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, const Type *RetTy,
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, RetTy, CallConv,
1795 isVarArg, Outs, Ins, DAG);
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);
1810 else if (isTailCall && !PerformTailCallOpt)
1811 // This is a sibcall. The memory operands are available in caller's
1812 // own caller's stack.
1819 // Lower arguments at fp - stackoffset + fpdiff.
1820 unsigned NumBytesCallerPushed =
1821 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1822 FPDiff = NumBytesCallerPushed - NumBytes;
1824 // Set the delta of movement of the returnaddr stackslot.
1825 // But only set if delta is greater than previous delta.
1826 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1827 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1830 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1832 SDValue RetAddrFrIdx;
1833 // Load return adress for tail calls.
1834 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
1837 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1838 SmallVector<SDValue, 8> MemOpChains;
1841 // Walk the register/memloc assignments, inserting copies/loads. In the case
1842 // of tail call optimization arguments are handle later.
1843 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1844 CCValAssign &VA = ArgLocs[i];
1845 EVT RegVT = VA.getLocVT();
1846 SDValue Arg = Outs[i].Val;
1847 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1848 bool isByVal = Flags.isByVal();
1850 // Promote the value if needed.
1851 switch (VA.getLocInfo()) {
1852 default: llvm_unreachable("Unknown loc info!");
1853 case CCValAssign::Full: break;
1854 case CCValAssign::SExt:
1855 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1857 case CCValAssign::ZExt:
1858 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1860 case CCValAssign::AExt:
1861 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1862 // Special case: passing MMX values in XMM registers.
1863 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1864 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1865 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1867 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1869 case CCValAssign::BCvt:
1870 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1872 case CCValAssign::Indirect: {
1873 // Store the argument.
1874 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1875 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1876 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1877 PseudoSourceValue::getFixedStack(FI), 0);
1883 if (VA.isRegLoc()) {
1884 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1886 if (!isTailCall || (isTailCall && isByVal)) {
1887 assert(VA.isMemLoc());
1888 if (StackPtr.getNode() == 0)
1889 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1891 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1892 dl, DAG, VA, Flags));
1897 if (!MemOpChains.empty())
1898 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1899 &MemOpChains[0], MemOpChains.size());
1901 // Build a sequence of copy-to-reg nodes chained together with token chain
1902 // and flag operands which copy the outgoing args into registers.
1904 // Tail call byval lowering might overwrite argument registers so in case of
1905 // tail call optimization the copies to registers are lowered later.
1907 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1908 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1909 RegsToPass[i].second, InFlag);
1910 InFlag = Chain.getValue(1);
1914 if (Subtarget->isPICStyleGOT()) {
1915 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1918 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1919 DAG.getNode(X86ISD::GlobalBaseReg,
1920 DebugLoc::getUnknownLoc(),
1923 InFlag = Chain.getValue(1);
1925 // If we are tail calling and generating PIC/GOT style code load the
1926 // address of the callee into ECX. The value in ecx is used as target of
1927 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1928 // for tail calls on PIC/GOT architectures. Normally we would just put the
1929 // address of GOT into ebx and then call target@PLT. But for tail calls
1930 // ebx would be restored (since ebx is callee saved) before jumping to the
1933 // Note: The actual moving to ECX is done further down.
1934 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1935 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1936 !G->getGlobal()->hasProtectedVisibility())
1937 Callee = LowerGlobalAddress(Callee, DAG);
1938 else if (isa<ExternalSymbolSDNode>(Callee))
1939 Callee = LowerExternalSymbol(Callee, DAG);
1943 if (Is64Bit && isVarArg) {
1944 // From AMD64 ABI document:
1945 // For calls that may call functions that use varargs or stdargs
1946 // (prototype-less calls or calls to functions containing ellipsis (...) in
1947 // the declaration) %al is used as hidden argument to specify the number
1948 // of SSE registers used. The contents of %al do not need to match exactly
1949 // the number of registers, but must be an ubound on the number of SSE
1950 // registers used and is in the range 0 - 8 inclusive.
1952 // FIXME: Verify this on Win64
1953 // Count the number of XMM registers allocated.
1954 static const unsigned XMMArgRegs[] = {
1955 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1956 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1958 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1959 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1960 && "SSE registers cannot be used when SSE is disabled");
1962 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1963 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1964 InFlag = Chain.getValue(1);
1968 // For tail calls lower the arguments to the 'real' stack slot.
1970 // Force all the incoming stack arguments to be loaded from the stack
1971 // before any new outgoing arguments are stored to the stack, because the
1972 // outgoing stack slots may alias the incoming argument stack slots, and
1973 // the alias isn't otherwise explicit. This is slightly more conservative
1974 // than necessary, because it means that each store effectively depends
1975 // on every argument instead of just those arguments it would clobber.
1976 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1978 SmallVector<SDValue, 8> MemOpChains2;
1981 // Do not flag preceeding copytoreg stuff together with the following stuff.
1983 if (PerformTailCallOpt) {
1984 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1985 CCValAssign &VA = ArgLocs[i];
1988 assert(VA.isMemLoc());
1989 SDValue Arg = Outs[i].Val;
1990 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1991 // Create frame index.
1992 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1993 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1994 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1995 FIN = DAG.getFrameIndex(FI, getPointerTy());
1997 if (Flags.isByVal()) {
1998 // Copy relative to framepointer.
1999 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
2000 if (StackPtr.getNode() == 0)
2001 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
2003 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
2005 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
2009 // Store relative to framepointer.
2010 MemOpChains2.push_back(
2011 DAG.getStore(ArgChain, dl, Arg, FIN,
2012 PseudoSourceValue::getFixedStack(FI), 0));
2017 if (!MemOpChains2.empty())
2018 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2019 &MemOpChains2[0], MemOpChains2.size());
2021 // Copy arguments to their registers.
2022 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2023 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2024 RegsToPass[i].second, InFlag);
2025 InFlag = Chain.getValue(1);
2029 // Store the return address to the appropriate stack slot.
2030 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2034 bool WasGlobalOrExternal = false;
2035 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2036 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2037 // In the 64-bit large code model, we have to make all calls
2038 // through a register, since the call instruction's 32-bit
2039 // pc-relative offset may not be large enough to hold the whole
2041 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2042 WasGlobalOrExternal = true;
2043 // If the callee is a GlobalAddress node (quite common, every direct call
2044 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2047 // We should use extra load for direct calls to dllimported functions in
2049 GlobalValue *GV = G->getGlobal();
2050 if (!GV->hasDLLImportLinkage()) {
2051 unsigned char OpFlags = 0;
2053 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2054 // external symbols most go through the PLT in PIC mode. If the symbol
2055 // has hidden or protected visibility, or if it is static or local, then
2056 // we don't need to use the PLT - we can directly call it.
2057 if (Subtarget->isTargetELF() &&
2058 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2059 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2060 OpFlags = X86II::MO_PLT;
2061 } else if (Subtarget->isPICStyleStubAny() &&
2062 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2063 Subtarget->getDarwinVers() < 9) {
2064 // PC-relative references to external symbols should go through $stub,
2065 // unless we're building with the leopard linker or later, which
2066 // automatically synthesizes these stubs.
2067 OpFlags = X86II::MO_DARWIN_STUB;
2070 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2071 G->getOffset(), OpFlags);
2073 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2074 WasGlobalOrExternal = true;
2075 unsigned char OpFlags = 0;
2077 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2078 // symbols should go through the PLT.
2079 if (Subtarget->isTargetELF() &&
2080 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2081 OpFlags = X86II::MO_PLT;
2082 } else if (Subtarget->isPICStyleStubAny() &&
2083 Subtarget->getDarwinVers() < 9) {
2084 // PC-relative references to external symbols should go through $stub,
2085 // unless we're building with the leopard linker or later, which
2086 // automatically synthesizes these stubs.
2087 OpFlags = X86II::MO_DARWIN_STUB;
2090 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2094 if (isTailCall && !WasGlobalOrExternal) {
2095 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
2097 Chain = DAG.getCopyToReg(Chain, dl,
2098 DAG.getRegister(Opc, getPointerTy()),
2100 Callee = DAG.getRegister(Opc, getPointerTy());
2101 // Add register as live out.
2102 MF.getRegInfo().addLiveOut(Opc);
2105 // Returns a chain & a flag for retval copy to use.
2106 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2107 SmallVector<SDValue, 8> Ops;
2110 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2111 DAG.getIntPtrConstant(0, true), InFlag);
2112 InFlag = Chain.getValue(1);
2115 Ops.push_back(Chain);
2116 Ops.push_back(Callee);
2119 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2121 // Add argument registers to the end of the list so that they are known live
2123 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2124 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2125 RegsToPass[i].second.getValueType()));
2127 // Add an implicit use GOT pointer in EBX.
2128 if (!isTailCall && Subtarget->isPICStyleGOT())
2129 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2131 // Add an implicit use of AL for x86 vararg functions.
2132 if (Is64Bit && isVarArg)
2133 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2135 if (InFlag.getNode())
2136 Ops.push_back(InFlag);
2139 // If this is the first return lowered for this function, add the regs
2140 // to the liveout set for the function.
2141 if (MF.getRegInfo().liveout_empty()) {
2142 SmallVector<CCValAssign, 16> RVLocs;
2143 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2145 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2146 for (unsigned i = 0; i != RVLocs.size(); ++i)
2147 if (RVLocs[i].isRegLoc())
2148 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2151 assert(((Callee.getOpcode() == ISD::Register &&
2152 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2153 cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
2154 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2155 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2156 "Expecting a global address, external symbol, or scratch register");
2158 return DAG.getNode(X86ISD::TC_RETURN, dl,
2159 NodeTys, &Ops[0], Ops.size());
2162 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2163 InFlag = Chain.getValue(1);
2165 // Create the CALLSEQ_END node.
2166 unsigned NumBytesForCalleeToPush;
2167 if (IsCalleePop(isVarArg, CallConv))
2168 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2169 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2170 // If this is is a call to a struct-return function, the callee
2171 // pops the hidden struct pointer, so we have to push it back.
2172 // This is common for Darwin/X86, Linux & Mingw32 targets.
2173 NumBytesForCalleeToPush = 4;
2175 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2177 // Returns a flag for retval copy to use.
2178 Chain = DAG.getCALLSEQ_END(Chain,
2179 DAG.getIntPtrConstant(NumBytes, true),
2180 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2183 InFlag = Chain.getValue(1);
2185 // Handle result values, copying them out of physregs into vregs that we
2187 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2188 Ins, dl, DAG, InVals);
2192 //===----------------------------------------------------------------------===//
2193 // Fast Calling Convention (tail call) implementation
2194 //===----------------------------------------------------------------------===//
2196 // Like std call, callee cleans arguments, convention except that ECX is
2197 // reserved for storing the tail called function address. Only 2 registers are
2198 // free for argument passing (inreg). Tail call optimization is performed
2200 // * tailcallopt is enabled
2201 // * caller/callee are fastcc
2202 // On X86_64 architecture with GOT-style position independent code only local
2203 // (within module) calls are supported at the moment.
2204 // To keep the stack aligned according to platform abi the function
2205 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2206 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2207 // If a tail called function callee has more arguments than the caller the
2208 // caller needs to make sure that there is room to move the RETADDR to. This is
2209 // achieved by reserving an area the size of the argument delta right after the
2210 // original REtADDR, but before the saved framepointer or the spilled registers
2211 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2223 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2224 /// for a 16 byte align requirement.
2225 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2226 SelectionDAG& DAG) {
2227 MachineFunction &MF = DAG.getMachineFunction();
2228 const TargetMachine &TM = MF.getTarget();
2229 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2230 unsigned StackAlignment = TFI.getStackAlignment();
2231 uint64_t AlignMask = StackAlignment - 1;
2232 int64_t Offset = StackSize;
2233 uint64_t SlotSize = TD->getPointerSize();
2234 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2235 // Number smaller than 12 so just add the difference.
2236 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2238 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2239 Offset = ((~AlignMask) & Offset) + StackAlignment +
2240 (StackAlignment-SlotSize);
2245 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2246 /// for tail call optimization. Targets which want to do tail call
2247 /// optimization should implement this function.
2249 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2251 CallingConv::ID CalleeCC,
2253 const SmallVectorImpl<ISD::OutputArg> &Outs,
2254 const SmallVectorImpl<ISD::InputArg> &Ins,
2255 SelectionDAG& DAG) const {
2256 if (CalleeCC != CallingConv::Fast &&
2257 CalleeCC != CallingConv::C)
2260 // If -tailcallopt is specified, make fastcc functions tail-callable.
2261 const Function *CallerF = DAG.getMachineFunction().getFunction();
2262 if (PerformTailCallOpt) {
2263 if (CalleeCC == CallingConv::Fast &&
2264 CallerF->getCallingConv() == CalleeCC)
2270 // Look for obvious safe cases to perform tail call optimization that does not
2271 // requite ABI changes. This is what gcc calls sibcall.
2273 // Do not tail call optimize vararg calls for now.
2277 // If the callee takes no arguments then go on to check the results of the
2279 if (!Outs.empty()) {
2280 // Check if stack adjustment is needed. For now, do not do this if any
2281 // argument is passed on the stack.
2282 SmallVector<CCValAssign, 16> ArgLocs;
2283 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2284 ArgLocs, *DAG.getContext());
2285 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2286 if (CCInfo.getNextStackOffset()) {
2287 MachineFunction &MF = DAG.getMachineFunction();
2288 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2290 if (Subtarget->isTargetWin64())
2291 // Win64 ABI has additional complications.
2294 // Check if the arguments are already laid out in the right way as
2295 // the caller's fixed stack objects.
2296 MachineFrameInfo *MFI = MF.getFrameInfo();
2297 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2298 CCValAssign &VA = ArgLocs[i];
2299 EVT RegVT = VA.getLocVT();
2300 SDValue Arg = Outs[i].Val;
2301 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2302 if (Flags.isByVal())
2303 return false; // TODO
2304 if (VA.getLocInfo() == CCValAssign::Indirect)
2306 if (!VA.isRegLoc()) {
2307 LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg);
2310 SDValue Ptr = Ld->getBasePtr();
2311 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2314 int FI = FINode->getIndex();
2315 if (!MFI->isFixedObjectIndex(FI))
2317 if (VA.getLocMemOffset() != MFI->getObjectOffset(FI))
2324 // If the caller does not return a value, then this is obviously safe.
2325 // This is one case where it's safe to perform this optimization even
2326 // if the return types do not match.
2327 const Type *CallerRetTy = CallerF->getReturnType();
2328 if (CallerRetTy->isVoidTy())
2331 // If the return types match, then it's safe.
2332 return CallerRetTy == RetTy;
2336 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2338 DenseMap<const Value *, unsigned> &vm,
2339 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2340 DenseMap<const AllocaInst *, int> &am
2342 , SmallSet<Instruction*, 8> &cil
2345 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2353 //===----------------------------------------------------------------------===//
2354 // Other Lowering Hooks
2355 //===----------------------------------------------------------------------===//
2358 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2359 MachineFunction &MF = DAG.getMachineFunction();
2360 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2361 int ReturnAddrIndex = FuncInfo->getRAIndex();
2363 if (ReturnAddrIndex == 0) {
2364 // Set up a frame object for the return address.
