
7952 (*MIB).addOperand(*argOpers[i]);
7953 MachineOperand newOp3 = *(argOpers[3]);
7955 newOp3.setImm(newOp3.getImm()+4);
7957 newOp3.setOffset(newOp3.getOffset()+4);
7958 (*MIB).addOperand(newOp3);
7959 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7961 // t3/4 are defined later, at the bottom of the loop
7962 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7963 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7964 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7965 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7966 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7967 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7969 // The subsequent operations should be using the destination registers of
7970 //the PHI instructions.
7972 t1 = F->getRegInfo().createVirtualRegister(RC);
7973 t2 = F->getRegInfo().createVirtualRegister(RC);
7974 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
7975 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
7977 t1 = dest1Oper.getReg();
7978 t2 = dest2Oper.getReg();
7981 int valArgIndx = lastAddrIndx + 1;
7982 assert((argOpers[valArgIndx]->isReg() ||
7983 argOpers[valArgIndx]->isImm()) &&
7985 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7986 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7987 if (argOpers[valArgIndx]->isReg())
7988 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7990 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7991 if (regOpcL != X86::MOV32rr)
7993 (*MIB).addOperand(*argOpers[valArgIndx]);
7994 assert(argOpers[valArgIndx + 1]->isReg() ==
7995 argOpers[valArgIndx]->isReg());
7996 assert(argOpers[valArgIndx + 1]->isImm() ==
7997 argOpers[valArgIndx]->isImm());
7998 if (argOpers[valArgIndx + 1]->isReg())
7999 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8001 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8002 if (regOpcH != X86::MOV32rr)
8004 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8006 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8008 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8011 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8013 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8016 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8017 for (int i=0; i <= lastAddrIndx; ++i)
8018 (*MIB).addOperand(*argOpers[i]);
8020 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8021 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8022 bInstr->memoperands_end());
8024 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8025 MIB.addReg(X86::EAX);
8026 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8027 MIB.addReg(X86::EDX);
8030 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8032 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8036 // private utility function
8038 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8039 MachineBasicBlock *MBB,
8040 unsigned cmovOpc) const {
8041 // For the atomic min/max operator, we generate
8044 // ld t1 = [min/max.addr]
8045 // mov t2 = [min/max.val]
8047 // cmov[cond] t2 = t1
8049 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8051 // fallthrough -->nextMBB
8053 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8054 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8055 MachineFunction::iterator MBBIter = MBB;
8058 /// First build the CFG
8059 MachineFunction *F = MBB->getParent();
8060 MachineBasicBlock *thisMBB = MBB;
8061 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8062 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8063 F->insert(MBBIter, newMBB);
8064 F->insert(MBBIter, nextMBB);
8066 // Move all successors of thisMBB to nextMBB
8067 nextMBB->transferSuccessors(thisMBB);
8069 // Update thisMBB to fall through to newMBB
8070 thisMBB->addSuccessor(newMBB);
8072 // newMBB jumps to newMBB and fall through to nextMBB
8073 newMBB->addSuccessor(nextMBB);
8074 newMBB->addSuccessor(newMBB);
8076 DebugLoc dl = mInstr->getDebugLoc();
8077 // Insert instructions into newMBB based on incoming instruction
8078 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8079 "unexpected number of operands");
8080 MachineOperand& destOper = mInstr->getOperand(0);
8081 MachineOperand* argOpers[2 + X86AddrNumOperands];
8082 int numArgs = mInstr->getNumOperands() - 1;
8083 for (int i=0; i < numArgs; ++i)
8084 argOpers[i] = &mInstr->getOperand(i+1);
8086 // x86 address has 4 operands: base, index, scale, and displacement
8087 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8088 int valArgIndx = lastAddrIndx + 1;
8090 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8091 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8092 for (int i=0; i <= lastAddrIndx; ++i)
8093 (*MIB).addOperand(*argOpers[i]);
8095 // We only support register and immediate values
8096 assert((argOpers[valArgIndx]->isReg() ||
8097 argOpers[valArgIndx]->isImm()) &&
8100 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8101 if (argOpers[valArgIndx]->isReg())
8102 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8104 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8105 (*MIB).addOperand(*argOpers[valArgIndx]);
8107 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8110 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8115 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8116 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8120 // Cmp and exchange if none has modified the memory location
8121 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8122 for (int i=0; i <= lastAddrIndx; ++i)
8123 (*MIB).addOperand(*argOpers[i]);
8125 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8126 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8127 mInstr->memoperands_end());
8129 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8130 MIB.addReg(X86::EAX);
8133 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8135 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8139 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8140 // all of this code can be replaced with that in the .td file.
8142 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8143 unsigned numArgs, bool memArg) const {
8145 MachineFunction *F = BB->getParent();
8146 DebugLoc dl = MI->getDebugLoc();
8147 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8151 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8153 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8155 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8157 for (unsigned i = 0; i < numArgs; ++i) {
8158 MachineOperand &Op = MI->getOperand(i+1);
8160 if (!(Op.isReg() && Op.isImplicit()))
8164 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8167 F->DeleteMachineInstr(MI);
8173 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8175 MachineBasicBlock *MBB) const {
8176 // Emit code to save XMM registers to the stack. The ABI says that the
8177 // number of registers to save is given in %al, so it's theoretically
8178 // possible to do an indirect jump trick to avoid saving all of them,
8179 // however this code takes a simpler approach and just executes all
8180 // of the stores if %al is non-zero. It's less code, and it's probably
8181 // easier on the hardware branch predictor, and stores aren't all that
8182 // expensive anyway.
8184 // Create the new basic blocks. One block contains all the XMM stores,
8185 // and one block is the final destination regardless of whether any
8186 // stores were performed.
8187 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8188 MachineFunction *F = MBB->getParent();
8189 MachineFunction::iterator MBBIter = MBB;
8191 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8192 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8193 F->insert(MBBIter, XMMSaveMBB);
8194 F->insert(MBBIter, EndMBB);
8197 // Move any original successors of MBB to the end block.
8198 EndMBB->transferSuccessors(MBB);
8199 // The original block will now fall through to the XMM save block.
8200 MBB->addSuccessor(XMMSaveMBB);
8201 // The XMMSaveMBB will fall through to the end block.
8202 XMMSaveMBB->addSuccessor(EndMBB);
8204 // Now add the instructions.
8205 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8206 DebugLoc DL = MI->getDebugLoc();
8208 unsigned CountReg = MI->getOperand(0).getReg();
8209 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8210 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8212 if (!Subtarget->isTargetWin64()) {
8213 // If %al is 0, branch around the XMM save block.
8214 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8215 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
8216 MBB->addSuccessor(EndMBB);
8219 // In the XMM save block, save all the XMM argument registers.
8220 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8221 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8222 MachineMemOperand *MMO =
8223 F->getMachineMemOperand(
8224 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8225 MachineMemOperand::MOStore, Offset,
8226 /*Size=*/16, /*Align=*/16);
8227 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8228 .addFrameIndex(RegSaveFrameIndex)
8229 .addImm(/*Scale=*/1)
8230 .addReg(/*IndexReg=*/0)
8231 .addImm(/*Disp=*/Offset)
8232 .addReg(/*Segment=*/0)
8233 .addReg(MI->getOperand(i).getReg())
8234 .addMemOperand(MMO);
8237 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8243 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8244 MachineBasicBlock *BB,
8245 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8246 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8247 DebugLoc DL = MI->getDebugLoc();
8249 // To "insert" a SELECT_CC instruction, we actually have to insert the
8250 // diamond control-flow pattern. The incoming instruction knows the
8251 // destination vreg to set, the condition code register to branch on, the
8252 // true/false values to select between, and a branch opcode to use.
