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 if (CalleeCC == CallingConv::Fast &&
2257 CallerF->getCallingConv() == CalleeCC)
2262 // Do not tail call optimize vararg calls for now.
2266 // Look for obvious safe cases to perform tail call optimization.
2267 // If the callee takes no arguments then go on to check the results of the
2269 if (!Outs.empty()) {
2270 // Check if stack adjustment is needed. For now, do not do this if any
2271 // argument is passed on the stack.
2272 SmallVector<CCValAssign, 16> ArgLocs;
2273 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2274 ArgLocs, *DAG.getContext());
2275 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2276 if (CCInfo.getNextStackOffset())
2280 // If the caller does not return a value, then this is obviously safe.
2281 // This is one case where it's safe to perform this optimization even
2282 // if the return types do not match.
2283 const Type *CallerRetTy = CallerF->getReturnType();
2284 if (CallerRetTy->isVoidTy())
2287 // If the return types match, then it's safe.
2288 // Don't tail call optimize recursive call.
2289 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2290 if (!G) return false; // FIXME: common external symbols?
2291 const Function *CalleeF = cast<Function>(G->getGlobal());
2292 const Type *CalleeRetTy = CalleeF->getReturnType();
2293 return CallerRetTy == CalleeRetTy;
2297 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2299 DenseMap<const Value *, unsigned> &vm,
2300 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2301 DenseMap<const AllocaInst *, int> &am
2303 , SmallSet<Instruction*, 8> &cil
2306 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2314 //===----------------------------------------------------------------------===//
2315 // Other Lowering Hooks
2316 //===----------------------------------------------------------------------===//
2319 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2320 MachineFunction &MF = DAG.getMachineFunction();
2321 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2322 int ReturnAddrIndex = FuncInfo->getRAIndex();
2324 if (ReturnAddrIndex == 0) {
2325 // Set up a frame object for the return address.
2326 uint64_t SlotSize = TD->getPointerSize();
2327 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2329 FuncInfo->setRAIndex(ReturnAddrIndex);
2332 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2336 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2337 bool hasSymbolicDisplacement) {
2338 // Offset should fit into 32 bit immediate field.
2339 if (!isInt32(Offset))
2342 // If we don't have a symbolic displacement - we don't have any extra
2344 if (!hasSymbolicDisplacement)
2347 // FIXME: Some tweaks might be needed for medium code model.
2348 if (M != CodeModel::Small && M != CodeModel::Kernel)
2351 // For small code model we assume that latest object is 16MB before end of 31
2352 // bits boundary. We may also accept pretty large negative constants knowing
2353 // that all objects are in the positive half of address space.
2354 if (M == CodeModel::Small && Offset < 16*1024*1024)
2357 // For kernel code model we know that all object resist in the negative half
2358 // of 32bits address space. We may not accept negative offsets, since they may
2359 // be just off and we may accept pretty large positive ones.
2360 if (M == CodeModel::Kernel && Offset > 0)
2366 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2367 /// specific condition code, returning the condition code and the LHS/RHS of the
2368 /// comparison to make.
2369 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2370 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2372 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2373 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2374 // X > -1 -> X == 0, jump !sign.
2375 RHS = DAG.getConstant(0, RHS.getValueType());
2376 return X86::COND_NS;
2377 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2378 // X < 0 -> X == 0, jump on sign.
2380 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2382 RHS = DAG.getConstant(0, RHS.getValueType());
2383 return X86::COND_LE;
2387 switch (SetCCOpcode) {
2388 default: llvm_unreachable("Invalid integer condition!");
2389 case ISD::SETEQ: return X86::COND_E;
2390 case ISD::SETGT: return X86::COND_G;
2391 case ISD::SETGE: return X86::COND_GE;
2392 case ISD::SETLT: return X86::COND_L;
2393 case ISD::SETLE: return X86::COND_LE;
2394 case ISD::SETNE: return X86::COND_NE;
2395 case ISD::SETULT: return X86::COND_B;
2396 case ISD::SETUGT: return X86::COND_A;
2397 case ISD::SETULE: return X86::COND_BE;
2398 case ISD::SETUGE: return X86::COND_AE;
2402 // First determine if it is required or is profitable to flip the operands.
2404 // If LHS is a foldable load, but RHS is not, flip the condition.
2405 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2406 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2407 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2408 std::swap(LHS, RHS);
2411 switch (SetCCOpcode) {
2417 std::swap(LHS, RHS);
2421 // On a floating point condition, the flags are set as follows:
2423 // 0 | 0 | 0 | X > Y
2424 // 0 | 0 | 1 | X < Y
2425 // 1 | 0 | 0 | X == Y
2426 // 1 | 1 | 1 | unordered
2427 switch (SetCCOpcode) {
2428 default: llvm_unreachable("Condcode should be pre-legalized away");
2430 case ISD::SETEQ: return X86::COND_E;
2431 case ISD::SETOLT: // flipped
2433 case ISD::SETGT: return X86::COND_A;
2434 case ISD::SETOLE: // flipped
2436 case ISD::SETGE: return X86::COND_AE;
2437 case ISD::SETUGT: // flipped
2439 case ISD::SETLT: return X86::COND_B;
2440 case ISD::SETUGE: // flipped
2442 case ISD::SETLE: return X86::COND_BE;
2444 case ISD::SETNE: return X86::COND_NE;
2445 case ISD::SETUO: return X86::COND_P;
2446 case ISD::SETO: return X86::COND_NP;
2448 case ISD::SETUNE: return X86::COND_INVALID;
2452 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2453 /// code. Current x86 isa includes the following FP cmov instructions:
2454 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2455 static bool hasFPCMov(unsigned X86CC) {
2471 /// isFPImmLegal - Returns true if the target can instruction select the
2472 /// specified FP immediate natively. If false, the legalizer will
2473 /// materialize the FP immediate as a load from a constant pool.
2474 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2475 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2476 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2482 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2483 /// the specified range (L, H].
2484 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2485 return (Val < 0) || (Val >= Low && Val < Hi);
2488 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2489 /// specified value.
2490 static bool isUndefOrEqual(int Val, int CmpVal) {
2491 if (Val < 0 || Val == CmpVal)
2496 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2497 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2498 /// the second operand.
2499 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2500 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2501 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2502 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2503 return (Mask[0] < 2 && Mask[1] < 2);
2507 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2508 SmallVector<int, 8> M;
2510 return ::isPSHUFDMask(M, N->getValueType(0));
2513 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2514 /// is suitable for input to PSHUFHW.
2515 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2516 if (VT != MVT::v8i16)
2519 // Lower quadword copied in order or undef.
2520 for (int i = 0; i != 4; ++i)
2521 if (Mask[i] >= 0 && Mask[i] != i)
2524 // Upper quadword shuffled.
2525 for (int i = 4; i != 8; ++i)
2526 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2532 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2533 SmallVector<int, 8> M;
2535 return ::isPSHUFHWMask(M, N->getValueType(0));
2538 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2539 /// is suitable for input to PSHUFLW.
2540 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2541 if (VT != MVT::v8i16)
2544 // Upper quadword copied in order.
2545 for (int i = 4; i != 8; ++i)
2546 if (Mask[i] >= 0 && Mask[i] != i)
2549 // Lower quadword shuffled.
2550 for (int i = 0; i != 4; ++i)
2557 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2558 SmallVector<int, 8> M;
2560 return ::isPSHUFLWMask(M, N->getValueType(0));
2563 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2564 /// is suitable for input to PALIGNR.
2565 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2567 int i, e = VT.getVectorNumElements();
2569 // Do not handle v2i64 / v2f64 shuffles with palignr.
2570 if (e < 4 || !hasSSSE3)
2573 for (i = 0; i != e; ++i)
2577 // All undef, not a palignr.
2581 // Determine if it's ok to perform a palignr with only the LHS, since we
2582 // don't have access to the actual shuffle elements to see if RHS is undef.
2583 bool Unary = Mask[i] < (int)e;
2584 bool NeedsUnary = false;
2586 int s = Mask[i] - i;
2588 // Check the rest of the elements to see if they are consecutive.
2589 for (++i; i != e; ++i) {
2594 Unary = Unary && (m < (int)e);
2595 NeedsUnary = NeedsUnary || (m < s);
2597 if (NeedsUnary && !Unary)
2599 if (Unary && m != ((s+i) & (e-1)))
2601 if (!Unary && m != (s+i))
2607 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2608 SmallVector<int, 8> M;
2610 return ::isPALIGNRMask(M, N->getValueType(0), true);
2613 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2614 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2615 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2616 int NumElems = VT.getVectorNumElements();
2617 if (NumElems != 2 && NumElems != 4)
2620 int Half = NumElems / 2;
2621 for (int i = 0; i < Half; ++i)
2622 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2624 for (int i = Half; i < NumElems; ++i)
2625 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2631 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2632 SmallVector<int, 8> M;
2634 return ::isSHUFPMask(M, N->getValueType(0));
2637 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2638 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2639 /// half elements to come from vector 1 (which would equal the dest.) and
2640 /// the upper half to come from vector 2.
2641 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2642 int NumElems = VT.getVectorNumElements();
2644 if (NumElems != 2 && NumElems != 4)
2647 int Half = NumElems / 2;
2648 for (int i = 0; i < Half; ++i)
2649 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2651 for (int i = Half; i < NumElems; ++i)
2652 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2657 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2658 SmallVector<int, 8> M;
2660 return isCommutedSHUFPMask(M, N->getValueType(0));
2663 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2664 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2665 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2666 if (N->getValueType(0).getVectorNumElements() != 4)
2669 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2670 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2671 isUndefOrEqual(N->getMaskElt(1), 7) &&
2672 isUndefOrEqual(N->getMaskElt(2), 2) &&
2673 isUndefOrEqual(N->getMaskElt(3), 3);
2676 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2677 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2679 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2680 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2685 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2686 isUndefOrEqual(N->getMaskElt(1), 3) &&
2687 isUndefOrEqual(N->getMaskElt(2), 2) &&
2688 isUndefOrEqual(N->getMaskElt(3), 3);
2691 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2692 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2693 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2694 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2696 if (NumElems != 2 && NumElems != 4)
2699 for (unsigned i = 0; i < NumElems/2; ++i)
2700 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2703 for (unsigned i = NumElems/2; i < NumElems; ++i)
2704 if (!isUndefOrEqual(N->getMaskElt(i), i))
2710 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2711 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2712 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2713 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2715 if (NumElems != 2 && NumElems != 4)
2718 for (unsigned i = 0; i < NumElems/2; ++i)
2719 if (!isUndefOrEqual(N->getMaskElt(i), i))
2722 for (unsigned i = 0; i < NumElems/2; ++i)
2723 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2729 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2730 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2731 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2732 bool V2IsSplat = false) {
2733 int NumElts = VT.getVectorNumElements();
2734 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2737 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2739 int BitI1 = Mask[i+1];
2740 if (!isUndefOrEqual(BitI, j))
2743 if (!isUndefOrEqual(BitI1, NumElts))
2746 if (!isUndefOrEqual(BitI1, j + NumElts))
2753 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2754 SmallVector<int, 8> M;
2756 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2759 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2760 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2761 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2762 bool V2IsSplat = false) {
2763 int NumElts = VT.getVectorNumElements();
2764 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2767 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2769 int BitI1 = Mask[i+1];
2770 if (!isUndefOrEqual(BitI, j + NumElts/2))
2773 if (isUndefOrEqual(BitI1, NumElts))
2776 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2783 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2784 SmallVector<int, 8> M;
2786 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2789 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2790 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2792 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2793 int NumElems = VT.getVectorNumElements();
2794 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2797 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2799 int BitI1 = Mask[i+1];
2800 if (!isUndefOrEqual(BitI, j))
2802 if (!isUndefOrEqual(BitI1, j))
2808 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2809 SmallVector<int, 8> M;
2811 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2814 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2815 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2817 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2818 int NumElems = VT.getVectorNumElements();
2819 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2822 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2824 int BitI1 = Mask[i+1];
2825 if (!isUndefOrEqual(BitI, j))
2827 if (!isUndefOrEqual(BitI1, j))
2833 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2834 SmallVector<int, 8> M;
2836 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2839 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2840 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2841 /// MOVSD, and MOVD, i.e. setting the lowest element.
2842 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2843 if (VT.getVectorElementType().getSizeInBits() < 32)
2846 int NumElts = VT.getVectorNumElements();
2848 if (!isUndefOrEqual(Mask[0], NumElts))
2851 for (int i = 1; i < NumElts; ++i)
2852 if (!isUndefOrEqual(Mask[i], i))
2858 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2859 SmallVector<int, 8> M;
2861 return ::isMOVLMask(M, N->getValueType(0));
2864 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2865 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2866 /// element of vector 2 and the other elements to come from vector 1 in order.
2867 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2868 bool V2IsSplat = false, bool V2IsUndef = false) {
2869 int NumOps = VT.getVectorNumElements();
2870 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2873 if (!isUndefOrEqual(Mask[0], 0))
2876 for (int i = 1; i < NumOps; ++i)
2877 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2878 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2879 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2885 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2886 bool V2IsUndef = false) {
2887 SmallVector<int, 8> M;
2889 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2892 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2893 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2894 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2895 if (N->getValueType(0).getVectorNumElements() != 4)
2898 // Expect 1, 1, 3, 3
2899 for (unsigned i = 0; i < 2; ++i) {
2900 int Elt = N->getMaskElt(i);
2901 if (Elt >= 0 && Elt != 1)
2906 for (unsigned i = 2; i < 4; ++i) {
2907 int Elt = N->getMaskElt(i);
2908 if (Elt >= 0 && Elt != 3)
2913 // Don't use movshdup if it can be done with a shufps.
2914 // FIXME: verify that matching u, u, 3, 3 is what we want.
2918 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2919 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2920 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2921 if (N->getValueType(0).getVectorNumElements() != 4)
2924 // Expect 0, 0, 2, 2
2925 for (unsigned i = 0; i < 2; ++i)
2926 if (N->getMaskElt(i) > 0)
2930 for (unsigned i = 2; i < 4; ++i) {
2931 int Elt = N->getMaskElt(i);
2932 if (Elt >= 0 && Elt != 2)
2937 // Don't use movsldup if it can be done with a shufps.
2941 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2942 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2943 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2944 int e = N->getValueType(0).getVectorNumElements() / 2;
2946 for (int i = 0; i < e; ++i)
2947 if (!isUndefOrEqual(N->getMaskElt(i), i))
2949 for (int i = 0; i < e; ++i)
2950 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2955 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2956 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
2957 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2958 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2959 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2961 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2963 for (int i = 0; i < NumOperands; ++i) {
2964 int Val = SVOp->getMaskElt(NumOperands-i-1);
2965 if (Val < 0) Val = 0;
2966 if (Val >= NumOperands) Val -= NumOperands;
2968 if (i != NumOperands - 1)
2974 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2975 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
2976 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2977 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2979 // 8 nodes, but we only care about the last 4.
2980 for (unsigned i = 7; i >= 4; --i) {
2981 int Val = SVOp->getMaskElt(i);
2990 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2991 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
2992 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2993 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2995 // 8 nodes, but we only care about the first 4.
2996 for (int i = 3; i >= 0; --i) {
2997 int Val = SVOp->getMaskElt(i);
3006 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3007 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3008 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3009 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3010 EVT VVT = N->getValueType(0);
3011 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3015 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3016 Val = SVOp->getMaskElt(i);
3020 return (Val - i) * EltSize;
3023 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3025 bool X86::isZeroNode(SDValue Elt) {
3026 return ((isa<ConstantSDNode>(Elt) &&
3027 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3028 (isa<ConstantFPSDNode>(Elt) &&
3029 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3032 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3033 /// their permute mask.
3034 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3035 SelectionDAG &DAG) {
3036 EVT VT = SVOp->getValueType(0);
3037 unsigned NumElems = VT.getVectorNumElements();
3038 SmallVector<int, 8> MaskVec;
3040 for (unsigned i = 0; i != NumElems; ++i) {
3041 int idx = SVOp->getMaskElt(i);
3043 MaskVec.push_back(idx);
3044 else if (idx < (int)NumElems)
3045 MaskVec.push_back(idx + NumElems);
3047 MaskVec.push_back(idx - NumElems);
3049 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3050 SVOp->getOperand(0), &MaskVec[0]);
3053 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3054 /// the two vector operands have swapped position.
3055 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3056 unsigned NumElems = VT.getVectorNumElements();
3057 for (unsigned i = 0; i != NumElems; ++i) {
3061 else if (idx < (int)NumElems)
3062 Mask[i] = idx + NumElems;
3064 Mask[i] = idx - NumElems;
3068 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3069 /// match movhlps. The lower half elements should come from upper half of
3070 /// V1 (and in order), and the upper half elements should come from the upper
3071 /// half of V2 (and in order).
3072 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3073 if (Op->getValueType(0).getVectorNumElements() != 4)
3075 for (unsigned i = 0, e = 2; i != e; ++i)
3076 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3078 for (unsigned i = 2; i != 4; ++i)
3079 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3084 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3085 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3087 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3088 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3090 N = N->getOperand(0).getNode();
3091 if (!ISD::isNON_EXTLoad(N))
3094 *LD = cast<LoadSDNode>(N);
3098 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3099 /// match movlp{s|d}. The lower half elements should come from lower half of
3100 /// V1 (and in order), and the upper half elements should come from the upper
3101 /// half of V2 (and in order). And since V1 will become the source of the
3102 /// MOVLP, it must be either a vector load or a scalar load to vector.
3103 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3104 ShuffleVectorSDNode *Op) {
3105 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3107 // Is V2 is a vector load, don't do this transformation. We will try to use
3108 // load folding shufps op.
3109 if (ISD::isNON_EXTLoad(V2))
3112 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3114 if (NumElems != 2 && NumElems != 4)
3116 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3117 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3119 for (unsigned i = NumElems/2; i != NumElems; ++i)
3120 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3125 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3127 static bool isSplatVector(SDNode *N) {
3128 if (N->getOpcode() != ISD::BUILD_VECTOR)
3131 SDValue SplatValue = N->getOperand(0);
3132 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3133 if (N->getOperand(i) != SplatValue)
3138 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3139 /// to an zero vector.
3140 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3141 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3142 SDValue V1 = N->getOperand(0);
3143 SDValue V2 = N->getOperand(1);
3144 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3145 for (unsigned i = 0; i != NumElems; ++i) {
3146 int Idx = N->getMaskElt(i);
3147 if (Idx >= (int)NumElems) {
3148 unsigned Opc = V2.getOpcode();
3149 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3151 if (Opc != ISD::BUILD_VECTOR ||
3152 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3154 } else if (Idx >= 0) {
3155 unsigned Opc = V1.getOpcode();
3156 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3158 if (Opc != ISD::BUILD_VECTOR ||
3159 !X86::isZeroNode(V1.getOperand(Idx)))
3166 /// getZeroVector - Returns a vector of specified type with all zero elements.
3168 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3170 assert(VT.isVector() && "Expected a vector type");
3172 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3173 // type. This ensures they get CSE'd.
3175 if (VT.getSizeInBits() == 64) { // MMX
3176 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3177 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3178 } else if (HasSSE2) { // SSE2
3179 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3180 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3182 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3183 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3185 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3188 /// getOnesVector - Returns a vector of specified type with all bits set.
3190 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3191 assert(VT.isVector() && "Expected a vector type");
3193 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3194 // type. This ensures they get CSE'd.
3195 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3197 if (VT.getSizeInBits() == 64) // MMX
3198 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3200 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3201 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3205 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3206 /// that point to V2 points to its first element.
3207 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3208 EVT VT = SVOp->getValueType(0);
3209 unsigned NumElems = VT.getVectorNumElements();
3211 bool Changed = false;
3212 SmallVector<int, 8> MaskVec;
3213 SVOp->getMask(MaskVec);
3215 for (unsigned i = 0; i != NumElems; ++i) {
3216 if (MaskVec[i] > (int)NumElems) {
3217 MaskVec[i] = NumElems;
3222 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3223 SVOp->getOperand(1), &MaskVec[0]);
3224 return SDValue(SVOp, 0);
3227 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3228 /// operation of specified width.
3229 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3231 unsigned NumElems = VT.getVectorNumElements();
3232 SmallVector<int, 8> Mask;
3233 Mask.push_back(NumElems);
3234 for (unsigned i = 1; i != NumElems; ++i)
3236 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3239 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3240 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3242 unsigned NumElems = VT.getVectorNumElements();
3243 SmallVector<int, 8> Mask;
3244 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3246 Mask.push_back(i + NumElems);
3248 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3251 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3252 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3254 unsigned NumElems = VT.getVectorNumElements();
3255 unsigned Half = NumElems/2;
3256 SmallVector<int, 8> Mask;
3257 for (unsigned i = 0; i != Half; ++i) {
3258 Mask.push_back(i + Half);
3259 Mask.push_back(i + NumElems + Half);
3261 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3264 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3265 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3267 if (SV->getValueType(0).getVectorNumElements() <= 4)
3268 return SDValue(SV, 0);
3270 EVT PVT = MVT::v4f32;
3271 EVT VT = SV->getValueType(0);
3272 DebugLoc dl = SV->getDebugLoc();
3273 SDValue V1 = SV->getOperand(0);
3274 int NumElems = VT.getVectorNumElements();
3275 int EltNo = SV->getSplatIndex();
3277 // unpack elements to the correct location
3278 while (NumElems > 4) {
3279 if (EltNo < NumElems/2) {
3280 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3282 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3283 EltNo -= NumElems/2;
3288 // Perform the splat.
3289 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3290 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3291 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3292 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3295 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3296 /// vector of zero or undef vector. This produces a shuffle where the low
3297 /// element of V2 is swizzled into the zero/undef vector, landing at element
3298 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3299 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3300 bool isZero, bool HasSSE2,
3301 SelectionDAG &DAG) {
3302 EVT VT = V2.getValueType();
3304 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3305 unsigned NumElems = VT.getVectorNumElements();
3306 SmallVector<int, 16> MaskVec;
3307 for (unsigned i = 0; i != NumElems; ++i)
3308 // If this is the insertion idx, put the low elt of V2 here.
3309 MaskVec.push_back(i == Idx ? NumElems : i);
3310 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3313 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3314 /// a shuffle that is zero.
3316 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3317 bool Low, SelectionDAG &DAG) {
3318 unsigned NumZeros = 0;
3319 for (int i = 0; i < NumElems; ++i) {
3320 unsigned Index = Low ? i : NumElems-i-1;
3321 int Idx = SVOp->getMaskElt(Index);
3326 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3327 if (Elt.getNode() && X86::isZeroNode(Elt))
3335 /// isVectorShift - Returns true if the shuffle can be implemented as a
3336 /// logical left or right shift of a vector.
3337 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3338 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3339 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3340 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3343 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3346 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3350 bool SeenV1 = false;
3351 bool SeenV2 = false;
3352 for (int i = NumZeros; i < NumElems; ++i) {
3353 int Val = isLeft ? (i - NumZeros) : i;
3354 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3366 if (SeenV1 && SeenV2)
3369 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3375 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3377 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3378 unsigned NumNonZero, unsigned NumZero,
3379 SelectionDAG &DAG, TargetLowering &TLI) {
3383 DebugLoc dl = Op.getDebugLoc();
3386 for (unsigned i = 0; i < 16; ++i) {
3387 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3388 if (ThisIsNonZero && First) {
3390 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3392 V = DAG.getUNDEF(MVT::v8i16);
3397 SDValue ThisElt(0, 0), LastElt(0, 0);
3398 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3399 if (LastIsNonZero) {
3400 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3401 MVT::i16, Op.getOperand(i-1));
3403 if (ThisIsNonZero) {
3404 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3405 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3406 ThisElt, DAG.getConstant(8, MVT::i8));
3408 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3412 if (ThisElt.getNode())
3413 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3414 DAG.getIntPtrConstant(i/2));
3418 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3421 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3423 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3424 unsigned NumNonZero, unsigned NumZero,
3425 SelectionDAG &DAG, TargetLowering &TLI) {
3429 DebugLoc dl = Op.getDebugLoc();
3432 for (unsigned i = 0; i < 8; ++i) {
3433 bool isNonZero = (NonZeros & (1 << i)) != 0;
3437 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3439 V = DAG.getUNDEF(MVT::v8i16);
3442 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3443 MVT::v8i16, V, Op.getOperand(i),
3444 DAG.getIntPtrConstant(i));
3451 /// getVShift - Return a vector logical shift node.