2365 uint64_t SlotSize = TD->getPointerSize();
2366 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2368 FuncInfo->setRAIndex(ReturnAddrIndex);
2371 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2375 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2376 bool hasSymbolicDisplacement) {
2377 // Offset should fit into 32 bit immediate field.
2378 if (!isInt32(Offset))
2381 // If we don't have a symbolic displacement - we don't have any extra
2383 if (!hasSymbolicDisplacement)
2386 // FIXME: Some tweaks might be needed for medium code model.
2387 if (M != CodeModel::Small && M != CodeModel::Kernel)
2390 // For small code model we assume that latest object is 16MB before end of 31
2391 // bits boundary. We may also accept pretty large negative constants knowing
2392 // that all objects are in the positive half of address space.
2393 if (M == CodeModel::Small && Offset < 16*1024*1024)
2396 // For kernel code model we know that all object resist in the negative half
2397 // of 32bits address space. We may not accept negative offsets, since they may
2398 // be just off and we may accept pretty large positive ones.
2399 if (M == CodeModel::Kernel && Offset > 0)
2405 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2406 /// specific condition code, returning the condition code and the LHS/RHS of the
2407 /// comparison to make.
2408 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2409 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2411 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2412 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2413 // X > -1 -> X == 0, jump !sign.
2414 RHS = DAG.getConstant(0, RHS.getValueType());
2415 return X86::COND_NS;
2416 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2417 // X < 0 -> X == 0, jump on sign.
2419 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2421 RHS = DAG.getConstant(0, RHS.getValueType());
2422 return X86::COND_LE;
2426 switch (SetCCOpcode) {
2427 default: llvm_unreachable("Invalid integer condition!");
2428 case ISD::SETEQ: return X86::COND_E;
2429 case ISD::SETGT: return X86::COND_G;
2430 case ISD::SETGE: return X86::COND_GE;
2431 case ISD::SETLT: return X86::COND_L;
2432 case ISD::SETLE: return X86::COND_LE;
2433 case ISD::SETNE: return X86::COND_NE;
2434 case ISD::SETULT: return X86::COND_B;
2435 case ISD::SETUGT: return X86::COND_A;
2436 case ISD::SETULE: return X86::COND_BE;
2437 case ISD::SETUGE: return X86::COND_AE;
2441 // First determine if it is required or is profitable to flip the operands.
2443 // If LHS is a foldable load, but RHS is not, flip the condition.
2444 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2445 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2446 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2447 std::swap(LHS, RHS);
2450 switch (SetCCOpcode) {
2456 std::swap(LHS, RHS);
2460 // On a floating point condition, the flags are set as follows:
2462 // 0 | 0 | 0 | X > Y
2463 // 0 | 0 | 1 | X < Y
2464 // 1 | 0 | 0 | X == Y
2465 // 1 | 1 | 1 | unordered
2466 switch (SetCCOpcode) {
2467 default: llvm_unreachable("Condcode should be pre-legalized away");
2469 case ISD::SETEQ: return X86::COND_E;
2470 case ISD::SETOLT: // flipped
2472 case ISD::SETGT: return X86::COND_A;
2473 case ISD::SETOLE: // flipped
2475 case ISD::SETGE: return X86::COND_AE;
2476 case ISD::SETUGT: // flipped
2478 case ISD::SETLT: return X86::COND_B;
2479 case ISD::SETUGE: // flipped
2481 case ISD::SETLE: return X86::COND_BE;
2483 case ISD::SETNE: return X86::COND_NE;
2484 case ISD::SETUO: return X86::COND_P;
2485 case ISD::SETO: return X86::COND_NP;
2487 case ISD::SETUNE: return X86::COND_INVALID;
2491 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2492 /// code. Current x86 isa includes the following FP cmov instructions:
2493 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2494 static bool hasFPCMov(unsigned X86CC) {
2510 /// isFPImmLegal - Returns true if the target can instruction select the
2511 /// specified FP immediate natively. If false, the legalizer will
2512 /// materialize the FP immediate as a load from a constant pool.
2513 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2514 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2515 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2521 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2522 /// the specified range (L, H].
2523 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2524 return (Val < 0) || (Val >= Low && Val < Hi);
2527 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2528 /// specified value.
2529 static bool isUndefOrEqual(int Val, int CmpVal) {
2530 if (Val < 0 || Val == CmpVal)
2535 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2536 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2537 /// the second operand.
2538 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2539 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2540 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2541 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2542 return (Mask[0] < 2 && Mask[1] < 2);
2546 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2547 SmallVector<int, 8> M;
2549 return ::isPSHUFDMask(M, N->getValueType(0));
2552 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2553 /// is suitable for input to PSHUFHW.
2554 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2555 if (VT != MVT::v8i16)
2558 // Lower quadword copied in order or undef.
2559 for (int i = 0; i != 4; ++i)
2560 if (Mask[i] >= 0 && Mask[i] != i)
2563 // Upper quadword shuffled.
2564 for (int i = 4; i != 8; ++i)
2565 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2571 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2572 SmallVector<int, 8> M;
2574 return ::isPSHUFHWMask(M, N->getValueType(0));
2577 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2578 /// is suitable for input to PSHUFLW.
2579 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2580 if (VT != MVT::v8i16)
2583 // Upper quadword copied in order.
2584 for (int i = 4; i != 8; ++i)
2585 if (Mask[i] >= 0 && Mask[i] != i)
2588 // Lower quadword shuffled.
2589 for (int i = 0; i != 4; ++i)
2596 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2597 SmallVector<int, 8> M;
2599 return ::isPSHUFLWMask(M, N->getValueType(0));
2602 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2603 /// is suitable for input to PALIGNR.
2604 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2606 int i, e = VT.getVectorNumElements();
2608 // Do not handle v2i64 / v2f64 shuffles with palignr.
2609 if (e < 4 || !hasSSSE3)
2612 for (i = 0; i != e; ++i)
2616 // All undef, not a palignr.
2620 // Determine if it's ok to perform a palignr with only the LHS, since we
2621 // don't have access to the actual shuffle elements to see if RHS is undef.
2622 bool Unary = Mask[i] < (int)e;
2623 bool NeedsUnary = false;
2625 int s = Mask[i] - i;
2627 // Check the rest of the elements to see if they are consecutive.
2628 for (++i; i != e; ++i) {
2633 Unary = Unary && (m < (int)e);
2634 NeedsUnary = NeedsUnary || (m < s);
2636 if (NeedsUnary && !Unary)
2638 if (Unary && m != ((s+i) & (e-1)))
2640 if (!Unary && m != (s+i))
2646 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2647 SmallVector<int, 8> M;
2649 return ::isPALIGNRMask(M, N->getValueType(0), true);
2652 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2653 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2654 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2655 int NumElems = VT.getVectorNumElements();
2656 if (NumElems != 2 && NumElems != 4)
2659 int Half = NumElems / 2;
2660 for (int i = 0; i < Half; ++i)
2661 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2663 for (int i = Half; i < NumElems; ++i)
2664 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2670 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2671 SmallVector<int, 8> M;
2673 return ::isSHUFPMask(M, N->getValueType(0));
2676 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2677 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2678 /// half elements to come from vector 1 (which would equal the dest.) and
2679 /// the upper half to come from vector 2.
2680 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2681 int NumElems = VT.getVectorNumElements();
2683 if (NumElems != 2 && NumElems != 4)
2686 int Half = NumElems / 2;
2687 for (int i = 0; i < Half; ++i)
2688 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2690 for (int i = Half; i < NumElems; ++i)
2691 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2696 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2697 SmallVector<int, 8> M;
2699 return isCommutedSHUFPMask(M, N->getValueType(0));
2702 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2703 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2704 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2705 if (N->getValueType(0).getVectorNumElements() != 4)
2708 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2709 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2710 isUndefOrEqual(N->getMaskElt(1), 7) &&
2711 isUndefOrEqual(N->getMaskElt(2), 2) &&
2712 isUndefOrEqual(N->getMaskElt(3), 3);
2715 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2716 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2718 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2719 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2724 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2725 isUndefOrEqual(N->getMaskElt(1), 3) &&
2726 isUndefOrEqual(N->getMaskElt(2), 2) &&
2727 isUndefOrEqual(N->getMaskElt(3), 3);
2730 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2731 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2732 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2733 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2735 if (NumElems != 2 && NumElems != 4)
2738 for (unsigned i = 0; i < NumElems/2; ++i)
2739 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2742 for (unsigned i = NumElems/2; i < NumElems; ++i)
2743 if (!isUndefOrEqual(N->getMaskElt(i), i))
2749 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2750 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2751 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2752 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2754 if (NumElems != 2 && NumElems != 4)
2757 for (unsigned i = 0; i < NumElems/2; ++i)
2758 if (!isUndefOrEqual(N->getMaskElt(i), i))
2761 for (unsigned i = 0; i < NumElems/2; ++i)
2762 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2768 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2769 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2770 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2771 bool V2IsSplat = false) {
2772 int NumElts = VT.getVectorNumElements();
2773 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2776 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2778 int BitI1 = Mask[i+1];
2779 if (!isUndefOrEqual(BitI, j))
2782 if (!isUndefOrEqual(BitI1, NumElts))
2785 if (!isUndefOrEqual(BitI1, j + NumElts))
2792 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2793 SmallVector<int, 8> M;
2795 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2798 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2799 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2800 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2801 bool V2IsSplat = false) {
2802 int NumElts = VT.getVectorNumElements();
2803 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2806 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2808 int BitI1 = Mask[i+1];
2809 if (!isUndefOrEqual(BitI, j + NumElts/2))
2812 if (isUndefOrEqual(BitI1, NumElts))
2815 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2822 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2823 SmallVector<int, 8> M;
2825 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2828 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2829 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2831 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2832 int NumElems = VT.getVectorNumElements();
2833 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2836 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2838 int BitI1 = Mask[i+1];
2839 if (!isUndefOrEqual(BitI, j))
2841 if (!isUndefOrEqual(BitI1, j))
2847 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2848 SmallVector<int, 8> M;
2850 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2853 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2854 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2856 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2857 int NumElems = VT.getVectorNumElements();
2858 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2861 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2863 int BitI1 = Mask[i+1];
2864 if (!isUndefOrEqual(BitI, j))
2866 if (!isUndefOrEqual(BitI1, j))
2872 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2873 SmallVector<int, 8> M;
2875 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2878 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2879 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2880 /// MOVSD, and MOVD, i.e. setting the lowest element.
2881 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2882 if (VT.getVectorElementType().getSizeInBits() < 32)
2885 int NumElts = VT.getVectorNumElements();
2887 if (!isUndefOrEqual(Mask[0], NumElts))
2890 for (int i = 1; i < NumElts; ++i)
2891 if (!isUndefOrEqual(Mask[i], i))
2897 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2898 SmallVector<int, 8> M;
2900 return ::isMOVLMask(M, N->getValueType(0));
2903 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2904 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2905 /// element of vector 2 and the other elements to come from vector 1 in order.
2906 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2907 bool V2IsSplat = false, bool V2IsUndef = false) {
2908 int NumOps = VT.getVectorNumElements();
2909 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2912 if (!isUndefOrEqual(Mask[0], 0))
2915 for (int i = 1; i < NumOps; ++i)
2916 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2917 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2918 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2924 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2925 bool V2IsUndef = false) {
2926 SmallVector<int, 8> M;
2928 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2931 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2932 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2933 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2934 if (N->getValueType(0).getVectorNumElements() != 4)
2937 // Expect 1, 1, 3, 3
2938 for (unsigned i = 0; i < 2; ++i) {
2939 int Elt = N->getMaskElt(i);
2940 if (Elt >= 0 && Elt != 1)
2945 for (unsigned i = 2; i < 4; ++i) {
2946 int Elt = N->getMaskElt(i);
2947 if (Elt >= 0 && Elt != 3)
2952 // Don't use movshdup if it can be done with a shufps.
2953 // FIXME: verify that matching u, u, 3, 3 is what we want.
2957 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2958 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2959 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2960 if (N->getValueType(0).getVectorNumElements() != 4)
2963 // Expect 0, 0, 2, 2
2964 for (unsigned i = 0; i < 2; ++i)
2965 if (N->getMaskElt(i) > 0)
2969 for (unsigned i = 2; i < 4; ++i) {
2970 int Elt = N->getMaskElt(i);
2971 if (Elt >= 0 && Elt != 2)
2976 // Don't use movsldup if it can be done with a shufps.
2980 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2981 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2982 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2983 int e = N->getValueType(0).getVectorNumElements() / 2;
2985 for (int i = 0; i < e; ++i)
2986 if (!isUndefOrEqual(N->getMaskElt(i), i))
2988 for (int i = 0; i < e; ++i)
2989 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2994 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2995 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
2996 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2997 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2998 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3000 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3002 for (int i = 0; i < NumOperands; ++i) {
3003 int Val = SVOp->getMaskElt(NumOperands-i-1);
3004 if (Val < 0) Val = 0;
3005 if (Val >= NumOperands) Val -= NumOperands;
3007 if (i != NumOperands - 1)
3013 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3014 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3015 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3016 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3018 // 8 nodes, but we only care about the last 4.
3019 for (unsigned i = 7; i >= 4; --i) {
3020 int Val = SVOp->getMaskElt(i);
3029 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3030 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3031 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3032 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3034 // 8 nodes, but we only care about the first 4.
3035 for (int i = 3; i >= 0; --i) {
3036 int Val = SVOp->getMaskElt(i);
3045 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3046 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3047 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3048 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3049 EVT VVT = N->getValueType(0);
3050 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3054 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3055 Val = SVOp->getMaskElt(i);
3059 return (Val - i) * EltSize;
3062 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3064 bool X86::isZeroNode(SDValue Elt) {
3065 return ((isa<ConstantSDNode>(Elt) &&
3066 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3067 (isa<ConstantFPSDNode>(Elt) &&
3068 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3071 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3072 /// their permute mask.
3073 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3074 SelectionDAG &DAG) {
3075 EVT VT = SVOp->getValueType(0);
3076 unsigned NumElems = VT.getVectorNumElements();
3077 SmallVector<int, 8> MaskVec;
3079 for (unsigned i = 0; i != NumElems; ++i) {
3080 int idx = SVOp->getMaskElt(i);
3082 MaskVec.push_back(idx);
3083 else if (idx < (int)NumElems)
3084 MaskVec.push_back(idx + NumElems);
3086 MaskVec.push_back(idx - NumElems);
3088 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3089 SVOp->getOperand(0), &MaskVec[0]);
3092 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3093 /// the two vector operands have swapped position.
3094 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3095 unsigned NumElems = VT.getVectorNumElements();
3096 for (unsigned i = 0; i != NumElems; ++i) {
3100 else if (idx < (int)NumElems)
3101 Mask[i] = idx + NumElems;
3103 Mask[i] = idx - NumElems;
3107 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3108 /// match movhlps. The lower half elements should come from upper half of
3109 /// V1 (and in order), and the upper half elements should come from the upper
3110 /// half of V2 (and in order).