8253 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8254 MachineFunction::iterator It = BB;
8260 // cmpTY ccX, r1, r2
8262 // fallthrough --> copy0MBB
8263 MachineBasicBlock *thisMBB = BB;
8264 MachineFunction *F = BB->getParent();
8265 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8266 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8268 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8269 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8270 F->insert(It, copy0MBB);
8271 F->insert(It, sinkMBB);
8272 // Update machine-CFG edges by first adding all successors of the current
8273 // block to the new block which will contain the Phi node for the select.
8274 // Also inform sdisel of the edge changes.
8275 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8276 E = BB->succ_end(); I != E; ++I) {
8277 EM->insert(std::make_pair(*I, sinkMBB));
8278 sinkMBB->addSuccessor(*I);
8280 // Next, remove all successors of the current block, and add the true
8281 // and fallthrough blocks as its successors.
8282 while (!BB->succ_empty())
8283 BB->removeSuccessor(BB->succ_begin());
8284 // Add the true and fallthrough blocks as its successors.
8285 BB->addSuccessor(copy0MBB);
8286 BB->addSuccessor(sinkMBB);
8289 // %FalseValue = ...
8290 // # fallthrough to sinkMBB
8293 // Update machine-CFG edges
8294 BB->addSuccessor(sinkMBB);
8297 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8300 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8301 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8302 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8304 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8310 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8311 MachineBasicBlock *BB,
8312 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8313 switch (MI->getOpcode()) {
8314 default: assert(false && "Unexpected instr type to insert");
8316 case X86::CMOV_V1I64:
8317 case X86::CMOV_FR32:
8318 case X86::CMOV_FR64:
8319 case X86::CMOV_V4F32:
8320 case X86::CMOV_V2F64:
8321 case X86::CMOV_V2I64:
8322 return EmitLoweredSelect(MI, BB, EM);
8324 case X86::FP32_TO_INT16_IN_MEM:
8325 case X86::FP32_TO_INT32_IN_MEM:
8326 case X86::FP32_TO_INT64_IN_MEM:
8327 case X86::FP64_TO_INT16_IN_MEM:
8328 case X86::FP64_TO_INT32_IN_MEM:
8329 case X86::FP64_TO_INT64_IN_MEM:
8330 case X86::FP80_TO_INT16_IN_MEM:
8331 case X86::FP80_TO_INT32_IN_MEM:
8332 case X86::FP80_TO_INT64_IN_MEM: {
8333 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8334 DebugLoc DL = MI->getDebugLoc();
8336 // Change the floating point control register to use "round towards zero"
8337 // mode when truncating to an integer value.
8338 MachineFunction *F = BB->getParent();
8339 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8340 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8342 // Load the old value of the high byte of the control word...
8344 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8345 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8348 // Set the high part to be round to zero...
8349 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8352 // Reload the modified control word now...
8353 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8355 // Restore the memory image of control word to original value
8356 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8359 // Get the X86 opcode to use.
8361 switch (MI->getOpcode()) {
8362 default: llvm_unreachable("illegal opcode!");
8363 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8364 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8365 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8366 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8367 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8368 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8369 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8370 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8371 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8375 MachineOperand &Op = MI->getOperand(0);
8377 AM.BaseType = X86AddressMode::RegBase;
8378 AM.Base.Reg = Op.getReg();
8380 AM.BaseType = X86AddressMode::FrameIndexBase;
8381 AM.Base.FrameIndex = Op.getIndex();
8383 Op = MI->getOperand(1);
8385 AM.Scale = Op.getImm();
8386 Op = MI->getOperand(2);
8388 AM.IndexReg = Op.getImm();
8389 Op = MI->getOperand(3);
8390 if (Op.isGlobal()) {
8391 AM.GV = Op.getGlobal();
8393 AM.Disp = Op.getImm();
8395 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8396 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8398 // Reload the original control word now.
8399 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8401 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8404 // String/text processing lowering.
8405 case X86::PCMPISTRM128REG:
8406 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8407 case X86::PCMPISTRM128MEM:
8408 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8409 case X86::PCMPESTRM128REG:
8410 return EmitPCMP(MI, BB, 5, false /* in mem */);
8411 case X86::PCMPESTRM128MEM:
8412 return EmitPCMP(MI, BB, 5, true /* in mem */);
8415 case X86::ATOMAND32:
8416 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8417 X86::AND32ri, X86::MOV32rm,
8418 X86::LCMPXCHG32, X86::MOV32rr,
8419 X86::NOT32r, X86::EAX,
8420 X86::GR32RegisterClass);
8422 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8423 X86::OR32ri, X86::MOV32rm,
8424 X86::LCMPXCHG32, X86::MOV32rr,
8425 X86::NOT32r, X86::EAX,
8426 X86::GR32RegisterClass);
8427 case X86::ATOMXOR32:
8428 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8429 X86::XOR32ri, X86::MOV32rm,
8430 X86::LCMPXCHG32, X86::MOV32rr,
8431 X86::NOT32r, X86::EAX,
8432 X86::GR32RegisterClass);
8433 case X86::ATOMNAND32:
8434 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8435 X86::AND32ri, X86::MOV32rm,
8436 X86::LCMPXCHG32, X86::MOV32rr,
8437 X86::NOT32r, X86::EAX,
8438 X86::GR32RegisterClass, true);
8439 case X86::ATOMMIN32:
8440 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8441 case X86::ATOMMAX32:
8442 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8443 case X86::ATOMUMIN32:
8444 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8445 case X86::ATOMUMAX32:
8446 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8448 case X86::ATOMAND16:
8449 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8450 X86::AND16ri, X86::MOV16rm,
8451 X86::LCMPXCHG16, X86::MOV16rr,
8452 X86::NOT16r, X86::AX,
8453 X86::GR16RegisterClass);
8455 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8456 X86::OR16ri, X86::MOV16rm,
8457 X86::LCMPXCHG16, X86::MOV16rr,
8458 X86::NOT16r, X86::AX,
8459 X86::GR16RegisterClass);
8460 case X86::ATOMXOR16:
8461 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8462 X86::XOR16ri, X86::MOV16rm,
8463 X86::LCMPXCHG16, X86::MOV16rr,
8464 X86::NOT16r, X86::AX,
8465 X86::GR16RegisterClass);
8466 case X86::ATOMNAND16:
8467 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8468 X86::AND16ri, X86::MOV16rm,
8469 X86::LCMPXCHG16, X86::MOV16rr,
8470 X86::NOT16r, X86::AX,
8471 X86::GR16RegisterClass, true);
8472 case X86::ATOMMIN16:
8473 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8474 case X86::ATOMMAX16:
8475 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8476 case X86::ATOMUMIN16:
8477 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8478 case X86::ATOMUMAX16:
8479 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8482 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8483 X86::AND8ri, X86::MOV8rm,
8484 X86::LCMPXCHG8, X86::MOV8rr,
8485 X86::NOT8r, X86::AL,
8486 X86::GR8RegisterClass);
8488 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8489 X86::OR8ri, X86::MOV8rm,
8490 X86::LCMPXCHG8, X86::MOV8rr,
8491 X86::NOT8r, X86::AL,
8492 X86::GR8RegisterClass);
8494 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8495 X86::XOR8ri, X86::MOV8rm,
8496 X86::LCMPXCHG8, X86::MOV8rr,
8497 X86::NOT8r, X86::AL,
8498 X86::GR8RegisterClass);
8499 case X86::ATOMNAND8:
8500 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8501 X86::AND8ri, X86::MOV8rm,
8502 X86::LCMPXCHG8, X86::MOV8rr,
8503 X86::NOT8r, X86::AL,
8504 X86::GR8RegisterClass, true);
8505 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8506 // This group is for 64-bit host.