3453 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3454 unsigned NumBits, SelectionDAG &DAG,
3455 const TargetLowering &TLI, DebugLoc dl) {
3456 bool isMMX = VT.getSizeInBits() == 64;
3457 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3458 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3459 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3460 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3461 DAG.getNode(Opc, dl, ShVT, SrcOp,
3462 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3466 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3467 SelectionDAG &DAG) {
3469 // Check if the scalar load can be widened into a vector load. And if
3470 // the address is "base + cst" see if the cst can be "absorbed" into
3471 // the shuffle mask.
3472 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3473 SDValue Ptr = LD->getBasePtr();
3474 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3476 EVT PVT = LD->getValueType(0);
3477 if (PVT != MVT::i32 && PVT != MVT::f32)
3482 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3483 FI = FINode->getIndex();
3485 } else if (Ptr.getOpcode() == ISD::ADD &&
3486 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3487 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3488 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3489 Offset = Ptr.getConstantOperandVal(1);
3490 Ptr = Ptr.getOperand(0);
3495 SDValue Chain = LD->getChain();
3496 // Make sure the stack object alignment is at least 16.
3497 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3498 if (DAG.InferPtrAlignment(Ptr) < 16) {
3499 if (MFI->isFixedObjectIndex(FI)) {
3500 // Can't change the alignment. FIXME: It's possible to compute
3501 // the exact stack offset and reference FI + adjust offset instead.
3502 // If someone *really* cares about this. That's the way to implement it.
3505 MFI->setObjectAlignment(FI, 16);
3509 // (Offset % 16) must be multiple of 4. Then address is then
3510 // Ptr + (Offset & ~15).
3513 if ((Offset % 16) & 3)
3515 int64_t StartOffset = Offset & ~15;
3517 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3518 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3520 int EltNo = (Offset - StartOffset) >> 2;
3521 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3522 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3523 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0);
3524 // Canonicalize it to a v4i32 shuffle.
3525 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3526 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3527 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3528 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3535 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3536 DebugLoc dl = Op.getDebugLoc();
3537 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3538 if (ISD::isBuildVectorAllZeros(Op.getNode())
3539 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3540 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3541 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3542 // eliminated on x86-32 hosts.
3543 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3546 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3547 return getOnesVector(Op.getValueType(), DAG, dl);
3548 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3551 EVT VT = Op.getValueType();
3552 EVT ExtVT = VT.getVectorElementType();
3553 unsigned EVTBits = ExtVT.getSizeInBits();
3555 unsigned NumElems = Op.getNumOperands();
3556 unsigned NumZero = 0;
3557 unsigned NumNonZero = 0;
3558 unsigned NonZeros = 0;
3559 bool IsAllConstants = true;
3560 SmallSet<SDValue, 8> Values;
3561 for (unsigned i = 0; i < NumElems; ++i) {
3562 SDValue Elt = Op.getOperand(i);
3563 if (Elt.getOpcode() == ISD::UNDEF)
3566 if (Elt.getOpcode() != ISD::Constant &&
3567 Elt.getOpcode() != ISD::ConstantFP)
3568 IsAllConstants = false;
3569 if (X86::isZeroNode(Elt))
3572 NonZeros |= (1 << i);
3577 if (NumNonZero == 0) {
3578 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3579 return DAG.getUNDEF(VT);
3582 // Special case for single non-zero, non-undef, element.
3583 if (NumNonZero == 1) {
3584 unsigned Idx = CountTrailingZeros_32(NonZeros);
3585 SDValue Item = Op.getOperand(Idx);
3587 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3588 // the value are obviously zero, truncate the value to i32 and do the
3589 // insertion that way. Only do this if the value is non-constant or if the
3590 // value is a constant being inserted into element 0. It is cheaper to do
3591 // a constant pool load than it is to do a movd + shuffle.
3592 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3593 (!IsAllConstants || Idx == 0)) {
3594 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3595 // Handle MMX and SSE both.
3596 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3597 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3599 // Truncate the value (which may itself be a constant) to i32, and
3600 // convert it to a vector with movd (S2V+shuffle to zero extend).
3601 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3602 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3603 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3604 Subtarget->hasSSE2(), DAG);
3606 // Now we have our 32-bit value zero extended in the low element of
3607 // a vector. If Idx != 0, swizzle it into place.
3609 SmallVector<int, 4> Mask;
3610 Mask.push_back(Idx);
3611 for (unsigned i = 1; i != VecElts; ++i)
3613 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3614 DAG.getUNDEF(Item.getValueType()),
3617 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3621 // If we have a constant or non-constant insertion into the low element of
3622 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3623 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3624 // depending on what the source datatype is.
3627 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3628 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3629 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3630 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3631 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3632 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3634 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3635 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3636 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3637 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3638 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3639 Subtarget->hasSSE2(), DAG);
3640 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3644 // Is it a vector logical left shift?
3645 if (NumElems == 2 && Idx == 1 &&
3646 X86::isZeroNode(Op.getOperand(0)) &&
3647 !X86::isZeroNode(Op.getOperand(1))) {
3648 unsigned NumBits = VT.getSizeInBits();
3649 return getVShift(true, VT,
3650 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3651 VT, Op.getOperand(1)),
3652 NumBits/2, DAG, *this, dl);
3655 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3658 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3659 // is a non-constant being inserted into an element other than the low one,
3660 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3661 // movd/movss) to move this into the low element, then shuffle it into
3663 if (EVTBits == 32) {
3664 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3666 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3667 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3668 Subtarget->hasSSE2(), DAG);
3669 SmallVector<int, 8> MaskVec;
3670 for (unsigned i = 0; i < NumElems; i++)
3671 MaskVec.push_back(i == Idx ? 0 : 1);
3672 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3676 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3677 if (Values.size() == 1) {
3678 if (EVTBits == 32) {
3679 // Instead of a shuffle like this:
3680 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3681 // Check if it's possible to issue this instead.
3682 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3683 unsigned Idx = CountTrailingZeros_32(NonZeros);
3684 SDValue Item = Op.getOperand(Idx);
3685 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3686 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3691 // A vector full of immediates; various special cases are already
3692 // handled, so this is best done with a single constant-pool load.
3696 // Let legalizer expand 2-wide build_vectors.
3697 if (EVTBits == 64) {
3698 if (NumNonZero == 1) {
3699 // One half is zero or undef.
3700 unsigned Idx = CountTrailingZeros_32(NonZeros);
3701 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3702 Op.getOperand(Idx));
3703 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3704 Subtarget->hasSSE2(), DAG);
3709 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3710 if (EVTBits == 8 && NumElems == 16) {
3711 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3713 if (V.getNode()) return V;
3716 if (EVTBits == 16 && NumElems == 8) {
3717 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3719 if (V.getNode()) return V;
3722 // If element VT is == 32 bits, turn it into a number of shuffles.
3723 SmallVector<SDValue, 8> V;
3725 if (NumElems == 4 && NumZero > 0) {
3726 for (unsigned i = 0; i < 4; ++i) {
3727 bool isZero = !(NonZeros & (1 << i));
3729 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3731 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3734 for (unsigned i = 0; i < 2; ++i) {
3735 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3738 V[i] = V[i*2]; // Must be a zero vector.
3741 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3744 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3747 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3752 SmallVector<int, 8> MaskVec;
3753 bool Reverse = (NonZeros & 0x3) == 2;
3754 for (unsigned i = 0; i < 2; ++i)
3755 MaskVec.push_back(Reverse ? 1-i : i);
3756 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3757 for (unsigned i = 0; i < 2; ++i)
3758 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3759 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3762 if (Values.size() > 2) {
3763 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3764 // values to be inserted is equal to the number of elements, in which case
3765 // use the unpack code below in the hopes of matching the consecutive elts
3766 // load merge pattern for shuffles.
3767 // FIXME: We could probably just check that here directly.
3768 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3769 getSubtarget()->hasSSE41()) {
3770 V[0] = DAG.getUNDEF(VT);
3771 for (unsigned i = 0; i < NumElems; ++i)
3772 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3773 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3774 Op.getOperand(i), DAG.getIntPtrConstant(i));
3777 // Expand into a number of unpckl*.
3779 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3780 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3781 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3782 for (unsigned i = 0; i < NumElems; ++i)
3783 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3785 while (NumElems != 0) {
3786 for (unsigned i = 0; i < NumElems; ++i)
3787 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3797 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3798 // We support concatenate two MMX registers and place them in a MMX
3799 // register. This is better than doing a stack convert.
3800 DebugLoc dl = Op.getDebugLoc();
3801 EVT ResVT = Op.getValueType();
3802 assert(Op.getNumOperands() == 2);
3803 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3804 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3806 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3807 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3808 InVec = Op.getOperand(1);
3809 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3810 unsigned NumElts = ResVT.getVectorNumElements();
3811 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3812 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3813 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3815 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3816 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3817 Mask[0] = 0; Mask[1] = 2;
3818 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3820 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3823 // v8i16 shuffles - Prefer shuffles in the following order:
3824 // 1. [all] pshuflw, pshufhw, optional move
3825 // 2. [ssse3] 1 x pshufb
3826 // 3. [ssse3] 2 x pshufb + 1 x por
3827 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3829 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3830 SelectionDAG &DAG, X86TargetLowering &TLI) {
3831 SDValue V1 = SVOp->getOperand(0);
3832 SDValue V2 = SVOp->getOperand(1);
3833 DebugLoc dl = SVOp->getDebugLoc();
3834 SmallVector<int, 8> MaskVals;
3836 // Determine if more than 1 of the words in each of the low and high quadwords
3837 // of the result come from the same quadword of one of the two inputs. Undef
3838 // mask values count as coming from any quadword, for better codegen.
3839 SmallVector<unsigned, 4> LoQuad(4);
3840 SmallVector<unsigned, 4> HiQuad(4);
3841 BitVector InputQuads(4);
3842 for (unsigned i = 0; i < 8; ++i) {
3843 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3844 int EltIdx = SVOp->getMaskElt(i);
3845 MaskVals.push_back(EltIdx);
3854 InputQuads.set(EltIdx / 4);
3857 int BestLoQuad = -1;
3858 unsigned MaxQuad = 1;
3859 for (unsigned i = 0; i < 4; ++i) {
3860 if (LoQuad[i] > MaxQuad) {
3862 MaxQuad = LoQuad[i];
3866 int BestHiQuad = -1;
3868 for (unsigned i = 0; i < 4; ++i) {
3869 if (HiQuad[i] > MaxQuad) {
3871 MaxQuad = HiQuad[i];
3875 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3876 // of the two input vectors, shuffle them into one input vector so only a
3877 // single pshufb instruction is necessary. If There are more than 2 input
3878 // quads, disable the next transformation since it does not help SSSE3.
3879 bool V1Used = InputQuads[0] || InputQuads[1];
3880 bool V2Used = InputQuads[2] || InputQuads[3];
3881 if (TLI.getSubtarget()->hasSSSE3()) {
3882 if (InputQuads.count() == 2 && V1Used && V2Used) {
3883 BestLoQuad = InputQuads.find_first();
3884 BestHiQuad = InputQuads.find_next(BestLoQuad);
3886 if (InputQuads.count() > 2) {
3892 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3893 // the shuffle mask. If a quad is scored as -1, that means that it contains
3894 // words from all 4 input quadwords.
3896 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3897 SmallVector<int, 8> MaskV;
3898 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3899 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3900 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3901 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3902 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3903 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3905 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3906 // source words for the shuffle, to aid later transformations.
3907 bool AllWordsInNewV = true;
3908 bool InOrder[2] = { true, true };
3909 for (unsigned i = 0; i != 8; ++i) {
3910 int idx = MaskVals[i];
3912 InOrder[i/4] = false;
3913 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3915 AllWordsInNewV = false;
3919 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3920 if (AllWordsInNewV) {
3921 for (int i = 0; i != 8; ++i) {
3922 int idx = MaskVals[i];
3925 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3926 if ((idx != i) && idx < 4)
3928 if ((idx != i) && idx > 3)
3937 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3938 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3939 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3940 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3941 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3945 // If we have SSSE3, and all words of the result are from 1 input vector,
3946 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3947 // is present, fall back to case 4.
3948 if (TLI.getSubtarget()->hasSSSE3()) {
3949 SmallVector<SDValue,16> pshufbMask;
3951 // If we have elements from both input vectors, set the high bit of the
3952 // shuffle mask element to zero out elements that come from V2 in the V1
3953 // mask, and elements that come from V1 in the V2 mask, so that the two
3954 // results can be OR'd together.
3955 bool TwoInputs = V1Used && V2Used;
3956 for (unsigned i = 0; i != 8; ++i) {
3957 int EltIdx = MaskVals[i] * 2;
3958 if (TwoInputs && (EltIdx >= 16)) {
3959 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3960 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3963 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3964 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3966 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3967 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3968 DAG.getNode(ISD::BUILD_VECTOR, dl,
3969 MVT::v16i8, &pshufbMask[0], 16));
3971 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3973 // Calculate the shuffle mask for the second input, shuffle it, and
3974 // OR it with the first shuffled input.
3976 for (unsigned i = 0; i != 8; ++i) {
3977 int EltIdx = MaskVals[i] * 2;
3979 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3980 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3983 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3984 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3986 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3987 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3988 DAG.getNode(ISD::BUILD_VECTOR, dl,
3989 MVT::v16i8, &pshufbMask[0], 16));
3990 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3991 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3994 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3995 // and update MaskVals with new element order.
3996 BitVector InOrder(8);
3997 if (BestLoQuad >= 0) {
3998 SmallVector<int, 8> MaskV;
3999 for (int i = 0; i != 4; ++i) {
4000 int idx = MaskVals[i];
4002 MaskV.push_back(-1);
4004 } else if ((idx / 4) == BestLoQuad) {
4005 MaskV.push_back(idx & 3);
4008 MaskV.push_back(-1);
4011 for (unsigned i = 4; i != 8; ++i)
4013 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4017 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4018 // and update MaskVals with the new element order.
4019 if (BestHiQuad >= 0) {
4020 SmallVector<int, 8> MaskV;
4021 for (unsigned i = 0; i != 4; ++i)
4023 for (unsigned i = 4; i != 8; ++i) {
4024 int idx = MaskVals[i];
4026 MaskV.push_back(-1);
4028 } else if ((idx / 4) == BestHiQuad) {
4029 MaskV.push_back((idx & 3) + 4);
4032 MaskV.push_back(-1);
4035 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4039 // In case BestHi & BestLo were both -1, which means each quadword has a word
4040 // from each of the four input quadwords, calculate the InOrder bitvector now
4041 // before falling through to the insert/extract cleanup.
4042 if (BestLoQuad == -1 && BestHiQuad == -1) {
4044 for (int i = 0; i != 8; ++i)
4045 if (MaskVals[i] < 0 || MaskVals[i] == i)
4049 // The other elements are put in the right place using pextrw and pinsrw.
4050 for (unsigned i = 0; i != 8; ++i) {
4053 int EltIdx = MaskVals[i];
4056 SDValue ExtOp = (EltIdx < 8)
4057 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4058 DAG.getIntPtrConstant(EltIdx))
4059 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4060 DAG.getIntPtrConstant(EltIdx - 8));
4061 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4062 DAG.getIntPtrConstant(i));
4067 // v16i8 shuffles - Prefer shuffles in the following order:
4068 // 1. [ssse3] 1 x pshufb
4069 // 2. [ssse3] 2 x pshufb + 1 x por
4070 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4072 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4073 SelectionDAG &DAG, X86TargetLowering &TLI) {
4074 SDValue V1 = SVOp->getOperand(0);
4075 SDValue V2 = SVOp->getOperand(1);
4076 DebugLoc dl = SVOp->getDebugLoc();
4077 SmallVector<int, 16> MaskVals;
4078 SVOp->getMask(MaskVals);
4080 // If we have SSSE3, case 1 is generated when all result bytes come from
4081 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4082 // present, fall back to case 3.
4083 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4086 for (unsigned i = 0; i < 16; ++i) {
4087 int EltIdx = MaskVals[i];
4096 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4097 if (TLI.getSubtarget()->hasSSSE3()) {
4098 SmallVector<SDValue,16> pshufbMask;
4100 // If all result elements are from one input vector, then only translate
4101 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4103 // Otherwise, we have elements from both input vectors, and must zero out
4104 // elements that come from V2 in the first mask, and V1 in the second mask
4105 // so that we can OR them together.
4106 bool TwoInputs = !(V1Only || V2Only);
4107 for (unsigned i = 0; i != 16; ++i) {
4108 int EltIdx = MaskVals[i];
4109 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4110 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4113 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4115 // If all the elements are from V2, assign it to V1 and return after
4116 // building the first pshufb.
4119 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4120 DAG.getNode(ISD::BUILD_VECTOR, dl,
4121 MVT::v16i8, &pshufbMask[0], 16));
4125 // Calculate the shuffle mask for the second input, shuffle it, and
4126 // OR it with the first shuffled input.
4128 for (unsigned i = 0; i != 16; ++i) {
4129 int EltIdx = MaskVals[i];
4131 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4134 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4136 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4137 DAG.getNode(ISD::BUILD_VECTOR, dl,
4138 MVT::v16i8, &pshufbMask[0], 16));
4139 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4142 // No SSSE3 - Calculate in place words and then fix all out of place words
4143 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4144 // the 16 different words that comprise the two doublequadword input vectors.
4145 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4146 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4147 SDValue NewV = V2Only ? V2 : V1;
4148 for (int i = 0; i != 8; ++i) {
4149 int Elt0 = MaskVals[i*2];
4150 int Elt1 = MaskVals[i*2+1];
4152 // This word of the result is all undef, skip it.
4153 if (Elt0 < 0 && Elt1 < 0)
4156 // This word of the result is already in the correct place, skip it.
4157 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4159 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4162 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4163 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4166 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4167 // using a single extract together, load it and store it.
4168 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4169 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4170 DAG.getIntPtrConstant(Elt1 / 2));
4171 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4172 DAG.getIntPtrConstant(i));
4176 // If Elt1 is defined, extract it from the appropriate source. If the
4177 // source byte is not also odd, shift the extracted word left 8 bits
4178 // otherwise clear the bottom 8 bits if we need to do an or.
4180 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4181 DAG.getIntPtrConstant(Elt1 / 2));
4182 if ((Elt1 & 1) == 0)
4183 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4184 DAG.getConstant(8, TLI.getShiftAmountTy()));
4186 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4187 DAG.getConstant(0xFF00, MVT::i16));
4189 // If Elt0 is defined, extract it from the appropriate source. If the
4190 // source byte is not also even, shift the extracted word right 8 bits. If
4191 // Elt1 was also defined, OR the extracted values together before
4192 // inserting them in the result.
4194 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4195 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4196 if ((Elt0 & 1) != 0)
4197 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4198 DAG.getConstant(8, TLI.getShiftAmountTy()));
4200 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4201 DAG.getConstant(0x00FF, MVT::i16));
4202 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4205 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4206 DAG.getIntPtrConstant(i));
4208 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4211 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4212 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4213 /// done when every pair / quad of shuffle mask elements point to elements in
4214 /// the right sequence. e.g.
4215 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4217 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4219 TargetLowering &TLI, DebugLoc dl) {
4220 EVT VT = SVOp->getValueType(0);
4221 SDValue V1 = SVOp->getOperand(0);
4222 SDValue V2 = SVOp->getOperand(1);
4223 unsigned NumElems = VT.getVectorNumElements();
4224 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4225 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4226 EVT MaskEltVT = MaskVT.getVectorElementType();
4228 switch (VT.getSimpleVT().SimpleTy) {
4229 default: assert(false && "Unexpected!");
4230 case MVT::v4f32: NewVT = MVT::v2f64; break;
4231 case MVT::v4i32: NewVT = MVT::v2i64; break;
4232 case MVT::v8i16: NewVT = MVT::v4i32; break;
4233 case MVT::v16i8: NewVT = MVT::v4i32; break;
4236 if (NewWidth == 2) {
4242 int Scale = NumElems / NewWidth;
4243 SmallVector<int, 8> MaskVec;
4244 for (unsigned i = 0; i < NumElems; i += Scale) {
4246 for (int j = 0; j < Scale; ++j) {
4247 int EltIdx = SVOp->getMaskElt(i+j);
4251 StartIdx = EltIdx - (EltIdx % Scale);
4252 if (EltIdx != StartIdx + j)
4256 MaskVec.push_back(-1);
4258 MaskVec.push_back(StartIdx / Scale);
4261 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4262 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4263 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4266 /// getVZextMovL - Return a zero-extending vector move low node.
4268 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4269 SDValue SrcOp, SelectionDAG &DAG,
4270 const X86Subtarget *Subtarget, DebugLoc dl) {
4271 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4272 LoadSDNode *LD = NULL;
4273 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4274 LD = dyn_cast<LoadSDNode>(SrcOp);
4276 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4278 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4279 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4280 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4281 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4282 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4284 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4285 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4286 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4287 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4295 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4296 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4297 DAG.getNode(ISD::BIT_CONVERT, dl,
4301 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4304 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4305 SDValue V1 = SVOp->getOperand(0);
4306 SDValue V2 = SVOp->getOperand(1);
4307 DebugLoc dl = SVOp->getDebugLoc();
4308 EVT VT = SVOp->getValueType(0);
4310 SmallVector<std::pair<int, int>, 8> Locs;
4312 SmallVector<int, 8> Mask1(4U, -1);
4313 SmallVector<int, 8> PermMask;
4314 SVOp->getMask(PermMask);
4318 for (unsigned i = 0; i != 4; ++i) {
4319 int Idx = PermMask[i];
4321 Locs[i] = std::make_pair(-1, -1);
4323 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4325 Locs[i] = std::make_pair(0, NumLo);
4329 Locs[i] = std::make_pair(1, NumHi);
4331 Mask1[2+NumHi] = Idx;
4337 if (NumLo <= 2 && NumHi <= 2) {
4338 // If no more than two elements come from either vector. This can be
4339 // implemented with two shuffles. First shuffle gather the elements.
4340 // The second shuffle, which takes the first shuffle as both of its
4341 // vector operands, put the elements into the right order.
4342 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4344 SmallVector<int, 8> Mask2(4U, -1);
4346 for (unsigned i = 0; i != 4; ++i) {
4347 if (Locs[i].first == -1)
4350 unsigned Idx = (i < 2) ? 0 : 4;
4351 Idx += Locs[i].first * 2 + Locs[i].second;
4356 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4357 } else if (NumLo == 3 || NumHi == 3) {
4358 // Otherwise, we must have three elements from one vector, call it X, and
4359 // one element from the other, call it Y. First, use a shufps to build an
4360 // intermediate vector with the one element from Y and the element from X
4361 // that will be in the same half in the final destination (the indexes don't
4362 // matter). Then, use a shufps to build the final vector, taking the half
4363 // containing the element from Y from the intermediate, and the other half
4366 // Normalize it so the 3 elements come from V1.
4367 CommuteVectorShuffleMask(PermMask, VT);
4371 // Find the element from V2.
4373 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4374 int Val = PermMask[HiIndex];
4381 Mask1[0] = PermMask[HiIndex];
4383 Mask1[2] = PermMask[HiIndex^1];
4385 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4388 Mask1[0] = PermMask[0];
4389 Mask1[1] = PermMask[1];
4390 Mask1[2] = HiIndex & 1 ? 6 : 4;
4391 Mask1[3] = HiIndex & 1 ? 4 : 6;
4392 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4394 Mask1[0] = HiIndex & 1 ? 2 : 0;
4395 Mask1[1] = HiIndex & 1 ? 0 : 2;
4396 Mask1[2] = PermMask[2];
4397 Mask1[3] = PermMask[3];
4402 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4406 // Break it into (shuffle shuffle_hi, shuffle_lo).