3111 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3112 if (Op->getValueType(0).getVectorNumElements() != 4)
3114 for (unsigned i = 0, e = 2; i != e; ++i)
3115 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3117 for (unsigned i = 2; i != 4; ++i)
3118 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3123 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3124 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3126 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3127 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3129 N = N->getOperand(0).getNode();
3130 if (!ISD::isNON_EXTLoad(N))
3133 *LD = cast<LoadSDNode>(N);
3137 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3138 /// match movlp{s|d}. The lower half elements should come from lower half of
3139 /// V1 (and in order), and the upper half elements should come from the upper
3140 /// half of V2 (and in order). And since V1 will become the source of the
3141 /// MOVLP, it must be either a vector load or a scalar load to vector.
3142 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3143 ShuffleVectorSDNode *Op) {
3144 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3146 // Is V2 is a vector load, don't do this transformation. We will try to use
3147 // load folding shufps op.
3148 if (ISD::isNON_EXTLoad(V2))
3151 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3153 if (NumElems != 2 && NumElems != 4)
3155 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3156 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3158 for (unsigned i = NumElems/2; i != NumElems; ++i)
3159 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3164 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3166 static bool isSplatVector(SDNode *N) {
3167 if (N->getOpcode() != ISD::BUILD_VECTOR)
3170 SDValue SplatValue = N->getOperand(0);
3171 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3172 if (N->getOperand(i) != SplatValue)
3177 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3178 /// to an zero vector.
3179 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3180 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3181 SDValue V1 = N->getOperand(0);
3182 SDValue V2 = N->getOperand(1);
3183 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3184 for (unsigned i = 0; i != NumElems; ++i) {
3185 int Idx = N->getMaskElt(i);
3186 if (Idx >= (int)NumElems) {
3187 unsigned Opc = V2.getOpcode();
3188 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3190 if (Opc != ISD::BUILD_VECTOR ||
3191 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3193 } else if (Idx >= 0) {
3194 unsigned Opc = V1.getOpcode();
3195 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3197 if (Opc != ISD::BUILD_VECTOR ||
3198 !X86::isZeroNode(V1.getOperand(Idx)))
3205 /// getZeroVector - Returns a vector of specified type with all zero elements.
3207 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3209 assert(VT.isVector() && "Expected a vector type");
3211 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3212 // type. This ensures they get CSE'd.
3214 if (VT.getSizeInBits() == 64) { // MMX
3215 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3216 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3217 } else if (HasSSE2) { // SSE2
3218 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3219 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3221 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3222 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3224 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3227 /// getOnesVector - Returns a vector of specified type with all bits set.
3229 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3230 assert(VT.isVector() && "Expected a vector type");
3232 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3233 // type. This ensures they get CSE'd.
3234 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3236 if (VT.getSizeInBits() == 64) // MMX
3237 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3239 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3240 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3244 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3245 /// that point to V2 points to its first element.
3246 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3247 EVT VT = SVOp->getValueType(0);
3248 unsigned NumElems = VT.getVectorNumElements();
3250 bool Changed = false;
3251 SmallVector<int, 8> MaskVec;
3252 SVOp->getMask(MaskVec);
3254 for (unsigned i = 0; i != NumElems; ++i) {
3255 if (MaskVec[i] > (int)NumElems) {
3256 MaskVec[i] = NumElems;
3261 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3262 SVOp->getOperand(1), &MaskVec[0]);
3263 return SDValue(SVOp, 0);
3266 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3267 /// operation of specified width.
3268 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3270 unsigned NumElems = VT.getVectorNumElements();
3271 SmallVector<int, 8> Mask;
3272 Mask.push_back(NumElems);
3273 for (unsigned i = 1; i != NumElems; ++i)
3275 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3278 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3279 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3281 unsigned NumElems = VT.getVectorNumElements();
3282 SmallVector<int, 8> Mask;
3283 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3285 Mask.push_back(i + NumElems);
3287 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3290 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3291 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3293 unsigned NumElems = VT.getVectorNumElements();
3294 unsigned Half = NumElems/2;
3295 SmallVector<int, 8> Mask;
3296 for (unsigned i = 0; i != Half; ++i) {
3297 Mask.push_back(i + Half);
3298 Mask.push_back(i + NumElems + Half);
3300 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3303 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3304 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3306 if (SV->getValueType(0).getVectorNumElements() <= 4)
3307 return SDValue(SV, 0);
3309 EVT PVT = MVT::v4f32;
3310 EVT VT = SV->getValueType(0);
3311 DebugLoc dl = SV->getDebugLoc();
3312 SDValue V1 = SV->getOperand(0);
3313 int NumElems = VT.getVectorNumElements();
3314 int EltNo = SV->getSplatIndex();
3316 // unpack elements to the correct location
3317 while (NumElems > 4) {
3318 if (EltNo < NumElems/2) {
3319 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3321 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3322 EltNo -= NumElems/2;
3327 // Perform the splat.
3328 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3329 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3330 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3331 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3334 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3335 /// vector of zero or undef vector. This produces a shuffle where the low
3336 /// element of V2 is swizzled into the zero/undef vector, landing at element
3337 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3338 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3339 bool isZero, bool HasSSE2,
3340 SelectionDAG &DAG) {
3341 EVT VT = V2.getValueType();
3343 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3344 unsigned NumElems = VT.getVectorNumElements();
3345 SmallVector<int, 16> MaskVec;
3346 for (unsigned i = 0; i != NumElems; ++i)
3347 // If this is the insertion idx, put the low elt of V2 here.
3348 MaskVec.push_back(i == Idx ? NumElems : i);
3349 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3352 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3353 /// a shuffle that is zero.
3355 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3356 bool Low, SelectionDAG &DAG) {
3357 unsigned NumZeros = 0;
3358 for (int i = 0; i < NumElems; ++i) {
3359 unsigned Index = Low ? i : NumElems-i-1;
3360 int Idx = SVOp->getMaskElt(Index);
3365 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3366 if (Elt.getNode() && X86::isZeroNode(Elt))
3374 /// isVectorShift - Returns true if the shuffle can be implemented as a
3375 /// logical left or right shift of a vector.
3376 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3377 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3378 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3379 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3382 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3385 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3389 bool SeenV1 = false;
3390 bool SeenV2 = false;
3391 for (int i = NumZeros; i < NumElems; ++i) {
3392 int Val = isLeft ? (i - NumZeros) : i;
3393 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3405 if (SeenV1 && SeenV2)
3408 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3414 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3416 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3417 unsigned NumNonZero, unsigned NumZero,
3418 SelectionDAG &DAG, TargetLowering &TLI) {
3422 DebugLoc dl = Op.getDebugLoc();
3425 for (unsigned i = 0; i < 16; ++i) {
3426 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3427 if (ThisIsNonZero && First) {
3429 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3431 V = DAG.getUNDEF(MVT::v8i16);
3436 SDValue ThisElt(0, 0), LastElt(0, 0);
3437 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3438 if (LastIsNonZero) {
3439 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3440 MVT::i16, Op.getOperand(i-1));
3442 if (ThisIsNonZero) {
3443 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3444 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3445 ThisElt, DAG.getConstant(8, MVT::i8));
3447 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3451 if (ThisElt.getNode())
3452 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3453 DAG.getIntPtrConstant(i/2));
3457 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3460 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3462 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3463 unsigned NumNonZero, unsigned NumZero,
3464 SelectionDAG &DAG, TargetLowering &TLI) {
3468 DebugLoc dl = Op.getDebugLoc();
3471 for (unsigned i = 0; i < 8; ++i) {
3472 bool isNonZero = (NonZeros & (1 << i)) != 0;
3476 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3478 V = DAG.getUNDEF(MVT::v8i16);
3481 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3482 MVT::v8i16, V, Op.getOperand(i),
3483 DAG.getIntPtrConstant(i));
3490 /// getVShift - Return a vector logical shift node.
3492 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3493 unsigned NumBits, SelectionDAG &DAG,
3494 const TargetLowering &TLI, DebugLoc dl) {
3495 bool isMMX = VT.getSizeInBits() == 64;
3496 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3497 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3498 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3499 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3500 DAG.getNode(Opc, dl, ShVT, SrcOp,
3501 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3505 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3506 SelectionDAG &DAG) {
3508 // Check if the scalar load can be widened into a vector load. And if
3509 // the address is "base + cst" see if the cst can be "absorbed" into
3510 // the shuffle mask.
3511 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3512 SDValue Ptr = LD->getBasePtr();
3513 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3515 EVT PVT = LD->getValueType(0);
3516 if (PVT != MVT::i32 && PVT != MVT::f32)
3521 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3522 FI = FINode->getIndex();
3524 } else if (Ptr.getOpcode() == ISD::ADD &&
3525 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3526 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3527 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3528 Offset = Ptr.getConstantOperandVal(1);
3529 Ptr = Ptr.getOperand(0);
3534 SDValue Chain = LD->getChain();
3535 // Make sure the stack object alignment is at least 16.
3536 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3537 if (DAG.InferPtrAlignment(Ptr) < 16) {
3538 if (MFI->isFixedObjectIndex(FI)) {
3539 // Can't change the alignment. FIXME: It's possible to compute
3540 // the exact stack offset and reference FI + adjust offset instead.
3541 // If someone *really* cares about this. That's the way to implement it.
3544 MFI->setObjectAlignment(FI, 16);
3548 // (Offset % 16) must be multiple of 4. Then address is then
3549 // Ptr + (Offset & ~15).
3552 if ((Offset % 16) & 3)
3554 int64_t StartOffset = Offset & ~15;
3556 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3557 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3559 int EltNo = (Offset - StartOffset) >> 2;
3560 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3561 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3562 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0);
3563 // Canonicalize it to a v4i32 shuffle.
3564 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3565 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3566 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3567 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3574 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3575 DebugLoc dl = Op.getDebugLoc();
3576 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3577 if (ISD::isBuildVectorAllZeros(Op.getNode())
3578 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3579 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3580 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3581 // eliminated on x86-32 hosts.
3582 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3585 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3586 return getOnesVector(Op.getValueType(), DAG, dl);
3587 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3590 EVT VT = Op.getValueType();
3591 EVT ExtVT = VT.getVectorElementType();
3592 unsigned EVTBits = ExtVT.getSizeInBits();
3594 unsigned NumElems = Op.getNumOperands();
3595 unsigned NumZero = 0;
3596 unsigned NumNonZero = 0;
3597 unsigned NonZeros = 0;
3598 bool IsAllConstants = true;
3599 SmallSet<SDValue, 8> Values;
3600 for (unsigned i = 0; i < NumElems; ++i) {
3601 SDValue Elt = Op.getOperand(i);
3602 if (Elt.getOpcode() == ISD::UNDEF)
3605 if (Elt.getOpcode() != ISD::Constant &&
3606 Elt.getOpcode() != ISD::ConstantFP)
3607 IsAllConstants = false;
3608 if (X86::isZeroNode(Elt))
3611 NonZeros |= (1 << i);
3616 if (NumNonZero == 0) {
3617 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3618 return DAG.getUNDEF(VT);
3621 // Special case for single non-zero, non-undef, element.
3622 if (NumNonZero == 1) {
3623 unsigned Idx = CountTrailingZeros_32(NonZeros);
3624 SDValue Item = Op.getOperand(Idx);
3626 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3627 // the value are obviously zero, truncate the value to i32 and do the
3628 // insertion that way. Only do this if the value is non-constant or if the
3629 // value is a constant being inserted into element 0. It is cheaper to do
3630 // a constant pool load than it is to do a movd + shuffle.
3631 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3632 (!IsAllConstants || Idx == 0)) {
3633 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3634 // Handle MMX and SSE both.
3635 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3636 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3638 // Truncate the value (which may itself be a constant) to i32, and
3639 // convert it to a vector with movd (S2V+shuffle to zero extend).
3640 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3641 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3642 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3643 Subtarget->hasSSE2(), DAG);
3645 // Now we have our 32-bit value zero extended in the low element of
3646 // a vector. If Idx != 0, swizzle it into place.
3648 SmallVector<int, 4> Mask;
3649 Mask.push_back(Idx);
3650 for (unsigned i = 1; i != VecElts; ++i)
3652 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3653 DAG.getUNDEF(Item.getValueType()),
3656 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3660 // If we have a constant or non-constant insertion into the low element of
3661 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3662 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3663 // depending on what the source datatype is.
3666 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3667 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3668 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3669 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3670 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3671 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3673 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3674 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3675 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3676 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3677 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3678 Subtarget->hasSSE2(), DAG);
3679 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3683 // Is it a vector logical left shift?
3684 if (NumElems == 2 && Idx == 1 &&
3685 X86::isZeroNode(Op.getOperand(0)) &&
3686 !X86::isZeroNode(Op.getOperand(1))) {
3687 unsigned NumBits = VT.getSizeInBits();
3688 return getVShift(true, VT,
3689 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3690 VT, Op.getOperand(1)),
3691 NumBits/2, DAG, *this, dl);
3694 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3697 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3698 // is a non-constant being inserted into an element other than the low one,
3699 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3700 // movd/movss) to move this into the low element, then shuffle it into
3702 if (EVTBits == 32) {
3703 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3705 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3706 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3707 Subtarget->hasSSE2(), DAG);
3708 SmallVector<int, 8> MaskVec;
3709 for (unsigned i = 0; i < NumElems; i++)
3710 MaskVec.push_back(i == Idx ? 0 : 1);
3711 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3715 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3716 if (Values.size() == 1) {
3717 if (EVTBits == 32) {
3718 // Instead of a shuffle like this:
3719 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3720 // Check if it's possible to issue this instead.
3721 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3722 unsigned Idx = CountTrailingZeros_32(NonZeros);
3723 SDValue Item = Op.getOperand(Idx);
3724 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3725 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3730 // A vector full of immediates; various special cases are already
3731 // handled, so this is best done with a single constant-pool load.
3735 // Let legalizer expand 2-wide build_vectors.
3736 if (EVTBits == 64) {
3737 if (NumNonZero == 1) {
3738 // One half is zero or undef.
3739 unsigned Idx = CountTrailingZeros_32(NonZeros);
3740 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3741 Op.getOperand(Idx));
3742 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3743 Subtarget->hasSSE2(), DAG);
3748 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3749 if (EVTBits == 8 && NumElems == 16) {
3750 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3752 if (V.getNode()) return V;
3755 if (EVTBits == 16 && NumElems == 8) {
3756 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3758 if (V.getNode()) return V;
3761 // If element VT is == 32 bits, turn it into a number of shuffles.
3762 SmallVector<SDValue, 8> V;
3764 if (NumElems == 4 && NumZero > 0) {
3765 for (unsigned i = 0; i < 4; ++i) {
3766 bool isZero = !(NonZeros & (1 << i));
3768 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3770 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3773 for (unsigned i = 0; i < 2; ++i) {
3774 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3777 V[i] = V[i*2]; // Must be a zero vector.
3780 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3783 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3786 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3791 SmallVector<int, 8> MaskVec;
3792 bool Reverse = (NonZeros & 0x3) == 2;
3793 for (unsigned i = 0; i < 2; ++i)
3794 MaskVec.push_back(Reverse ? 1-i : i);
3795 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3796 for (unsigned i = 0; i < 2; ++i)
3797 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3798 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3801 if (Values.size() > 2) {
3802 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3803 // values to be inserted is equal to the number of elements, in which case
3804 // use the unpack code below in the hopes of matching the consecutive elts
3805 // load merge pattern for shuffles.
3806 // FIXME: We could probably just check that here directly.