8507 case X86::ATOMAND64:
8508 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8509 X86::AND64ri32, X86::MOV64rm,
8510 X86::LCMPXCHG64, X86::MOV64rr,
8511 X86::NOT64r, X86::RAX,
8512 X86::GR64RegisterClass);
8514 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8515 X86::OR64ri32, X86::MOV64rm,
8516 X86::LCMPXCHG64, X86::MOV64rr,
8517 X86::NOT64r, X86::RAX,
8518 X86::GR64RegisterClass);
8519 case X86::ATOMXOR64:
8520 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8521 X86::XOR64ri32, X86::MOV64rm,
8522 X86::LCMPXCHG64, X86::MOV64rr,
8523 X86::NOT64r, X86::RAX,
8524 X86::GR64RegisterClass);
8525 case X86::ATOMNAND64:
8526 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8527 X86::AND64ri32, X86::MOV64rm,
8528 X86::LCMPXCHG64, X86::MOV64rr,
8529 X86::NOT64r, X86::RAX,
8530 X86::GR64RegisterClass, true);
8531 case X86::ATOMMIN64:
8532 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8533 case X86::ATOMMAX64:
8534 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8535 case X86::ATOMUMIN64:
8536 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8537 case X86::ATOMUMAX64:
8538 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8540 // This group does 64-bit operations on a 32-bit host.
8541 case X86::ATOMAND6432:
8542 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8543 X86::AND32rr, X86::AND32rr,
8544 X86::AND32ri, X86::AND32ri,
8546 case X86::ATOMOR6432:
8547 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8548 X86::OR32rr, X86::OR32rr,
8549 X86::OR32ri, X86::OR32ri,
8551 case X86::ATOMXOR6432:
8552 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8553 X86::XOR32rr, X86::XOR32rr,
8554 X86::XOR32ri, X86::XOR32ri,
8556 case X86::ATOMNAND6432:
8557 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8558 X86::AND32rr, X86::AND32rr,
8559 X86::AND32ri, X86::AND32ri,
8561 case X86::ATOMADD6432:
8562 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8563 X86::ADD32rr, X86::ADC32rr,
8564 X86::ADD32ri, X86::ADC32ri,
8566 case X86::ATOMSUB6432:
8567 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8568 X86::SUB32rr, X86::SBB32rr,
8569 X86::SUB32ri, X86::SBB32ri,
8571 case X86::ATOMSWAP6432:
8572 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8573 X86::MOV32rr, X86::MOV32rr,
8574 X86::MOV32ri, X86::MOV32ri,
8576 case X86::VASTART_SAVE_XMM_REGS:
8577 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8581 //===----------------------------------------------------------------------===//
8582 // X86 Optimization Hooks
8583 //===----------------------------------------------------------------------===//
8585 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8589 const SelectionDAG &DAG,
8590 unsigned Depth) const {
8591 unsigned Opc = Op.getOpcode();
8592 assert((Opc >= ISD::BUILTIN_OP_END ||
8593 Opc == ISD::INTRINSIC_WO_CHAIN ||
8594 Opc == ISD::INTRINSIC_W_CHAIN ||
8595 Opc == ISD::INTRINSIC_VOID) &&
8596 "Should use MaskedValueIsZero if you don't know whether Op"
8597 " is a target node!");
8599 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8611 // These nodes' second result is a boolean.
8612 if (Op.getResNo() == 0)
8616 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8617 Mask.getBitWidth() - 1);
8622 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8623 /// node is a GlobalAddress + offset.
8624 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8625 GlobalValue* &GA, int64_t &Offset) const{
8626 if (N->getOpcode() == X86ISD::Wrapper) {
8627 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8628 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8629 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8633 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8636 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8637 EVT EltVT, LoadSDNode *&LDBase,
8638 unsigned &LastLoadedElt,
8639 SelectionDAG &DAG, MachineFrameInfo *MFI,
8640 const TargetLowering &TLI) {
8642 LastLoadedElt = -1U;
8643 for (unsigned i = 0; i < NumElems; ++i) {
8644 if (N->getMaskElt(i) < 0) {
8650 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8651 if (!Elt.getNode() ||
8652 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8655 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8657 LDBase = cast<LoadSDNode>(Elt.getNode());
8661 if (Elt.getOpcode() == ISD::UNDEF)
8664 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8665 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8672 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8673 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8674 /// if the load addresses are consecutive, non-overlapping, and in the right
8675 /// order. In the case of v2i64, it will see if it can rewrite the
8676 /// shuffle to be an appropriate build vector so it can take advantage of
8677 // performBuildVectorCombine.
8678 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8679 const TargetLowering &TLI) {
8680 DebugLoc dl = N->getDebugLoc();
8681 EVT VT = N->getValueType(0);
8682 EVT EltVT = VT.getVectorElementType();
8683 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8684 unsigned NumElems = VT.getVectorNumElements();
8686 if (VT.getSizeInBits() != 128)
8689 // Try to combine a vector_shuffle into a 128-bit load.
8690 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8691 LoadSDNode *LD = NULL;
8692 unsigned LastLoadedElt;
8693 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8697 if (LastLoadedElt == NumElems - 1) {
8698 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8699 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8700 LD->getSrcValue(), LD->getSrcValueOffset(),
8702 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8703 LD->getSrcValue(), LD->getSrcValueOffset(),
8704 LD->isVolatile(), LD->getAlignment());
8705 } else if (NumElems == 4 && LastLoadedElt == 1) {
8706 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8707 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8708 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8709 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8714 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8715 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8716 const X86Subtarget *Subtarget) {
8717 DebugLoc DL = N->getDebugLoc();
8718 SDValue Cond = N->getOperand(0);
8719 // Get the LHS/RHS of the select.
8720 SDValue LHS = N->getOperand(1);
8721 SDValue RHS = N->getOperand(2);
8723 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8724 // instructions have the peculiarity that if either operand is a NaN,
8725 // they chose what we call the RHS operand (and as such are not symmetric).
8726 // It happens that this matches the semantics of the common C idiom
8727 // x<y?x:y and related forms, so we can recognize these cases.
8728 if (Subtarget->hasSSE2() &&
8729 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8730 Cond.getOpcode() == ISD::SETCC) {
8731 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8733 unsigned Opcode = 0;
8734 // Check for x CC y ? x : y.
8735 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8739 // This can be a min if we can prove that at least one of the operands
8741 if (!FiniteOnlyFPMath()) {
8742 if (DAG.isKnownNeverNaN(RHS)) {
8743 // Put the potential NaN in the RHS so that SSE will preserve it.
8744 std::swap(LHS, RHS);
8745 } else if (!DAG.isKnownNeverNaN(LHS))
8748 Opcode = X86ISD::FMIN;
8751 // This can be a min if we can prove that at least one of the operands
8753 if (!FiniteOnlyFPMath()) {
8754 if (DAG.isKnownNeverNaN(LHS)) {
8755 // Put the potential NaN in the RHS so that SSE will preserve it.
8756 std::swap(LHS, RHS);
8757 } else if (!DAG.isKnownNeverNaN(RHS))
8760 Opcode = X86ISD::FMIN;
8763 // This can be a min, but if either operand is a NaN we need it to
8764 // preserve the original LHS.