4408 SmallVector<int,8> LoMask(4U, -1);
4409 SmallVector<int,8> HiMask(4U, -1);
4411 SmallVector<int,8> *MaskPtr = &LoMask;
4412 unsigned MaskIdx = 0;
4415 for (unsigned i = 0; i != 4; ++i) {
4422 int Idx = PermMask[i];
4424 Locs[i] = std::make_pair(-1, -1);
4425 } else if (Idx < 4) {
4426 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4427 (*MaskPtr)[LoIdx] = Idx;
4430 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4431 (*MaskPtr)[HiIdx] = Idx;
4436 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4437 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4438 SmallVector<int, 8> MaskOps;
4439 for (unsigned i = 0; i != 4; ++i) {
4440 if (Locs[i].first == -1) {
4441 MaskOps.push_back(-1);
4443 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4444 MaskOps.push_back(Idx);
4447 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4451 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4452 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4453 SDValue V1 = Op.getOperand(0);
4454 SDValue V2 = Op.getOperand(1);
4455 EVT VT = Op.getValueType();
4456 DebugLoc dl = Op.getDebugLoc();
4457 unsigned NumElems = VT.getVectorNumElements();
4458 bool isMMX = VT.getSizeInBits() == 64;
4459 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4460 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4461 bool V1IsSplat = false;
4462 bool V2IsSplat = false;
4464 if (isZeroShuffle(SVOp))
4465 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4467 // Promote splats to v4f32.
4468 if (SVOp->isSplat()) {
4469 if (isMMX || NumElems < 4)
4471 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4474 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4476 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4477 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4478 if (NewOp.getNode())
4479 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4480 LowerVECTOR_SHUFFLE(NewOp, DAG));
4481 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4482 // FIXME: Figure out a cleaner way to do this.
4483 // Try to make use of movq to zero out the top part.
4484 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4485 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4486 if (NewOp.getNode()) {
4487 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4488 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4489 DAG, Subtarget, dl);
4491 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4492 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4493 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4494 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4495 DAG, Subtarget, dl);
4499 if (X86::isPSHUFDMask(SVOp))
4502 // Check if this can be converted into a logical shift.
4503 bool isLeft = false;
4506 bool isShift = getSubtarget()->hasSSE2() &&
4507 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4508 if (isShift && ShVal.hasOneUse()) {
4509 // If the shifted value has multiple uses, it may be cheaper to use
4510 // v_set0 + movlhps or movhlps, etc.
4511 EVT EltVT = VT.getVectorElementType();
4512 ShAmt *= EltVT.getSizeInBits();
4513 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4516 if (X86::isMOVLMask(SVOp)) {
4519 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4520 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4525 // FIXME: fold these into legal mask.
4526 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4527 X86::isMOVSLDUPMask(SVOp) ||
4528 X86::isMOVHLPSMask(SVOp) ||
4529 X86::isMOVLHPSMask(SVOp) ||
4530 X86::isMOVLPMask(SVOp)))
4533 if (ShouldXformToMOVHLPS(SVOp) ||
4534 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4535 return CommuteVectorShuffle(SVOp, DAG);
4538 // No better options. Use a vshl / vsrl.
4539 EVT EltVT = VT.getVectorElementType();
4540 ShAmt *= EltVT.getSizeInBits();
4541 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4544 bool Commuted = false;
4545 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4546 // 1,1,1,1 -> v8i16 though.
4547 V1IsSplat = isSplatVector(V1.getNode());
4548 V2IsSplat = isSplatVector(V2.getNode());
4550 // Canonicalize the splat or undef, if present, to be on the RHS.
4551 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4552 Op = CommuteVectorShuffle(SVOp, DAG);
4553 SVOp = cast<ShuffleVectorSDNode>(Op);
4554 V1 = SVOp->getOperand(0);
4555 V2 = SVOp->getOperand(1);
4556 std::swap(V1IsSplat, V2IsSplat);
4557 std::swap(V1IsUndef, V2IsUndef);
4561 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4562 // Shuffling low element of v1 into undef, just return v1.
4565 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4566 // the instruction selector will not match, so get a canonical MOVL with
4567 // swapped operands to undo the commute.
4568 return getMOVL(DAG, dl, VT, V2, V1);
4571 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4572 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4573 X86::isUNPCKLMask(SVOp) ||
4574 X86::isUNPCKHMask(SVOp))
4578 // Normalize mask so all entries that point to V2 points to its first
4579 // element then try to match unpck{h|l} again. If match, return a
4580 // new vector_shuffle with the corrected mask.
4581 SDValue NewMask = NormalizeMask(SVOp, DAG);
4582 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4583 if (NSVOp != SVOp) {
4584 if (X86::isUNPCKLMask(NSVOp, true)) {
4586 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4593 // Commute is back and try unpck* again.
4594 // FIXME: this seems wrong.
4595 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4596 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4597 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4598 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4599 X86::isUNPCKLMask(NewSVOp) ||
4600 X86::isUNPCKHMask(NewSVOp))
4604 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4606 // Normalize the node to match x86 shuffle ops if needed
4607 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4608 return CommuteVectorShuffle(SVOp, DAG);
4610 // Check for legal shuffle and return?
4611 SmallVector<int, 16> PermMask;
4612 SVOp->getMask(PermMask);
4613 if (isShuffleMaskLegal(PermMask, VT))
4616 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4617 if (VT == MVT::v8i16) {
4618 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4619 if (NewOp.getNode())
4623 if (VT == MVT::v16i8) {
4624 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4625 if (NewOp.getNode())
4629 // Handle all 4 wide cases with a number of shuffles except for MMX.
4630 if (NumElems == 4 && !isMMX)
4631 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4637 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4638 SelectionDAG &DAG) {
4639 EVT VT = Op.getValueType();
4640 DebugLoc dl = Op.getDebugLoc();
4641 if (VT.getSizeInBits() == 8) {
4642 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4643 Op.getOperand(0), Op.getOperand(1));
4644 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4645 DAG.getValueType(VT));
4646 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4647 } else if (VT.getSizeInBits() == 16) {
4648 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4649 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4651 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4652 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4653 DAG.getNode(ISD::BIT_CONVERT, dl,
4657 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4658 Op.getOperand(0), Op.getOperand(1));
4659 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4660 DAG.getValueType(VT));
4661 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4662 } else if (VT == MVT::f32) {
4663 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4664 // the result back to FR32 register. It's only worth matching if the
4665 // result has a single use which is a store or a bitcast to i32. And in
4666 // the case of a store, it's not worth it if the index is a constant 0,
4667 // because a MOVSSmr can be used instead, which is smaller and faster.
4668 if (!Op.hasOneUse())
4670 SDNode *User = *Op.getNode()->use_begin();
4671 if ((User->getOpcode() != ISD::STORE ||
4672 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4673 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4674 (User->getOpcode() != ISD::BIT_CONVERT ||
4675 User->getValueType(0) != MVT::i32))
4677 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4678 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4681 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4682 } else if (VT == MVT::i32) {
4683 // ExtractPS works with constant index.
4684 if (isa<ConstantSDNode>(Op.getOperand(1)))
4692 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4693 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4696 if (Subtarget->hasSSE41()) {
4697 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4702 EVT VT = Op.getValueType();
4703 DebugLoc dl = Op.getDebugLoc();
4704 // TODO: handle v16i8.
4705 if (VT.getSizeInBits() == 16) {
4706 SDValue Vec = Op.getOperand(0);
4707 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4709 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4710 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4711 DAG.getNode(ISD::BIT_CONVERT, dl,
4714 // Transform it so it match pextrw which produces a 32-bit result.
4715 EVT EltVT = MVT::i32;
4716 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4717 Op.getOperand(0), Op.getOperand(1));
4718 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4719 DAG.getValueType(VT));
4720 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4721 } else if (VT.getSizeInBits() == 32) {
4722 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4726 // SHUFPS the element to the lowest double word, then movss.
4727 int Mask[4] = { Idx, -1, -1, -1 };
4728 EVT VVT = Op.getOperand(0).getValueType();
4729 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4730 DAG.getUNDEF(VVT), Mask);
4731 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4732 DAG.getIntPtrConstant(0));
4733 } else if (VT.getSizeInBits() == 64) {
4734 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4735 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4736 // to match extract_elt for f64.
4737 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4741 // UNPCKHPD the element to the lowest double word, then movsd.
4742 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4743 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4744 int Mask[2] = { 1, -1 };
4745 EVT VVT = Op.getOperand(0).getValueType();
4746 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4747 DAG.getUNDEF(VVT), Mask);
4748 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4749 DAG.getIntPtrConstant(0));
4756 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4757 EVT VT = Op.getValueType();
4758 EVT EltVT = VT.getVectorElementType();
4759 DebugLoc dl = Op.getDebugLoc();
4761 SDValue N0 = Op.getOperand(0);
4762 SDValue N1 = Op.getOperand(1);
4763 SDValue N2 = Op.getOperand(2);
4765 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4766 isa<ConstantSDNode>(N2)) {
4767 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4769 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4771 if (N1.getValueType() != MVT::i32)
4772 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4773 if (N2.getValueType() != MVT::i32)
4774 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4775 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4776 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4777 // Bits [7:6] of the constant are the source select. This will always be
4778 // zero here. The DAG Combiner may combine an extract_elt index into these
4779 // bits. For example (insert (extract, 3), 2) could be matched by putting
4780 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4781 // Bits [5:4] of the constant are the destination select. This is the
4782 // value of the incoming immediate.
4783 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4784 // combine either bitwise AND or insert of float 0.0 to set these bits.
4785 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4786 // Create this as a scalar to vector..
4787 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4788 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4789 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4790 // PINSR* works with constant index.
4797 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4798 EVT VT = Op.getValueType();
4799 EVT EltVT = VT.getVectorElementType();
4801 if (Subtarget->hasSSE41())
4802 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4804 if (EltVT == MVT::i8)
4807 DebugLoc dl = Op.getDebugLoc();
4808 SDValue N0 = Op.getOperand(0);
4809 SDValue N1 = Op.getOperand(1);
4810 SDValue N2 = Op.getOperand(2);
4812 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4813 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4814 // as its second argument.
4815 if (N1.getValueType() != MVT::i32)
4816 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4817 if (N2.getValueType() != MVT::i32)
4818 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4819 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4825 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4826 DebugLoc dl = Op.getDebugLoc();
4827 if (Op.getValueType() == MVT::v2f32)
4828 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4829 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4830 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4831 Op.getOperand(0))));
4833 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4834 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4836 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4837 EVT VT = MVT::v2i32;
4838 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4845 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4846 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4849 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4850 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4851 // one of the above mentioned nodes. It has to be wrapped because otherwise
4852 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4853 // be used to form addressing mode. These wrapped nodes will be selected
4856 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4857 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4859 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4861 unsigned char OpFlag = 0;
4862 unsigned WrapperKind = X86ISD::Wrapper;
4863 CodeModel::Model M = getTargetMachine().getCodeModel();
4865 if (Subtarget->isPICStyleRIPRel() &&
4866 (M == CodeModel::Small || M == CodeModel::Kernel))
4867 WrapperKind = X86ISD::WrapperRIP;
4868 else if (Subtarget->isPICStyleGOT())
4869 OpFlag = X86II::MO_GOTOFF;
4870 else if (Subtarget->isPICStyleStubPIC())
4871 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4873 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4875 CP->getOffset(), OpFlag);
4876 DebugLoc DL = CP->getDebugLoc();
4877 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4878 // With PIC, the address is actually $g + Offset.
4880 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4881 DAG.getNode(X86ISD::GlobalBaseReg,
4882 DebugLoc::getUnknownLoc(), getPointerTy()),
4889 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4890 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4892 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4894 unsigned char OpFlag = 0;
4895 unsigned WrapperKind = X86ISD::Wrapper;
4896 CodeModel::Model M = getTargetMachine().getCodeModel();
4898 if (Subtarget->isPICStyleRIPRel() &&
4899 (M == CodeModel::Small || M == CodeModel::Kernel))
4900 WrapperKind = X86ISD::WrapperRIP;
4901 else if (Subtarget->isPICStyleGOT())
4902 OpFlag = X86II::MO_GOTOFF;
4903 else if (Subtarget->isPICStyleStubPIC())
4904 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4906 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4908 DebugLoc DL = JT->getDebugLoc();
4909 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4911 // With PIC, the address is actually $g + Offset.
4913 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4914 DAG.getNode(X86ISD::GlobalBaseReg,
4915 DebugLoc::getUnknownLoc(), getPointerTy()),
4923 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4924 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4926 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4928 unsigned char OpFlag = 0;
4929 unsigned WrapperKind = X86ISD::Wrapper;
4930 CodeModel::Model M = getTargetMachine().getCodeModel();
4932 if (Subtarget->isPICStyleRIPRel() &&
4933 (M == CodeModel::Small || M == CodeModel::Kernel))
4934 WrapperKind = X86ISD::WrapperRIP;
4935 else if (Subtarget->isPICStyleGOT())
4936 OpFlag = X86II::MO_GOTOFF;
4937 else if (Subtarget->isPICStyleStubPIC())
4938 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4940 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4942 DebugLoc DL = Op.getDebugLoc();
4943 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4946 // With PIC, the address is actually $g + Offset.
4947 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4948 !Subtarget->is64Bit()) {
4949 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4950 DAG.getNode(X86ISD::GlobalBaseReg,
4951 DebugLoc::getUnknownLoc(),
4960 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
4961 // Create the TargetBlockAddressAddress node.
4962 unsigned char OpFlags =
4963 Subtarget->ClassifyBlockAddressReference();
4964 CodeModel::Model M = getTargetMachine().getCodeModel();
4965 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4966 DebugLoc dl = Op.getDebugLoc();
4967 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
4968 /*isTarget=*/true, OpFlags);
4970 if (Subtarget->isPICStyleRIPRel() &&
4971 (M == CodeModel::Small || M == CodeModel::Kernel))
4972 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4974 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4976 // With PIC, the address is actually $g + Offset.
4977 if (isGlobalRelativeToPICBase(OpFlags)) {
4978 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4979 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4987 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4989 SelectionDAG &DAG) const {
4990 // Create the TargetGlobalAddress node, folding in the constant
4991 // offset if it is legal.
4992 unsigned char OpFlags =
4993 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4994 CodeModel::Model M = getTargetMachine().getCodeModel();
4996 if (OpFlags == X86II::MO_NO_FLAG &&
4997 X86::isOffsetSuitableForCodeModel(Offset, M)) {
4998 // A direct static reference to a global.
4999 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5002 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5005 if (Subtarget->isPICStyleRIPRel() &&
5006 (M == CodeModel::Small || M == CodeModel::Kernel))
5007 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5009 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5011 // With PIC, the address is actually $g + Offset.
5012 if (isGlobalRelativeToPICBase(OpFlags)) {
5013 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5014 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5018 // For globals that require a load from a stub to get the address, emit the
5020 if (isGlobalStubReference(OpFlags))
5021 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5022 PseudoSourceValue::getGOT(), 0);
5024 // If there was a non-zero offset that we didn't fold, create an explicit
5027 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5028 DAG.getConstant(Offset, getPointerTy()));
5034 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5035 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5036 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5037 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5041 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5042 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5043 unsigned char OperandFlags) {
5044 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5045 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5046 DebugLoc dl = GA->getDebugLoc();
5047 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5048 GA->getValueType(0),
5052 SDValue Ops[] = { Chain, TGA, *InFlag };
5053 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5055 SDValue Ops[] = { Chain, TGA };
5056 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5059 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5060 MFI->setHasCalls(true);
5062 SDValue Flag = Chain.getValue(1);
5063 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5066 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5068 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5071 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5072 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5073 DAG.getNode(X86ISD::GlobalBaseReg,
5074 DebugLoc::getUnknownLoc(),
5076 InFlag = Chain.getValue(1);
5078 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5081 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5083 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5085 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5086 X86::RAX, X86II::MO_TLSGD);
5089 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5090 // "local exec" model.
5091 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5092 const EVT PtrVT, TLSModel::Model model,
5094 DebugLoc dl = GA->getDebugLoc();
5095 // Get the Thread Pointer
5096 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5097 DebugLoc::getUnknownLoc(), PtrVT,
5098 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5101 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5104 unsigned char OperandFlags = 0;
5105 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5107 unsigned WrapperKind = X86ISD::Wrapper;
5108 if (model == TLSModel::LocalExec) {
5109 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5110 } else if (is64Bit) {
5111 assert(model == TLSModel::InitialExec);
5112 OperandFlags = X86II::MO_GOTTPOFF;
5113 WrapperKind = X86ISD::WrapperRIP;
5115 assert(model == TLSModel::InitialExec);
5116 OperandFlags = X86II::MO_INDNTPOFF;
5119 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5121 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5122 GA->getOffset(), OperandFlags);
5123 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5125 if (model == TLSModel::InitialExec)
5126 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5127 PseudoSourceValue::getGOT(), 0);
5129 // The address of the thread local variable is the add of the thread
5130 // pointer with the offset of the variable.
5131 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5135 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5136 // TODO: implement the "local dynamic" model
5137 // TODO: implement the "initial exec"model for pic executables
5138 assert(Subtarget->isTargetELF() &&
5139 "TLS not implemented for non-ELF targets");
5140 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5141 const GlobalValue *GV = GA->getGlobal();
5143 // If GV is an alias then use the aliasee for determining
5144 // thread-localness.
5145 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5146 GV = GA->resolveAliasedGlobal(false);
5148 TLSModel::Model model = getTLSModel(GV,
5149 getTargetMachine().getRelocationModel());
5152 case TLSModel::GeneralDynamic:
5153 case TLSModel::LocalDynamic: // not implemented
5154 if (Subtarget->is64Bit())
5155 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5156 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5158 case TLSModel::InitialExec:
5159 case TLSModel::LocalExec:
5160 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5161 Subtarget->is64Bit());
5164 llvm_unreachable("Unreachable");
5169 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5170 /// take a 2 x i32 value to shift plus a shift amount.
5171 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5172 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5173 EVT VT = Op.getValueType();
5174 unsigned VTBits = VT.getSizeInBits();
5175 DebugLoc dl = Op.getDebugLoc();
5176 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5177 SDValue ShOpLo = Op.getOperand(0);
5178 SDValue ShOpHi = Op.getOperand(1);
5179 SDValue ShAmt = Op.getOperand(2);
5180 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5181 DAG.getConstant(VTBits - 1, MVT::i8))
5182 : DAG.getConstant(0, VT);
5185 if (Op.getOpcode() == ISD::SHL_PARTS) {
5186 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5187 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5189 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5190 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5193 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5194 DAG.getConstant(VTBits, MVT::i8));
5195 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
5196 AndNode, DAG.getConstant(0, MVT::i8));
5199 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5200 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5201 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5203 if (Op.getOpcode() == ISD::SHL_PARTS) {
5204 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5205 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5207 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5208 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5211 SDValue Ops[2] = { Lo, Hi };
5212 return DAG.getMergeValues(Ops, 2, dl);
5215 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5216 EVT SrcVT = Op.getOperand(0).getValueType();
5218 if (SrcVT.isVector()) {
5219 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5225 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5226 "Unknown SINT_TO_FP to lower!");
5228 // These are really Legal; return the operand so the caller accepts it as
5230 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5232 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5233 Subtarget->is64Bit()) {
5237 DebugLoc dl = Op.getDebugLoc();
5238 unsigned Size = SrcVT.getSizeInBits()/8;
5239 MachineFunction &MF = DAG.getMachineFunction();
5240 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5241 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5242 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5244 PseudoSourceValue::getFixedStack(SSFI), 0);
5245 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5248 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5250 SelectionDAG &DAG) {
5252 DebugLoc dl = Op.getDebugLoc();
5254 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5256 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5258 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5259 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5260 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5261 Tys, Ops, array_lengthof(Ops));
5264 Chain = Result.getValue(1);
5265 SDValue InFlag = Result.getValue(2);
5267 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5268 // shouldn't be necessary except that RFP cannot be live across
5269 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5270 MachineFunction &MF = DAG.getMachineFunction();
5271 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5272 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5273 Tys = DAG.getVTList(MVT::Other);
5275 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5277 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5278 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5279 PseudoSourceValue::getFixedStack(SSFI), 0);
5285 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5286 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5287 // This algorithm is not obvious. Here it is in C code, more or less:
5289 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5290 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5291 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5293 // Copy ints to xmm registers.
5294 __m128i xh = _mm_cvtsi32_si128( hi );
5295 __m128i xl = _mm_cvtsi32_si128( lo );
5297 // Combine into low half of a single xmm register.
5298 __m128i x = _mm_unpacklo_epi32( xh, xl );
5302 // Merge in appropriate exponents to give the integer bits the right
5304 x = _mm_unpacklo_epi32( x, exp );
5306 // Subtract away the biases to deal with the IEEE-754 double precision
5308 d = _mm_sub_pd( (__m128d) x, bias );
5310 // All conversions up to here are exact. The correctly rounded result is
5311 // calculated using the current rounding mode using the following
5313 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5314 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5315 // store doesn't really need to be here (except
5316 // maybe to zero the other double)
5321 DebugLoc dl = Op.getDebugLoc();
5322 LLVMContext *Context = DAG.getContext();
5324 // Build some magic constants.
5325 std::vector<Constant*> CV0;
5326 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5327 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5328 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5329 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5330 Constant *C0 = ConstantVector::get(CV0);
5331 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5333 std::vector<Constant*> CV1;
5335 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5337 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5338 Constant *C1 = ConstantVector::get(CV1);
5339 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5341 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5342 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5344 DAG.getIntPtrConstant(1)));
5345 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5346 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5348 DAG.getIntPtrConstant(0)));
5349 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5350 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5351 PseudoSourceValue::getConstantPool(), 0,
5353 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5354 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5355 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5356 PseudoSourceValue::getConstantPool(), 0,
5358 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5360 // Add the halves; easiest way is to swap them into another reg first.
5361 int ShufMask[2] = { 1, -1 };
5362 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5363 DAG.getUNDEF(MVT::v2f64), ShufMask);
5364 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5365 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5366 DAG.getIntPtrConstant(0));
5369 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5370 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5371 DebugLoc dl = Op.getDebugLoc();
5372 // FP constant to bias correct the final result.
5373 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5376 // Load the 32-bit value into an XMM register.
5377 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5378 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5380 DAG.getIntPtrConstant(0)));
5382 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5383 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5384 DAG.getIntPtrConstant(0));
5386 // Or the load with the bias.
5387 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5388 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5389 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5391 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5392 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5393 MVT::v2f64, Bias)));
5394 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5395 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5396 DAG.getIntPtrConstant(0));
5398 // Subtract the bias.
5399 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5401 // Handle final rounding.
5402 EVT DestVT = Op.getValueType();
5404 if (DestVT.bitsLT(MVT::f64)) {
5405 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5406 DAG.getIntPtrConstant(0));
5407 } else if (DestVT.bitsGT(MVT::f64)) {
5408 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5411 // Handle final rounding.
5415 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5416 SDValue N0 = Op.getOperand(0);
5417 DebugLoc dl = Op.getDebugLoc();
5419 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5420 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5421 // the optimization here.
5422 if (DAG.SignBitIsZero(N0))
5423 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5425 EVT SrcVT = N0.getValueType();
5426 if (SrcVT == MVT::i64) {
5427 // We only handle SSE2 f64 target here; caller can expand the rest.
5428 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5431 return LowerUINT_TO_FP_i64(Op, DAG);
5432 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5433 return LowerUINT_TO_FP_i32(Op, DAG);
5436 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5438 // Make a 64-bit buffer, and use it to build an FILD.
5439 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5440 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5441 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5442 getPointerTy(), StackSlot, WordOff);
5443 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5444 StackSlot, NULL, 0);
5445 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5446 OffsetSlot, NULL, 0);
5447 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5450 std::pair<SDValue,SDValue> X86TargetLowering::
5451 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5452 DebugLoc dl = Op.getDebugLoc();
5454 EVT DstTy = Op.getValueType();
5457 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5461 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5462 DstTy.getSimpleVT() >= MVT::i16 &&
5463 "Unknown FP_TO_SINT to lower!");
5465 // These are really Legal.