3807 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3808 getSubtarget()->hasSSE41()) {
3809 V[0] = DAG.getUNDEF(VT);
3810 for (unsigned i = 0; i < NumElems; ++i)
3811 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3812 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3813 Op.getOperand(i), DAG.getIntPtrConstant(i));
3816 // Expand into a number of unpckl*.
3818 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3819 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3820 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3821 for (unsigned i = 0; i < NumElems; ++i)
3822 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3824 while (NumElems != 0) {
3825 for (unsigned i = 0; i < NumElems; ++i)
3826 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3836 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3837 // We support concatenate two MMX registers and place them in a MMX
3838 // register. This is better than doing a stack convert.
3839 DebugLoc dl = Op.getDebugLoc();
3840 EVT ResVT = Op.getValueType();
3841 assert(Op.getNumOperands() == 2);
3842 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3843 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3845 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3846 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3847 InVec = Op.getOperand(1);
3848 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3849 unsigned NumElts = ResVT.getVectorNumElements();
3850 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3851 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3852 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3854 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3855 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3856 Mask[0] = 0; Mask[1] = 2;
3857 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3859 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3862 // v8i16 shuffles - Prefer shuffles in the following order:
3863 // 1. [all] pshuflw, pshufhw, optional move
3864 // 2. [ssse3] 1 x pshufb
3865 // 3. [ssse3] 2 x pshufb + 1 x por
3866 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3868 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3869 SelectionDAG &DAG, X86TargetLowering &TLI) {
3870 SDValue V1 = SVOp->getOperand(0);
3871 SDValue V2 = SVOp->getOperand(1);
3872 DebugLoc dl = SVOp->getDebugLoc();
3873 SmallVector<int, 8> MaskVals;
3875 // Determine if more than 1 of the words in each of the low and high quadwords
3876 // of the result come from the same quadword of one of the two inputs. Undef
3877 // mask values count as coming from any quadword, for better codegen.
3878 SmallVector<unsigned, 4> LoQuad(4);
3879 SmallVector<unsigned, 4> HiQuad(4);
3880 BitVector InputQuads(4);
3881 for (unsigned i = 0; i < 8; ++i) {
3882 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3883 int EltIdx = SVOp->getMaskElt(i);
3884 MaskVals.push_back(EltIdx);
3893 InputQuads.set(EltIdx / 4);
3896 int BestLoQuad = -1;
3897 unsigned MaxQuad = 1;
3898 for (unsigned i = 0; i < 4; ++i) {
3899 if (LoQuad[i] > MaxQuad) {
3901 MaxQuad = LoQuad[i];
3905 int BestHiQuad = -1;
3907 for (unsigned i = 0; i < 4; ++i) {
3908 if (HiQuad[i] > MaxQuad) {
3910 MaxQuad = HiQuad[i];
3914 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3915 // of the two input vectors, shuffle them into one input vector so only a
3916 // single pshufb instruction is necessary. If There are more than 2 input
3917 // quads, disable the next transformation since it does not help SSSE3.
3918 bool V1Used = InputQuads[0] || InputQuads[1];
3919 bool V2Used = InputQuads[2] || InputQuads[3];
3920 if (TLI.getSubtarget()->hasSSSE3()) {
3921 if (InputQuads.count() == 2 && V1Used && V2Used) {
3922 BestLoQuad = InputQuads.find_first();
3923 BestHiQuad = InputQuads.find_next(BestLoQuad);
3925 if (InputQuads.count() > 2) {
3931 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3932 // the shuffle mask. If a quad is scored as -1, that means that it contains
3933 // words from all 4 input quadwords.
3935 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3936 SmallVector<int, 8> MaskV;
3937 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3938 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3939 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3940 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3941 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3942 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3944 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3945 // source words for the shuffle, to aid later transformations.
3946 bool AllWordsInNewV = true;
3947 bool InOrder[2] = { true, true };
3948 for (unsigned i = 0; i != 8; ++i) {
3949 int idx = MaskVals[i];
3951 InOrder[i/4] = false;
3952 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3954 AllWordsInNewV = false;
3958 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3959 if (AllWordsInNewV) {
3960 for (int i = 0; i != 8; ++i) {
3961 int idx = MaskVals[i];
3964 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3965 if ((idx != i) && idx < 4)
3967 if ((idx != i) && idx > 3)
3976 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3977 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3978 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3979 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3980 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3984 // If we have SSSE3, and all words of the result are from 1 input vector,
3985 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3986 // is present, fall back to case 4.
3987 if (TLI.getSubtarget()->hasSSSE3()) {
3988 SmallVector<SDValue,16> pshufbMask;
3990 // If we have elements from both input vectors, set the high bit of the
3991 // shuffle mask element to zero out elements that come from V2 in the V1
3992 // mask, and elements that come from V1 in the V2 mask, so that the two
3993 // results can be OR'd together.
3994 bool TwoInputs = V1Used && V2Used;
3995 for (unsigned i = 0; i != 8; ++i) {
3996 int EltIdx = MaskVals[i] * 2;
3997 if (TwoInputs && (EltIdx >= 16)) {
3998 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3999 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4002 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4003 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4005 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4006 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4007 DAG.getNode(ISD::BUILD_VECTOR, dl,
4008 MVT::v16i8, &pshufbMask[0], 16));
4010 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4012 // Calculate the shuffle mask for the second input, shuffle it, and
4013 // OR it with the first shuffled input.
4015 for (unsigned i = 0; i != 8; ++i) {
4016 int EltIdx = MaskVals[i] * 2;
4018 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4019 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4022 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4023 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4025 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4026 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4027 DAG.getNode(ISD::BUILD_VECTOR, dl,
4028 MVT::v16i8, &pshufbMask[0], 16));
4029 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4030 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4033 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4034 // and update MaskVals with new element order.
4035 BitVector InOrder(8);
4036 if (BestLoQuad >= 0) {
4037 SmallVector<int, 8> MaskV;
4038 for (int i = 0; i != 4; ++i) {
4039 int idx = MaskVals[i];
4041 MaskV.push_back(-1);
4043 } else if ((idx / 4) == BestLoQuad) {
4044 MaskV.push_back(idx & 3);
4047 MaskV.push_back(-1);
4050 for (unsigned i = 4; i != 8; ++i)
4052 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4056 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4057 // and update MaskVals with the new element order.
4058 if (BestHiQuad >= 0) {
4059 SmallVector<int, 8> MaskV;
4060 for (unsigned i = 0; i != 4; ++i)
4062 for (unsigned i = 4; i != 8; ++i) {
4063 int idx = MaskVals[i];
4065 MaskV.push_back(-1);
4067 } else if ((idx / 4) == BestHiQuad) {
4068 MaskV.push_back((idx & 3) + 4);
4071 MaskV.push_back(-1);
4074 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4078 // In case BestHi & BestLo were both -1, which means each quadword has a word
4079 // from each of the four input quadwords, calculate the InOrder bitvector now
4080 // before falling through to the insert/extract cleanup.
4081 if (BestLoQuad == -1 && BestHiQuad == -1) {
4083 for (int i = 0; i != 8; ++i)
4084 if (MaskVals[i] < 0 || MaskVals[i] == i)
4088 // The other elements are put in the right place using pextrw and pinsrw.
4089 for (unsigned i = 0; i != 8; ++i) {
4092 int EltIdx = MaskVals[i];
4095 SDValue ExtOp = (EltIdx < 8)
4096 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4097 DAG.getIntPtrConstant(EltIdx))
4098 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4099 DAG.getIntPtrConstant(EltIdx - 8));
4100 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4101 DAG.getIntPtrConstant(i));
4106 // v16i8 shuffles - Prefer shuffles in the following order:
4107 // 1. [ssse3] 1 x pshufb
4108 // 2. [ssse3] 2 x pshufb + 1 x por
4109 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4111 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4112 SelectionDAG &DAG, X86TargetLowering &TLI) {
4113 SDValue V1 = SVOp->getOperand(0);
4114 SDValue V2 = SVOp->getOperand(1);
4115 DebugLoc dl = SVOp->getDebugLoc();
4116 SmallVector<int, 16> MaskVals;
4117 SVOp->getMask(MaskVals);
4119 // If we have SSSE3, case 1 is generated when all result bytes come from
4120 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4121 // present, fall back to case 3.
4122 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4125 for (unsigned i = 0; i < 16; ++i) {
4126 int EltIdx = MaskVals[i];
4135 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4136 if (TLI.getSubtarget()->hasSSSE3()) {
4137 SmallVector<SDValue,16> pshufbMask;
4139 // If all result elements are from one input vector, then only translate
4140 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4142 // Otherwise, we have elements from both input vectors, and must zero out
4143 // elements that come from V2 in the first mask, and V1 in the second mask
4144 // so that we can OR them together.
4145 bool TwoInputs = !(V1Only || V2Only);
4146 for (unsigned i = 0; i != 16; ++i) {
4147 int EltIdx = MaskVals[i];
4148 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4149 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4152 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4154 // If all the elements are from V2, assign it to V1 and return after
4155 // building the first pshufb.
4158 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4159 DAG.getNode(ISD::BUILD_VECTOR, dl,
4160 MVT::v16i8, &pshufbMask[0], 16));
4164 // Calculate the shuffle mask for the second input, shuffle it, and
4165 // OR it with the first shuffled input.
4167 for (unsigned i = 0; i != 16; ++i) {
4168 int EltIdx = MaskVals[i];
4170 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4173 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4175 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4176 DAG.getNode(ISD::BUILD_VECTOR, dl,
4177 MVT::v16i8, &pshufbMask[0], 16));
4178 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4181 // No SSSE3 - Calculate in place words and then fix all out of place words
4182 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4183 // the 16 different words that comprise the two doublequadword input vectors.
4184 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4185 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4186 SDValue NewV = V2Only ? V2 : V1;
4187 for (int i = 0; i != 8; ++i) {
4188 int Elt0 = MaskVals[i*2];
4189 int Elt1 = MaskVals[i*2+1];
4191 // This word of the result is all undef, skip it.
4192 if (Elt0 < 0 && Elt1 < 0)
4195 // This word of the result is already in the correct place, skip it.
4196 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4198 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4201 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4202 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4205 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4206 // using a single extract together, load it and store it.
4207 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4208 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4209 DAG.getIntPtrConstant(Elt1 / 2));
4210 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4211 DAG.getIntPtrConstant(i));
4215 // If Elt1 is defined, extract it from the appropriate source. If the
4216 // source byte is not also odd, shift the extracted word left 8 bits
4217 // otherwise clear the bottom 8 bits if we need to do an or.
4219 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4220 DAG.getIntPtrConstant(Elt1 / 2));
4221 if ((Elt1 & 1) == 0)
4222 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4223 DAG.getConstant(8, TLI.getShiftAmountTy()));
4225 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4226 DAG.getConstant(0xFF00, MVT::i16));
4228 // If Elt0 is defined, extract it from the appropriate source. If the
4229 // source byte is not also even, shift the extracted word right 8 bits. If
4230 // Elt1 was also defined, OR the extracted values together before
4231 // inserting them in the result.
4233 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4234 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4235 if ((Elt0 & 1) != 0)
4236 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4237 DAG.getConstant(8, TLI.getShiftAmountTy()));
4239 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4240 DAG.getConstant(0x00FF, MVT::i16));
4241 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4244 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4245 DAG.getIntPtrConstant(i));
4247 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4250 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4251 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4252 /// done when every pair / quad of shuffle mask elements point to elements in
4253 /// the right sequence. e.g.
4254 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4256 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4258 TargetLowering &TLI, DebugLoc dl) {
4259 EVT VT = SVOp->getValueType(0);
4260 SDValue V1 = SVOp->getOperand(0);
4261 SDValue V2 = SVOp->getOperand(1);
4262 unsigned NumElems = VT.getVectorNumElements();
4263 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4264 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4265 EVT MaskEltVT = MaskVT.getVectorElementType();
4267 switch (VT.getSimpleVT().SimpleTy) {
4268 default: assert(false && "Unexpected!");
4269 case MVT::v4f32: NewVT = MVT::v2f64; break;
4270 case MVT::v4i32: NewVT = MVT::v2i64; break;
4271 case MVT::v8i16: NewVT = MVT::v4i32; break;
4272 case MVT::v16i8: NewVT = MVT::v4i32; break;
4275 if (NewWidth == 2) {
4281 int Scale = NumElems / NewWidth;
4282 SmallVector<int, 8> MaskVec;
4283 for (unsigned i = 0; i < NumElems; i += Scale) {
4285 for (int j = 0; j < Scale; ++j) {
4286 int EltIdx = SVOp->getMaskElt(i+j);
4290 StartIdx = EltIdx - (EltIdx % Scale);
4291 if (EltIdx != StartIdx + j)
4295 MaskVec.push_back(-1);
4297 MaskVec.push_back(StartIdx / Scale);
4300 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4301 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4302 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4305 /// getVZextMovL - Return a zero-extending vector move low node.
4307 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4308 SDValue SrcOp, SelectionDAG &DAG,
4309 const X86Subtarget *Subtarget, DebugLoc dl) {
4310 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4311 LoadSDNode *LD = NULL;
4312 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4313 LD = dyn_cast<LoadSDNode>(SrcOp);
4315 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4317 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4318 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4319 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4320 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4321 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4323 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4324 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4325 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4326 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4334 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4335 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4336 DAG.getNode(ISD::BIT_CONVERT, dl,
4340 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4343 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4344 SDValue V1 = SVOp->getOperand(0);
4345 SDValue V2 = SVOp->getOperand(1);
4346 DebugLoc dl = SVOp->getDebugLoc();
4347 EVT VT = SVOp->getValueType(0);
4349 SmallVector<std::pair<int, int>, 8> Locs;
4351 SmallVector<int, 8> Mask1(4U, -1);
4352 SmallVector<int, 8> PermMask;
4353 SVOp->getMask(PermMask);
4357 for (unsigned i = 0; i != 4; ++i) {
4358 int Idx = PermMask[i];
4360 Locs[i] = std::make_pair(-1, -1);
4362 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4364 Locs[i] = std::make_pair(0, NumLo);
4368 Locs[i] = std::make_pair(1, NumHi);
4370 Mask1[2+NumHi] = Idx;
4376 if (NumLo <= 2 && NumHi <= 2) {
4377 // If no more than two elements come from either vector. This can be
4378 // implemented with two shuffles. First shuffle gather the elements.
4379 // The second shuffle, which takes the first shuffle as both of its
4380 // vector operands, put the elements into the right order.
4381 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4383 SmallVector<int, 8> Mask2(4U, -1);
4385 for (unsigned i = 0; i != 4; ++i) {
4386 if (Locs[i].first == -1)
4389 unsigned Idx = (i < 2) ? 0 : 4;
4390 Idx += Locs[i].first * 2 + Locs[i].second;
4395 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4396 } else if (NumLo == 3 || NumHi == 3) {
4397 // Otherwise, we must have three elements from one vector, call it X, and
4398 // one element from the other, call it Y. First, use a shufps to build an
4399 // intermediate vector with the one element from Y and the element from X
4400 // that will be in the same half in the final destination (the indexes don't
4401 // matter). Then, use a shufps to build the final vector, taking the half
4402 // containing the element from Y from the intermediate, and the other half
4405 // Normalize it so the 3 elements come from V1.