8765 std::swap(LHS, RHS);
8769 Opcode = X86ISD::FMIN;
8773 // This can be a max if we can prove that at least one of the operands
8775 if (!FiniteOnlyFPMath()) {
8776 if (DAG.isKnownNeverNaN(LHS)) {
8777 // Put the potential NaN in the RHS so that SSE will preserve it.
8778 std::swap(LHS, RHS);
8779 } else if (!DAG.isKnownNeverNaN(RHS))
8782 Opcode = X86ISD::FMAX;
8785 // This can be a max if we can prove that at least one of the operands
8787 if (!FiniteOnlyFPMath()) {
8788 if (DAG.isKnownNeverNaN(RHS)) {
8789 // Put the potential NaN in the RHS so that SSE will preserve it.
8790 std::swap(LHS, RHS);
8791 } else if (!DAG.isKnownNeverNaN(LHS))
8794 Opcode = X86ISD::FMAX;
8797 // This can be a max, but if either operand is a NaN we need it to
8798 // preserve the original LHS.
8799 std::swap(LHS, RHS);
8803 Opcode = X86ISD::FMAX;
8806 // Check for x CC y ? y : x -- a min/max with reversed arms.
8807 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8811 // This can be a min if we can prove that at least one of the operands
8813 if (!FiniteOnlyFPMath()) {
8814 if (DAG.isKnownNeverNaN(RHS)) {
8815 // Put the potential NaN in the RHS so that SSE will preserve it.
8816 std::swap(LHS, RHS);
8817 } else if (!DAG.isKnownNeverNaN(LHS))
8820 Opcode = X86ISD::FMIN;
8823 // This can be a min if we can prove that at least one of the operands
8825 if (!FiniteOnlyFPMath()) {
8826 if (DAG.isKnownNeverNaN(LHS)) {
8827 // Put the potential NaN in the RHS so that SSE will preserve it.
8828 std::swap(LHS, RHS);
8829 } else if (!DAG.isKnownNeverNaN(RHS))
8832 Opcode = X86ISD::FMIN;
8835 // This can be a min, but if either operand is a NaN we need it to
8836 // preserve the original LHS.
8837 std::swap(LHS, RHS);
8841 Opcode = X86ISD::FMIN;
8845 // This can be a max if we can prove that at least one of the operands
8847 if (!FiniteOnlyFPMath()) {
8848 if (DAG.isKnownNeverNaN(LHS)) {
8849 // Put the potential NaN in the RHS so that SSE will preserve it.
8850 std::swap(LHS, RHS);
8851 } else if (!DAG.isKnownNeverNaN(RHS))
8854 Opcode = X86ISD::FMAX;
8857 // This can be a max if we can prove that at least one of the operands
8859 if (!FiniteOnlyFPMath()) {
8860 if (DAG.isKnownNeverNaN(RHS)) {
8861 // Put the potential NaN in the RHS so that SSE will preserve it.
8862 std::swap(LHS, RHS);
8863 } else if (!DAG.isKnownNeverNaN(LHS))
8866 Opcode = X86ISD::FMAX;
8869 // This can be a max, but if either operand is a NaN we need it to
8870 // preserve the original LHS.
8871 std::swap(LHS, RHS);
8875 Opcode = X86ISD::FMAX;
8881 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8884 // If this is a select between two integer constants, try to do some
8886 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8887 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8888 // Don't do this for crazy integer types.
8889 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8890 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8891 // so that TrueC (the true value) is larger than FalseC.
8892 bool NeedsCondInvert = false;
8894 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8895 // Efficiently invertible.
8896 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8897 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8898 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8899 NeedsCondInvert = true;
8900 std::swap(TrueC, FalseC);
8903 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8904 if (FalseC->getAPIntValue() == 0 &&
8905 TrueC->getAPIntValue().isPowerOf2()) {
8906 if (NeedsCondInvert) // Invert the condition if needed.
8907 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8908 DAG.getConstant(1, Cond.getValueType()));
8910 // Zero extend the condition if needed.
8911 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8913 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8914 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8915 DAG.getConstant(ShAmt, MVT::i8));
8918 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8919 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8920 if (NeedsCondInvert) // Invert the condition if needed.
8921 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8922 DAG.getConstant(1, Cond.getValueType()));
8924 // Zero extend the condition if needed.
8925 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8926 FalseC->getValueType(0), Cond);
8927 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8928 SDValue(FalseC, 0));
8931 // Optimize cases that will turn into an LEA instruction. This requires
8932 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8933 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8934 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8935 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8937 bool isFastMultiplier = false;
8939 switch ((unsigned char)Diff) {
8941 case 1: // result = add base, cond
8942 case 2: // result = lea base( , cond*2)
8943 case 3: // result = lea base(cond, cond*2)
8944 case 4: // result = lea base( , cond*4)
8945 case 5: // result = lea base(cond, cond*4)
8946 case 8: // result = lea base( , cond*8)
8947 case 9: // result = lea base(cond, cond*8)
8948 isFastMultiplier = true;
8953 if (isFastMultiplier) {
8954 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8955 if (NeedsCondInvert) // Invert the condition if needed.
8956 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8957 DAG.getConstant(1, Cond.getValueType()));
8959 // Zero extend the condition if needed.
8960 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8962 // Scale the condition by the difference.
8964 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8965 DAG.getConstant(Diff, Cond.getValueType()));
8967 // Add the base if non-zero.
8968 if (FalseC->getAPIntValue() != 0)
8969 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8970 SDValue(FalseC, 0));
8980 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8981 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8982 TargetLowering::DAGCombinerInfo &DCI) {
8983 DebugLoc DL = N->getDebugLoc();
8985 // If the flag operand isn't dead, don't touch this CMOV.
8986 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8989 // If this is a select between two integer constants, try to do some
8990 // optimizations. Note that the operands are ordered the opposite of SELECT
8992 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8993 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8994 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8995 // larger than FalseC (the false value).
8996 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8998 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8999 CC = X86::GetOppositeBranchCondition(CC);
9000 std::swap(TrueC, FalseC);
9003 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9004 // This is efficient for any integer data type (including i8/i16) and
9006 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9007 SDValue Cond = N->getOperand(3);
9008 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9009 DAG.getConstant(CC, MVT::i8), Cond);
9011 // Zero extend the condition if needed.
9012 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9014 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9015 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9016 DAG.getConstant(ShAmt, MVT::i8));
9017 if (N->getNumValues() == 2) // Dead flag value?
9018 return DCI.CombineTo(N, Cond, SDValue());
9022 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9023 // for any integer data type, including i8/i16.
9024 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9025 SDValue Cond = N->getOperand(3);
9026 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9027 DAG.getConstant(CC, MVT::i8), Cond);
9029 // Zero extend the condition if needed.
9030 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9031 FalseC->getValueType(0), Cond);
9032 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9033 SDValue(FalseC, 0));
9035 if (N->getNumValues() == 2) // Dead flag value?
9036 return DCI.CombineTo(N, Cond, SDValue());
9040 // Optimize cases that will turn into an LEA instruction. This requires
9041 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9042 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9043 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9044 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9046 bool isFastMultiplier = false;
9048 switch ((unsigned char)Diff) {
9050 case 1: // result = add base, cond
9051 case 2: // result = lea base( , cond*2)
9052 case 3: // result = lea base(cond, cond*2)
9053 case 4: // result = lea base( , cond*4)
9054 case 5: // result = lea base(cond, cond*4)
9055 case 8: // result = lea base( , cond*8)
9056 case 9: // result = lea base(cond, cond*8)
9057 isFastMultiplier = true;
9062 if (isFastMultiplier) {
9063 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9064 SDValue Cond = N->getOperand(3);
9065 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9066 DAG.getConstant(CC, MVT::i8), Cond);
9067 // Zero extend the condition if needed.