5466 if (DstTy == MVT::i32 &&
5467 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5468 return std::make_pair(SDValue(), SDValue());
5469 if (Subtarget->is64Bit() &&
5470 DstTy == MVT::i64 &&
5471 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5472 return std::make_pair(SDValue(), SDValue());
5474 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5476 MachineFunction &MF = DAG.getMachineFunction();
5477 unsigned MemSize = DstTy.getSizeInBits()/8;
5478 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5479 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5482 switch (DstTy.getSimpleVT().SimpleTy) {
5483 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5484 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5485 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5486 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5489 SDValue Chain = DAG.getEntryNode();
5490 SDValue Value = Op.getOperand(0);
5491 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5492 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5493 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5494 PseudoSourceValue::getFixedStack(SSFI), 0);
5495 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5497 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5499 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5500 Chain = Value.getValue(1);
5501 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5502 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5505 // Build the FP_TO_INT*_IN_MEM
5506 SDValue Ops[] = { Chain, Value, StackSlot };
5507 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5509 return std::make_pair(FIST, StackSlot);
5512 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5513 if (Op.getValueType().isVector()) {
5514 if (Op.getValueType() == MVT::v2i32 &&
5515 Op.getOperand(0).getValueType() == MVT::v2f64) {
5521 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5522 SDValue FIST = Vals.first, StackSlot = Vals.second;
5523 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5524 if (FIST.getNode() == 0) return Op;
5527 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5528 FIST, StackSlot, NULL, 0);
5531 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5532 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5533 SDValue FIST = Vals.first, StackSlot = Vals.second;
5534 assert(FIST.getNode() && "Unexpected failure");
5537 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5538 FIST, StackSlot, NULL, 0);
5541 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5542 LLVMContext *Context = DAG.getContext();
5543 DebugLoc dl = Op.getDebugLoc();
5544 EVT VT = Op.getValueType();
5547 EltVT = VT.getVectorElementType();
5548 std::vector<Constant*> CV;
5549 if (EltVT == MVT::f64) {
5550 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5554 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5560 Constant *C = ConstantVector::get(CV);
5561 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5562 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5563 PseudoSourceValue::getConstantPool(), 0,
5565 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5568 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5569 LLVMContext *Context = DAG.getContext();
5570 DebugLoc dl = Op.getDebugLoc();
5571 EVT VT = Op.getValueType();
5574 EltVT = VT.getVectorElementType();
5575 std::vector<Constant*> CV;
5576 if (EltVT == MVT::f64) {
5577 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5581 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5587 Constant *C = ConstantVector::get(CV);
5588 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5589 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5590 PseudoSourceValue::getConstantPool(), 0,
5592 if (VT.isVector()) {
5593 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5594 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5595 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5597 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5599 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5603 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5604 LLVMContext *Context = DAG.getContext();
5605 SDValue Op0 = Op.getOperand(0);
5606 SDValue Op1 = Op.getOperand(1);
5607 DebugLoc dl = Op.getDebugLoc();
5608 EVT VT = Op.getValueType();
5609 EVT SrcVT = Op1.getValueType();
5611 // If second operand is smaller, extend it first.
5612 if (SrcVT.bitsLT(VT)) {
5613 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5616 // And if it is bigger, shrink it first.
5617 if (SrcVT.bitsGT(VT)) {
5618 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5622 // At this point the operands and the result should have the same
5623 // type, and that won't be f80 since that is not custom lowered.
5625 // First get the sign bit of second operand.
5626 std::vector<Constant*> CV;
5627 if (SrcVT == MVT::f64) {
5628 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5629 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5631 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5632 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5633 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5634 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5636 Constant *C = ConstantVector::get(CV);
5637 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5638 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5639 PseudoSourceValue::getConstantPool(), 0,
5641 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5643 // Shift sign bit right or left if the two operands have different types.
5644 if (SrcVT.bitsGT(VT)) {
5645 // Op0 is MVT::f32, Op1 is MVT::f64.
5646 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5647 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5648 DAG.getConstant(32, MVT::i32));
5649 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5650 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5651 DAG.getIntPtrConstant(0));
5654 // Clear first operand sign bit.
5656 if (VT == MVT::f64) {
5657 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5658 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5660 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5661 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5662 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5663 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5665 C = ConstantVector::get(CV);
5666 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5667 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5668 PseudoSourceValue::getConstantPool(), 0,
5670 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5672 // Or the value with the sign bit.
5673 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5676 /// Emit nodes that will be selected as "test Op0,Op0", or something
5678 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5679 SelectionDAG &DAG) {
5680 DebugLoc dl = Op.getDebugLoc();
5682 // CF and OF aren't always set the way we want. Determine which
5683 // of these we need.
5684 bool NeedCF = false;
5685 bool NeedOF = false;
5687 case X86::COND_A: case X86::COND_AE:
5688 case X86::COND_B: case X86::COND_BE:
5691 case X86::COND_G: case X86::COND_GE:
5692 case X86::COND_L: case X86::COND_LE:
5693 case X86::COND_O: case X86::COND_NO:
5699 // See if we can use the EFLAGS value from the operand instead of
5700 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5701 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5702 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5703 unsigned Opcode = 0;
5704 unsigned NumOperands = 0;
5705 switch (Op.getNode()->getOpcode()) {
5707 // Due to an isel shortcoming, be conservative if this add is likely to
5708 // be selected as part of a load-modify-store instruction. When the root
5709 // node in a match is a store, isel doesn't know how to remap non-chain
5710 // non-flag uses of other nodes in the match, such as the ADD in this
5711 // case. This leads to the ADD being left around and reselected, with
5712 // the result being two adds in the output.
5713 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5714 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5715 if (UI->getOpcode() == ISD::STORE)
5717 if (ConstantSDNode *C =
5718 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5719 // An add of one will be selected as an INC.
5720 if (C->getAPIntValue() == 1) {
5721 Opcode = X86ISD::INC;
5725 // An add of negative one (subtract of one) will be selected as a DEC.
5726 if (C->getAPIntValue().isAllOnesValue()) {
5727 Opcode = X86ISD::DEC;
5732 // Otherwise use a regular EFLAGS-setting add.
5733 Opcode = X86ISD::ADD;
5737 // If the primary and result isn't used, don't bother using X86ISD::AND,
5738 // because a TEST instruction will be better.
5739 bool NonFlagUse = false;
5740 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5741 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5743 unsigned UOpNo = UI.getOperandNo();
5744 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5745 // Look pass truncate.
5746 UOpNo = User->use_begin().getOperandNo();
5747 User = *User->use_begin();
5749 if (User->getOpcode() != ISD::BRCOND &&
5750 User->getOpcode() != ISD::SETCC &&
5751 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5763 // Due to the ISEL shortcoming noted above, be conservative if this op is
5764 // likely to be selected as part of a load-modify-store instruction.
5765 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5766 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5767 if (UI->getOpcode() == ISD::STORE)
5769 // Otherwise use a regular EFLAGS-setting instruction.
5770 switch (Op.getNode()->getOpcode()) {
5771 case ISD::SUB: Opcode = X86ISD::SUB; break;
5772 case ISD::OR: Opcode = X86ISD::OR; break;
5773 case ISD::XOR: Opcode = X86ISD::XOR; break;
5774 case ISD::AND: Opcode = X86ISD::AND; break;
5775 default: llvm_unreachable("unexpected operator!");
5786 return SDValue(Op.getNode(), 1);
5792 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5793 SmallVector<SDValue, 4> Ops;
5794 for (unsigned i = 0; i != NumOperands; ++i)
5795 Ops.push_back(Op.getOperand(i));
5796 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5797 DAG.ReplaceAllUsesWith(Op, New);
5798 return SDValue(New.getNode(), 1);
5802 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5803 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5804 DAG.getConstant(0, Op.getValueType()));
5807 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5809 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5810 SelectionDAG &DAG) {
5811 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5812 if (C->getAPIntValue() == 0)
5813 return EmitTest(Op0, X86CC, DAG);
5815 DebugLoc dl = Op0.getDebugLoc();
5816 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5819 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5820 /// if it's possible.
5821 static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
5822 DebugLoc dl, SelectionDAG &DAG) {
5824 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5825 if (ConstantSDNode *Op010C =
5826 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5827 if (Op010C->getZExtValue() == 1) {
5828 LHS = Op0.getOperand(0);
5829 RHS = Op0.getOperand(1).getOperand(1);
5831 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5832 if (ConstantSDNode *Op000C =
5833 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5834 if (Op000C->getZExtValue() == 1) {
5835 LHS = Op0.getOperand(1);
5836 RHS = Op0.getOperand(0).getOperand(1);
5838 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5839 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5840 SDValue AndLHS = Op0.getOperand(0);
5841 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5842 LHS = AndLHS.getOperand(0);
5843 RHS = AndLHS.getOperand(1);
5847 if (LHS.getNode()) {
5848 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5849 // instruction. Since the shift amount is in-range-or-undefined, we know
5850 // that doing a bittest on the i16 value is ok. We extend to i32 because
5851 // the encoding for the i16 version is larger than the i32 version.
5852 if (LHS.getValueType() == MVT::i8)
5853 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5855 // If the operand types disagree, extend the shift amount to match. Since
5856 // BT ignores high bits (like shifts) we can use anyextend.
5857 if (LHS.getValueType() != RHS.getValueType())
5858 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5860 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5861 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5862 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5863 DAG.getConstant(Cond, MVT::i8), BT);
5869 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5870 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5871 SDValue Op0 = Op.getOperand(0);
5872 SDValue Op1 = Op.getOperand(1);
5873 DebugLoc dl = Op.getDebugLoc();
5874 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5876 // Optimize to BT if possible.
5877 // Lower (X & (1 << N)) == 0 to BT(X, N).
5878 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5879 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5880 if (Op0.getOpcode() == ISD::AND &&
5882 Op1.getOpcode() == ISD::Constant &&
5883 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5884 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5885 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5886 if (NewSetCC.getNode())
5890 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5891 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5892 if (X86CC == X86::COND_INVALID)
5895 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5897 // Use sbb x, x to materialize carry bit into a GPR.
5898 if (X86CC == X86::COND_B)
5899 return DAG.getNode(ISD::AND, dl, MVT::i8,
5900 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5901 DAG.getConstant(X86CC, MVT::i8), Cond),
5902 DAG.getConstant(1, MVT::i8));
5904 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5905 DAG.getConstant(X86CC, MVT::i8), Cond);
5908 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5910 SDValue Op0 = Op.getOperand(0);
5911 SDValue Op1 = Op.getOperand(1);
5912 SDValue CC = Op.getOperand(2);
5913 EVT VT = Op.getValueType();
5914 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5915 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5916 DebugLoc dl = Op.getDebugLoc();
5920 EVT VT0 = Op0.getValueType();
5921 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5922 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5925 switch (SetCCOpcode) {
5928 case ISD::SETEQ: SSECC = 0; break;
5930 case ISD::SETGT: Swap = true; // Fallthrough
5932 case ISD::SETOLT: SSECC = 1; break;
5934 case ISD::SETGE: Swap = true; // Fallthrough
5936 case ISD::SETOLE: SSECC = 2; break;
5937 case ISD::SETUO: SSECC = 3; break;
5939 case ISD::SETNE: SSECC = 4; break;
5940 case ISD::SETULE: Swap = true;
5941 case ISD::SETUGE: SSECC = 5; break;
5942 case ISD::SETULT: Swap = true;
5943 case ISD::SETUGT: SSECC = 6; break;
5944 case ISD::SETO: SSECC = 7; break;
5947 std::swap(Op0, Op1);
5949 // In the two special cases we can't handle, emit two comparisons.
5951 if (SetCCOpcode == ISD::SETUEQ) {
5953 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5954 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5955 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5957 else if (SetCCOpcode == ISD::SETONE) {
5959 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5960 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5961 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5963 llvm_unreachable("Illegal FP comparison");
5965 // Handle all other FP comparisons here.
5966 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5969 // We are handling one of the integer comparisons here. Since SSE only has
5970 // GT and EQ comparisons for integer, swapping operands and multiple
5971 // operations may be required for some comparisons.
5972 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5973 bool Swap = false, Invert = false, FlipSigns = false;
5975 switch (VT.getSimpleVT().SimpleTy) {
5978 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5980 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5982 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5983 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5986 switch (SetCCOpcode) {
5988 case ISD::SETNE: Invert = true;
5989 case ISD::SETEQ: Opc = EQOpc; break;
5990 case ISD::SETLT: Swap = true;
5991 case ISD::SETGT: Opc = GTOpc; break;
5992 case ISD::SETGE: Swap = true;
5993 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5994 case ISD::SETULT: Swap = true;
5995 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5996 case ISD::SETUGE: Swap = true;
5997 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6000 std::swap(Op0, Op1);
6002 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6003 // bits of the inputs before performing those operations.
6005 EVT EltVT = VT.getVectorElementType();
6006 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6008 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6009 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6011 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6012 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6015 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6017 // If the logical-not of the result is required, perform that now.
6019 Result = DAG.getNOT(dl, Result, VT);
6024 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6025 static bool isX86LogicalCmp(SDValue Op) {
6026 unsigned Opc = Op.getNode()->getOpcode();
6027 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6029 if (Op.getResNo() == 1 &&
6030 (Opc == X86ISD::ADD ||
6031 Opc == X86ISD::SUB ||
6032 Opc == X86ISD::SMUL ||
6033 Opc == X86ISD::UMUL ||
6034 Opc == X86ISD::INC ||
6035 Opc == X86ISD::DEC ||
6036 Opc == X86ISD::OR ||
6037 Opc == X86ISD::XOR ||
6038 Opc == X86ISD::AND))
6044 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6045 bool addTest = true;
6046 SDValue Cond = Op.getOperand(0);
6047 DebugLoc dl = Op.getDebugLoc();
6050 if (Cond.getOpcode() == ISD::SETCC) {
6051 SDValue NewCond = LowerSETCC(Cond, DAG);
6052 if (NewCond.getNode())
6056 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6057 SDValue Op1 = Op.getOperand(1);
6058 SDValue Op2 = Op.getOperand(2);
6059 if (Cond.getOpcode() == X86ISD::SETCC &&
6060 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6061 SDValue Cmp = Cond.getOperand(1);
6062 if (Cmp.getOpcode() == X86ISD::CMP) {
6063 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6064 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6065 ConstantSDNode *RHSC =
6066 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6067 if (N1C && N1C->isAllOnesValue() &&
6068 N2C && N2C->isNullValue() &&
6069 RHSC && RHSC->isNullValue()) {
6070 SDValue CmpOp0 = Cmp.getOperand(0);
6071 Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
6072 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6073 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6074 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6079 // Look pass (and (setcc_carry (cmp ...)), 1).
6080 if (Cond.getOpcode() == ISD::AND &&
6081 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6082 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6083 if (C && C->getAPIntValue() == 1)
6084 Cond = Cond.getOperand(0);
6087 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6088 // setting operand in place of the X86ISD::SETCC.
6089 if (Cond.getOpcode() == X86ISD::SETCC ||
6090 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6091 CC = Cond.getOperand(0);
6093 SDValue Cmp = Cond.getOperand(1);
6094 unsigned Opc = Cmp.getOpcode();
6095 EVT VT = Op.getValueType();
6097 bool IllegalFPCMov = false;
6098 if (VT.isFloatingPoint() && !VT.isVector() &&
6099 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6100 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6102 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6103 Opc == X86ISD::BT) { // FIXME
6110 // Look pass the truncate.
6111 if (Cond.getOpcode() == ISD::TRUNCATE)
6112 Cond = Cond.getOperand(0);
6114 // We know the result of AND is compared against zero. Try to match
6116 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6117 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6118 if (NewSetCC.getNode()) {
6119 CC = NewSetCC.getOperand(0);
6120 Cond = NewSetCC.getOperand(1);
6127 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6128 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6131 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6132 // condition is true.
6133 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6134 SDValue Ops[] = { Op2, Op1, CC, Cond };
6135 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6138 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6139 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6140 // from the AND / OR.
6141 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6142 Opc = Op.getOpcode();
6143 if (Opc != ISD::OR && Opc != ISD::AND)
6145 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6146 Op.getOperand(0).hasOneUse() &&
6147 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6148 Op.getOperand(1).hasOneUse());
6151 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6152 // 1 and that the SETCC node has a single use.
6153 static bool isXor1OfSetCC(SDValue Op) {
6154 if (Op.getOpcode() != ISD::XOR)
6156 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6157 if (N1C && N1C->getAPIntValue() == 1) {
6158 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6159 Op.getOperand(0).hasOneUse();
6164 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6165 bool addTest = true;
6166 SDValue Chain = Op.getOperand(0);
6167 SDValue Cond = Op.getOperand(1);
6168 SDValue Dest = Op.getOperand(2);
6169 DebugLoc dl = Op.getDebugLoc();
6172 if (Cond.getOpcode() == ISD::SETCC) {
6173 SDValue NewCond = LowerSETCC(Cond, DAG);
6174 if (NewCond.getNode())
6178 // FIXME: LowerXALUO doesn't handle these!!
6179 else if (Cond.getOpcode() == X86ISD::ADD ||
6180 Cond.getOpcode() == X86ISD::SUB ||
6181 Cond.getOpcode() == X86ISD::SMUL ||
6182 Cond.getOpcode() == X86ISD::UMUL)
6183 Cond = LowerXALUO(Cond, DAG);
6186 // Look pass (and (setcc_carry (cmp ...)), 1).
6187 if (Cond.getOpcode() == ISD::AND &&
6188 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6189 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6190 if (C && C->getAPIntValue() == 1)
6191 Cond = Cond.getOperand(0);
6194 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6195 // setting operand in place of the X86ISD::SETCC.
6196 if (Cond.getOpcode() == X86ISD::SETCC ||
6197 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6198 CC = Cond.getOperand(0);
6200 SDValue Cmp = Cond.getOperand(1);
6201 unsigned Opc = Cmp.getOpcode();
6202 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6203 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6207 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6211 // These can only come from an arithmetic instruction with overflow,
6212 // e.g. SADDO, UADDO.
6213 Cond = Cond.getNode()->getOperand(1);
6220 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6221 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6222 if (CondOpc == ISD::OR) {
6223 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6224 // two branches instead of an explicit OR instruction with a
6226 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6227 isX86LogicalCmp(Cmp)) {
6228 CC = Cond.getOperand(0).getOperand(0);
6229 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6230 Chain, Dest, CC, Cmp);
6231 CC = Cond.getOperand(1).getOperand(0);
6235 } else { // ISD::AND
6236 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6237 // two branches instead of an explicit AND instruction with a
6238 // separate test. However, we only do this if this block doesn't
6239 // have a fall-through edge, because this requires an explicit
6240 // jmp when the condition is false.
6241 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6242 isX86LogicalCmp(Cmp) &&
6243 Op.getNode()->hasOneUse()) {
6244 X86::CondCode CCode =
6245 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6246 CCode = X86::GetOppositeBranchCondition(CCode);
6247 CC = DAG.getConstant(CCode, MVT::i8);
6248 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6249 // Look for an unconditional branch following this conditional branch.
6250 // We need this because we need to reverse the successors in order
6251 // to implement FCMP_OEQ.
6252 if (User.getOpcode() == ISD::BR) {
6253 SDValue FalseBB = User.getOperand(1);
6255 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6256 assert(NewBR == User);
6259 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6260 Chain, Dest, CC, Cmp);
6261 X86::CondCode CCode =
6262 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6263 CCode = X86::GetOppositeBranchCondition(CCode);
6264 CC = DAG.getConstant(CCode, MVT::i8);
6270 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6271 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6272 // It should be transformed during dag combiner except when the condition
6273 // is set by a arithmetics with overflow node.
6274 X86::CondCode CCode =
6275 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6276 CCode = X86::GetOppositeBranchCondition(CCode);
6277 CC = DAG.getConstant(CCode, MVT::i8);
6278 Cond = Cond.getOperand(0).getOperand(1);
6284 // Look pass the truncate.
6285 if (Cond.getOpcode() == ISD::TRUNCATE)
6286 Cond = Cond.getOperand(0);
6288 // We know the result of AND is compared against zero. Try to match
6290 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6291 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6292 if (NewSetCC.getNode()) {
6293 CC = NewSetCC.getOperand(0);
6294 Cond = NewSetCC.getOperand(1);
6301 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6302 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6304 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6305 Chain, Dest, CC, Cond);
6309 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6310 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6311 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6312 // that the guard pages used by the OS virtual memory manager are allocated in
6313 // correct sequence.
6315 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6316 SelectionDAG &DAG) {
6317 assert(Subtarget->isTargetCygMing() &&
6318 "This should be used only on Cygwin/Mingw targets");
6319 DebugLoc dl = Op.getDebugLoc();
6322 SDValue Chain = Op.getOperand(0);
6323 SDValue Size = Op.getOperand(1);
6324 // FIXME: Ensure alignment here
6328 EVT IntPtr = getPointerTy();
6329 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6331 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6333 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6334 Flag = Chain.getValue(1);
6336 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6337 SDValue Ops[] = { Chain,
6338 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6339 DAG.getRegister(X86::EAX, IntPtr),
6340 DAG.getRegister(X86StackPtr, SPTy),
6342 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6343 Flag = Chain.getValue(1);
6345 Chain = DAG.getCALLSEQ_END(Chain,
6346 DAG.getIntPtrConstant(0, true),
6347 DAG.getIntPtrConstant(0, true),
6350 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6352 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6353 return DAG.getMergeValues(Ops1, 2, dl);
6357 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6359 SDValue Dst, SDValue Src,
6360 SDValue Size, unsigned Align,
6362 uint64_t DstSVOff) {
6363 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6365 // If not DWORD aligned or size is more than the threshold, call the library.
6366 // The libc version is likely to be faster for these cases. It can use the
6367 // address value and run time information about the CPU.
6368 if ((Align & 3) != 0 ||
6370 ConstantSize->getZExtValue() >
6371 getSubtarget()->getMaxInlineSizeThreshold()) {
6372 SDValue InFlag(0, 0);
6374 // Check to see if there is a specialized entry-point for memory zeroing.
6375 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6377 if (const char *bzeroEntry = V &&
6378 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6379 EVT IntPtr = getPointerTy();
6380 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6381 TargetLowering::ArgListTy Args;
6382 TargetLowering::ArgListEntry Entry;
6384 Entry.Ty = IntPtrTy;
6385 Args.push_back(Entry);
6387 Args.push_back(Entry);
6388 std::pair<SDValue,SDValue> CallResult =
6389 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6390 false, false, false, false,
6391 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6392 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6393 DAG.GetOrdering(Chain.getNode()));
6394 return CallResult.second;
6397 // Otherwise have the target-independent code call memset.
6401 uint64_t SizeVal = ConstantSize->getZExtValue();
6402 SDValue InFlag(0, 0);
6405 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6406 unsigned BytesLeft = 0;
6407 bool TwoRepStos = false;
6410 uint64_t Val = ValC->getZExtValue() & 255;
6412 // If the value is a constant, then we can potentially use larger sets.
6413 switch (Align & 3) {
6414 case 2: // WORD aligned
6417 Val = (Val << 8) | Val;
6419 case 0: // DWORD aligned
6422 Val = (Val << 8) | Val;
6423 Val = (Val << 16) | Val;
6424 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6427 Val = (Val << 32) | Val;
6430 default: // Byte aligned
6433 Count = DAG.getIntPtrConstant(SizeVal);
6437 if (AVT.bitsGT(MVT::i8)) {
6438 unsigned UBytes = AVT.getSizeInBits() / 8;
6439 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6440 BytesLeft = SizeVal % UBytes;
6443 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6445 InFlag = Chain.getValue(1);
6448 Count = DAG.getIntPtrConstant(SizeVal);
6449 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6450 InFlag = Chain.getValue(1);
6453 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6456 InFlag = Chain.getValue(1);
6457 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6460 InFlag = Chain.getValue(1);
6462 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6463 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6464 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6467 InFlag = Chain.getValue(1);
6469 EVT CVT = Count.getValueType();
6470 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6471 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6472 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6475 InFlag = Chain.getValue(1);
6476 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6477 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6478 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6479 } else if (BytesLeft) {
6480 // Handle the last 1 - 7 bytes.