4406 CommuteVectorShuffleMask(PermMask, VT);
4410 // Find the element from V2.
4412 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4413 int Val = PermMask[HiIndex];
4420 Mask1[0] = PermMask[HiIndex];
4422 Mask1[2] = PermMask[HiIndex^1];
4424 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4427 Mask1[0] = PermMask[0];
4428 Mask1[1] = PermMask[1];
4429 Mask1[2] = HiIndex & 1 ? 6 : 4;
4430 Mask1[3] = HiIndex & 1 ? 4 : 6;
4431 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4433 Mask1[0] = HiIndex & 1 ? 2 : 0;
4434 Mask1[1] = HiIndex & 1 ? 0 : 2;
4435 Mask1[2] = PermMask[2];
4436 Mask1[3] = PermMask[3];
4441 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4445 // Break it into (shuffle shuffle_hi, shuffle_lo).
4447 SmallVector<int,8> LoMask(4U, -1);
4448 SmallVector<int,8> HiMask(4U, -1);
4450 SmallVector<int,8> *MaskPtr = &LoMask;
4451 unsigned MaskIdx = 0;
4454 for (unsigned i = 0; i != 4; ++i) {
4461 int Idx = PermMask[i];
4463 Locs[i] = std::make_pair(-1, -1);
4464 } else if (Idx < 4) {
4465 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4466 (*MaskPtr)[LoIdx] = Idx;
4469 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4470 (*MaskPtr)[HiIdx] = Idx;
4475 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4476 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4477 SmallVector<int, 8> MaskOps;
4478 for (unsigned i = 0; i != 4; ++i) {
4479 if (Locs[i].first == -1) {
4480 MaskOps.push_back(-1);
4482 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4483 MaskOps.push_back(Idx);
4486 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4490 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4491 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4492 SDValue V1 = Op.getOperand(0);
4493 SDValue V2 = Op.getOperand(1);
4494 EVT VT = Op.getValueType();
4495 DebugLoc dl = Op.getDebugLoc();
4496 unsigned NumElems = VT.getVectorNumElements();
4497 bool isMMX = VT.getSizeInBits() == 64;
4498 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4499 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4500 bool V1IsSplat = false;
4501 bool V2IsSplat = false;
4503 if (isZeroShuffle(SVOp))
4504 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4506 // Promote splats to v4f32.
4507 if (SVOp->isSplat()) {
4508 if (isMMX || NumElems < 4)
4510 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4513 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4515 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4516 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4517 if (NewOp.getNode())
4518 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4519 LowerVECTOR_SHUFFLE(NewOp, DAG));
4520 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4521 // FIXME: Figure out a cleaner way to do this.
4522 // Try to make use of movq to zero out the top part.
4523 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4524 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4525 if (NewOp.getNode()) {
4526 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4527 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4528 DAG, Subtarget, dl);
4530 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4531 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4532 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4533 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4534 DAG, Subtarget, dl);
4538 if (X86::isPSHUFDMask(SVOp))
4541 // Check if this can be converted into a logical shift.
4542 bool isLeft = false;
4545 bool isShift = getSubtarget()->hasSSE2() &&
4546 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4547 if (isShift && ShVal.hasOneUse()) {
4548 // If the shifted value has multiple uses, it may be cheaper to use
4549 // v_set0 + movlhps or movhlps, etc.
4550 EVT EltVT = VT.getVectorElementType();
4551 ShAmt *= EltVT.getSizeInBits();
4552 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4555 if (X86::isMOVLMask(SVOp)) {
4558 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4559 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4564 // FIXME: fold these into legal mask.
4565 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4566 X86::isMOVSLDUPMask(SVOp) ||
4567 X86::isMOVHLPSMask(SVOp) ||
4568 X86::isMOVLHPSMask(SVOp) ||
4569 X86::isMOVLPMask(SVOp)))
4572 if (ShouldXformToMOVHLPS(SVOp) ||
4573 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4574 return CommuteVectorShuffle(SVOp, DAG);
4577 // No better options. Use a vshl / vsrl.
4578 EVT EltVT = VT.getVectorElementType();
4579 ShAmt *= EltVT.getSizeInBits();
4580 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4583 bool Commuted = false;
4584 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4585 // 1,1,1,1 -> v8i16 though.
4586 V1IsSplat = isSplatVector(V1.getNode());
4587 V2IsSplat = isSplatVector(V2.getNode());
4589 // Canonicalize the splat or undef, if present, to be on the RHS.
4590 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4591 Op = CommuteVectorShuffle(SVOp, DAG);
4592 SVOp = cast<ShuffleVectorSDNode>(Op);
4593 V1 = SVOp->getOperand(0);
4594 V2 = SVOp->getOperand(1);
4595 std::swap(V1IsSplat, V2IsSplat);
4596 std::swap(V1IsUndef, V2IsUndef);
4600 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4601 // Shuffling low element of v1 into undef, just return v1.
4604 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4605 // the instruction selector will not match, so get a canonical MOVL with
4606 // swapped operands to undo the commute.
4607 return getMOVL(DAG, dl, VT, V2, V1);
4610 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4611 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4612 X86::isUNPCKLMask(SVOp) ||
4613 X86::isUNPCKHMask(SVOp))
4617 // Normalize mask so all entries that point to V2 points to its first
4618 // element then try to match unpck{h|l} again. If match, return a
4619 // new vector_shuffle with the corrected mask.
4620 SDValue NewMask = NormalizeMask(SVOp, DAG);
4621 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4622 if (NSVOp != SVOp) {
4623 if (X86::isUNPCKLMask(NSVOp, true)) {
4625 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4632 // Commute is back and try unpck* again.
4633 // FIXME: this seems wrong.
4634 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4635 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4636 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4637 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4638 X86::isUNPCKLMask(NewSVOp) ||
4639 X86::isUNPCKHMask(NewSVOp))
4643 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4645 // Normalize the node to match x86 shuffle ops if needed
4646 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4647 return CommuteVectorShuffle(SVOp, DAG);
4649 // Check for legal shuffle and return?
4650 SmallVector<int, 16> PermMask;
4651 SVOp->getMask(PermMask);
4652 if (isShuffleMaskLegal(PermMask, VT))
4655 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4656 if (VT == MVT::v8i16) {
4657 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4658 if (NewOp.getNode())
4662 if (VT == MVT::v16i8) {
4663 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4664 if (NewOp.getNode())
4668 // Handle all 4 wide cases with a number of shuffles except for MMX.
4669 if (NumElems == 4 && !isMMX)
4670 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4676 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4677 SelectionDAG &DAG) {
4678 EVT VT = Op.getValueType();
4679 DebugLoc dl = Op.getDebugLoc();
4680 if (VT.getSizeInBits() == 8) {
4681 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4682 Op.getOperand(0), Op.getOperand(1));
4683 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4684 DAG.getValueType(VT));
4685 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4686 } else if (VT.getSizeInBits() == 16) {
4687 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4688 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4690 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4691 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4692 DAG.getNode(ISD::BIT_CONVERT, dl,
4696 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4697 Op.getOperand(0), Op.getOperand(1));
4698 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4699 DAG.getValueType(VT));
4700 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4701 } else if (VT == MVT::f32) {
4702 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4703 // the result back to FR32 register. It's only worth matching if the
4704 // result has a single use which is a store or a bitcast to i32. And in
4705 // the case of a store, it's not worth it if the index is a constant 0,
4706 // because a MOVSSmr can be used instead, which is smaller and faster.
4707 if (!Op.hasOneUse())
4709 SDNode *User = *Op.getNode()->use_begin();
4710 if ((User->getOpcode() != ISD::STORE ||
4711 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4712 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4713 (User->getOpcode() != ISD::BIT_CONVERT ||
4714 User->getValueType(0) != MVT::i32))
4716 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4717 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4720 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4721 } else if (VT == MVT::i32) {
4722 // ExtractPS works with constant index.
4723 if (isa<ConstantSDNode>(Op.getOperand(1)))
4731 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4732 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4735 if (Subtarget->hasSSE41()) {
4736 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4741 EVT VT = Op.getValueType();
4742 DebugLoc dl = Op.getDebugLoc();
4743 // TODO: handle v16i8.
4744 if (VT.getSizeInBits() == 16) {
4745 SDValue Vec = Op.getOperand(0);
4746 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4748 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4749 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4750 DAG.getNode(ISD::BIT_CONVERT, dl,
4753 // Transform it so it match pextrw which produces a 32-bit result.
4754 EVT EltVT = MVT::i32;
4755 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4756 Op.getOperand(0), Op.getOperand(1));
4757 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4758 DAG.getValueType(VT));
4759 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4760 } else if (VT.getSizeInBits() == 32) {
4761 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4765 // SHUFPS the element to the lowest double word, then movss.
4766 int Mask[4] = { Idx, -1, -1, -1 };
4767 EVT VVT = Op.getOperand(0).getValueType();
4768 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4769 DAG.getUNDEF(VVT), Mask);
4770 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4771 DAG.getIntPtrConstant(0));
4772 } else if (VT.getSizeInBits() == 64) {
4773 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4774 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4775 // to match extract_elt for f64.
4776 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4780 // UNPCKHPD the element to the lowest double word, then movsd.
4781 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4782 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4783 int Mask[2] = { 1, -1 };
4784 EVT VVT = Op.getOperand(0).getValueType();
4785 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4786 DAG.getUNDEF(VVT), Mask);
4787 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4788 DAG.getIntPtrConstant(0));
4795 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4796 EVT VT = Op.getValueType();
4797 EVT EltVT = VT.getVectorElementType();
4798 DebugLoc dl = Op.getDebugLoc();
4800 SDValue N0 = Op.getOperand(0);
4801 SDValue N1 = Op.getOperand(1);
4802 SDValue N2 = Op.getOperand(2);
4804 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4805 isa<ConstantSDNode>(N2)) {
4806 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4808 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4810 if (N1.getValueType() != MVT::i32)
4811 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4812 if (N2.getValueType() != MVT::i32)
4813 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4814 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4815 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4816 // Bits [7:6] of the constant are the source select. This will always be
4817 // zero here. The DAG Combiner may combine an extract_elt index into these
4818 // bits. For example (insert (extract, 3), 2) could be matched by putting
4819 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4820 // Bits [5:4] of the constant are the destination select. This is the
4821 // value of the incoming immediate.
4822 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4823 // combine either bitwise AND or insert of float 0.0 to set these bits.
4824 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4825 // Create this as a scalar to vector..
4826 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4827 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4828 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4829 // PINSR* works with constant index.
4836 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4837 EVT VT = Op.getValueType();
4838 EVT EltVT = VT.getVectorElementType();
4840 if (Subtarget->hasSSE41())
4841 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4843 if (EltVT == MVT::i8)
4846 DebugLoc dl = Op.getDebugLoc();
4847 SDValue N0 = Op.getOperand(0);
4848 SDValue N1 = Op.getOperand(1);
4849 SDValue N2 = Op.getOperand(2);
4851 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4852 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4853 // as its second argument.
4854 if (N1.getValueType() != MVT::i32)
4855 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4856 if (N2.getValueType() != MVT::i32)
4857 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4858 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4864 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4865 DebugLoc dl = Op.getDebugLoc();
4866 if (Op.getValueType() == MVT::v2f32)
4867 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4868 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4869 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4870 Op.getOperand(0))));
4872 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4873 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4875 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4876 EVT VT = MVT::v2i32;
4877 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4884 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4885 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4888 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4889 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4890 // one of the above mentioned nodes. It has to be wrapped because otherwise
4891 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4892 // be used to form addressing mode. These wrapped nodes will be selected
4895 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4896 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4898 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4900 unsigned char OpFlag = 0;
4901 unsigned WrapperKind = X86ISD::Wrapper;
4902 CodeModel::Model M = getTargetMachine().getCodeModel();
4904 if (Subtarget->isPICStyleRIPRel() &&
4905 (M == CodeModel::Small || M == CodeModel::Kernel))
4906 WrapperKind = X86ISD::WrapperRIP;
4907 else if (Subtarget->isPICStyleGOT())
4908 OpFlag = X86II::MO_GOTOFF;
4909 else if (Subtarget->isPICStyleStubPIC())
4910 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4912 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4914 CP->getOffset(), OpFlag);
4915 DebugLoc DL = CP->getDebugLoc();
4916 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4917 // With PIC, the address is actually $g + Offset.
4919 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4920 DAG.getNode(X86ISD::GlobalBaseReg,
4921 DebugLoc::getUnknownLoc(), getPointerTy()),
4928 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4929 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4931 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4933 unsigned char OpFlag = 0;
4934 unsigned WrapperKind = X86ISD::Wrapper;
4935 CodeModel::Model M = getTargetMachine().getCodeModel();
4937 if (Subtarget->isPICStyleRIPRel() &&
4938 (M == CodeModel::Small || M == CodeModel::Kernel))
4939 WrapperKind = X86ISD::WrapperRIP;
4940 else if (Subtarget->isPICStyleGOT())
4941 OpFlag = X86II::MO_GOTOFF;
4942 else if (Subtarget->isPICStyleStubPIC())
4943 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4945 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4947 DebugLoc DL = JT->getDebugLoc();
4948 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4950 // With PIC, the address is actually $g + Offset.
4952 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4953 DAG.getNode(X86ISD::GlobalBaseReg,
4954 DebugLoc::getUnknownLoc(), getPointerTy()),
4962 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4963 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4965 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4967 unsigned char OpFlag = 0;
4968 unsigned WrapperKind = X86ISD::Wrapper;
4969 CodeModel::Model M = getTargetMachine().getCodeModel();
4971 if (Subtarget->isPICStyleRIPRel() &&
4972 (M == CodeModel::Small || M == CodeModel::Kernel))
4973 WrapperKind = X86ISD::WrapperRIP;
4974 else if (Subtarget->isPICStyleGOT())
4975 OpFlag = X86II::MO_GOTOFF;
4976 else if (Subtarget->isPICStyleStubPIC())
4977 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4979 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4981 DebugLoc DL = Op.getDebugLoc();
4982 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4985 // With PIC, the address is actually $g + Offset.
4986 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4987 !Subtarget->is64Bit()) {
4988 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4989 DAG.getNode(X86ISD::GlobalBaseReg,
4990 DebugLoc::getUnknownLoc(),
4999 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
5000 // Create the TargetBlockAddressAddress node.
5001 unsigned char OpFlags =
5002 Subtarget->ClassifyBlockAddressReference();
5003 CodeModel::Model M = getTargetMachine().getCodeModel();
5004 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5005 DebugLoc dl = Op.getDebugLoc();
5006 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5007 /*isTarget=*/true, OpFlags);
5009 if (Subtarget->isPICStyleRIPRel() &&
5010 (M == CodeModel::Small || M == CodeModel::Kernel))
5011 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5013 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5015 // With PIC, the address is actually $g + Offset.
5016 if (isGlobalRelativeToPICBase(OpFlags)) {
5017 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5018 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5026 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5028 SelectionDAG &DAG) const {
5029 // Create the TargetGlobalAddress node, folding in the constant
5030 // offset if it is legal.
5031 unsigned char OpFlags =
5032 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5033 CodeModel::Model M = getTargetMachine().getCodeModel();
5035 if (OpFlags == X86II::MO_NO_FLAG &&
5036 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5037 // A direct static reference to a global.