9068 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9070 // Scale the condition by the difference.
9072 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9073 DAG.getConstant(Diff, Cond.getValueType()));
9075 // Add the base if non-zero.
9076 if (FalseC->getAPIntValue() != 0)
9077 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9078 SDValue(FalseC, 0));
9079 if (N->getNumValues() == 2) // Dead flag value?
9080 return DCI.CombineTo(N, Cond, SDValue());
9090 /// PerformMulCombine - Optimize a single multiply with constant into two
9091 /// in order to implement it with two cheaper instructions, e.g.
9092 /// LEA + SHL, LEA + LEA.
9093 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9094 TargetLowering::DAGCombinerInfo &DCI) {
9095 if (DAG.getMachineFunction().
9096 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
9099 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9102 EVT VT = N->getValueType(0);
9106 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9109 uint64_t MulAmt = C->getZExtValue();
9110 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9113 uint64_t MulAmt1 = 0;
9114 uint64_t MulAmt2 = 0;
9115 if ((MulAmt % 9) == 0) {
9117 MulAmt2 = MulAmt / 9;
9118 } else if ((MulAmt % 5) == 0) {
9120 MulAmt2 = MulAmt / 5;
9121 } else if ((MulAmt % 3) == 0) {
9123 MulAmt2 = MulAmt / 3;
9126 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9127 DebugLoc DL = N->getDebugLoc();
9129 if (isPowerOf2_64(MulAmt2) &&
9130 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9131 // If second multiplifer is pow2, issue it first. We want the multiply by
9132 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9134 std::swap(MulAmt1, MulAmt2);
9137 if (isPowerOf2_64(MulAmt1))
9138 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9139 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9141 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9142 DAG.getConstant(MulAmt1, VT));
9144 if (isPowerOf2_64(MulAmt2))
9145 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9146 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9148 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9149 DAG.getConstant(MulAmt2, VT));
9151 // Do not add new nodes to DAG combiner worklist.
9152 DCI.CombineTo(N, NewMul, false);
9157 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9158 SDValue N0 = N->getOperand(0);
9159 SDValue N1 = N->getOperand(1);
9160 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9161 EVT VT = N0.getValueType();
9163 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9164 // since the result of setcc_c is all zero's or all ones.
9165 if (N1C && N0.getOpcode() == ISD::AND &&
9166 N0.getOperand(1).getOpcode() == ISD::Constant) {
9167 SDValue N00 = N0.getOperand(0);
9168 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9169 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9170 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9171 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9172 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9173 APInt ShAmt = N1C->getAPIntValue();
9174 Mask = Mask.shl(ShAmt);
9176 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9177 N00, DAG.getConstant(Mask, VT));
9184 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9186 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9187 const X86Subtarget *Subtarget) {
9188 EVT VT = N->getValueType(0);
9189 if (!VT.isVector() && VT.isInteger() &&
9190 N->getOpcode() == ISD::SHL)
9191 return PerformSHLCombine(N, DAG);
9193 // On X86 with SSE2 support, we can transform this to a vector shift if
9194 // all elements are shifted by the same amount. We can't do this in legalize
9195 // because the a constant vector is typically transformed to a constant pool
9196 // so we have no knowledge of the shift amount.
9197 if (!Subtarget->hasSSE2())
9200 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9203 SDValue ShAmtOp = N->getOperand(1);
9204 EVT EltVT = VT.getVectorElementType();
9205 DebugLoc DL = N->getDebugLoc();
9206 SDValue BaseShAmt = SDValue();
9207 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9208 unsigned NumElts = VT.getVectorNumElements();
9210 for (; i != NumElts; ++i) {
9211 SDValue Arg = ShAmtOp.getOperand(i);
9212 if (Arg.getOpcode() == ISD::UNDEF) continue;
9216 for (; i != NumElts; ++i) {
9217 SDValue Arg = ShAmtOp.getOperand(i);
9218 if (Arg.getOpcode() == ISD::UNDEF) continue;
9219 if (Arg != BaseShAmt) {
9223 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9224 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9225 SDValue InVec = ShAmtOp.getOperand(0);
9226 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9227 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9229 for (; i != NumElts; ++i) {
9230 SDValue Arg = InVec.getOperand(i);
9231 if (Arg.getOpcode() == ISD::UNDEF) continue;
9235 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9236 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9237 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9238 if (C->getZExtValue() == SplatIdx)
9239 BaseShAmt = InVec.getOperand(1);
9242 if (BaseShAmt.getNode() == 0)
9243 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9244 DAG.getIntPtrConstant(0));
9248 // The shift amount is an i32.
9249 if (EltVT.bitsGT(MVT::i32))
9250 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9251 else if (EltVT.bitsLT(MVT::i32))
9252 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9254 // The shift amount is identical so we can do a vector shift.
9255 SDValue ValOp = N->getOperand(0);
9256 switch (N->getOpcode()) {
9258 llvm_unreachable("Unknown shift opcode!");
9261 if (VT == MVT::v2i64)
9262 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9263 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9265 if (VT == MVT::v4i32)
9266 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9267 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9269 if (VT == MVT::v8i16)
9270 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9271 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9275 if (VT == MVT::v4i32)
9276 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9277 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9279 if (VT == MVT::v8i16)
9280 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9281 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9285 if (VT == MVT::v2i64)
9286 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9287 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9289 if (VT == MVT::v4i32)
9290 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9291 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9293 if (VT == MVT::v8i16)
9294 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9295 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9302 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9303 const X86Subtarget *Subtarget) {
9304 EVT VT = N->getValueType(0);
9305 if (VT != MVT::i64 || !Subtarget->is64Bit())
9308 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9309 SDValue N0 = N->getOperand(0);
9310 SDValue N1 = N->getOperand(1);
9311 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9313 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9316 SDValue ShAmt0 = N0.getOperand(1);
9317 if (ShAmt0.getValueType() != MVT::i8)
9319 SDValue ShAmt1 = N1.getOperand(1);
9320 if (ShAmt1.getValueType() != MVT::i8)
9322 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9323 ShAmt0 = ShAmt0.getOperand(0);
9324 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9325 ShAmt1 = ShAmt1.getOperand(0);
9327 DebugLoc DL = N->getDebugLoc();
9328 unsigned Opc = X86ISD::SHLD;
9329 SDValue Op0 = N0.getOperand(0);
9330 SDValue Op1 = N1.getOperand(0);
9331 if (ShAmt0.getOpcode() == ISD::SUB) {
9333 std::swap(Op0, Op1);
9334 std::swap(ShAmt0, ShAmt1);
9337 if (ShAmt1.getOpcode() == ISD::SUB) {
9338 SDValue Sum = ShAmt1.getOperand(0);
9339 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9340 if (SumC->getSExtValue() == 64 &&
9341 ShAmt1.getOperand(1) == ShAmt0)
9342 return DAG.getNode(Opc, DL, VT,
9344 DAG.getNode(ISD::TRUNCATE, DL,
9347 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9348 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9350 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9351 return DAG.getNode(Opc, DL, VT,
9352 N0.getOperand(0), N1.getOperand(0),
9353 DAG.getNode(ISD::TRUNCATE, DL,
9360 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9361 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9362 const X86Subtarget *Subtarget) {
9363 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9364 // the FP state in cases where an emms may be missing.
9365 // A preferable solution to the general problem is to figure out the right
9366 // places to insert EMMS. This qualifies as a quick hack.