6481 unsigned Offset = SizeVal - BytesLeft;
6482 EVT AddrVT = Dst.getValueType();
6483 EVT SizeVT = Size.getValueType();
6485 Chain = DAG.getMemset(Chain, dl,
6486 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6487 DAG.getConstant(Offset, AddrVT)),
6489 DAG.getConstant(BytesLeft, SizeVT),
6490 Align, DstSV, DstSVOff + Offset);
6493 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6498 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6499 SDValue Chain, SDValue Dst, SDValue Src,
6500 SDValue Size, unsigned Align,
6502 const Value *DstSV, uint64_t DstSVOff,
6503 const Value *SrcSV, uint64_t SrcSVOff) {
6504 // This requires the copy size to be a constant, preferrably
6505 // within a subtarget-specific limit.
6506 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6509 uint64_t SizeVal = ConstantSize->getZExtValue();
6510 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6513 /// If not DWORD aligned, call the library.
6514 if ((Align & 3) != 0)
6519 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6522 unsigned UBytes = AVT.getSizeInBits() / 8;
6523 unsigned CountVal = SizeVal / UBytes;
6524 SDValue Count = DAG.getIntPtrConstant(CountVal);
6525 unsigned BytesLeft = SizeVal % UBytes;
6527 SDValue InFlag(0, 0);
6528 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6531 InFlag = Chain.getValue(1);
6532 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6535 InFlag = Chain.getValue(1);
6536 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6539 InFlag = Chain.getValue(1);
6541 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6542 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6543 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6544 array_lengthof(Ops));
6546 SmallVector<SDValue, 4> Results;
6547 Results.push_back(RepMovs);
6549 // Handle the last 1 - 7 bytes.
6550 unsigned Offset = SizeVal - BytesLeft;
6551 EVT DstVT = Dst.getValueType();
6552 EVT SrcVT = Src.getValueType();
6553 EVT SizeVT = Size.getValueType();
6554 Results.push_back(DAG.getMemcpy(Chain, dl,
6555 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6556 DAG.getConstant(Offset, DstVT)),
6557 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6558 DAG.getConstant(Offset, SrcVT)),
6559 DAG.getConstant(BytesLeft, SizeVT),
6560 Align, AlwaysInline,
6561 DstSV, DstSVOff + Offset,
6562 SrcSV, SrcSVOff + Offset));
6565 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6566 &Results[0], Results.size());
6569 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6570 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6571 DebugLoc dl = Op.getDebugLoc();
6573 if (!Subtarget->is64Bit()) {
6574 // vastart just stores the address of the VarArgsFrameIndex slot into the
6575 // memory location argument.
6576 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6577 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6581 // gp_offset (0 - 6 * 8)
6582 // fp_offset (48 - 48 + 8 * 16)
6583 // overflow_arg_area (point to parameters coming in memory).
6585 SmallVector<SDValue, 8> MemOps;
6586 SDValue FIN = Op.getOperand(1);
6588 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6589 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6591 MemOps.push_back(Store);
6594 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6595 FIN, DAG.getIntPtrConstant(4));
6596 Store = DAG.getStore(Op.getOperand(0), dl,
6597 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6599 MemOps.push_back(Store);
6601 // Store ptr to overflow_arg_area
6602 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6603 FIN, DAG.getIntPtrConstant(4));
6604 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6605 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6606 MemOps.push_back(Store);
6608 // Store ptr to reg_save_area.
6609 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6610 FIN, DAG.getIntPtrConstant(8));
6611 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6612 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6613 MemOps.push_back(Store);
6614 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6615 &MemOps[0], MemOps.size());
6618 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6619 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6620 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6621 SDValue Chain = Op.getOperand(0);
6622 SDValue SrcPtr = Op.getOperand(1);
6623 SDValue SrcSV = Op.getOperand(2);
6625 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6629 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6630 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6631 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6632 SDValue Chain = Op.getOperand(0);
6633 SDValue DstPtr = Op.getOperand(1);
6634 SDValue SrcPtr = Op.getOperand(2);
6635 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6636 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6637 DebugLoc dl = Op.getDebugLoc();
6639 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6640 DAG.getIntPtrConstant(24), 8, false,
6641 DstSV, 0, SrcSV, 0);
6645 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6646 DebugLoc dl = Op.getDebugLoc();
6647 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6649 default: return SDValue(); // Don't custom lower most intrinsics.
6650 // Comparison intrinsics.
6651 case Intrinsic::x86_sse_comieq_ss:
6652 case Intrinsic::x86_sse_comilt_ss:
6653 case Intrinsic::x86_sse_comile_ss:
6654 case Intrinsic::x86_sse_comigt_ss:
6655 case Intrinsic::x86_sse_comige_ss:
6656 case Intrinsic::x86_sse_comineq_ss:
6657 case Intrinsic::x86_sse_ucomieq_ss:
6658 case Intrinsic::x86_sse_ucomilt_ss:
6659 case Intrinsic::x86_sse_ucomile_ss:
6660 case Intrinsic::x86_sse_ucomigt_ss:
6661 case Intrinsic::x86_sse_ucomige_ss:
6662 case Intrinsic::x86_sse_ucomineq_ss:
6663 case Intrinsic::x86_sse2_comieq_sd:
6664 case Intrinsic::x86_sse2_comilt_sd:
6665 case Intrinsic::x86_sse2_comile_sd:
6666 case Intrinsic::x86_sse2_comigt_sd:
6667 case Intrinsic::x86_sse2_comige_sd:
6668 case Intrinsic::x86_sse2_comineq_sd:
6669 case Intrinsic::x86_sse2_ucomieq_sd:
6670 case Intrinsic::x86_sse2_ucomilt_sd:
6671 case Intrinsic::x86_sse2_ucomile_sd:
6672 case Intrinsic::x86_sse2_ucomigt_sd:
6673 case Intrinsic::x86_sse2_ucomige_sd:
6674 case Intrinsic::x86_sse2_ucomineq_sd: {
6676 ISD::CondCode CC = ISD::SETCC_INVALID;
6679 case Intrinsic::x86_sse_comieq_ss:
6680 case Intrinsic::x86_sse2_comieq_sd:
6684 case Intrinsic::x86_sse_comilt_ss:
6685 case Intrinsic::x86_sse2_comilt_sd:
6689 case Intrinsic::x86_sse_comile_ss:
6690 case Intrinsic::x86_sse2_comile_sd:
6694 case Intrinsic::x86_sse_comigt_ss:
6695 case Intrinsic::x86_sse2_comigt_sd:
6699 case Intrinsic::x86_sse_comige_ss:
6700 case Intrinsic::x86_sse2_comige_sd:
6704 case Intrinsic::x86_sse_comineq_ss:
6705 case Intrinsic::x86_sse2_comineq_sd:
6709 case Intrinsic::x86_sse_ucomieq_ss:
6710 case Intrinsic::x86_sse2_ucomieq_sd:
6711 Opc = X86ISD::UCOMI;
6714 case Intrinsic::x86_sse_ucomilt_ss:
6715 case Intrinsic::x86_sse2_ucomilt_sd:
6716 Opc = X86ISD::UCOMI;
6719 case Intrinsic::x86_sse_ucomile_ss:
6720 case Intrinsic::x86_sse2_ucomile_sd:
6721 Opc = X86ISD::UCOMI;
6724 case Intrinsic::x86_sse_ucomigt_ss:
6725 case Intrinsic::x86_sse2_ucomigt_sd:
6726 Opc = X86ISD::UCOMI;
6729 case Intrinsic::x86_sse_ucomige_ss:
6730 case Intrinsic::x86_sse2_ucomige_sd:
6731 Opc = X86ISD::UCOMI;
6734 case Intrinsic::x86_sse_ucomineq_ss:
6735 case Intrinsic::x86_sse2_ucomineq_sd:
6736 Opc = X86ISD::UCOMI;
6741 SDValue LHS = Op.getOperand(1);
6742 SDValue RHS = Op.getOperand(2);
6743 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6744 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6745 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6746 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6747 DAG.getConstant(X86CC, MVT::i8), Cond);
6748 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6750 // ptest intrinsics. The intrinsic these come from are designed to return
6751 // an integer value, not just an instruction so lower it to the ptest
6752 // pattern and a setcc for the result.
6753 case Intrinsic::x86_sse41_ptestz:
6754 case Intrinsic::x86_sse41_ptestc:
6755 case Intrinsic::x86_sse41_ptestnzc:{
6758 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6759 case Intrinsic::x86_sse41_ptestz:
6761 X86CC = X86::COND_E;
6763 case Intrinsic::x86_sse41_ptestc:
6765 X86CC = X86::COND_B;
6767 case Intrinsic::x86_sse41_ptestnzc:
6769 X86CC = X86::COND_A;
6773 SDValue LHS = Op.getOperand(1);
6774 SDValue RHS = Op.getOperand(2);
6775 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6776 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6777 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6778 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6781 // Fix vector shift instructions where the last operand is a non-immediate
6783 case Intrinsic::x86_sse2_pslli_w:
6784 case Intrinsic::x86_sse2_pslli_d:
6785 case Intrinsic::x86_sse2_pslli_q:
6786 case Intrinsic::x86_sse2_psrli_w:
6787 case Intrinsic::x86_sse2_psrli_d:
6788 case Intrinsic::x86_sse2_psrli_q:
6789 case Intrinsic::x86_sse2_psrai_w:
6790 case Intrinsic::x86_sse2_psrai_d:
6791 case Intrinsic::x86_mmx_pslli_w:
6792 case Intrinsic::x86_mmx_pslli_d:
6793 case Intrinsic::x86_mmx_pslli_q:
6794 case Intrinsic::x86_mmx_psrli_w:
6795 case Intrinsic::x86_mmx_psrli_d:
6796 case Intrinsic::x86_mmx_psrli_q:
6797 case Intrinsic::x86_mmx_psrai_w:
6798 case Intrinsic::x86_mmx_psrai_d: {
6799 SDValue ShAmt = Op.getOperand(2);
6800 if (isa<ConstantSDNode>(ShAmt))
6803 unsigned NewIntNo = 0;
6804 EVT ShAmtVT = MVT::v4i32;
6806 case Intrinsic::x86_sse2_pslli_w:
6807 NewIntNo = Intrinsic::x86_sse2_psll_w;
6809 case Intrinsic::x86_sse2_pslli_d:
6810 NewIntNo = Intrinsic::x86_sse2_psll_d;
6812 case Intrinsic::x86_sse2_pslli_q:
6813 NewIntNo = Intrinsic::x86_sse2_psll_q;
6815 case Intrinsic::x86_sse2_psrli_w:
6816 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6818 case Intrinsic::x86_sse2_psrli_d:
6819 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6821 case Intrinsic::x86_sse2_psrli_q:
6822 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6824 case Intrinsic::x86_sse2_psrai_w:
6825 NewIntNo = Intrinsic::x86_sse2_psra_w;
6827 case Intrinsic::x86_sse2_psrai_d:
6828 NewIntNo = Intrinsic::x86_sse2_psra_d;
6831 ShAmtVT = MVT::v2i32;
6833 case Intrinsic::x86_mmx_pslli_w:
6834 NewIntNo = Intrinsic::x86_mmx_psll_w;
6836 case Intrinsic::x86_mmx_pslli_d:
6837 NewIntNo = Intrinsic::x86_mmx_psll_d;
6839 case Intrinsic::x86_mmx_pslli_q:
6840 NewIntNo = Intrinsic::x86_mmx_psll_q;
6842 case Intrinsic::x86_mmx_psrli_w:
6843 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6845 case Intrinsic::x86_mmx_psrli_d:
6846 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6848 case Intrinsic::x86_mmx_psrli_q:
6849 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6851 case Intrinsic::x86_mmx_psrai_w:
6852 NewIntNo = Intrinsic::x86_mmx_psra_w;
6854 case Intrinsic::x86_mmx_psrai_d:
6855 NewIntNo = Intrinsic::x86_mmx_psra_d;
6857 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6863 // The vector shift intrinsics with scalars uses 32b shift amounts but
6864 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6868 ShOps[1] = DAG.getConstant(0, MVT::i32);
6869 if (ShAmtVT == MVT::v4i32) {
6870 ShOps[2] = DAG.getUNDEF(MVT::i32);
6871 ShOps[3] = DAG.getUNDEF(MVT::i32);
6872 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6874 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6877 EVT VT = Op.getValueType();
6878 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6879 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6880 DAG.getConstant(NewIntNo, MVT::i32),
6881 Op.getOperand(1), ShAmt);
6886 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6887 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6888 DebugLoc dl = Op.getDebugLoc();
6891 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6893 DAG.getConstant(TD->getPointerSize(),
6894 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6895 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6896 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6901 // Just load the return address.
6902 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6903 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6904 RetAddrFI, NULL, 0);
6907 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6908 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6909 MFI->setFrameAddressIsTaken(true);
6910 EVT VT = Op.getValueType();
6911 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6912 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6913 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6914 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6916 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6920 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6921 SelectionDAG &DAG) {
6922 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6925 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6927 MachineFunction &MF = DAG.getMachineFunction();
6928 SDValue Chain = Op.getOperand(0);
6929 SDValue Offset = Op.getOperand(1);
6930 SDValue Handler = Op.getOperand(2);
6931 DebugLoc dl = Op.getDebugLoc();
6933 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6935 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6937 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6938 DAG.getIntPtrConstant(-TD->getPointerSize()));
6939 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6940 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6941 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6942 MF.getRegInfo().addLiveOut(StoreAddrReg);
6944 return DAG.getNode(X86ISD::EH_RETURN, dl,
6946 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6949 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6950 SelectionDAG &DAG) {
6951 SDValue Root = Op.getOperand(0);
6952 SDValue Trmp = Op.getOperand(1); // trampoline
6953 SDValue FPtr = Op.getOperand(2); // nested function
6954 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6955 DebugLoc dl = Op.getDebugLoc();
6957 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6959 const X86InstrInfo *TII =
6960 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6962 if (Subtarget->is64Bit()) {
6963 SDValue OutChains[6];
6965 // Large code-model.
6967 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6968 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6970 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6971 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6973 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6975 // Load the pointer to the nested function into R11.
6976 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6977 SDValue Addr = Trmp;
6978 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6981 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6982 DAG.getConstant(2, MVT::i64));
6983 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6985 // Load the 'nest' parameter value into R10.
6986 // R10 is specified in X86CallingConv.td
6987 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6988 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6989 DAG.getConstant(10, MVT::i64));
6990 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6991 Addr, TrmpAddr, 10);
6993 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6994 DAG.getConstant(12, MVT::i64));
6995 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6997 // Jump to the nested function.
6998 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6999 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7000 DAG.getConstant(20, MVT::i64));
7001 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7002 Addr, TrmpAddr, 20);
7004 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7005 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7006 DAG.getConstant(22, MVT::i64));
7007 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7011 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7012 return DAG.getMergeValues(Ops, 2, dl);
7014 const Function *Func =
7015 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7016 CallingConv::ID CC = Func->getCallingConv();
7021 llvm_unreachable("Unsupported calling convention");
7022 case CallingConv::C:
7023 case CallingConv::X86_StdCall: {
7024 // Pass 'nest' parameter in ECX.
7025 // Must be kept in sync with X86CallingConv.td
7028 // Check that ECX wasn't needed by an 'inreg' parameter.
7029 const FunctionType *FTy = Func->getFunctionType();
7030 const AttrListPtr &Attrs = Func->getAttributes();
7032 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7033 unsigned InRegCount = 0;
7036 for (FunctionType::param_iterator I = FTy->param_begin(),
7037 E = FTy->param_end(); I != E; ++I, ++Idx)
7038 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7039 // FIXME: should only count parameters that are lowered to integers.
7040 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7042 if (InRegCount > 2) {
7043 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7048 case CallingConv::X86_FastCall:
7049 case CallingConv::Fast:
7050 // Pass 'nest' parameter in EAX.
7051 // Must be kept in sync with X86CallingConv.td
7056 SDValue OutChains[4];
7059 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7060 DAG.getConstant(10, MVT::i32));
7061 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7063 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
7064 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7065 OutChains[0] = DAG.getStore(Root, dl,
7066 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7069 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7070 DAG.getConstant(1, MVT::i32));
7071 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
7073 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
7074 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7075 DAG.getConstant(5, MVT::i32));
7076 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7077 TrmpAddr, 5, false, 1);
7079 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7080 DAG.getConstant(6, MVT::i32));
7081 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
7084 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7085 return DAG.getMergeValues(Ops, 2, dl);
7089 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7091 The rounding mode is in bits 11:10 of FPSR, and has the following
7098 FLT_ROUNDS, on the other hand, expects the following:
7105 To perform the conversion, we do:
7106 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7109 MachineFunction &MF = DAG.getMachineFunction();
7110 const TargetMachine &TM = MF.getTarget();
7111 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7112 unsigned StackAlignment = TFI.getStackAlignment();
7113 EVT VT = Op.getValueType();
7114 DebugLoc dl = Op.getDebugLoc();
7116 // Save FP Control Word to stack slot
7117 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7118 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7120 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7121 DAG.getEntryNode(), StackSlot);
7123 // Load FP Control Word from stack slot
7124 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
7126 // Transform as necessary
7128 DAG.getNode(ISD::SRL, dl, MVT::i16,
7129 DAG.getNode(ISD::AND, dl, MVT::i16,
7130 CWD, DAG.getConstant(0x800, MVT::i16)),
7131 DAG.getConstant(11, MVT::i8));
7133 DAG.getNode(ISD::SRL, dl, MVT::i16,
7134 DAG.getNode(ISD::AND, dl, MVT::i16,
7135 CWD, DAG.getConstant(0x400, MVT::i16)),
7136 DAG.getConstant(9, MVT::i8));
7139 DAG.getNode(ISD::AND, dl, MVT::i16,
7140 DAG.getNode(ISD::ADD, dl, MVT::i16,
7141 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7142 DAG.getConstant(1, MVT::i16)),
7143 DAG.getConstant(3, MVT::i16));
7146 return DAG.getNode((VT.getSizeInBits() < 16 ?
7147 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7150 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7151 EVT VT = Op.getValueType();
7153 unsigned NumBits = VT.getSizeInBits();
7154 DebugLoc dl = Op.getDebugLoc();
7156 Op = Op.getOperand(0);
7157 if (VT == MVT::i8) {
7158 // Zero extend to i32 since there is not an i8 bsr.
7160 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7163 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7164 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7165 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7167 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7170 DAG.getConstant(NumBits+NumBits-1, OpVT),
7171 DAG.getConstant(X86::COND_E, MVT::i8),
7174 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7176 // Finally xor with NumBits-1.
7177 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7180 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7184 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7185 EVT VT = Op.getValueType();
7187 unsigned NumBits = VT.getSizeInBits();
7188 DebugLoc dl = Op.getDebugLoc();
7190 Op = Op.getOperand(0);
7191 if (VT == MVT::i8) {
7193 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7196 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7197 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7198 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7200 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7203 DAG.getConstant(NumBits, OpVT),
7204 DAG.getConstant(X86::COND_E, MVT::i8),
7207 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7210 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7214 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7215 EVT VT = Op.getValueType();
7216 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7217 DebugLoc dl = Op.getDebugLoc();
7219 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7220 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7221 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7222 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7223 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7225 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7226 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7227 // return AloBlo + AloBhi + AhiBlo;
7229 SDValue A = Op.getOperand(0);
7230 SDValue B = Op.getOperand(1);
7232 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7233 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7234 A, DAG.getConstant(32, MVT::i32));
7235 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7236 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7237 B, DAG.getConstant(32, MVT::i32));
7238 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7239 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7241 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7242 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7244 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7245 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7247 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7248 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7249 AloBhi, DAG.getConstant(32, MVT::i32));
7250 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7251 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7252 AhiBlo, DAG.getConstant(32, MVT::i32));
7253 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7254 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7259 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7260 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7261 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7262 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7263 // has only one use.
7264 SDNode *N = Op.getNode();
7265 SDValue LHS = N->getOperand(0);
7266 SDValue RHS = N->getOperand(1);
7267 unsigned BaseOp = 0;
7269 DebugLoc dl = Op.getDebugLoc();
7271 switch (Op.getOpcode()) {
7272 default: llvm_unreachable("Unknown ovf instruction!");
7274 // A subtract of one will be selected as a INC. Note that INC doesn't
7275 // set CF, so we can't do this for UADDO.
7276 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7277 if (C->getAPIntValue() == 1) {
7278 BaseOp = X86ISD::INC;
7282 BaseOp = X86ISD::ADD;
7286 BaseOp = X86ISD::ADD;
7290 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7291 // set CF, so we can't do this for USUBO.
7292 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7293 if (C->getAPIntValue() == 1) {
7294 BaseOp = X86ISD::DEC;
7298 BaseOp = X86ISD::SUB;
7302 BaseOp = X86ISD::SUB;
7306 BaseOp = X86ISD::SMUL;
7310 BaseOp = X86ISD::UMUL;
7315 // Also sets EFLAGS.
7316 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7317 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7320 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7321 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7323 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7327 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7328 EVT T = Op.getValueType();
7329 DebugLoc dl = Op.getDebugLoc();
7332 switch(T.getSimpleVT().SimpleTy) {
7334 assert(false && "Invalid value type!");
7335 case MVT::i8: Reg = X86::AL; size = 1; break;
7336 case MVT::i16: Reg = X86::AX; size = 2; break;
7337 case MVT::i32: Reg = X86::EAX; size = 4; break;
7339 assert(Subtarget->is64Bit() && "Node not type legal!");
7340 Reg = X86::RAX; size = 8;
7343 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7344 Op.getOperand(2), SDValue());
7345 SDValue Ops[] = { cpIn.getValue(0),
7348 DAG.getTargetConstant(size, MVT::i8),
7350 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7351 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7353 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7357 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7358 SelectionDAG &DAG) {
7359 assert(Subtarget->is64Bit() && "Result not type legalized?");
7360 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7361 SDValue TheChain = Op.getOperand(0);
7362 DebugLoc dl = Op.getDebugLoc();
7363 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7364 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7365 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7367 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7368 DAG.getConstant(32, MVT::i8));
7370 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7373 return DAG.getMergeValues(Ops, 2, dl);
7376 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7377 SDNode *Node = Op.getNode();
7378 DebugLoc dl = Node->getDebugLoc();
7379 EVT T = Node->getValueType(0);
7380 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7381 DAG.getConstant(0, T), Node->getOperand(2));
7382 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7383 cast<AtomicSDNode>(Node)->getMemoryVT(),
7384 Node->getOperand(0),
7385 Node->getOperand(1), negOp,
7386 cast<AtomicSDNode>(Node)->getSrcValue(),
7387 cast<AtomicSDNode>(Node)->getAlignment());
7390 /// LowerOperation - Provide custom lowering hooks for some operations.
7392 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7393 switch (Op.getOpcode()) {
7394 default: llvm_unreachable("Should not custom lower this!");
7395 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7396 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7397 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7398 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7399 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7400 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7401 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7402 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7403 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7404 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7405 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7406 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7407 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7408 case ISD::SHL_PARTS:
7409 case ISD::SRA_PARTS:
7410 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7411 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7412 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7413 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7414 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7415 case ISD::FABS: return LowerFABS(Op, DAG);
7416 case ISD::FNEG: return LowerFNEG(Op, DAG);
7417 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7418 case ISD::SETCC: return LowerSETCC(Op, DAG);
7419 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7420 case ISD::SELECT: return LowerSELECT(Op, DAG);
7421 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7422 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7423 case ISD::VASTART: return LowerVASTART(Op, DAG);
7424 case ISD::VAARG: return LowerVAARG(Op, DAG);
7425 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7426 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7427 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7428 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7429 case ISD::FRAME_TO_ARGS_OFFSET:
7430 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7431 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7432 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7433 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7434 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7435 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7436 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7437 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7443 case ISD::UMULO: return LowerXALUO(Op, DAG);
7444 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7448 void X86TargetLowering::
7449 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7450 SelectionDAG &DAG, unsigned NewOp) {
7451 EVT T = Node->getValueType(0);
7452 DebugLoc dl = Node->getDebugLoc();
7453 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7455 SDValue Chain = Node->getOperand(0);
7456 SDValue In1 = Node->getOperand(1);
7457 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7458 Node->getOperand(2), DAG.getIntPtrConstant(0));
7459 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7460 Node->getOperand(2), DAG.getIntPtrConstant(1));
7461 SDValue Ops[] = { Chain, In1, In2L, In2H };
7462 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7464 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7465 cast<MemSDNode>(Node)->getMemOperand());
7466 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7467 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7468 Results.push_back(Result.getValue(2));
7471 /// ReplaceNodeResults - Replace a node with an illegal result type
7472 /// with a new node built out of custom code.