5038 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5041 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5044 if (Subtarget->isPICStyleRIPRel() &&
5045 (M == CodeModel::Small || M == CodeModel::Kernel))
5046 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5048 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5050 // With PIC, the address is actually $g + Offset.
5051 if (isGlobalRelativeToPICBase(OpFlags)) {
5052 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5053 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5057 // For globals that require a load from a stub to get the address, emit the
5059 if (isGlobalStubReference(OpFlags))
5060 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5061 PseudoSourceValue::getGOT(), 0);
5063 // If there was a non-zero offset that we didn't fold, create an explicit
5066 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5067 DAG.getConstant(Offset, getPointerTy()));
5073 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5074 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5075 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5076 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5080 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5081 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5082 unsigned char OperandFlags) {
5083 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5084 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5085 DebugLoc dl = GA->getDebugLoc();
5086 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5087 GA->getValueType(0),
5091 SDValue Ops[] = { Chain, TGA, *InFlag };
5092 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5094 SDValue Ops[] = { Chain, TGA };
5095 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5098 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5099 MFI->setHasCalls(true);
5101 SDValue Flag = Chain.getValue(1);
5102 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5105 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5107 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5110 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5111 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5112 DAG.getNode(X86ISD::GlobalBaseReg,
5113 DebugLoc::getUnknownLoc(),
5115 InFlag = Chain.getValue(1);
5117 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5120 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5122 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5124 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5125 X86::RAX, X86II::MO_TLSGD);
5128 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5129 // "local exec" model.
5130 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5131 const EVT PtrVT, TLSModel::Model model,
5133 DebugLoc dl = GA->getDebugLoc();
5134 // Get the Thread Pointer
5135 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5136 DebugLoc::getUnknownLoc(), PtrVT,
5137 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5140 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5143 unsigned char OperandFlags = 0;
5144 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5146 unsigned WrapperKind = X86ISD::Wrapper;
5147 if (model == TLSModel::LocalExec) {
5148 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5149 } else if (is64Bit) {
5150 assert(model == TLSModel::InitialExec);
5151 OperandFlags = X86II::MO_GOTTPOFF;
5152 WrapperKind = X86ISD::WrapperRIP;
5154 assert(model == TLSModel::InitialExec);
5155 OperandFlags = X86II::MO_INDNTPOFF;
5158 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5160 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5161 GA->getOffset(), OperandFlags);
5162 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5164 if (model == TLSModel::InitialExec)
5165 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5166 PseudoSourceValue::getGOT(), 0);
5168 // The address of the thread local variable is the add of the thread
5169 // pointer with the offset of the variable.
5170 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5174 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5175 // TODO: implement the "local dynamic" model
5176 // TODO: implement the "initial exec"model for pic executables
5177 assert(Subtarget->isTargetELF() &&
5178 "TLS not implemented for non-ELF targets");
5179 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5180 const GlobalValue *GV = GA->getGlobal();
5182 // If GV is an alias then use the aliasee for determining
5183 // thread-localness.
5184 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5185 GV = GA->resolveAliasedGlobal(false);
5187 TLSModel::Model model = getTLSModel(GV,
5188 getTargetMachine().getRelocationModel());
5191 case TLSModel::GeneralDynamic:
5192 case TLSModel::LocalDynamic: // not implemented
5193 if (Subtarget->is64Bit())
5194 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5195 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5197 case TLSModel::InitialExec:
5198 case TLSModel::LocalExec:
5199 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5200 Subtarget->is64Bit());
5203 llvm_unreachable("Unreachable");
5208 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5209 /// take a 2 x i32 value to shift plus a shift amount.
5210 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5211 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5212 EVT VT = Op.getValueType();
5213 unsigned VTBits = VT.getSizeInBits();
5214 DebugLoc dl = Op.getDebugLoc();
5215 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5216 SDValue ShOpLo = Op.getOperand(0);
5217 SDValue ShOpHi = Op.getOperand(1);
5218 SDValue ShAmt = Op.getOperand(2);
5219 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5220 DAG.getConstant(VTBits - 1, MVT::i8))
5221 : DAG.getConstant(0, VT);
5224 if (Op.getOpcode() == ISD::SHL_PARTS) {
5225 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5226 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5228 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5229 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5232 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5233 DAG.getConstant(VTBits, MVT::i8));
5234 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
5235 AndNode, DAG.getConstant(0, MVT::i8));
5238 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5239 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5240 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5242 if (Op.getOpcode() == ISD::SHL_PARTS) {
5243 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5244 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5246 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5247 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5250 SDValue Ops[2] = { Lo, Hi };
5251 return DAG.getMergeValues(Ops, 2, dl);
5254 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5255 EVT SrcVT = Op.getOperand(0).getValueType();
5257 if (SrcVT.isVector()) {
5258 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5264 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5265 "Unknown SINT_TO_FP to lower!");
5267 // These are really Legal; return the operand so the caller accepts it as
5269 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5271 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5272 Subtarget->is64Bit()) {
5276 DebugLoc dl = Op.getDebugLoc();
5277 unsigned Size = SrcVT.getSizeInBits()/8;
5278 MachineFunction &MF = DAG.getMachineFunction();
5279 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5280 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5281 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5283 PseudoSourceValue::getFixedStack(SSFI), 0);
5284 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5287 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5289 SelectionDAG &DAG) {
5291 DebugLoc dl = Op.getDebugLoc();
5293 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5295 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5297 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5298 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5299 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5300 Tys, Ops, array_lengthof(Ops));
5303 Chain = Result.getValue(1);
5304 SDValue InFlag = Result.getValue(2);
5306 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5307 // shouldn't be necessary except that RFP cannot be live across
5308 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5309 MachineFunction &MF = DAG.getMachineFunction();
5310 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5311 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5312 Tys = DAG.getVTList(MVT::Other);
5314 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5316 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5317 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5318 PseudoSourceValue::getFixedStack(SSFI), 0);
5324 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5325 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5326 // This algorithm is not obvious. Here it is in C code, more or less:
5328 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5329 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5330 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5332 // Copy ints to xmm registers.
5333 __m128i xh = _mm_cvtsi32_si128( hi );
5334 __m128i xl = _mm_cvtsi32_si128( lo );
5336 // Combine into low half of a single xmm register.
5337 __m128i x = _mm_unpacklo_epi32( xh, xl );
5341 // Merge in appropriate exponents to give the integer bits the right
5343 x = _mm_unpacklo_epi32( x, exp );
5345 // Subtract away the biases to deal with the IEEE-754 double precision
5347 d = _mm_sub_pd( (__m128d) x, bias );
5349 // All conversions up to here are exact. The correctly rounded result is
5350 // calculated using the current rounding mode using the following
5352 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5353 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5354 // store doesn't really need to be here (except
5355 // maybe to zero the other double)
5360 DebugLoc dl = Op.getDebugLoc();
5361 LLVMContext *Context = DAG.getContext();
5363 // Build some magic constants.
5364 std::vector<Constant*> CV0;
5365 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5366 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5367 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5368 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5369 Constant *C0 = ConstantVector::get(CV0);
5370 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5372 std::vector<Constant*> CV1;
5374 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5376 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5377 Constant *C1 = ConstantVector::get(CV1);
5378 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5380 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5381 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5383 DAG.getIntPtrConstant(1)));
5384 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5385 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5387 DAG.getIntPtrConstant(0)));
5388 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5389 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5390 PseudoSourceValue::getConstantPool(), 0,
5392 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5393 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5394 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5395 PseudoSourceValue::getConstantPool(), 0,
5397 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5399 // Add the halves; easiest way is to swap them into another reg first.
5400 int ShufMask[2] = { 1, -1 };
5401 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5402 DAG.getUNDEF(MVT::v2f64), ShufMask);
5403 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5404 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5405 DAG.getIntPtrConstant(0));
5408 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5409 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5410 DebugLoc dl = Op.getDebugLoc();
5411 // FP constant to bias correct the final result.
5412 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5415 // Load the 32-bit value into an XMM register.
5416 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5417 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5419 DAG.getIntPtrConstant(0)));
5421 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5422 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5423 DAG.getIntPtrConstant(0));
5425 // Or the load with the bias.
5426 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5427 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5428 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5430 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5431 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5432 MVT::v2f64, Bias)));
5433 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5434 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5435 DAG.getIntPtrConstant(0));
5437 // Subtract the bias.
5438 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5440 // Handle final rounding.
5441 EVT DestVT = Op.getValueType();
5443 if (DestVT.bitsLT(MVT::f64)) {
5444 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5445 DAG.getIntPtrConstant(0));
5446 } else if (DestVT.bitsGT(MVT::f64)) {
5447 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5450 // Handle final rounding.
5454 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5455 SDValue N0 = Op.getOperand(0);
5456 DebugLoc dl = Op.getDebugLoc();
5458 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5459 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5460 // the optimization here.
5461 if (DAG.SignBitIsZero(N0))
5462 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5464 EVT SrcVT = N0.getValueType();
5465 if (SrcVT == MVT::i64) {
5466 // We only handle SSE2 f64 target here; caller can expand the rest.
5467 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5470 return LowerUINT_TO_FP_i64(Op, DAG);
5471 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5472 return LowerUINT_TO_FP_i32(Op, DAG);
5475 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5477 // Make a 64-bit buffer, and use it to build an FILD.
5478 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5479 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5480 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5481 getPointerTy(), StackSlot, WordOff);
5482 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5483 StackSlot, NULL, 0);
5484 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5485 OffsetSlot, NULL, 0);
5486 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5489 std::pair<SDValue,SDValue> X86TargetLowering::
5490 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5491 DebugLoc dl = Op.getDebugLoc();
5493 EVT DstTy = Op.getValueType();
5496 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5500 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5501 DstTy.getSimpleVT() >= MVT::i16 &&
5502 "Unknown FP_TO_SINT to lower!");
5504 // These are really Legal.
5505 if (DstTy == MVT::i32 &&
5506 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5507 return std::make_pair(SDValue(), SDValue());
5508 if (Subtarget->is64Bit() &&
5509 DstTy == MVT::i64 &&
5510 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5511 return std::make_pair(SDValue(), SDValue());
5513 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5515 MachineFunction &MF = DAG.getMachineFunction();
5516 unsigned MemSize = DstTy.getSizeInBits()/8;
5517 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5518 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5521 switch (DstTy.getSimpleVT().SimpleTy) {
5522 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5523 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5524 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5525 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5528 SDValue Chain = DAG.getEntryNode();
5529 SDValue Value = Op.getOperand(0);
5530 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5531 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5532 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5533 PseudoSourceValue::getFixedStack(SSFI), 0);
5534 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5536 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5538 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5539 Chain = Value.getValue(1);
5540 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5541 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5544 // Build the FP_TO_INT*_IN_MEM
5545 SDValue Ops[] = { Chain, Value, StackSlot };
5546 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5548 return std::make_pair(FIST, StackSlot);
5551 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5552 if (Op.getValueType().isVector()) {
5553 if (Op.getValueType() == MVT::v2i32 &&
5554 Op.getOperand(0).getValueType() == MVT::v2f64) {
5560 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5561 SDValue FIST = Vals.first, StackSlot = Vals.second;
5562 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5563 if (FIST.getNode() == 0) return Op;
5566 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5567 FIST, StackSlot, NULL, 0);
5570 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5571 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5572 SDValue FIST = Vals.first, StackSlot = Vals.second;
5573 assert(FIST.getNode() && "Unexpected failure");
5576 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5577 FIST, StackSlot, NULL, 0);
5580 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5581 LLVMContext *Context = DAG.getContext();
5582 DebugLoc dl = Op.getDebugLoc();
5583 EVT VT = Op.getValueType();
5586 EltVT = VT.getVectorElementType();
5587 std::vector<Constant*> CV;
5588 if (EltVT == MVT::f64) {
5589 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5593 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5599 Constant *C = ConstantVector::get(CV);
5600 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5601 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5602 PseudoSourceValue::getConstantPool(), 0,
5604 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5607 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5608 LLVMContext *Context = DAG.getContext();
5609 DebugLoc dl = Op.getDebugLoc();
5610 EVT VT = Op.getValueType();
5613 EltVT = VT.getVectorElementType();
5614 std::vector<Constant*> CV;
5615 if (EltVT == MVT::f64) {
5616 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5620 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5626 Constant *C = ConstantVector::get(CV);
5627 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5628 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5629 PseudoSourceValue::getConstantPool(), 0,
5631 if (VT.isVector()) {
5632 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5633 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5634 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5636 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5638 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5642 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5643 LLVMContext *Context = DAG.getContext();
5644 SDValue Op0 = Op.getOperand(0);
5645 SDValue Op1 = Op.getOperand(1);
5646 DebugLoc dl = Op.getDebugLoc();
5647 EVT VT = Op.getValueType();
5648 EVT SrcVT = Op1.getValueType();
5650 // If second operand is smaller, extend it first.
5651 if (SrcVT.bitsLT(VT)) {
5652 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5655 // And if it is bigger, shrink it first.
5656 if (SrcVT.bitsGT(VT)) {
5657 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5661 // At this point the operands and the result should have the same
5662 // type, and that won't be f80 since that is not custom lowered.
5664 // First get the sign bit of second operand.
5665 std::vector<Constant*> CV;
5666 if (SrcVT == MVT::f64) {
5667 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5668 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5670 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5671 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5672 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5673 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5675 Constant *C = ConstantVector::get(CV);
5676 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5677 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5678 PseudoSourceValue::getConstantPool(), 0,
5680 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5682 // Shift sign bit right or left if the two operands have different types.
5683 if (SrcVT.bitsGT(VT)) {
5684 // Op0 is MVT::f32, Op1 is MVT::f64.
5685 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5686 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5687 DAG.getConstant(32, MVT::i32));
5688 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5689 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5690 DAG.getIntPtrConstant(0));
5693 // Clear first operand sign bit.
5695 if (VT == MVT::f64) {
5696 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5697 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5699 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5700 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5701 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5702 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5704 C = ConstantVector::get(CV);
5705 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5706 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5707 PseudoSourceValue::getConstantPool(), 0,
5709 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5711 // Or the value with the sign bit.
5712 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5715 /// Emit nodes that will be selected as "test Op0,Op0", or something
5717 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5718 SelectionDAG &DAG) {
5719 DebugLoc dl = Op.getDebugLoc();
5721 // CF and OF aren't always set the way we want. Determine which
5722 // of these we need.
5723 bool NeedCF = false;
5724 bool NeedOF = false;
5726 case X86::COND_A: case X86::COND_AE:
5727 case X86::COND_B: case X86::COND_BE:
5730 case X86::COND_G: case X86::COND_GE:
5731 case X86::COND_L: case X86::COND_LE:
5732 case X86::COND_O: case X86::COND_NO:
5738 // See if we can use the EFLAGS value from the operand instead of
5739 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5740 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5741 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5742 unsigned Opcode = 0;
5743 unsigned NumOperands = 0;
5744 switch (Op.getNode()->getOpcode()) {
5746 // Due to an isel shortcoming, be conservative if this add is likely to
5747 // be selected as part of a load-modify-store instruction. When the root
5748 // node in a match is a store, isel doesn't know how to remap non-chain
5749 // non-flag uses of other nodes in the match, such as the ADD in this
5750 // case. This leads to the ADD being left around and reselected, with
5751 // the result being two adds in the output.