9368 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9369 StoreSDNode *St = cast<StoreSDNode>(N);
9370 EVT VT = St->getValue().getValueType();
9371 if (VT.getSizeInBits() != 64)
9374 const Function *F = DAG.getMachineFunction().getFunction();
9375 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9376 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9377 && Subtarget->hasSSE2();
9378 if ((VT.isVector() ||
9379 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9380 isa<LoadSDNode>(St->getValue()) &&
9381 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9382 St->getChain().hasOneUse() && !St->isVolatile()) {
9383 SDNode* LdVal = St->getValue().getNode();
9385 int TokenFactorIndex = -1;
9386 SmallVector<SDValue, 8> Ops;
9387 SDNode* ChainVal = St->getChain().getNode();
9388 // Must be a store of a load. We currently handle two cases: the load
9389 // is a direct child, and it's under an intervening TokenFactor. It is
9390 // possible to dig deeper under nested TokenFactors.
9391 if (ChainVal == LdVal)
9392 Ld = cast<LoadSDNode>(St->getChain());
9393 else if (St->getValue().hasOneUse() &&
9394 ChainVal->getOpcode() == ISD::TokenFactor) {
9395 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9396 if (ChainVal->getOperand(i).getNode() == LdVal) {
9397 TokenFactorIndex = i;
9398 Ld = cast<LoadSDNode>(St->getValue());
9400 Ops.push_back(ChainVal->getOperand(i));
9404 if (!Ld || !ISD::isNormalLoad(Ld))
9407 // If this is not the MMX case, i.e. we are just turning i64 load/store
9408 // into f64 load/store, avoid the transformation if there are multiple
9409 // uses of the loaded value.
9410 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9413 DebugLoc LdDL = Ld->getDebugLoc();
9414 DebugLoc StDL = N->getDebugLoc();
9415 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9416 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9418 if (Subtarget->is64Bit() || F64IsLegal) {
9419 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9420 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9421 Ld->getBasePtr(), Ld->getSrcValue(),
9422 Ld->getSrcValueOffset(), Ld->isVolatile(),
9423 Ld->getAlignment());
9424 SDValue NewChain = NewLd.getValue(1);
9425 if (TokenFactorIndex != -1) {
9426 Ops.push_back(NewChain);
9427 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9430 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9431 St->getSrcValue(), St->getSrcValueOffset(),
9432 St->isVolatile(), St->getAlignment());
9435 // Otherwise, lower to two pairs of 32-bit loads / stores.
9436 SDValue LoAddr = Ld->getBasePtr();
9437 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9438 DAG.getConstant(4, MVT::i32));
9440 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9441 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9442 Ld->isVolatile(), Ld->getAlignment());
9443 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9444 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9446 MinAlign(Ld->getAlignment(), 4));
9448 SDValue NewChain = LoLd.getValue(1);
9449 if (TokenFactorIndex != -1) {
9450 Ops.push_back(LoLd);
9451 Ops.push_back(HiLd);
9452 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9456 LoAddr = St->getBasePtr();
9457 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9458 DAG.getConstant(4, MVT::i32));
9460 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9461 St->getSrcValue(), St->getSrcValueOffset(),
9462 St->isVolatile(), St->getAlignment());
9463 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9465 St->getSrcValueOffset() + 4,
9467 MinAlign(St->getAlignment(), 4));
9468 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9473 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9474 /// X86ISD::FXOR nodes.
9475 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9476 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9477 // F[X]OR(0.0, x) -> x
9478 // F[X]OR(x, 0.0) -> x
9479 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9480 if (C->getValueAPF().isPosZero())
9481 return N->getOperand(1);
9482 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9483 if (C->getValueAPF().isPosZero())
9484 return N->getOperand(0);
9488 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9489 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9490 // FAND(0.0, x) -> 0.0
9491 // FAND(x, 0.0) -> 0.0
9492 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9493 if (C->getValueAPF().isPosZero())
9494 return N->getOperand(0);
9495 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9496 if (C->getValueAPF().isPosZero())
9497 return N->getOperand(1);
9501 static SDValue PerformBTCombine(SDNode *N,
9503 TargetLowering::DAGCombinerInfo &DCI) {
9504 // BT ignores high bits in the bit index operand.
9505 SDValue Op1 = N->getOperand(1);
9506 if (Op1.hasOneUse()) {
9507 unsigned BitWidth = Op1.getValueSizeInBits();
9508 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9509 APInt KnownZero, KnownOne;
9510 TargetLowering::TargetLoweringOpt TLO(DAG);
9511 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9512 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9513 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9514 DCI.CommitTargetLoweringOpt(TLO);
9519 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9520 SDValue Op = N->getOperand(0);
9521 if (Op.getOpcode() == ISD::BIT_CONVERT)
9522 Op = Op.getOperand(0);
9523 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9524 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9525 VT.getVectorElementType().getSizeInBits() ==
9526 OpVT.getVectorElementType().getSizeInBits()) {
9527 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9532 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9533 // Locked instructions, in turn, have implicit fence semantics (all memory
9534 // operations are flushed before issuing the locked instruction, and the
9535 // are not buffered), so we can fold away the common pattern of
9536 // fence-atomic-fence.
9537 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9538 SDValue atomic = N->getOperand(0);
9539 switch (atomic.getOpcode()) {
9540 case ISD::ATOMIC_CMP_SWAP:
9541 case ISD::ATOMIC_SWAP:
9542 case ISD::ATOMIC_LOAD_ADD:
9543 case ISD::ATOMIC_LOAD_SUB:
9544 case ISD::ATOMIC_LOAD_AND:
9545 case ISD::ATOMIC_LOAD_OR:
9546 case ISD::ATOMIC_LOAD_XOR:
9547 case ISD::ATOMIC_LOAD_NAND:
9548 case ISD::ATOMIC_LOAD_MIN:
9549 case ISD::ATOMIC_LOAD_MAX:
9550 case ISD::ATOMIC_LOAD_UMIN:
9551 case ISD::ATOMIC_LOAD_UMAX:
9557 SDValue fence = atomic.getOperand(0);
9558 if (fence.getOpcode() != ISD::MEMBARRIER)
9561 switch (atomic.getOpcode()) {
9562 case ISD::ATOMIC_CMP_SWAP:
9563 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9564 atomic.getOperand(1), atomic.getOperand(2),
9565 atomic.getOperand(3));
9566 case ISD::ATOMIC_SWAP:
9567 case ISD::ATOMIC_LOAD_ADD:
9568 case ISD::ATOMIC_LOAD_SUB:
9569 case ISD::ATOMIC_LOAD_AND:
9570 case ISD::ATOMIC_LOAD_OR:
9571 case ISD::ATOMIC_LOAD_XOR:
9572 case ISD::ATOMIC_LOAD_NAND:
9573 case ISD::ATOMIC_LOAD_MIN:
9574 case ISD::ATOMIC_LOAD_MAX:
9575 case ISD::ATOMIC_LOAD_UMIN:
9576 case ISD::ATOMIC_LOAD_UMAX:
9577 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9578 atomic.getOperand(1), atomic.getOperand(2));
9584 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9585 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9586 // (and (i32 x86isd::setcc_carry), 1)
9587 // This eliminates the zext. This transformation is necessary because
9588 // ISD::SETCC is always legalized to i8.