7473 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7474 SmallVectorImpl<SDValue>&Results,
7475 SelectionDAG &DAG) {
7476 DebugLoc dl = N->getDebugLoc();
7477 switch (N->getOpcode()) {
7479 assert(false && "Do not know how to custom type legalize this operation!");
7481 case ISD::FP_TO_SINT: {
7482 std::pair<SDValue,SDValue> Vals =
7483 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7484 SDValue FIST = Vals.first, StackSlot = Vals.second;
7485 if (FIST.getNode() != 0) {
7486 EVT VT = N->getValueType(0);
7487 // Return a load from the stack slot.
7488 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7492 case ISD::READCYCLECOUNTER: {
7493 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7494 SDValue TheChain = N->getOperand(0);
7495 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7496 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7498 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7500 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7501 SDValue Ops[] = { eax, edx };
7502 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7503 Results.push_back(edx.getValue(1));
7510 EVT WidenVT = getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
7511 Results.push_back(DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()));
7514 case ISD::ATOMIC_CMP_SWAP: {
7515 EVT T = N->getValueType(0);
7516 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7517 SDValue cpInL, cpInH;
7518 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7519 DAG.getConstant(0, MVT::i32));
7520 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7521 DAG.getConstant(1, MVT::i32));
7522 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7523 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7525 SDValue swapInL, swapInH;
7526 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7527 DAG.getConstant(0, MVT::i32));
7528 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7529 DAG.getConstant(1, MVT::i32));
7530 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7532 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7533 swapInL.getValue(1));
7534 SDValue Ops[] = { swapInH.getValue(0),
7536 swapInH.getValue(1) };
7537 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7538 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7539 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7540 MVT::i32, Result.getValue(1));
7541 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7542 MVT::i32, cpOutL.getValue(2));
7543 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7544 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7545 Results.push_back(cpOutH.getValue(1));
7548 case ISD::ATOMIC_LOAD_ADD:
7549 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7551 case ISD::ATOMIC_LOAD_AND:
7552 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7554 case ISD::ATOMIC_LOAD_NAND:
7555 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7557 case ISD::ATOMIC_LOAD_OR:
7558 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7560 case ISD::ATOMIC_LOAD_SUB:
7561 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7563 case ISD::ATOMIC_LOAD_XOR:
7564 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7566 case ISD::ATOMIC_SWAP:
7567 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7572 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7574 default: return NULL;
7575 case X86ISD::BSF: return "X86ISD::BSF";
7576 case X86ISD::BSR: return "X86ISD::BSR";
7577 case X86ISD::SHLD: return "X86ISD::SHLD";
7578 case X86ISD::SHRD: return "X86ISD::SHRD";
7579 case X86ISD::FAND: return "X86ISD::FAND";
7580 case X86ISD::FOR: return "X86ISD::FOR";
7581 case X86ISD::FXOR: return "X86ISD::FXOR";
7582 case X86ISD::FSRL: return "X86ISD::FSRL";
7583 case X86ISD::FILD: return "X86ISD::FILD";
7584 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7585 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7586 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7587 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7588 case X86ISD::FLD: return "X86ISD::FLD";
7589 case X86ISD::FST: return "X86ISD::FST";
7590 case X86ISD::CALL: return "X86ISD::CALL";
7591 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7592 case X86ISD::BT: return "X86ISD::BT";
7593 case X86ISD::CMP: return "X86ISD::CMP";
7594 case X86ISD::COMI: return "X86ISD::COMI";
7595 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7596 case X86ISD::SETCC: return "X86ISD::SETCC";
7597 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7598 case X86ISD::CMOV: return "X86ISD::CMOV";
7599 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7600 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7601 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7602 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7603 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7604 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7605 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7606 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7607 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7608 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7609 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7610 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7611 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7612 case X86ISD::FMAX: return "X86ISD::FMAX";
7613 case X86ISD::FMIN: return "X86ISD::FMIN";
7614 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7615 case X86ISD::FRCP: return "X86ISD::FRCP";
7616 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7617 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7618 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7619 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7620 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7621 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7622 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7623 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7624 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7625 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7626 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7627 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7628 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7629 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7630 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7631 case X86ISD::VSHL: return "X86ISD::VSHL";
7632 case X86ISD::VSRL: return "X86ISD::VSRL";
7633 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7634 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7635 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7636 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7637 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7638 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7639 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7640 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7641 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7642 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7643 case X86ISD::ADD: return "X86ISD::ADD";
7644 case X86ISD::SUB: return "X86ISD::SUB";
7645 case X86ISD::SMUL: return "X86ISD::SMUL";
7646 case X86ISD::UMUL: return "X86ISD::UMUL";
7647 case X86ISD::INC: return "X86ISD::INC";
7648 case X86ISD::DEC: return "X86ISD::DEC";
7649 case X86ISD::OR: return "X86ISD::OR";
7650 case X86ISD::XOR: return "X86ISD::XOR";
7651 case X86ISD::AND: return "X86ISD::AND";
7652 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7653 case X86ISD::PTEST: return "X86ISD::PTEST";
7654 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7658 // isLegalAddressingMode - Return true if the addressing mode represented
7659 // by AM is legal for this target, for a load/store of the specified type.
7660 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7661 const Type *Ty) const {
7662 // X86 supports extremely general addressing modes.
7663 CodeModel::Model M = getTargetMachine().getCodeModel();
7665 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7666 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7671 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7673 // If a reference to this global requires an extra load, we can't fold it.
7674 if (isGlobalStubReference(GVFlags))
7677 // If BaseGV requires a register for the PIC base, we cannot also have a
7678 // BaseReg specified.
7679 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7682 // If lower 4G is not available, then we must use rip-relative addressing.
7683 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7693 // These scales always work.
7698 // These scales are formed with basereg+scalereg. Only accept if there is
7703 default: // Other stuff never works.
7711 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7712 if (!Ty1->isInteger() || !Ty2->isInteger())
7714 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7715 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7716 if (NumBits1 <= NumBits2)
7718 return Subtarget->is64Bit() || NumBits1 < 64;
7721 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7722 if (!VT1.isInteger() || !VT2.isInteger())
7724 unsigned NumBits1 = VT1.getSizeInBits();
7725 unsigned NumBits2 = VT2.getSizeInBits();
7726 if (NumBits1 <= NumBits2)
7728 return Subtarget->is64Bit() || NumBits1 < 64;
7731 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7732 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7733 return Ty1->isInteger(32) && Ty2->isInteger(64) && Subtarget->is64Bit();
7736 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7737 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7738 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7741 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7742 // i16 instructions are longer (0x66 prefix) and potentially slower.
7743 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7746 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7747 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7748 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7749 /// are assumed to be legal.
7751 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7753 // Only do shuffles on 128-bit vector types for now.
7754 if (VT.getSizeInBits() == 64)
7757 // FIXME: pshufb, blends, shifts.
7758 return (VT.getVectorNumElements() == 2 ||
7759 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7760 isMOVLMask(M, VT) ||
7761 isSHUFPMask(M, VT) ||
7762 isPSHUFDMask(M, VT) ||
7763 isPSHUFHWMask(M, VT) ||
7764 isPSHUFLWMask(M, VT) ||
7765 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7766 isUNPCKLMask(M, VT) ||
7767 isUNPCKHMask(M, VT) ||
7768 isUNPCKL_v_undef_Mask(M, VT) ||
7769 isUNPCKH_v_undef_Mask(M, VT));
7773 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7775 unsigned NumElts = VT.getVectorNumElements();
7776 // FIXME: This collection of masks seems suspect.
7779 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7780 return (isMOVLMask(Mask, VT) ||
7781 isCommutedMOVLMask(Mask, VT, true) ||
7782 isSHUFPMask(Mask, VT) ||
7783 isCommutedSHUFPMask(Mask, VT));
7788 //===----------------------------------------------------------------------===//
7789 // X86 Scheduler Hooks
7790 //===----------------------------------------------------------------------===//
7792 // private utility function
7794 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7795 MachineBasicBlock *MBB,
7803 TargetRegisterClass *RC,
7804 bool invSrc) const {
7805 // For the atomic bitwise operator, we generate
7808 // ld t1 = [bitinstr.addr]
7809 // op t2 = t1, [bitinstr.val]
7811 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7813 // fallthrough -->nextMBB
7814 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7815 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7816 MachineFunction::iterator MBBIter = MBB;
7819 /// First build the CFG
7820 MachineFunction *F = MBB->getParent();
7821 MachineBasicBlock *thisMBB = MBB;
7822 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7823 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7824 F->insert(MBBIter, newMBB);
7825 F->insert(MBBIter, nextMBB);
7827 // Move all successors to thisMBB to nextMBB
7828 nextMBB->transferSuccessors(thisMBB);
7830 // Update thisMBB to fall through to newMBB
7831 thisMBB->addSuccessor(newMBB);
7833 // newMBB jumps to itself and fall through to nextMBB
7834 newMBB->addSuccessor(nextMBB);
7835 newMBB->addSuccessor(newMBB);
7837 // Insert instructions into newMBB based on incoming instruction
7838 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7839 "unexpected number of operands");
7840 DebugLoc dl = bInstr->getDebugLoc();
7841 MachineOperand& destOper = bInstr->getOperand(0);
7842 MachineOperand* argOpers[2 + X86AddrNumOperands];
7843 int numArgs = bInstr->getNumOperands() - 1;
7844 for (int i=0; i < numArgs; ++i)
7845 argOpers[i] = &bInstr->getOperand(i+1);
7847 // x86 address has 4 operands: base, index, scale, and displacement
7848 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7849 int valArgIndx = lastAddrIndx + 1;
7851 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7852 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7853 for (int i=0; i <= lastAddrIndx; ++i)
7854 (*MIB).addOperand(*argOpers[i]);
7856 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7858 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7863 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7864 assert((argOpers[valArgIndx]->isReg() ||
7865 argOpers[valArgIndx]->isImm()) &&
7867 if (argOpers[valArgIndx]->isReg())
7868 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7870 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7872 (*MIB).addOperand(*argOpers[valArgIndx]);
7874 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7877 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7878 for (int i=0; i <= lastAddrIndx; ++i)
7879 (*MIB).addOperand(*argOpers[i]);
7881 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7882 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7883 bInstr->memoperands_end());
7885 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7889 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7891 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7895 // private utility function: 64 bit atomics on 32 bit host.
7897 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7898 MachineBasicBlock *MBB,
7903 bool invSrc) const {
7904 // For the atomic bitwise operator, we generate
7905 // thisMBB (instructions are in pairs, except cmpxchg8b)
7906 // ld t1,t2 = [bitinstr.addr]
7908 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7909 // op t5, t6 <- out1, out2, [bitinstr.val]
7910 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7911 // mov ECX, EBX <- t5, t6
7912 // mov EAX, EDX <- t1, t2
7913 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7914 // mov t3, t4 <- EAX, EDX
7916 // result in out1, out2
7917 // fallthrough -->nextMBB
7919 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7920 const unsigned LoadOpc = X86::MOV32rm;
7921 const unsigned copyOpc = X86::MOV32rr;
7922 const unsigned NotOpc = X86::NOT32r;
7923 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7924 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7925 MachineFunction::iterator MBBIter = MBB;
7928 /// First build the CFG
7929 MachineFunction *F = MBB->getParent();
7930 MachineBasicBlock *thisMBB = MBB;
7931 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7932 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7933 F->insert(MBBIter, newMBB);
7934 F->insert(MBBIter, nextMBB);
7936 // Move all successors to thisMBB to nextMBB
7937 nextMBB->transferSuccessors(thisMBB);
7939 // Update thisMBB to fall through to newMBB
7940 thisMBB->addSuccessor(newMBB);
7942 // newMBB jumps to itself and fall through to nextMBB
7943 newMBB->addSuccessor(nextMBB);
7944 newMBB->addSuccessor(newMBB);
7946 DebugLoc dl = bInstr->getDebugLoc();
7947 // Insert instructions into newMBB based on incoming instruction
7948 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7949 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7950 "unexpected number of operands");
7951 MachineOperand& dest1Oper = bInstr->getOperand(0);
7952 MachineOperand& dest2Oper = bInstr->getOperand(1);
7953 MachineOperand* argOpers[2 + X86AddrNumOperands];
7954 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7955 argOpers[i] = &bInstr->getOperand(i+2);
7957 // x86 address has 5 operands: base, index, scale, displacement, and segment.
7958 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7960 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7961 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7962 for (int i=0; i <= lastAddrIndx; ++i)
7963 (*MIB).addOperand(*argOpers[i]);
7964 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7965 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7966 // add 4 to displacement.
7967 for (int i=0; i <= lastAddrIndx-2; ++i)
7968 (*MIB).addOperand(*argOpers[i]);
7969 MachineOperand newOp3 = *(argOpers[3]);
7971 newOp3.setImm(newOp3.getImm()+4);
7973 newOp3.setOffset(newOp3.getOffset()+4);
7974 (*MIB).addOperand(newOp3);
7975 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7977 // t3/4 are defined later, at the bottom of the loop
7978 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7979 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7980 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7981 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7982 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7983 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7985 // The subsequent operations should be using the destination registers of
7986 //the PHI instructions.
7988 t1 = F->getRegInfo().createVirtualRegister(RC);
7989 t2 = F->getRegInfo().createVirtualRegister(RC);
7990 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
7991 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
7993 t1 = dest1Oper.getReg();
7994 t2 = dest2Oper.getReg();
7997 int valArgIndx = lastAddrIndx + 1;
7998 assert((argOpers[valArgIndx]->isReg() ||
7999 argOpers[valArgIndx]->isImm()) &&
8001 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8002 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8003 if (argOpers[valArgIndx]->isReg())
8004 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8006 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8007 if (regOpcL != X86::MOV32rr)
8009 (*MIB).addOperand(*argOpers[valArgIndx]);
8010 assert(argOpers[valArgIndx + 1]->isReg() ==
8011 argOpers[valArgIndx]->isReg());
8012 assert(argOpers[valArgIndx + 1]->isImm() ==
8013 argOpers[valArgIndx]->isImm());
8014 if (argOpers[valArgIndx + 1]->isReg())
8015 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8017 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8018 if (regOpcH != X86::MOV32rr)
8020 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8022 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8024 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8027 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8029 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8032 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8033 for (int i=0; i <= lastAddrIndx; ++i)
8034 (*MIB).addOperand(*argOpers[i]);
8036 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8037 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8038 bInstr->memoperands_end());
8040 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8041 MIB.addReg(X86::EAX);
8042 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8043 MIB.addReg(X86::EDX);
8046 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8048 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8052 // private utility function
8054 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8055 MachineBasicBlock *MBB,
8056 unsigned cmovOpc) const {
8057 // For the atomic min/max operator, we generate
8060 // ld t1 = [min/max.addr]
8061 // mov t2 = [min/max.val]
8063 // cmov[cond] t2 = t1
8065 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8067 // fallthrough -->nextMBB
8069 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8070 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8071 MachineFunction::iterator MBBIter = MBB;
8074 /// First build the CFG
8075 MachineFunction *F = MBB->getParent();
8076 MachineBasicBlock *thisMBB = MBB;
8077 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8078 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8079 F->insert(MBBIter, newMBB);
8080 F->insert(MBBIter, nextMBB);
8082 // Move all successors of thisMBB to nextMBB
8083 nextMBB->transferSuccessors(thisMBB);
8085 // Update thisMBB to fall through to newMBB
8086 thisMBB->addSuccessor(newMBB);
8088 // newMBB jumps to newMBB and fall through to nextMBB
8089 newMBB->addSuccessor(nextMBB);
8090 newMBB->addSuccessor(newMBB);
8092 DebugLoc dl = mInstr->getDebugLoc();
8093 // Insert instructions into newMBB based on incoming instruction
8094 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8095 "unexpected number of operands");
8096 MachineOperand& destOper = mInstr->getOperand(0);
8097 MachineOperand* argOpers[2 + X86AddrNumOperands];
8098 int numArgs = mInstr->getNumOperands() - 1;
8099 for (int i=0; i < numArgs; ++i)
8100 argOpers[i] = &mInstr->getOperand(i+1);
8102 // x86 address has 4 operands: base, index, scale, and displacement
8103 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8104 int valArgIndx = lastAddrIndx + 1;
8106 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8107 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8108 for (int i=0; i <= lastAddrIndx; ++i)
8109 (*MIB).addOperand(*argOpers[i]);
8111 // We only support register and immediate values
8112 assert((argOpers[valArgIndx]->isReg() ||
8113 argOpers[valArgIndx]->isImm()) &&
8116 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8117 if (argOpers[valArgIndx]->isReg())
8118 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8120 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8121 (*MIB).addOperand(*argOpers[valArgIndx]);
8123 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8126 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8131 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8132 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8136 // Cmp and exchange if none has modified the memory location
8137 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8138 for (int i=0; i <= lastAddrIndx; ++i)
8139 (*MIB).addOperand(*argOpers[i]);
8141 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8142 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8143 mInstr->memoperands_end());
8145 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8146 MIB.addReg(X86::EAX);
8149 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8151 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8155 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8156 // all of this code can be replaced with that in the .td file.
8158 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8159 unsigned numArgs, bool memArg) const {
8161 MachineFunction *F = BB->getParent();
8162 DebugLoc dl = MI->getDebugLoc();
8163 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8167 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8169 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8171 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8173 for (unsigned i = 0; i < numArgs; ++i) {
8174 MachineOperand &Op = MI->getOperand(i+1);
8176 if (!(Op.isReg() && Op.isImplicit()))
8180 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8183 F->DeleteMachineInstr(MI);
8189 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8191 MachineBasicBlock *MBB) const {
8192 // Emit code to save XMM registers to the stack. The ABI says that the
8193 // number of registers to save is given in %al, so it's theoretically
8194 // possible to do an indirect jump trick to avoid saving all of them,
8195 // however this code takes a simpler approach and just executes all
8196 // of the stores if %al is non-zero. It's less code, and it's probably
8197 // easier on the hardware branch predictor, and stores aren't all that
8198 // expensive anyway.
8200 // Create the new basic blocks. One block contains all the XMM stores,
8201 // and one block is the final destination regardless of whether any
8202 // stores were performed.
8203 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8204 MachineFunction *F = MBB->getParent();
8205 MachineFunction::iterator MBBIter = MBB;
8207 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8208 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8209 F->insert(MBBIter, XMMSaveMBB);
8210 F->insert(MBBIter, EndMBB);
8213 // Move any original successors of MBB to the end block.
8214 EndMBB->transferSuccessors(MBB);
8215 // The original block will now fall through to the XMM save block.
8216 MBB->addSuccessor(XMMSaveMBB);
8217 // The XMMSaveMBB will fall through to the end block.
8218 XMMSaveMBB->addSuccessor(EndMBB);
8220 // Now add the instructions.
8221 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8222 DebugLoc DL = MI->getDebugLoc();
8224 unsigned CountReg = MI->getOperand(0).getReg();
8225 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8226 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8228 if (!Subtarget->isTargetWin64()) {
8229 // If %al is 0, branch around the XMM save block.
8230 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8231 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
8232 MBB->addSuccessor(EndMBB);
8235 // In the XMM save block, save all the XMM argument registers.
8236 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8237 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8238 MachineMemOperand *MMO =
8239 F->getMachineMemOperand(
8240 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8241 MachineMemOperand::MOStore, Offset,
8242 /*Size=*/16, /*Align=*/16);
8243 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8244 .addFrameIndex(RegSaveFrameIndex)
8245 .addImm(/*Scale=*/1)
8246 .addReg(/*IndexReg=*/0)
8247 .addImm(/*Disp=*/Offset)
8248 .addReg(/*Segment=*/0)
8249 .addReg(MI->getOperand(i).getReg())
8250 .addMemOperand(MMO);
8253 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8259 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8260 MachineBasicBlock *BB,
8261 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8262 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8263 DebugLoc DL = MI->getDebugLoc();
8265 // To "insert" a SELECT_CC instruction, we actually have to insert the
8266 // diamond control-flow pattern. The incoming instruction knows the
8267 // destination vreg to set, the condition code register to branch on, the
8268 // true/false values to select between, and a branch opcode to use.
8269 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8270 MachineFunction::iterator It = BB;
8276 // cmpTY ccX, r1, r2
8278 // fallthrough --> copy0MBB
8279 MachineBasicBlock *thisMBB = BB;
8280 MachineFunction *F = BB->getParent();
8281 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8282 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8284 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8285 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8286 F->insert(It, copy0MBB);
8287 F->insert(It, sinkMBB);
8288 // Update machine-CFG edges by first adding all successors of the current
8289 // block to the new block which will contain the Phi node for the select.
8290 // Also inform sdisel of the edge changes.
8291 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8292 E = BB->succ_end(); I != E; ++I) {
8293 EM->insert(std::make_pair(*I, sinkMBB));
8294 sinkMBB->addSuccessor(*I);
8296 // Next, remove all successors of the current block, and add the true
8297 // and fallthrough blocks as its successors.
8298 while (!BB->succ_empty())
8299 BB->removeSuccessor(BB->succ_begin());
8300 // Add the true and fallthrough blocks as its successors.
8301 BB->addSuccessor(copy0MBB);
8302 BB->addSuccessor(sinkMBB);
8305 // %FalseValue = ...
8306 // # fallthrough to sinkMBB
8309 // Update machine-CFG edges
8310 BB->addSuccessor(sinkMBB);
8313 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8316 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8317 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8318 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8320 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8326 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8327 MachineBasicBlock *BB,
8328 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8329 switch (MI->getOpcode()) {
8330 default: assert(false && "Unexpected instr type to insert");
8332 case X86::CMOV_V1I64:
8333 case X86::CMOV_FR32:
8334 case X86::CMOV_FR64:
8335 case X86::CMOV_V4F32:
8336 case X86::CMOV_V2F64:
8337 case X86::CMOV_V2I64:
8338 return EmitLoweredSelect(MI, BB, EM);
8340 case X86::FP32_TO_INT16_IN_MEM:
8341 case X86::FP32_TO_INT32_IN_MEM:
8342 case X86::FP32_TO_INT64_IN_MEM:
8343 case X86::FP64_TO_INT16_IN_MEM:
8344 case X86::FP64_TO_INT32_IN_MEM:
8345 case X86::FP64_TO_INT64_IN_MEM:
8346 case X86::FP80_TO_INT16_IN_MEM:
8347 case X86::FP80_TO_INT32_IN_MEM:
8348 case X86::FP80_TO_INT64_IN_MEM: {
8349 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8350 DebugLoc DL = MI->getDebugLoc();
8352 // Change the floating point control register to use "round towards zero"
8353 // mode when truncating to an integer value.
8354 MachineFunction *F = BB->getParent();
8355 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8356 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8358 // Load the old value of the high byte of the control word...
8360 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8361 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8364 // Set the high part to be round to zero...
8365 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8368 // Reload the modified control word now...
8369 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8371 // Restore the memory image of control word to original value
8372 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8375 // Get the X86 opcode to use.
8377 switch (MI->getOpcode()) {
8378 default: llvm_unreachable("illegal opcode!");
8379 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8380 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8381 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8382 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8383 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8384 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8385 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8386 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8387 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8391 MachineOperand &Op = MI->getOperand(0);
8393 AM.BaseType = X86AddressMode::RegBase;
8394 AM.Base.Reg = Op.getReg();
8396 AM.BaseType = X86AddressMode::FrameIndexBase;
8397 AM.Base.FrameIndex = Op.getIndex();
8399 Op = MI->getOperand(1);
8401 AM.Scale = Op.getImm();
8402 Op = MI->getOperand(2);
8404 AM.IndexReg = Op.getImm();
8405 Op = MI->getOperand(3);
8406 if (Op.isGlobal()) {
8407 AM.GV = Op.getGlobal();
8409 AM.Disp = Op.getImm();
8411 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8412 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8414 // Reload the original control word now.