5752 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5753 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5754 if (UI->getOpcode() == ISD::STORE)
5756 if (ConstantSDNode *C =
5757 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5758 // An add of one will be selected as an INC.
5759 if (C->getAPIntValue() == 1) {
5760 Opcode = X86ISD::INC;
5764 // An add of negative one (subtract of one) will be selected as a DEC.
5765 if (C->getAPIntValue().isAllOnesValue()) {
5766 Opcode = X86ISD::DEC;
5771 // Otherwise use a regular EFLAGS-setting add.
5772 Opcode = X86ISD::ADD;
5776 // If the primary and result isn't used, don't bother using X86ISD::AND,
5777 // because a TEST instruction will be better.
5778 bool NonFlagUse = false;
5779 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5780 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5782 unsigned UOpNo = UI.getOperandNo();
5783 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5784 // Look pass truncate.
5785 UOpNo = User->use_begin().getOperandNo();
5786 User = *User->use_begin();
5788 if (User->getOpcode() != ISD::BRCOND &&
5789 User->getOpcode() != ISD::SETCC &&
5790 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5802 // Due to the ISEL shortcoming noted above, be conservative if this op is
5803 // likely to be selected as part of a load-modify-store instruction.
5804 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5805 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5806 if (UI->getOpcode() == ISD::STORE)
5808 // Otherwise use a regular EFLAGS-setting instruction.
5809 switch (Op.getNode()->getOpcode()) {
5810 case ISD::SUB: Opcode = X86ISD::SUB; break;
5811 case ISD::OR: Opcode = X86ISD::OR; break;
5812 case ISD::XOR: Opcode = X86ISD::XOR; break;
5813 case ISD::AND: Opcode = X86ISD::AND; break;
5814 default: llvm_unreachable("unexpected operator!");
5825 return SDValue(Op.getNode(), 1);
5831 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5832 SmallVector<SDValue, 4> Ops;
5833 for (unsigned i = 0; i != NumOperands; ++i)
5834 Ops.push_back(Op.getOperand(i));
5835 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5836 DAG.ReplaceAllUsesWith(Op, New);
5837 return SDValue(New.getNode(), 1);
5841 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5842 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5843 DAG.getConstant(0, Op.getValueType()));
5846 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5848 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5849 SelectionDAG &DAG) {
5850 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5851 if (C->getAPIntValue() == 0)
5852 return EmitTest(Op0, X86CC, DAG);
5854 DebugLoc dl = Op0.getDebugLoc();
5855 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5858 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5859 /// if it's possible.
5860 static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
5861 DebugLoc dl, SelectionDAG &DAG) {
5863 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5864 if (ConstantSDNode *Op010C =
5865 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5866 if (Op010C->getZExtValue() == 1) {
5867 LHS = Op0.getOperand(0);
5868 RHS = Op0.getOperand(1).getOperand(1);
5870 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5871 if (ConstantSDNode *Op000C =
5872 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5873 if (Op000C->getZExtValue() == 1) {
5874 LHS = Op0.getOperand(1);
5875 RHS = Op0.getOperand(0).getOperand(1);
5877 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5878 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5879 SDValue AndLHS = Op0.getOperand(0);
5880 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5881 LHS = AndLHS.getOperand(0);
5882 RHS = AndLHS.getOperand(1);
5886 if (LHS.getNode()) {
5887 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5888 // instruction. Since the shift amount is in-range-or-undefined, we know
5889 // that doing a bittest on the i16 value is ok. We extend to i32 because
5890 // the encoding for the i16 version is larger than the i32 version.
5891 if (LHS.getValueType() == MVT::i8)
5892 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5894 // If the operand types disagree, extend the shift amount to match. Since
5895 // BT ignores high bits (like shifts) we can use anyextend.
5896 if (LHS.getValueType() != RHS.getValueType())
5897 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5899 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5900 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5901 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5902 DAG.getConstant(Cond, MVT::i8), BT);
5908 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5909 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5910 SDValue Op0 = Op.getOperand(0);
5911 SDValue Op1 = Op.getOperand(1);
5912 DebugLoc dl = Op.getDebugLoc();
5913 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5915 // Optimize to BT if possible.
5916 // Lower (X & (1 << N)) == 0 to BT(X, N).
5917 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5918 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5919 if (Op0.getOpcode() == ISD::AND &&
5921 Op1.getOpcode() == ISD::Constant &&
5922 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5923 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5924 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5925 if (NewSetCC.getNode())
5929 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5930 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5931 if (X86CC == X86::COND_INVALID)
5934 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5936 // Use sbb x, x to materialize carry bit into a GPR.
5937 if (X86CC == X86::COND_B)
5938 return DAG.getNode(ISD::AND, dl, MVT::i8,
5939 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5940 DAG.getConstant(X86CC, MVT::i8), Cond),
5941 DAG.getConstant(1, MVT::i8));
5943 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5944 DAG.getConstant(X86CC, MVT::i8), Cond);
5947 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5949 SDValue Op0 = Op.getOperand(0);
5950 SDValue Op1 = Op.getOperand(1);
5951 SDValue CC = Op.getOperand(2);
5952 EVT VT = Op.getValueType();
5953 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5954 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5955 DebugLoc dl = Op.getDebugLoc();
5959 EVT VT0 = Op0.getValueType();
5960 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5961 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5964 switch (SetCCOpcode) {
5967 case ISD::SETEQ: SSECC = 0; break;
5969 case ISD::SETGT: Swap = true; // Fallthrough
5971 case ISD::SETOLT: SSECC = 1; break;
5973 case ISD::SETGE: Swap = true; // Fallthrough
5975 case ISD::SETOLE: SSECC = 2; break;
5976 case ISD::SETUO: SSECC = 3; break;
5978 case ISD::SETNE: SSECC = 4; break;
5979 case ISD::SETULE: Swap = true;
5980 case ISD::SETUGE: SSECC = 5; break;
5981 case ISD::SETULT: Swap = true;
5982 case ISD::SETUGT: SSECC = 6; break;
5983 case ISD::SETO: SSECC = 7; break;
5986 std::swap(Op0, Op1);
5988 // In the two special cases we can't handle, emit two comparisons.
5990 if (SetCCOpcode == ISD::SETUEQ) {
5992 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5993 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5994 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5996 else if (SetCCOpcode == ISD::SETONE) {
5998 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5999 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6000 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6002 llvm_unreachable("Illegal FP comparison");
6004 // Handle all other FP comparisons here.
6005 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6008 // We are handling one of the integer comparisons here. Since SSE only has
6009 // GT and EQ comparisons for integer, swapping operands and multiple
6010 // operations may be required for some comparisons.
6011 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6012 bool Swap = false, Invert = false, FlipSigns = false;
6014 switch (VT.getSimpleVT().SimpleTy) {
6017 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6019 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6021 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6022 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6025 switch (SetCCOpcode) {
6027 case ISD::SETNE: Invert = true;
6028 case ISD::SETEQ: Opc = EQOpc; break;
6029 case ISD::SETLT: Swap = true;
6030 case ISD::SETGT: Opc = GTOpc; break;
6031 case ISD::SETGE: Swap = true;
6032 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6033 case ISD::SETULT: Swap = true;
6034 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6035 case ISD::SETUGE: Swap = true;
6036 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6039 std::swap(Op0, Op1);
6041 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6042 // bits of the inputs before performing those operations.
6044 EVT EltVT = VT.getVectorElementType();
6045 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6047 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6048 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6050 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6051 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6054 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6056 // If the logical-not of the result is required, perform that now.
6058 Result = DAG.getNOT(dl, Result, VT);
6063 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6064 static bool isX86LogicalCmp(SDValue Op) {
6065 unsigned Opc = Op.getNode()->getOpcode();
6066 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6068 if (Op.getResNo() == 1 &&
6069 (Opc == X86ISD::ADD ||
6070 Opc == X86ISD::SUB ||
6071 Opc == X86ISD::SMUL ||
6072 Opc == X86ISD::UMUL ||
6073 Opc == X86ISD::INC ||
6074 Opc == X86ISD::DEC ||
6075 Opc == X86ISD::OR ||
6076 Opc == X86ISD::XOR ||
6077 Opc == X86ISD::AND))
6083 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6084 bool addTest = true;
6085 SDValue Cond = Op.getOperand(0);
6086 DebugLoc dl = Op.getDebugLoc();
6089 if (Cond.getOpcode() == ISD::SETCC) {
6090 SDValue NewCond = LowerSETCC(Cond, DAG);
6091 if (NewCond.getNode())
6095 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6096 SDValue Op1 = Op.getOperand(1);
6097 SDValue Op2 = Op.getOperand(2);
6098 if (Cond.getOpcode() == X86ISD::SETCC &&
6099 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6100 SDValue Cmp = Cond.getOperand(1);
6101 if (Cmp.getOpcode() == X86ISD::CMP) {
6102 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6103 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6104 ConstantSDNode *RHSC =
6105 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6106 if (N1C && N1C->isAllOnesValue() &&
6107 N2C && N2C->isNullValue() &&
6108 RHSC && RHSC->isNullValue()) {
6109 SDValue CmpOp0 = Cmp.getOperand(0);
6110 Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
6111 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6112 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6113 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6118 // Look pass (and (setcc_carry (cmp ...)), 1).
6119 if (Cond.getOpcode() == ISD::AND &&
6120 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6121 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6122 if (C && C->getAPIntValue() == 1)
6123 Cond = Cond.getOperand(0);
6126 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6127 // setting operand in place of the X86ISD::SETCC.
6128 if (Cond.getOpcode() == X86ISD::SETCC ||
6129 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6130 CC = Cond.getOperand(0);
6132 SDValue Cmp = Cond.getOperand(1);
6133 unsigned Opc = Cmp.getOpcode();
6134 EVT VT = Op.getValueType();
6136 bool IllegalFPCMov = false;
6137 if (VT.isFloatingPoint() && !VT.isVector() &&
6138 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6139 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6141 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6142 Opc == X86ISD::BT) { // FIXME
6149 // Look pass the truncate.
6150 if (Cond.getOpcode() == ISD::TRUNCATE)
6151 Cond = Cond.getOperand(0);
6153 // We know the result of AND is compared against zero. Try to match
6155 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6156 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6157 if (NewSetCC.getNode()) {
6158 CC = NewSetCC.getOperand(0);
6159 Cond = NewSetCC.getOperand(1);
6166 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6167 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6170 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6171 // condition is true.
6172 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6173 SDValue Ops[] = { Op2, Op1, CC, Cond };
6174 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6177 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6178 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6179 // from the AND / OR.
6180 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6181 Opc = Op.getOpcode();
6182 if (Opc != ISD::OR && Opc != ISD::AND)
6184 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6185 Op.getOperand(0).hasOneUse() &&
6186 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6187 Op.getOperand(1).hasOneUse());
6190 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6191 // 1 and that the SETCC node has a single use.
6192 static bool isXor1OfSetCC(SDValue Op) {
6193 if (Op.getOpcode() != ISD::XOR)
6195 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6196 if (N1C && N1C->getAPIntValue() == 1) {
6197 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6198 Op.getOperand(0).hasOneUse();
6203 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6204 bool addTest = true;
6205 SDValue Chain = Op.getOperand(0);
6206 SDValue Cond = Op.getOperand(1);
6207 SDValue Dest = Op.getOperand(2);
6208 DebugLoc dl = Op.getDebugLoc();
6211 if (Cond.getOpcode() == ISD::SETCC) {
6212 SDValue NewCond = LowerSETCC(Cond, DAG);
6213 if (NewCond.getNode())
6217 // FIXME: LowerXALUO doesn't handle these!!
6218 else if (Cond.getOpcode() == X86ISD::ADD ||
6219 Cond.getOpcode() == X86ISD::SUB ||
6220 Cond.getOpcode() == X86ISD::SMUL ||
6221 Cond.getOpcode() == X86ISD::UMUL)
6222 Cond = LowerXALUO(Cond, DAG);
6225 // Look pass (and (setcc_carry (cmp ...)), 1).
6226 if (Cond.getOpcode() == ISD::AND &&
6227 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6228 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6229 if (C && C->getAPIntValue() == 1)
6230 Cond = Cond.getOperand(0);
6233 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6234 // setting operand in place of the X86ISD::SETCC.
6235 if (Cond.getOpcode() == X86ISD::SETCC ||
6236 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6237 CC = Cond.getOperand(0);
6239 SDValue Cmp = Cond.getOperand(1);
6240 unsigned Opc = Cmp.getOpcode();
6241 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6242 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6246 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6250 // These can only come from an arithmetic instruction with overflow,
6251 // e.g. SADDO, UADDO.
6252 Cond = Cond.getNode()->getOperand(1);
6259 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6260 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6261 if (CondOpc == ISD::OR) {
6262 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6263 // two branches instead of an explicit OR instruction with a
6265 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6266 isX86LogicalCmp(Cmp)) {
6267 CC = Cond.getOperand(0).getOperand(0);
6268 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6269 Chain, Dest, CC, Cmp);
6270 CC = Cond.getOperand(1).getOperand(0);
6274 } else { // ISD::AND
6275 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6276 // two branches instead of an explicit AND instruction with a
6277 // separate test. However, we only do this if this block doesn't
6278 // have a fall-through edge, because this requires an explicit
6279 // jmp when the condition is false.
6280 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6281 isX86LogicalCmp(Cmp) &&
6282 Op.getNode()->hasOneUse()) {
6283 X86::CondCode CCode =
6284 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6285 CCode = X86::GetOppositeBranchCondition(CCode);
6286 CC = DAG.getConstant(CCode, MVT::i8);
6287 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6288 // Look for an unconditional branch following this conditional branch.
6289 // We need this because we need to reverse the successors in order
6290 // to implement FCMP_OEQ.
6291 if (User.getOpcode() == ISD::BR) {
6292 SDValue FalseBB = User.getOperand(1);
6294 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6295 assert(NewBR == User);
6298 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6299 Chain, Dest, CC, Cmp);
6300 X86::CondCode CCode =
6301 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6302 CCode = X86::GetOppositeBranchCondition(CCode);
6303 CC = DAG.getConstant(CCode, MVT::i8);
6309 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6310 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6311 // It should be transformed during dag combiner except when the condition
6312 // is set by a arithmetics with overflow node.
6313 X86::CondCode CCode =
6314 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6315 CCode = X86::GetOppositeBranchCondition(CCode);
6316 CC = DAG.getConstant(CCode, MVT::i8);
6317 Cond = Cond.getOperand(0).getOperand(1);
6323 // Look pass the truncate.
6324 if (Cond.getOpcode() == ISD::TRUNCATE)
6325 Cond = Cond.getOperand(0);
6327 // We know the result of AND is compared against zero. Try to match
6329 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6330 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6331 if (NewSetCC.getNode()) {
6332 CC = NewSetCC.getOperand(0);
6333 Cond = NewSetCC.getOperand(1);
6340 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6341 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6343 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6344 Chain, Dest, CC, Cond);
6348 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6349 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6350 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6351 // that the guard pages used by the OS virtual memory manager are allocated in
6352 // correct sequence.