9589 DebugLoc dl = N->getDebugLoc();
9590 SDValue N0 = N->getOperand(0);
9591 EVT VT = N->getValueType(0);
9592 if (N0.getOpcode() == ISD::AND &&
9594 N0.getOperand(0).hasOneUse()) {
9595 SDValue N00 = N0.getOperand(0);
9596 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9598 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9599 if (!C || C->getZExtValue() != 1)
9601 return DAG.getNode(ISD::AND, dl, VT,
9602 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9603 N00.getOperand(0), N00.getOperand(1)),
9604 DAG.getConstant(1, VT));
9610 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9611 DAGCombinerInfo &DCI) const {
9612 SelectionDAG &DAG = DCI.DAG;
9613 switch (N->getOpcode()) {
9615 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9616 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9617 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9618 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9621 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9622 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9623 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9625 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9626 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9627 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9628 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9629 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9630 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9636 //===----------------------------------------------------------------------===//
9637 // X86 Inline Assembly Support
9638 //===----------------------------------------------------------------------===//
9640 static bool LowerToBSwap(CallInst *CI) {
9641 // FIXME: this should verify that we are targetting a 486 or better. If not,
9642 // we will turn this bswap into something that will be lowered to logical ops
9643 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9644 // so don't worry about this.
9646 // Verify this is a simple bswap.
9647 if (CI->getNumOperands() != 2 ||
9648 CI->getType() != CI->getOperand(1)->getType() ||
9649 !CI->getType()->isInteger())
9652 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9653 if (!Ty || Ty->getBitWidth() % 16 != 0)
9656 // Okay, we can do this xform, do so now.
9657 const Type *Tys[] = { Ty };
9658 Module *M = CI->getParent()->getParent()->getParent();
9659 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9661 Value *Op = CI->getOperand(1);
9662 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9664 CI->replaceAllUsesWith(Op);
9665 CI->eraseFromParent();
9669 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9670 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9671 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9673 std::string AsmStr = IA->getAsmString();
9675 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9676 SmallVector<StringRef, 4> AsmPieces;
9677 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9679 switch (AsmPieces.size()) {
9680 default: return false;
9682 AsmStr = AsmPieces[0];
9684 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9687 if (AsmPieces.size() == 2 &&
9688 (AsmPieces[0] == "bswap" ||
9689 AsmPieces[0] == "bswapq" ||
9690 AsmPieces[0] == "bswapl") &&
9691 (AsmPieces[1] == "$0" ||
9692 AsmPieces[1] == "${0:q}")) {
9693 // No need to check constraints, nothing other than the equivalent of
9694 // "=r,0" would be valid here.
9695 return LowerToBSwap(CI);
9697 // rorw $$8, ${0:w} --> llvm.bswap.i16
9698 if (CI->getType()->isInteger(16) &&
9699 AsmPieces.size() == 3 &&
9700 AsmPieces[0] == "rorw" &&
9701 AsmPieces[1] == "$$8," &&
9702 AsmPieces[2] == "${0:w}" &&
9703 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9704 return LowerToBSwap(CI);
9708 if (CI->getType()->isInteger(64) &&
9709 Constraints.size() >= 2 &&
9710 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9711 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9712 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9713 SmallVector<StringRef, 4> Words;
9714 SplitString(AsmPieces[0], Words, " \t");
9715 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9717 SplitString(AsmPieces[1], Words, " \t");
9718 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9720 SplitString(AsmPieces[2], Words, " \t,");
9721 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9722 Words[2] == "%edx") {
9723 return LowerToBSwap(CI);
9735 /// getConstraintType - Given a constraint letter, return the type of
9736 /// constraint it is for this target.
9737 X86TargetLowering::ConstraintType
9738 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9739 if (Constraint.size() == 1) {
9740 switch (Constraint[0]) {
9752 return C_RegisterClass;
9760 return TargetLowering::getConstraintType(Constraint);
9763 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9764 /// with another that has more specific requirements based on the type of the
9765 /// corresponding operand.
9766 const char *X86TargetLowering::
9767 LowerXConstraint(EVT ConstraintVT) const {
9768 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9769 // 'f' like normal targets.
9770 if (ConstraintVT.isFloatingPoint()) {
9771 if (Subtarget->hasSSE2())
9773 if (Subtarget->hasSSE1())
9777 return TargetLowering::LowerXConstraint(ConstraintVT);
9780 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9781 /// vector. If it is invalid, don't add anything to Ops.
9782 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9785 std::vector<SDValue>&Ops,
9786 SelectionDAG &DAG) const {
9787 SDValue Result(0, 0);
9789 switch (Constraint) {
9792 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9793 if (C->getZExtValue() <= 31) {
9794 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9800 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9801 if (C->getZExtValue() <= 63) {
9802 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9808 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9809 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9810 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9816 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9817 if (C->getZExtValue() <= 255) {
9818 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9824 // 32-bit signed value
9825 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9826 const ConstantInt *CI = C->getConstantIntValue();
9827 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9828 C->getSExtValue())) {
9829 // Widen to 64 bits here to get it sign extended.
9830 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9833 // FIXME gcc accepts some relocatable values here too, but only in certain
9834 // memory models; it's complicated.
9839 // 32-bit unsigned value
9840 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9841 const ConstantInt *CI = C->getConstantIntValue();
9842 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9843 C->getZExtValue())) {
9844 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9848 // FIXME gcc accepts some relocatable values here too, but only in certain
9849 // memory models; it's complicated.
9853 // Literal immediates are always ok.
9854 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9855 // Widen to 64 bits here to get it sign extended.
9856 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9860 // If we are in non-pic codegen mode, we allow the address of a global (with
9861 // an optional displacement) to be used with 'i'.
9862 GlobalAddressSDNode *GA = 0;
9865 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9867 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9868 Offset += GA->getOffset();
9870 } else if (Op.getOpcode() == ISD::ADD) {
9871 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9872 Offset += C->getZExtValue();
9873 Op = Op.getOperand(0);
9876 } else if (Op.getOpcode() == ISD::SUB) {
9877 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9878 Offset += -C->getZExtValue();
9879 Op = Op.getOperand(0);
9884 // Otherwise, this isn't something we can handle, reject it.
9888 GlobalValue *GV = GA->getGlobal();
9889 // If we require an extra load to get this address, as in PIC mode, we
9891 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9892 getTargetMachine())))
9896 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9898 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9904 if (Result.getNode()) {
9905 Ops.push_back(Result);
9908 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9912 std::vector<unsigned> X86TargetLowering::
9913 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9915 if (Constraint.size() == 1) {
9916 // FIXME: not handling fp-stack yet!