8415 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8417 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8420 // String/text processing lowering.
8421 case X86::PCMPISTRM128REG:
8422 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8423 case X86::PCMPISTRM128MEM:
8424 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8425 case X86::PCMPESTRM128REG:
8426 return EmitPCMP(MI, BB, 5, false /* in mem */);
8427 case X86::PCMPESTRM128MEM:
8428 return EmitPCMP(MI, BB, 5, true /* in mem */);
8431 case X86::ATOMAND32:
8432 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8433 X86::AND32ri, X86::MOV32rm,
8434 X86::LCMPXCHG32, X86::MOV32rr,
8435 X86::NOT32r, X86::EAX,
8436 X86::GR32RegisterClass);
8438 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8439 X86::OR32ri, X86::MOV32rm,
8440 X86::LCMPXCHG32, X86::MOV32rr,
8441 X86::NOT32r, X86::EAX,
8442 X86::GR32RegisterClass);
8443 case X86::ATOMXOR32:
8444 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8445 X86::XOR32ri, X86::MOV32rm,
8446 X86::LCMPXCHG32, X86::MOV32rr,
8447 X86::NOT32r, X86::EAX,
8448 X86::GR32RegisterClass);
8449 case X86::ATOMNAND32:
8450 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8451 X86::AND32ri, X86::MOV32rm,
8452 X86::LCMPXCHG32, X86::MOV32rr,
8453 X86::NOT32r, X86::EAX,
8454 X86::GR32RegisterClass, true);
8455 case X86::ATOMMIN32:
8456 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8457 case X86::ATOMMAX32:
8458 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8459 case X86::ATOMUMIN32:
8460 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8461 case X86::ATOMUMAX32:
8462 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8464 case X86::ATOMAND16:
8465 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8466 X86::AND16ri, X86::MOV16rm,
8467 X86::LCMPXCHG16, X86::MOV16rr,
8468 X86::NOT16r, X86::AX,
8469 X86::GR16RegisterClass);
8471 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8472 X86::OR16ri, X86::MOV16rm,
8473 X86::LCMPXCHG16, X86::MOV16rr,
8474 X86::NOT16r, X86::AX,
8475 X86::GR16RegisterClass);
8476 case X86::ATOMXOR16:
8477 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8478 X86::XOR16ri, X86::MOV16rm,
8479 X86::LCMPXCHG16, X86::MOV16rr,
8480 X86::NOT16r, X86::AX,
8481 X86::GR16RegisterClass);
8482 case X86::ATOMNAND16:
8483 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8484 X86::AND16ri, X86::MOV16rm,
8485 X86::LCMPXCHG16, X86::MOV16rr,
8486 X86::NOT16r, X86::AX,
8487 X86::GR16RegisterClass, true);
8488 case X86::ATOMMIN16:
8489 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8490 case X86::ATOMMAX16:
8491 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8492 case X86::ATOMUMIN16:
8493 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8494 case X86::ATOMUMAX16:
8495 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8498 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8499 X86::AND8ri, X86::MOV8rm,
8500 X86::LCMPXCHG8, X86::MOV8rr,
8501 X86::NOT8r, X86::AL,
8502 X86::GR8RegisterClass);
8504 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8505 X86::OR8ri, X86::MOV8rm,
8506 X86::LCMPXCHG8, X86::MOV8rr,
8507 X86::NOT8r, X86::AL,
8508 X86::GR8RegisterClass);
8510 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8511 X86::XOR8ri, X86::MOV8rm,
8512 X86::LCMPXCHG8, X86::MOV8rr,
8513 X86::NOT8r, X86::AL,
8514 X86::GR8RegisterClass);
8515 case X86::ATOMNAND8:
8516 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8517 X86::AND8ri, X86::MOV8rm,
8518 X86::LCMPXCHG8, X86::MOV8rr,
8519 X86::NOT8r, X86::AL,
8520 X86::GR8RegisterClass, true);
8521 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8522 // This group is for 64-bit host.
8523 case X86::ATOMAND64:
8524 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8525 X86::AND64ri32, X86::MOV64rm,
8526 X86::LCMPXCHG64, X86::MOV64rr,
8527 X86::NOT64r, X86::RAX,
8528 X86::GR64RegisterClass);
8530 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8531 X86::OR64ri32, X86::MOV64rm,
8532 X86::LCMPXCHG64, X86::MOV64rr,
8533 X86::NOT64r, X86::RAX,
8534 X86::GR64RegisterClass);
8535 case X86::ATOMXOR64:
8536 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8537 X86::XOR64ri32, X86::MOV64rm,
8538 X86::LCMPXCHG64, X86::MOV64rr,
8539 X86::NOT64r, X86::RAX,
8540 X86::GR64RegisterClass);
8541 case X86::ATOMNAND64:
8542 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8543 X86::AND64ri32, X86::MOV64rm,
8544 X86::LCMPXCHG64, X86::MOV64rr,
8545 X86::NOT64r, X86::RAX,
8546 X86::GR64RegisterClass, true);
8547 case X86::ATOMMIN64:
8548 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8549 case X86::ATOMMAX64:
8550 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8551 case X86::ATOMUMIN64:
8552 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8553 case X86::ATOMUMAX64:
8554 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8556 // This group does 64-bit operations on a 32-bit host.
8557 case X86::ATOMAND6432:
8558 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8559 X86::AND32rr, X86::AND32rr,
8560 X86::AND32ri, X86::AND32ri,
8562 case X86::ATOMOR6432:
8563 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8564 X86::OR32rr, X86::OR32rr,
8565 X86::OR32ri, X86::OR32ri,
8567 case X86::ATOMXOR6432:
8568 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8569 X86::XOR32rr, X86::XOR32rr,
8570 X86::XOR32ri, X86::XOR32ri,
8572 case X86::ATOMNAND6432:
8573 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8574 X86::AND32rr, X86::AND32rr,
8575 X86::AND32ri, X86::AND32ri,
8577 case X86::ATOMADD6432:
8578 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8579 X86::ADD32rr, X86::ADC32rr,
8580 X86::ADD32ri, X86::ADC32ri,
8582 case X86::ATOMSUB6432:
8583 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8584 X86::SUB32rr, X86::SBB32rr,
8585 X86::SUB32ri, X86::SBB32ri,
8587 case X86::ATOMSWAP6432:
8588 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8589 X86::MOV32rr, X86::MOV32rr,
8590 X86::MOV32ri, X86::MOV32ri,
8592 case X86::VASTART_SAVE_XMM_REGS:
8593 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8597 //===----------------------------------------------------------------------===//
8598 // X86 Optimization Hooks
8599 //===----------------------------------------------------------------------===//
8601 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8605 const SelectionDAG &DAG,
8606 unsigned Depth) const {
8607 unsigned Opc = Op.getOpcode();
8608 assert((Opc >= ISD::BUILTIN_OP_END ||
8609 Opc == ISD::INTRINSIC_WO_CHAIN ||
8610 Opc == ISD::INTRINSIC_W_CHAIN ||
8611 Opc == ISD::INTRINSIC_VOID) &&
8612 "Should use MaskedValueIsZero if you don't know whether Op"
8613 " is a target node!");
8615 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8627 // These nodes' second result is a boolean.
8628 if (Op.getResNo() == 0)
8632 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8633 Mask.getBitWidth() - 1);
8638 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8639 /// node is a GlobalAddress + offset.
8640 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8641 GlobalValue* &GA, int64_t &Offset) const{
8642 if (N->getOpcode() == X86ISD::Wrapper) {
8643 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8644 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8645 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8649 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8652 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8653 EVT EltVT, LoadSDNode *&LDBase,
8654 unsigned &LastLoadedElt,
8655 SelectionDAG &DAG, MachineFrameInfo *MFI,
8656 const TargetLowering &TLI) {
8658 LastLoadedElt = -1U;
8659 for (unsigned i = 0; i < NumElems; ++i) {
8660 if (N->getMaskElt(i) < 0) {
8666 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8667 if (!Elt.getNode() ||
8668 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8671 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8673 LDBase = cast<LoadSDNode>(Elt.getNode());
8677 if (Elt.getOpcode() == ISD::UNDEF)
8680 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8681 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8688 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8689 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8690 /// if the load addresses are consecutive, non-overlapping, and in the right
8691 /// order. In the case of v2i64, it will see if it can rewrite the
8692 /// shuffle to be an appropriate build vector so it can take advantage of
8693 // performBuildVectorCombine.
8694 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8695 const TargetLowering &TLI) {
8696 DebugLoc dl = N->getDebugLoc();
8697 EVT VT = N->getValueType(0);
8698 EVT EltVT = VT.getVectorElementType();
8699 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8700 unsigned NumElems = VT.getVectorNumElements();
8702 if (VT.getSizeInBits() != 128)
8705 // Try to combine a vector_shuffle into a 128-bit load.
8706 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8707 LoadSDNode *LD = NULL;
8708 unsigned LastLoadedElt;
8709 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8713 if (LastLoadedElt == NumElems - 1) {
8714 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8715 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8716 LD->getSrcValue(), LD->getSrcValueOffset(),
8718 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8719 LD->getSrcValue(), LD->getSrcValueOffset(),
8720 LD->isVolatile(), LD->getAlignment());
8721 } else if (NumElems == 4 && LastLoadedElt == 1) {
8722 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8723 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8724 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8725 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8730 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8731 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8732 const X86Subtarget *Subtarget) {
8733 DebugLoc DL = N->getDebugLoc();
8734 SDValue Cond = N->getOperand(0);
8735 // Get the LHS/RHS of the select.
8736 SDValue LHS = N->getOperand(1);
8737 SDValue RHS = N->getOperand(2);
8739 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8740 // instructions have the peculiarity that if either operand is a NaN,
8741 // they chose what we call the RHS operand (and as such are not symmetric).
8742 // It happens that this matches the semantics of the common C idiom
8743 // x<y?x:y and related forms, so we can recognize these cases.
8744 if (Subtarget->hasSSE2() &&
8745 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8746 Cond.getOpcode() == ISD::SETCC) {
8747 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8749 unsigned Opcode = 0;
8750 // Check for x CC y ? x : y.
8751 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8755 // This can be a min if we can prove that at least one of the operands
8757 if (!FiniteOnlyFPMath()) {
8758 if (DAG.isKnownNeverNaN(RHS)) {
8759 // Put the potential NaN in the RHS so that SSE will preserve it.
8760 std::swap(LHS, RHS);
8761 } else if (!DAG.isKnownNeverNaN(LHS))
8764 Opcode = X86ISD::FMIN;
8767 // This can be a min if we can prove that at least one of the operands
8769 if (!FiniteOnlyFPMath()) {
8770 if (DAG.isKnownNeverNaN(LHS)) {
8771 // Put the potential NaN in the RHS so that SSE will preserve it.
8772 std::swap(LHS, RHS);
8773 } else if (!DAG.isKnownNeverNaN(RHS))
8776 Opcode = X86ISD::FMIN;
8779 // This can be a min, but if either operand is a NaN we need it to
8780 // preserve the original LHS.
8781 std::swap(LHS, RHS);
8785 Opcode = X86ISD::FMIN;
8789 // This can be a max if we can prove that at least one of the operands
8791 if (!FiniteOnlyFPMath()) {
8792 if (DAG.isKnownNeverNaN(LHS)) {
8793 // Put the potential NaN in the RHS so that SSE will preserve it.
8794 std::swap(LHS, RHS);
8795 } else if (!DAG.isKnownNeverNaN(RHS))
8798 Opcode = X86ISD::FMAX;
8801 // This can be a max if we can prove that at least one of the operands
8803 if (!FiniteOnlyFPMath()) {
8804 if (DAG.isKnownNeverNaN(RHS)) {
8805 // Put the potential NaN in the RHS so that SSE will preserve it.
8806 std::swap(LHS, RHS);
8807 } else if (!DAG.isKnownNeverNaN(LHS))
8810 Opcode = X86ISD::FMAX;
8813 // This can be a max, but if either operand is a NaN we need it to
8814 // preserve the original LHS.
8815 std::swap(LHS, RHS);
8819 Opcode = X86ISD::FMAX;
8822 // Check for x CC y ? y : x -- a min/max with reversed arms.
8823 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8827 // This can be a min if we can prove that at least one of the operands
8829 if (!FiniteOnlyFPMath()) {
8830 if (DAG.isKnownNeverNaN(RHS)) {
8831 // Put the potential NaN in the RHS so that SSE will preserve it.
8832 std::swap(LHS, RHS);
8833 } else if (!DAG.isKnownNeverNaN(LHS))
8836 Opcode = X86ISD::FMIN;
8839 // This can be a min if we can prove that at least one of the operands
8841 if (!FiniteOnlyFPMath()) {
8842 if (DAG.isKnownNeverNaN(LHS)) {
8843 // Put the potential NaN in the RHS so that SSE will preserve it.
8844 std::swap(LHS, RHS);
8845 } else if (!DAG.isKnownNeverNaN(RHS))
8848 Opcode = X86ISD::FMIN;
8851 // This can be a min, but if either operand is a NaN we need it to
8852 // preserve the original LHS.
8853 std::swap(LHS, RHS);
8857 Opcode = X86ISD::FMIN;
8861 // This can be a max if we can prove that at least one of the operands
8863 if (!FiniteOnlyFPMath()) {
8864 if (DAG.isKnownNeverNaN(LHS)) {
8865 // Put the potential NaN in the RHS so that SSE will preserve it.
8866 std::swap(LHS, RHS);
8867 } else if (!DAG.isKnownNeverNaN(RHS))
8870 Opcode = X86ISD::FMAX;
8873 // This can be a max if we can prove that at least one of the operands
8875 if (!FiniteOnlyFPMath()) {
8876 if (DAG.isKnownNeverNaN(RHS)) {
8877 // Put the potential NaN in the RHS so that SSE will preserve it.
8878 std::swap(LHS, RHS);
8879 } else if (!DAG.isKnownNeverNaN(LHS))
8882 Opcode = X86ISD::FMAX;
8885 // This can be a max, but if either operand is a NaN we need it to
8886 // preserve the original LHS.
8887 std::swap(LHS, RHS);
8891 Opcode = X86ISD::FMAX;
8897 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8900 // If this is a select between two integer constants, try to do some
8902 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8903 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8904 // Don't do this for crazy integer types.
8905 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8906 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8907 // so that TrueC (the true value) is larger than FalseC.
8908 bool NeedsCondInvert = false;
8910 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8911 // Efficiently invertible.
8912 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8913 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8914 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8915 NeedsCondInvert = true;
8916 std::swap(TrueC, FalseC);
8919 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8920 if (FalseC->getAPIntValue() == 0 &&
8921 TrueC->getAPIntValue().isPowerOf2()) {
8922 if (NeedsCondInvert) // Invert the condition if needed.
8923 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8924 DAG.getConstant(1, Cond.getValueType()));
8926 // Zero extend the condition if needed.
8927 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8929 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8930 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8931 DAG.getConstant(ShAmt, MVT::i8));
8934 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8935 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8936 if (NeedsCondInvert) // Invert the condition if needed.
8937 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8938 DAG.getConstant(1, Cond.getValueType()));
8940 // Zero extend the condition if needed.
8941 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8942 FalseC->getValueType(0), Cond);
8943 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8944 SDValue(FalseC, 0));
8947 // Optimize cases that will turn into an LEA instruction. This requires
8948 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8949 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8950 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8951 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8953 bool isFastMultiplier = false;
8955 switch ((unsigned char)Diff) {
8957 case 1: // result = add base, cond
8958 case 2: // result = lea base( , cond*2)
8959 case 3: // result = lea base(cond, cond*2)
8960 case 4: // result = lea base( , cond*4)
8961 case 5: // result = lea base(cond, cond*4)
8962 case 8: // result = lea base( , cond*8)
8963 case 9: // result = lea base(cond, cond*8)
8964 isFastMultiplier = true;
8969 if (isFastMultiplier) {
8970 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8971 if (NeedsCondInvert) // Invert the condition if needed.
8972 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8973 DAG.getConstant(1, Cond.getValueType()));
8975 // Zero extend the condition if needed.
8976 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8978 // Scale the condition by the difference.
8980 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8981 DAG.getConstant(Diff, Cond.getValueType()));
8983 // Add the base if non-zero.
8984 if (FalseC->getAPIntValue() != 0)
8985 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8986 SDValue(FalseC, 0));
8996 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8997 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8998 TargetLowering::DAGCombinerInfo &DCI) {
8999 DebugLoc DL = N->getDebugLoc();
9001 // If the flag operand isn't dead, don't touch this CMOV.
9002 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9005 // If this is a select between two integer constants, try to do some
9006 // optimizations. Note that the operands are ordered the opposite of SELECT
9008 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9009 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9010 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9011 // larger than FalseC (the false value).
9012 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9014 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9015 CC = X86::GetOppositeBranchCondition(CC);
9016 std::swap(TrueC, FalseC);
9019 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9020 // This is efficient for any integer data type (including i8/i16) and
9022 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9023 SDValue Cond = N->getOperand(3);
9024 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9025 DAG.getConstant(CC, MVT::i8), Cond);
9027 // Zero extend the condition if needed.
9028 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9030 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9031 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9032 DAG.getConstant(ShAmt, MVT::i8));
9033 if (N->getNumValues() == 2) // Dead flag value?
9034 return DCI.CombineTo(N, Cond, SDValue());
9038 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9039 // for any integer data type, including i8/i16.
9040 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9041 SDValue Cond = N->getOperand(3);
9042 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9043 DAG.getConstant(CC, MVT::i8), Cond);
9045 // Zero extend the condition if needed.
9046 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9047 FalseC->getValueType(0), Cond);
9048 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9049 SDValue(FalseC, 0));
9051 if (N->getNumValues() == 2) // Dead flag value?
9052 return DCI.CombineTo(N, Cond, SDValue());
9056 // Optimize cases that will turn into an LEA instruction. This requires
9057 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9058 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9059 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9060 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9062 bool isFastMultiplier = false;
9064 switch ((unsigned char)Diff) {
9066 case 1: // result = add base, cond
9067 case 2: // result = lea base( , cond*2)
9068 case 3: // result = lea base(cond, cond*2)
9069 case 4: // result = lea base( , cond*4)
9070 case 5: // result = lea base(cond, cond*4)
9071 case 8: // result = lea base( , cond*8)
9072 case 9: // result = lea base(cond, cond*8)
9073 isFastMultiplier = true;
9078 if (isFastMultiplier) {
9079 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9080 SDValue Cond = N->getOperand(3);
9081 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9082 DAG.getConstant(CC, MVT::i8), Cond);
9083 // Zero extend the condition if needed.
9084 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9086 // Scale the condition by the difference.
9088 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9089 DAG.getConstant(Diff, Cond.getValueType()));
9091 // Add the base if non-zero.
9092 if (FalseC->getAPIntValue() != 0)
9093 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9094 SDValue(FalseC, 0));
9095 if (N->getNumValues() == 2) // Dead flag value?
9096 return DCI.CombineTo(N, Cond, SDValue());
9106 /// PerformMulCombine - Optimize a single multiply with constant into two
9107 /// in order to implement it with two cheaper instructions, e.g.
9108 /// LEA + SHL, LEA + LEA.
9109 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9110 TargetLowering::DAGCombinerInfo &DCI) {
9111 if (DAG.getMachineFunction().
9112 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
9115 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9118 EVT VT = N->getValueType(0);
9122 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9125 uint64_t MulAmt = C->getZExtValue();
9126 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9129 uint64_t MulAmt1 = 0;
9130 uint64_t MulAmt2 = 0;
9131 if ((MulAmt % 9) == 0) {
9133 MulAmt2 = MulAmt / 9;
9134 } else if ((MulAmt % 5) == 0) {
9136 MulAmt2 = MulAmt / 5;
9137 } else if ((MulAmt % 3) == 0) {
9139 MulAmt2 = MulAmt / 3;
9142 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9143 DebugLoc DL = N->getDebugLoc();
9145 if (isPowerOf2_64(MulAmt2) &&
9146 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9147 // If second multiplifer is pow2, issue it first. We want the multiply by
9148 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9150 std::swap(MulAmt1, MulAmt2);
9153 if (isPowerOf2_64(MulAmt1))
9154 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9155 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9157 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9158 DAG.getConstant(MulAmt1, VT));
9160 if (isPowerOf2_64(MulAmt2))
9161 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9162 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9164 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9165 DAG.getConstant(MulAmt2, VT));
9167 // Do not add new nodes to DAG combiner worklist.
9168 DCI.CombineTo(N, NewMul, false);
9173 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9174 SDValue N0 = N->getOperand(0);
9175 SDValue N1 = N->getOperand(1);
9176 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9177 EVT VT = N0.getValueType();
9179 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9180 // since the result of setcc_c is all zero's or all ones.
9181 if (N1C && N0.getOpcode() == ISD::AND &&
9182 N0.getOperand(1).getOpcode() == ISD::Constant) {
9183 SDValue N00 = N0.getOperand(0);
9184 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9185 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9186 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9187 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9188 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9189 APInt ShAmt = N1C->getAPIntValue();
9190 Mask = Mask.shl(ShAmt);
9192 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9193 N00, DAG.getConstant(Mask, VT));
9200 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9202 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9203 const X86Subtarget *Subtarget) {
9204 EVT VT = N->getValueType(0);
9205 if (!VT.isVector() && VT.isInteger() &&
9206 N->getOpcode() == ISD::SHL)
9207 return PerformSHLCombine(N, DAG);
9209 // On X86 with SSE2 support, we can transform this to a vector shift if
9210 // all elements are shifted by the same amount. We can't do this in legalize
9211 // because the a constant vector is typically transformed to a constant pool
9212 // so we have no knowledge of the shift amount.
9213 if (!Subtarget->hasSSE2())
9216 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9219 SDValue ShAmtOp = N->getOperand(1);
9220 EVT EltVT = VT.getVectorElementType();
9221 DebugLoc DL = N->getDebugLoc();
9222 SDValue BaseShAmt = SDValue();
9223 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9224 unsigned NumElts = VT.getVectorNumElements();
9226 for (; i != NumElts; ++i) {
9227 SDValue Arg = ShAmtOp.getOperand(i);
9228 if (Arg.getOpcode() == ISD::UNDEF) continue;
9232 for (; i != NumElts; ++i) {
9233 SDValue Arg = ShAmtOp.getOperand(i);
9234 if (Arg.getOpcode() == ISD::UNDEF) continue;
9235 if (Arg != BaseShAmt) {
9239 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9240 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9241 SDValue InVec = ShAmtOp.getOperand(0);
9242 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9243 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9245 for (; i != NumElts; ++i) {
9246 SDValue Arg = InVec.getOperand(i);
9247 if (Arg.getOpcode() == ISD::UNDEF) continue;
9251 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9252 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9253 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9254 if (C->getZExtValue() == SplatIdx)
9255 BaseShAmt = InVec.getOperand(1);
9258 if (BaseShAmt.getNode() == 0)
9259 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9260 DAG.getIntPtrConstant(0));
9264 // The shift amount is an i32.
9265 if (EltVT.bitsGT(MVT::i32))
9266 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9267 else if (EltVT.bitsLT(MVT::i32))
9268 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9270 // The shift amount is identical so we can do a vector shift.
9271 SDValue ValOp = N->getOperand(0);
9272 switch (N->getOpcode()) {
9274 llvm_unreachable("Unknown shift opcode!");
9277 if (VT == MVT::v2i64)
9278 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9279 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9281 if (VT == MVT::v4i32)
9282 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9283 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9285 if (VT == MVT::v8i16)
9286 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9287 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9291 if (VT == MVT::v4i32)
9292 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9293 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9295 if (VT == MVT::v8i16)
9296 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9297 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9301 if (VT == MVT::v2i64)
9302 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9303 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9305 if (VT == MVT::v4i32)
9306 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9307 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9309 if (VT == MVT::v8i16)
9310 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9311 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9318 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9319 const X86Subtarget *Subtarget) {
9320 EVT VT = N->getValueType(0);
9321 if (VT != MVT::i64 || !Subtarget->is64Bit())
9324 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9325 SDValue N0 = N->getOperand(0);
9326 SDValue N1 = N->getOperand(1);
9327 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9329 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9332 SDValue ShAmt0 = N0.getOperand(1);
9333 if (ShAmt0.getValueType() != MVT::i8)
9335 SDValue ShAmt1 = N1.getOperand(1);
9336 if (ShAmt1.getValueType() != MVT::i8)
9338 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9339 ShAmt0 = ShAmt0.getOperand(0);
9340 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9341 ShAmt1 = ShAmt1.getOperand(0);
9343 DebugLoc DL = N->getDebugLoc();
9344 unsigned Opc = X86ISD::SHLD;
9345 SDValue Op0 = N0.getOperand(0);
9346 SDValue Op1 = N1.getOperand(0);
9347 if (ShAmt0.getOpcode() == ISD::SUB) {
9349 std::swap(Op0, Op1);
9350 std::swap(ShAmt0, ShAmt1);
9353 if (ShAmt1.getOpcode() == ISD::SUB) {
9354 SDValue Sum = ShAmt1.getOperand(0);
9355 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9356 if (SumC->getSExtValue() == 64 &&
9357 ShAmt1.getOperand(1) == ShAmt0)
9358 return DAG.getNode(Opc, DL, VT,
9360 DAG.getNode(ISD::TRUNCATE, DL,
9363 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9364 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9366 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9367 return DAG.getNode(Opc, DL, VT,
9368 N0.getOperand(0), N1.getOperand(0),
9369 DAG.getNode(ISD::TRUNCATE, DL,
9376 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9377 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9378 const X86Subtarget *Subtarget) {
9379 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9380 // the FP state in cases where an emms may be missing.
9381 // A preferable solution to the general problem is to figure out the right
9382 // places to insert EMMS. This qualifies as a quick hack.
9384 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9385 StoreSDNode *St = cast<StoreSDNode>(N);
9386 EVT VT = St->getValue().getValueType();
9387 if (VT.getSizeInBits() != 64)
9390 const Function *F = DAG.getMachineFunction().getFunction();
9391 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9392 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9393 && Subtarget->hasSSE2();
9394 if ((VT.isVector() ||
9395 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9396 isa<LoadSDNode>(St->getValue()) &&
9397 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9398 St->getChain().hasOneUse() && !St->isVolatile()) {
9399 SDNode* LdVal = St->getValue().getNode();
9401 int TokenFactorIndex = -1;
9402 SmallVector<SDValue, 8> Ops;
9403 SDNode* ChainVal = St->getChain().getNode();
9404 // Must be a store of a load. We currently handle two cases: the load
9405 // is a direct child, and it's under an intervening TokenFactor. It is
9406 // possible to dig deeper under nested TokenFactors.
9407 if (ChainVal == LdVal)
9408 Ld = cast<LoadSDNode>(St->getChain());
9409 else if (St->getValue().hasOneUse() &&
9410 ChainVal->getOpcode() == ISD::TokenFactor) {
9411 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9412 if (ChainVal->getOperand(i).getNode() == LdVal) {
9413 TokenFactorIndex = i;
9414 Ld = cast<LoadSDNode>(St->getValue());
9416 Ops.push_back(ChainVal->getOperand(i));
9420 if (!Ld || !ISD::isNormalLoad(Ld))
9423 // If this is not the MMX case, i.e. we are just turning i64 load/store
9424 // into f64 load/store, avoid the transformation if there are multiple
9425 // uses of the loaded value.
9426 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9429 DebugLoc LdDL = Ld->getDebugLoc();
9430 DebugLoc StDL = N->getDebugLoc();
9431 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9432 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9434 if (Subtarget->is64Bit() || F64IsLegal) {
9435 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9436 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9437 Ld->getBasePtr(), Ld->getSrcValue(),
9438 Ld->getSrcValueOffset(), Ld->isVolatile(),
9439 Ld->getAlignment());
9440 SDValue NewChain = NewLd.getValue(1);
9441 if (TokenFactorIndex != -1) {
9442 Ops.push_back(NewChain);
9443 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9446 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9447 St->getSrcValue(), St->getSrcValueOffset(),
9448 St->isVolatile(), St->getAlignment());
9451 // Otherwise, lower to two pairs of 32-bit loads / stores.
9452 SDValue LoAddr = Ld->getBasePtr();
9453 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9454 DAG.getConstant(4, MVT::i32));
9456 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9457 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9458 Ld->isVolatile(), Ld->getAlignment());
9459 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9460 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9462 MinAlign(Ld->getAlignment(), 4));
9464 SDValue NewChain = LoLd.getValue(1);
9465 if (TokenFactorIndex != -1) {
9466 Ops.push_back(LoLd);
9467 Ops.push_back(HiLd);
9468 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9472 LoAddr = St->getBasePtr();
9473 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9474 DAG.getConstant(4, MVT::i32));
9476 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9477 St->getSrcValue(), St->getSrcValueOffset(),
9478 St->isVolatile(), St->getAlignment());
9479 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9481 St->getSrcValueOffset() + 4,
9483 MinAlign(St->getAlignment(), 4));
9484 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9489 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9490 /// X86ISD::FXOR nodes.
9491 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9492 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9493 // F[X]OR(0.0, x) -> x
9494 // F[X]OR(x, 0.0) -> x
9495 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9496 if (C->getValueAPF().isPosZero())
9497 return N->getOperand(1);
9498 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9499 if (C->getValueAPF().isPosZero())
9500 return N->getOperand(0);
9504 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9505 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9506 // FAND(0.0, x) -> 0.0
9507 // FAND(x, 0.0) -> 0.0
9508 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9509 if (C->getValueAPF().isPosZero())
9510 return N->getOperand(0);
9511 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9512 if (C->getValueAPF().isPosZero())
9513 return N->getOperand(1);
9517 static SDValue PerformBTCombine(SDNode *N,
9519 TargetLowering::DAGCombinerInfo &DCI) {
9520 // BT ignores high bits in the bit index operand.
9521 SDValue Op1 = N->getOperand(1);
9522 if (Op1.hasOneUse()) {
9523 unsigned BitWidth = Op1.getValueSizeInBits();
9524 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9525 APInt KnownZero, KnownOne;
9526 TargetLowering::TargetLoweringOpt TLO(DAG);
9527 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9528 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9529 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9530 DCI.CommitTargetLoweringOpt(TLO);
9535 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9536 SDValue Op = N->getOperand(0);
9537 if (Op.getOpcode() == ISD::BIT_CONVERT)
9538 Op = Op.getOperand(0);
9539 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9540 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9541 VT.getVectorElementType().getSizeInBits() ==
9542 OpVT.getVectorElementType().getSizeInBits()) {
9543 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9548 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9549 // Locked instructions, in turn, have implicit fence semantics (all memory
9550 // operations are flushed before issuing the locked instruction, and the
9551 // are not buffered), so we can fold away the common pattern of
9552 // fence-atomic-fence.
9553 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9554 SDValue atomic = N->getOperand(0);
9555 switch (atomic.getOpcode()) {
9556 case ISD::ATOMIC_CMP_SWAP:
9557 case ISD::ATOMIC_SWAP:
9558 case ISD::ATOMIC_LOAD_ADD:
9559 case ISD::ATOMIC_LOAD_SUB:
9560 case ISD::ATOMIC_LOAD_AND:
9561 case ISD::ATOMIC_LOAD_OR:
9562 case ISD::ATOMIC_LOAD_XOR:
9563 case ISD::ATOMIC_LOAD_NAND:
9564 case ISD::ATOMIC_LOAD_MIN:
9565 case ISD::ATOMIC_LOAD_MAX:
9566 case ISD::ATOMIC_LOAD_UMIN:
9567 case ISD::ATOMIC_LOAD_UMAX:
9573 SDValue fence = atomic.getOperand(0);
9574 if (fence.getOpcode() != ISD::MEMBARRIER)
9577 switch (atomic.getOpcode()) {
9578 case ISD::ATOMIC_CMP_SWAP:
9579 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9580 atomic.getOperand(1), atomic.getOperand(2),
9581 atomic.getOperand(3));
9582 case ISD::ATOMIC_SWAP:
9583 case ISD::ATOMIC_LOAD_ADD:
9584 case ISD::ATOMIC_LOAD_SUB:
9585 case ISD::ATOMIC_LOAD_AND:
9586 case ISD::ATOMIC_LOAD_OR:
9587 case ISD::ATOMIC_LOAD_XOR:
9588 case ISD::ATOMIC_LOAD_NAND:
9589 case ISD::ATOMIC_LOAD_MIN:
9590 case ISD::ATOMIC_LOAD_MAX:
9591 case ISD::ATOMIC_LOAD_UMIN:
9592 case ISD::ATOMIC_LOAD_UMAX:
9593 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9594 atomic.getOperand(1), atomic.getOperand(2));
9600 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9601 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9602 // (and (i32 x86isd::setcc_carry), 1)
9603 // This eliminates the zext. This transformation is necessary because
9604 // ISD::SETCC is always legalized to i8.
9605 DebugLoc dl = N->getDebugLoc();
9606 SDValue N0 = N->getOperand(0);
9607 EVT VT = N->getValueType(0);
9608 if (N0.getOpcode() == ISD::AND &&
9610 N0.getOperand(0).hasOneUse()) {
9611 SDValue N00 = N0.getOperand(0);
9612 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9614 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9615 if (!C || C->getZExtValue() != 1)
9617 return DAG.getNode(ISD::AND, dl, VT,
9618 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9619 N00.getOperand(0), N00.getOperand(1)),
9620 DAG.getConstant(1, VT));
9626 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9627 DAGCombinerInfo &DCI) const {
9628 SelectionDAG &DAG = DCI.DAG;
9629 switch (N->getOpcode()) {
9631 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9632 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9633 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9634 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9637 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9638 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9639 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9641 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9642 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9643 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9644 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9645 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9646 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9652 //===----------------------------------------------------------------------===//
9653 // X86 Inline Assembly Support
9654 //===----------------------------------------------------------------------===//
9656 static bool LowerToBSwap(CallInst *CI) {
9657 // FIXME: this should verify that we are targetting a 486 or better. If not,
9658 // we will turn this bswap into something that will be lowered to logical ops
9659 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9660 // so don't worry about this.
9662 // Verify this is a simple bswap.
9663 if (CI->getNumOperands() != 2 ||
9664 CI->getType() != CI->getOperand(1)->getType() ||
9665 !CI->getType()->isInteger())
9668 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9669 if (!Ty || Ty->getBitWidth() % 16 != 0)
9672 // Okay, we can do this xform, do so now.
9673 const Type *Tys[] = { Ty };
9674 Module *M = CI->getParent()->getParent()->getParent();
9675 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9677 Value *Op = CI->getOperand(1);
9678 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9680 CI->replaceAllUsesWith(Op);
9681 CI->eraseFromParent();
9685 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9686 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9687 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9689 std::string AsmStr = IA->getAsmString();
9691 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9692 SmallVector<StringRef, 4> AsmPieces;
9693 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9695 switch (AsmPieces.size()) {
9696 default: return false;
9698 AsmStr = AsmPieces[0];
9700 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9703 if (AsmPieces.size() == 2 &&
9704 (AsmPieces[0] == "bswap" ||
9705 AsmPieces[0] == "bswapq" ||
9706 AsmPieces[0] == "bswapl") &&
9707 (AsmPieces[1] == "$0" ||
9708 AsmPieces[1] == "${0:q}")) {
9709 // No need to check constraints, nothing other than the equivalent of
9710 // "=r,0" would be valid here.
9711 return LowerToBSwap(CI);
9713 // rorw $$8, ${0:w} --> llvm.bswap.i16
9714 if (CI->getType()->isInteger(16) &&
9715 AsmPieces.size() == 3 &&
9716 AsmPieces[0] == "rorw" &&
9717 AsmPieces[1] == "$$8," &&
9718 AsmPieces[2] == "${0:w}" &&
9719 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9720 return LowerToBSwap(CI);
9724 if (CI->getType()->isInteger(64) &&
9725 Constraints.size() >= 2 &&
9726 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9727 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9728 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9729 SmallVector<StringRef, 4> Words;
9730 SplitString(AsmPieces[0], Words, " \t");
9731 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9733 SplitString(AsmPieces[1], Words, " \t");
9734 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9736 SplitString(AsmPieces[2], Words, " \t,");
9737 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9738 Words[2] == "%edx") {
9739 return LowerToBSwap(CI);
9751 /// getConstraintType - Given a constraint letter, return the type of
9752 /// constraint it is for this target.
9753 X86TargetLowering::ConstraintType
9754 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9755 if (Constraint.size() == 1) {
9756 switch (Constraint[0]) {
9768 return C_RegisterClass;
9776 return TargetLowering::getConstraintType(Constraint);
9779 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9780 /// with another that has more specific requirements based on the type of the
9781 /// corresponding operand.
9782 const char *X86TargetLowering::
9783 LowerXConstraint(EVT ConstraintVT) const {
9784 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9785 // 'f' like normal targets.
9786 if (ConstraintVT.isFloatingPoint()) {
9787 if (Subtarget->hasSSE2())
9789 if (Subtarget->hasSSE1())
9793 return TargetLowering::LowerXConstraint(ConstraintVT);
9796 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9797 /// vector. If it is invalid, don't add anything to Ops.
9798 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9801 std::vector<SDValue>&Ops,
9802 SelectionDAG &DAG) const {
9803 SDValue Result(0, 0);
9805 switch (Constraint) {
9808 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9809 if (C->getZExtValue() <= 31) {
9810 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9816 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9817 if (C->getZExtValue() <= 63) {
9818 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9824 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9825 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9826 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9832 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9833 if (C->getZExtValue() <= 255) {
9834 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9840 // 32-bit signed value
9841 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9842 const ConstantInt *CI = C->getConstantIntValue();
9843 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9844 C->getSExtValue())) {
9845 // Widen to 64 bits here to get it sign extended.
9846 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9849 // FIXME gcc accepts some relocatable values here too, but only in certain
9850 // memory models; it's complicated.
9855 // 32-bit unsigned value
9856 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9857 const ConstantInt *CI = C->getConstantIntValue();
9858 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9859 C->getZExtValue())) {
9860 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9864 // FIXME gcc accepts some relocatable values here too, but only in certain
9865 // memory models; it's complicated.
9869 // Literal immediates are always ok.
9870 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9871 // Widen to 64 bits here to get it sign extended.
9872 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9876 // If we are in non-pic codegen mode, we allow the address of a global (with
9877 // an optional displacement) to be used with 'i'.
9878 GlobalAddressSDNode *GA = 0;
9881 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9883 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9884 Offset += GA->getOffset();
9886 } else if (Op.getOpcode() == ISD::ADD) {
9887 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9888 Offset += C->getZExtValue();
9889 Op = Op.getOperand(0);
9892 } else if (Op.getOpcode() == ISD::SUB) {
9893 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9894 Offset += -C->getZExtValue();
9895 Op = Op.getOperand(0);
9900 // Otherwise, this isn't something we can handle, reject it.
9904 GlobalValue *GV = GA->getGlobal();
9905 // If we require an extra load to get this address, as in PIC mode, we
9907 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9908 getTargetMachine())))
9912 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9914 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9920 if (Result.getNode()) {
9921 Ops.push_back(Result);
9924 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9928 std::vector<unsigned> X86TargetLowering::
9929 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9931 if (Constraint.size() == 1) {
9932 // FIXME: not handling fp-stack yet!
9933 switch (Constraint[0]) { // GCC X86 Constraint Letters
9934 default: break; // Unknown constraint letter
9935 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9936 if (Subtarget->is64Bit()) {
9938 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9939 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9940 X86::R10D,X86::R11D,X86::R12D,
9941 X86::R13D,X86::R14D,X86::R15D,
9942 X86::EBP, X86::ESP, 0);
9943 else if (VT == MVT::i16)
9944 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9945 X86::SI, X86::DI, X86::R8W,X86::R9W,
9946 X86::R10W,X86::R11W,X86::R12W,
9947 X86::R13W,X86::R14W,X86::R15W,
9948 X86::BP, X86::SP, 0);
9949 else if (VT == MVT::i8)
9950 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9951 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9952 X86::R10B,X86::R11B,X86::R12B,
9953 X86::R13B,X86::R14B,X86::R15B,
9954 X86::BPL, X86::SPL, 0);
9956 else if (VT == MVT::i64)
9957 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9958 X86::RSI, X86::RDI, X86::R8, X86::R9,
9959 X86::R10, X86::R11, X86::R12,
9960 X86::R13, X86::R14, X86::R15,
9961 X86::RBP, X86::RSP, 0);
9965 // 32-bit fallthrough
9968 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9969 else if (VT == MVT::i16)
9970 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9971 else if (VT == MVT::i8)
9972 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9973 else if (VT == MVT::i64)
9974 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9979 return std::vector<unsigned>();
9982 std::pair<unsigned, const TargetRegisterClass*>
9983 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9985 // First, see if this is a constraint that directly corresponds to an LLVM
9987 if (Constraint.size() == 1) {
9988 // GCC Constraint Letters
9989 switch (Constraint[0]) {
9991 case 'r': // GENERAL_REGS
9992 case 'l': // INDEX_REGS
9994 return std::make_pair(0U, X86::GR8RegisterClass);
9996 return std::make_pair(0U, X86::GR16RegisterClass);
9997 if (VT == MVT::i32 || !Subtarget->is64Bit())
9998 return std::make_pair(0U, X86::GR32RegisterClass);
9999 return std::make_pair(0U, X86::GR64RegisterClass);
10000 case 'R': // LEGACY_REGS
10002 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10003 if (VT == MVT::i16)
10004 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10005 if (VT == MVT::i32 || !Subtarget->is64Bit())
10006 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10007 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10008 case 'f': // FP Stack registers.
10009 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10010 // value to the correct fpstack register class.
10011 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10012 return std::make_pair(0U, X86::RFP32RegisterClass);
10013 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10014 return std::make_pair(0U, X86::RFP64RegisterClass);
10015 return std::make_pair(0U, X86::RFP80RegisterClass);
10016 case 'y': // MMX_REGS if MMX allowed.
10017 if (!Subtarget->hasMMX()) break;
10018 return std::make_pair(0U, X86::VR64RegisterClass);
10019 case 'Y': // SSE_REGS if SSE2 allowed
10020 if (!Subtarget->hasSSE2()) break;
10022 case 'x': // SSE_REGS if SSE1 allowed
10023 if (!Subtarget->hasSSE1()) break;
10025 switch (VT.getSimpleVT().SimpleTy) {
10027 // Scalar SSE types.
10030 return std::make_pair(0U, X86::FR32RegisterClass);
10033 return std::make_pair(0U, X86::FR64RegisterClass);
10041 return std::make_pair(0U, X86::VR128RegisterClass);
10047 // Use the default implementation in TargetLowering to convert the register
10048 // constraint into a member of a register class.
10049 std::pair<unsigned, const TargetRegisterClass*> Res;
10050 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10052 // Not found as a standard register?
10053 if (Res.second == 0) {
10054 // Map st(0) -> st(7) -> ST0
10055 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10056 tolower(Constraint[1]) == 's' &&
10057 tolower(Constraint[2]) == 't' &&
10058 Constraint[3] == '(' &&
10059 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10060 Constraint[5] == ')' &&
10061 Constraint[6] == '}') {
10063 Res.first = X86::ST0+Constraint[4]-'0';
10064 Res.second = X86::RFP80RegisterClass;
10068 // GCC allows "st(0)" to be called just plain "st".
10069 if (StringRef("{st}").equals_lower(Constraint)) {
10070 Res.first = X86::ST0;
10071 Res.second = X86::RFP80RegisterClass;
10076 if (StringRef("{flags}").equals_lower(Constraint)) {
10077 Res.first = X86::EFLAGS;
10078 Res.second = X86::CCRRegisterClass;
10082 // 'A' means EAX + EDX.
10083 if (Constraint == "A") {
10084 Res.first = X86::EAX;
10085 Res.second = X86::GR32_ADRegisterClass;
10091 // Otherwise, check to see if this is a register class of the wrong value
10092 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10093 // turn into {ax},{dx}.
10094 if (Res.second->hasType(VT))
10095 return Res; // Correct type already, nothing to do.
10097 // All of the single-register GCC register classes map their values onto
10098 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10099 // really want an 8-bit or 32-bit register, map to the appropriate register
10100 // class and return the appropriate register.
10101 if (Res.second == X86::GR16RegisterClass) {
10102 if (VT == MVT::i8) {
10103 unsigned DestReg = 0;
10104 switch (Res.first) {
10106 case X86::AX: DestReg = X86::AL; break;
10107 case X86::DX: DestReg = X86::DL; break;
10108 case X86::CX: DestReg = X86::CL; break;
10109 case X86::BX: DestReg = X86::BL; break;
10112 Res.first = DestReg;
10113 Res.second = X86::GR8RegisterClass;
10115 } else if (VT == MVT::i32) {
10116 unsigned DestReg = 0;
10117 switch (Res.first) {
10119 case X86::AX: DestReg = X86::EAX; break;
10120 case X86::DX: DestReg = X86::EDX; break;
10121 case X86::CX: DestReg = X86::ECX; break;
10122 case X86::BX: DestReg = X86::EBX; break;
10123 case X86::SI: DestReg = X86::ESI; break;
10124 case X86::DI: DestReg = X86::EDI; break;
10125 case X86::BP: DestReg = X86::EBP; break;
10126 case X86::SP: DestReg = X86::ESP; break;
10129 Res.first = DestReg;
10130 Res.second = X86::GR32RegisterClass;
10132 } else if (VT == MVT::i64) {
10133 unsigned DestReg = 0;
10134 switch (Res.first) {
10136 case X86::AX: DestReg = X86::RAX; break;
10137 case X86::DX: DestReg = X86::RDX; break;
10138 case X86::CX: DestReg = X86::RCX; break;
10139 case X86::BX: DestReg = X86::RBX; break;
10140 case X86::SI: DestReg = X86::RSI; break;
10141 case X86::DI: DestReg = X86::RDI; break;
10142 case X86::BP: DestReg = X86::RBP; break;
10143 case X86::SP: DestReg = X86::RSP; break;
10146 Res.first = DestReg;
10147 Res.second = X86::GR64RegisterClass;
10150 } else if (Res.second == X86::FR32RegisterClass ||
10151 Res.second == X86::FR64RegisterClass ||
10152 Res.second == X86::VR128RegisterClass) {
10153 // Handle references to XMM physical registers that got mapped into the
10154 // wrong class. This can happen with constraints like {xmm0} where the
10155 // target independent register mapper will just pick the first match it can
10156 // find, ignoring the required type.
10157 if (VT == MVT::f32)
10158 Res.second = X86::FR32RegisterClass;
10159 else if (VT == MVT::f64)
10160 Res.second = X86::FR64RegisterClass;
10161 else if (X86::VR128RegisterClass->hasType(VT))
10162 Res.second = X86::VR128RegisterClass;
10168 //===----------------------------------------------------------------------===//
10169 // X86 Widen vector type
10170 //===----------------------------------------------------------------------===//
10172 /// getWidenVectorType: given a vector type, returns the type to widen
10173 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10174 /// If there is no vector type that we want to widen to, returns MVT::Other
10175 /// When and where to widen is target dependent based on the cost of
10176 /// scalarizing vs using the wider vector type.
10178 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10179 assert(VT.isVector());
10180 if (isTypeLegal(VT))
10183 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10184 // type based on element type. This would speed up our search (though
10185 // it may not be worth it since the size of the list is relatively
10187 EVT EltVT = VT.getVectorElementType();
10188 unsigned NElts = VT.getVectorNumElements();
10190 // On X86, it make sense to widen any vector wider than 1
10194 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10195 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10196 EVT SVT = (MVT::SimpleValueType)nVT;
10198 if (isTypeLegal(SVT) &&
10199 SVT.getVectorElementType() == EltVT &&
10200 SVT.getVectorNumElements() > NElts)