9917 switch (Constraint[0]) { // GCC X86 Constraint Letters
9918 default: break; // Unknown constraint letter
9919 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9920 if (Subtarget->is64Bit()) {
9922 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9923 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9924 X86::R10D,X86::R11D,X86::R12D,
9925 X86::R13D,X86::R14D,X86::R15D,
9926 X86::EBP, X86::ESP, 0);
9927 else if (VT == MVT::i16)
9928 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9929 X86::SI, X86::DI, X86::R8W,X86::R9W,
9930 X86::R10W,X86::R11W,X86::R12W,
9931 X86::R13W,X86::R14W,X86::R15W,
9932 X86::BP, X86::SP, 0);
9933 else if (VT == MVT::i8)
9934 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9935 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9936 X86::R10B,X86::R11B,X86::R12B,
9937 X86::R13B,X86::R14B,X86::R15B,
9938 X86::BPL, X86::SPL, 0);
9940 else if (VT == MVT::i64)
9941 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9942 X86::RSI, X86::RDI, X86::R8, X86::R9,
9943 X86::R10, X86::R11, X86::R12,
9944 X86::R13, X86::R14, X86::R15,
9945 X86::RBP, X86::RSP, 0);
9949 // 32-bit fallthrough
9952 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9953 else if (VT == MVT::i16)
9954 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9955 else if (VT == MVT::i8)
9956 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9957 else if (VT == MVT::i64)
9958 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9963 return std::vector<unsigned>();
9966 std::pair<unsigned, const TargetRegisterClass*>
9967 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9969 // First, see if this is a constraint that directly corresponds to an LLVM
9971 if (Constraint.size() == 1) {
9972 // GCC Constraint Letters
9973 switch (Constraint[0]) {
9975 case 'r': // GENERAL_REGS
9976 case 'l': // INDEX_REGS
9978 return std::make_pair(0U, X86::GR8RegisterClass);
9980 return std::make_pair(0U, X86::GR16RegisterClass);
9981 if (VT == MVT::i32 || !Subtarget->is64Bit())
9982 return std::make_pair(0U, X86::GR32RegisterClass);
9983 return std::make_pair(0U, X86::GR64RegisterClass);
9984 case 'R': // LEGACY_REGS
9986 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
9988 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
9989 if (VT == MVT::i32 || !Subtarget->is64Bit())
9990 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
9991 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
9992 case 'f': // FP Stack registers.
9993 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
9994 // value to the correct fpstack register class.
9995 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9996 return std::make_pair(0U, X86::RFP32RegisterClass);
9997 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9998 return std::make_pair(0U, X86::RFP64RegisterClass);
9999 return std::make_pair(0U, X86::RFP80RegisterClass);
10000 case 'y': // MMX_REGS if MMX allowed.
10001 if (!Subtarget->hasMMX()) break;
10002 return std::make_pair(0U, X86::VR64RegisterClass);
10003 case 'Y': // SSE_REGS if SSE2 allowed
10004 if (!Subtarget->hasSSE2()) break;
10006 case 'x': // SSE_REGS if SSE1 allowed
10007 if (!Subtarget->hasSSE1()) break;
10009 switch (VT.getSimpleVT().SimpleTy) {
10011 // Scalar SSE types.
10014 return std::make_pair(0U, X86::FR32RegisterClass);
10017 return std::make_pair(0U, X86::FR64RegisterClass);
10025 return std::make_pair(0U, X86::VR128RegisterClass);
10031 // Use the default implementation in TargetLowering to convert the register
10032 // constraint into a member of a register class.
10033 std::pair<unsigned, const TargetRegisterClass*> Res;
10034 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10036 // Not found as a standard register?
10037 if (Res.second == 0) {
10038 // Map st(0) -> st(7) -> ST0
10039 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10040 tolower(Constraint[1]) == 's' &&
10041 tolower(Constraint[2]) == 't' &&
10042 Constraint[3] == '(' &&
10043 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10044 Constraint[5] == ')' &&
10045 Constraint[6] == '}') {
10047 Res.first = X86::ST0+Constraint[4]-'0';
10048 Res.second = X86::RFP80RegisterClass;
10052 // GCC allows "st(0)" to be called just plain "st".
10053 if (StringRef("{st}").equals_lower(Constraint)) {
10054 Res.first = X86::ST0;
10055 Res.second = X86::RFP80RegisterClass;
10060 if (StringRef("{flags}").equals_lower(Constraint)) {
10061 Res.first = X86::EFLAGS;
10062 Res.second = X86::CCRRegisterClass;
10066 // 'A' means EAX + EDX.
10067 if (Constraint == "A") {
10068 Res.first = X86::EAX;
10069 Res.second = X86::GR32_ADRegisterClass;
10075 // Otherwise, check to see if this is a register class of the wrong value
10076 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10077 // turn into {ax},{dx}.
10078 if (Res.second->hasType(VT))
10079 return Res; // Correct type already, nothing to do.
10081 // All of the single-register GCC register classes map their values onto
10082 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10083 // really want an 8-bit or 32-bit register, map to the appropriate register
10084 // class and return the appropriate register.
10085 if (Res.second == X86::GR16RegisterClass) {
10086 if (VT == MVT::i8) {
10087 unsigned DestReg = 0;
10088 switch (Res.first) {
10090 case X86::AX: DestReg = X86::AL; break;
10091 case X86::DX: DestReg = X86::DL; break;
10092 case X86::CX: DestReg = X86::CL; break;
10093 case X86::BX: DestReg = X86::BL; break;
10096 Res.first = DestReg;
10097 Res.second = X86::GR8RegisterClass;
10099 } else if (VT == MVT::i32) {
10100 unsigned DestReg = 0;
10101 switch (Res.first) {
10103 case X86::AX: DestReg = X86::EAX; break;
10104 case X86::DX: DestReg = X86::EDX; break;
10105 case X86::CX: DestReg = X86::ECX; break;
10106 case X86::BX: DestReg = X86::EBX; break;
10107 case X86::SI: DestReg = X86::ESI; break;
10108 case X86::DI: DestReg = X86::EDI; break;
10109 case X86::BP: DestReg = X86::EBP; break;
10110 case X86::SP: DestReg = X86::ESP; break;
10113 Res.first = DestReg;
10114 Res.second = X86::GR32RegisterClass;
10116 } else if (VT == MVT::i64) {
10117 unsigned DestReg = 0;
10118 switch (Res.first) {
10120 case X86::AX: DestReg = X86::RAX; break;
10121 case X86::DX: DestReg = X86::RDX; break;
10122 case X86::CX: DestReg = X86::RCX; break;
10123 case X86::BX: DestReg = X86::RBX; break;
10124 case X86::SI: DestReg = X86::RSI; break;
10125 case X86::DI: DestReg = X86::RDI; break;
10126 case X86::BP: DestReg = X86::RBP; break;
10127 case X86::SP: DestReg = X86::RSP; break;
10130 Res.first = DestReg;
10131 Res.second = X86::GR64RegisterClass;
10134 } else if (Res.second == X86::FR32RegisterClass ||
10135 Res.second == X86::FR64RegisterClass ||
10136 Res.second == X86::VR128RegisterClass) {
10137 // Handle references to XMM physical registers that got mapped into the
10138 // wrong class. This can happen with constraints like {xmm0} where the
10139 // target independent register mapper will just pick the first match it can
10140 // find, ignoring the required type.
10141 if (VT == MVT::f32)
10142 Res.second = X86::FR32RegisterClass;
10143 else if (VT == MVT::f64)
10144 Res.second = X86::FR64RegisterClass;
10145 else if (X86::VR128RegisterClass->hasType(VT))
10146 Res.second = X86::VR128RegisterClass;
10152 //===----------------------------------------------------------------------===//
10153 // X86 Widen vector type
10154 //===----------------------------------------------------------------------===//
10156 /// getWidenVectorType: given a vector type, returns the type to widen
10157 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10158 /// If there is no vector type that we want to widen to, returns MVT::Other
10159 /// When and where to widen is target dependent based on the cost of
10160 /// scalarizing vs using the wider vector type.
10162 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10163 assert(VT.isVector());
10164 if (isTypeLegal(VT))
10167 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10168 // type based on element type. This would speed up our search (though
10169 // it may not be worth it since the size of the list is relatively
10171 EVT EltVT = VT.getVectorElementType();
10172 unsigned NElts = VT.getVectorNumElements();
10174 // On X86, it make sense to widen any vector wider than 1
10178 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10179 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10180 EVT SVT = (MVT::SimpleValueType)nVT;
10182 if (isTypeLegal(SVT) &&
10183 SVT.getVectorElementType() == EltVT &&
10184 SVT.getVectorNumElements() > NElts)