6354 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6355 SelectionDAG &DAG) {
6356 assert(Subtarget->isTargetCygMing() &&
6357 "This should be used only on Cygwin/Mingw targets");
6358 DebugLoc dl = Op.getDebugLoc();
6361 SDValue Chain = Op.getOperand(0);
6362 SDValue Size = Op.getOperand(1);
6363 // FIXME: Ensure alignment here
6367 EVT IntPtr = getPointerTy();
6368 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6370 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6372 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6373 Flag = Chain.getValue(1);
6375 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6376 SDValue Ops[] = { Chain,
6377 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6378 DAG.getRegister(X86::EAX, IntPtr),
6379 DAG.getRegister(X86StackPtr, SPTy),
6381 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6382 Flag = Chain.getValue(1);
6384 Chain = DAG.getCALLSEQ_END(Chain,
6385 DAG.getIntPtrConstant(0, true),
6386 DAG.getIntPtrConstant(0, true),
6389 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6391 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6392 return DAG.getMergeValues(Ops1, 2, dl);
6396 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6398 SDValue Dst, SDValue Src,
6399 SDValue Size, unsigned Align,
6401 uint64_t DstSVOff) {
6402 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6404 // If not DWORD aligned or size is more than the threshold, call the library.
6405 // The libc version is likely to be faster for these cases. It can use the
6406 // address value and run time information about the CPU.
6407 if ((Align & 3) != 0 ||
6409 ConstantSize->getZExtValue() >
6410 getSubtarget()->getMaxInlineSizeThreshold()) {
6411 SDValue InFlag(0, 0);
6413 // Check to see if there is a specialized entry-point for memory zeroing.
6414 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6416 if (const char *bzeroEntry = V &&
6417 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6418 EVT IntPtr = getPointerTy();
6419 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6420 TargetLowering::ArgListTy Args;
6421 TargetLowering::ArgListEntry Entry;
6423 Entry.Ty = IntPtrTy;
6424 Args.push_back(Entry);
6426 Args.push_back(Entry);
6427 std::pair<SDValue,SDValue> CallResult =
6428 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6429 false, false, false, false,
6430 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6431 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6432 DAG.GetOrdering(Chain.getNode()));
6433 return CallResult.second;
6436 // Otherwise have the target-independent code call memset.
6440 uint64_t SizeVal = ConstantSize->getZExtValue();
6441 SDValue InFlag(0, 0);
6444 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6445 unsigned BytesLeft = 0;
6446 bool TwoRepStos = false;
6449 uint64_t Val = ValC->getZExtValue() & 255;
6451 // If the value is a constant, then we can potentially use larger sets.
6452 switch (Align & 3) {
6453 case 2: // WORD aligned
6456 Val = (Val << 8) | Val;
6458 case 0: // DWORD aligned
6461 Val = (Val << 8) | Val;
6462 Val = (Val << 16) | Val;
6463 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6466 Val = (Val << 32) | Val;
6469 default: // Byte aligned
6472 Count = DAG.getIntPtrConstant(SizeVal);
6476 if (AVT.bitsGT(MVT::i8)) {
6477 unsigned UBytes = AVT.getSizeInBits() / 8;
6478 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6479 BytesLeft = SizeVal % UBytes;
6482 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6484 InFlag = Chain.getValue(1);
6487 Count = DAG.getIntPtrConstant(SizeVal);
6488 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6489 InFlag = Chain.getValue(1);
6492 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6495 InFlag = Chain.getValue(1);
6496 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6499 InFlag = Chain.getValue(1);
6501 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6502 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6503 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6506 InFlag = Chain.getValue(1);
6508 EVT CVT = Count.getValueType();
6509 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6510 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6511 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6514 InFlag = Chain.getValue(1);
6515 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6516 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6517 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6518 } else if (BytesLeft) {
6519 // Handle the last 1 - 7 bytes.
6520 unsigned Offset = SizeVal - BytesLeft;
6521 EVT AddrVT = Dst.getValueType();
6522 EVT SizeVT = Size.getValueType();
6524 Chain = DAG.getMemset(Chain, dl,
6525 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6526 DAG.getConstant(Offset, AddrVT)),
6528 DAG.getConstant(BytesLeft, SizeVT),
6529 Align, DstSV, DstSVOff + Offset);
6532 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6537 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6538 SDValue Chain, SDValue Dst, SDValue Src,
6539 SDValue Size, unsigned Align,
6541 const Value *DstSV, uint64_t DstSVOff,
6542 const Value *SrcSV, uint64_t SrcSVOff) {
6543 // This requires the copy size to be a constant, preferrably
6544 // within a subtarget-specific limit.
6545 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6548 uint64_t SizeVal = ConstantSize->getZExtValue();
6549 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6552 /// If not DWORD aligned, call the library.
6553 if ((Align & 3) != 0)
6558 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6561 unsigned UBytes = AVT.getSizeInBits() / 8;
6562 unsigned CountVal = SizeVal / UBytes;
6563 SDValue Count = DAG.getIntPtrConstant(CountVal);
6564 unsigned BytesLeft = SizeVal % UBytes;
6566 SDValue InFlag(0, 0);
6567 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6570 InFlag = Chain.getValue(1);
6571 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6574 InFlag = Chain.getValue(1);
6575 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6578 InFlag = Chain.getValue(1);
6580 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6581 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6582 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6583 array_lengthof(Ops));
6585 SmallVector<SDValue, 4> Results;
6586 Results.push_back(RepMovs);
6588 // Handle the last 1 - 7 bytes.
6589 unsigned Offset = SizeVal - BytesLeft;
6590 EVT DstVT = Dst.getValueType();
6591 EVT SrcVT = Src.getValueType();
6592 EVT SizeVT = Size.getValueType();
6593 Results.push_back(DAG.getMemcpy(Chain, dl,
6594 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6595 DAG.getConstant(Offset, DstVT)),
6596 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6597 DAG.getConstant(Offset, SrcVT)),
6598 DAG.getConstant(BytesLeft, SizeVT),
6599 Align, AlwaysInline,
6600 DstSV, DstSVOff + Offset,
6601 SrcSV, SrcSVOff + Offset));
6604 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6605 &Results[0], Results.size());
6608 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6609 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6610 DebugLoc dl = Op.getDebugLoc();
6612 if (!Subtarget->is64Bit()) {
6613 // vastart just stores the address of the VarArgsFrameIndex slot into the
6614 // memory location argument.
6615 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6616 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6620 // gp_offset (0 - 6 * 8)
6621 // fp_offset (48 - 48 + 8 * 16)
6622 // overflow_arg_area (point to parameters coming in memory).
6624 SmallVector<SDValue, 8> MemOps;
6625 SDValue FIN = Op.getOperand(1);
6627 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6628 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6630 MemOps.push_back(Store);
6633 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6634 FIN, DAG.getIntPtrConstant(4));
6635 Store = DAG.getStore(Op.getOperand(0), dl,
6636 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6638 MemOps.push_back(Store);
6640 // Store ptr to overflow_arg_area
6641 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6642 FIN, DAG.getIntPtrConstant(4));
6643 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6644 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6645 MemOps.push_back(Store);
6647 // Store ptr to reg_save_area.
6648 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6649 FIN, DAG.getIntPtrConstant(8));
6650 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6651 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6652 MemOps.push_back(Store);
6653 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6654 &MemOps[0], MemOps.size());
6657 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6658 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6659 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6660 SDValue Chain = Op.getOperand(0);
6661 SDValue SrcPtr = Op.getOperand(1);
6662 SDValue SrcSV = Op.getOperand(2);
6664 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6668 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6669 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6670 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6671 SDValue Chain = Op.getOperand(0);
6672 SDValue DstPtr = Op.getOperand(1);
6673 SDValue SrcPtr = Op.getOperand(2);
6674 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6675 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6676 DebugLoc dl = Op.getDebugLoc();
6678 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6679 DAG.getIntPtrConstant(24), 8, false,
6680 DstSV, 0, SrcSV, 0);
6684 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6685 DebugLoc dl = Op.getDebugLoc();
6686 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6688 default: return SDValue(); // Don't custom lower most intrinsics.
6689 // Comparison intrinsics.
6690 case Intrinsic::x86_sse_comieq_ss:
6691 case Intrinsic::x86_sse_comilt_ss:
6692 case Intrinsic::x86_sse_comile_ss:
6693 case Intrinsic::x86_sse_comigt_ss:
6694 case Intrinsic::x86_sse_comige_ss:
6695 case Intrinsic::x86_sse_comineq_ss:
6696 case Intrinsic::x86_sse_ucomieq_ss:
6697 case Intrinsic::x86_sse_ucomilt_ss:
6698 case Intrinsic::x86_sse_ucomile_ss:
6699 case Intrinsic::x86_sse_ucomigt_ss:
6700 case Intrinsic::x86_sse_ucomige_ss:
6701 case Intrinsic::x86_sse_ucomineq_ss:
6702 case Intrinsic::x86_sse2_comieq_sd:
6703 case Intrinsic::x86_sse2_comilt_sd:
6704 case Intrinsic::x86_sse2_comile_sd:
6705 case Intrinsic::x86_sse2_comigt_sd:
6706 case Intrinsic::x86_sse2_comige_sd:
6707 case Intrinsic::x86_sse2_comineq_sd:
6708 case Intrinsic::x86_sse2_ucomieq_sd:
6709 case Intrinsic::x86_sse2_ucomilt_sd:
6710 case Intrinsic::x86_sse2_ucomile_sd:
6711 case Intrinsic::x86_sse2_ucomigt_sd:
6712 case Intrinsic::x86_sse2_ucomige_sd:
6713 case Intrinsic::x86_sse2_ucomineq_sd: {
6715 ISD::CondCode CC = ISD::SETCC_INVALID;
6718 case Intrinsic::x86_sse_comieq_ss:
6719 case Intrinsic::x86_sse2_comieq_sd:
6723 case Intrinsic::x86_sse_comilt_ss:
6724 case Intrinsic::x86_sse2_comilt_sd:
6728 case Intrinsic::x86_sse_comile_ss:
6729 case Intrinsic::x86_sse2_comile_sd:
6733 case Intrinsic::x86_sse_comigt_ss:
6734 case Intrinsic::x86_sse2_comigt_sd:
6738 case Intrinsic::x86_sse_comige_ss:
6739 case Intrinsic::x86_sse2_comige_sd:
6743 case Intrinsic::x86_sse_comineq_ss:
6744 case Intrinsic::x86_sse2_comineq_sd:
6748 case Intrinsic::x86_sse_ucomieq_ss:
6749 case Intrinsic::x86_sse2_ucomieq_sd:
6750 Opc = X86ISD::UCOMI;
6753 case Intrinsic::x86_sse_ucomilt_ss:
6754 case Intrinsic::x86_sse2_ucomilt_sd:
6755 Opc = X86ISD::UCOMI;
6758 case Intrinsic::x86_sse_ucomile_ss:
6759 case Intrinsic::x86_sse2_ucomile_sd:
6760 Opc = X86ISD::UCOMI;
6763 case Intrinsic::x86_sse_ucomigt_ss:
6764 case Intrinsic::x86_sse2_ucomigt_sd:
6765 Opc = X86ISD::UCOMI;
6768 case Intrinsic::x86_sse_ucomige_ss:
6769 case Intrinsic::x86_sse2_ucomige_sd:
6770 Opc = X86ISD::UCOMI;
6773 case Intrinsic::x86_sse_ucomineq_ss:
6774 case Intrinsic::x86_sse2_ucomineq_sd:
6775 Opc = X86ISD::UCOMI;
6780 SDValue LHS = Op.getOperand(1);
6781 SDValue RHS = Op.getOperand(2);
6782 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6783 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6784 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6785 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6786 DAG.getConstant(X86CC, MVT::i8), Cond);
6787 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6789 // ptest intrinsics. The intrinsic these come from are designed to return
6790 // an integer value, not just an instruction so lower it to the ptest
6791 // pattern and a setcc for the result.
6792 case Intrinsic::x86_sse41_ptestz:
6793 case Intrinsic::x86_sse41_ptestc:
6794 case Intrinsic::x86_sse41_ptestnzc:{
6797 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6798 case Intrinsic::x86_sse41_ptestz:
6800 X86CC = X86::COND_E;
6802 case Intrinsic::x86_sse41_ptestc:
6804 X86CC = X86::COND_B;
6806 case Intrinsic::x86_sse41_ptestnzc:
6808 X86CC = X86::COND_A;
6812 SDValue LHS = Op.getOperand(1);
6813 SDValue RHS = Op.getOperand(2);
6814 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6815 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6816 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6817 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6820 // Fix vector shift instructions where the last operand is a non-immediate
6822 case Intrinsic::x86_sse2_pslli_w:
6823 case Intrinsic::x86_sse2_pslli_d:
6824 case Intrinsic::x86_sse2_pslli_q:
6825 case Intrinsic::x86_sse2_psrli_w:
6826 case Intrinsic::x86_sse2_psrli_d:
6827 case Intrinsic::x86_sse2_psrli_q:
6828 case Intrinsic::x86_sse2_psrai_w:
6829 case Intrinsic::x86_sse2_psrai_d:
6830 case Intrinsic::x86_mmx_pslli_w:
6831 case Intrinsic::x86_mmx_pslli_d:
6832 case Intrinsic::x86_mmx_pslli_q:
6833 case Intrinsic::x86_mmx_psrli_w:
6834 case Intrinsic::x86_mmx_psrli_d:
6835 case Intrinsic::x86_mmx_psrli_q:
6836 case Intrinsic::x86_mmx_psrai_w:
6837 case Intrinsic::x86_mmx_psrai_d: {
6838 SDValue ShAmt = Op.getOperand(2);
6839 if (isa<ConstantSDNode>(ShAmt))
6842 unsigned NewIntNo = 0;
6843 EVT ShAmtVT = MVT::v4i32;
6845 case Intrinsic::x86_sse2_pslli_w:
6846 NewIntNo = Intrinsic::x86_sse2_psll_w;
6848 case Intrinsic::x86_sse2_pslli_d:
6849 NewIntNo = Intrinsic::x86_sse2_psll_d;
6851 case Intrinsic::x86_sse2_pslli_q:
6852 NewIntNo = Intrinsic::x86_sse2_psll_q;
6854 case Intrinsic::x86_sse2_psrli_w:
6855 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6857 case Intrinsic::x86_sse2_psrli_d:
6858 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6860 case Intrinsic::x86_sse2_psrli_q:
6861 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6863 case Intrinsic::x86_sse2_psrai_w:
6864 NewIntNo = Intrinsic::x86_sse2_psra_w;
6866 case Intrinsic::x86_sse2_psrai_d:
6867 NewIntNo = Intrinsic::x86_sse2_psra_d;
6870 ShAmtVT = MVT::v2i32;
6872 case Intrinsic::x86_mmx_pslli_w:
6873 NewIntNo = Intrinsic::x86_mmx_psll_w;
6875 case Intrinsic::x86_mmx_pslli_d:
6876 NewIntNo = Intrinsic::x86_mmx_psll_d;
6878 case Intrinsic::x86_mmx_pslli_q:
6879 NewIntNo = Intrinsic::x86_mmx_psll_q;
6881 case Intrinsic::x86_mmx_psrli_w:
6882 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6884 case Intrinsic::x86_mmx_psrli_d:
6885 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6887 case Intrinsic::x86_mmx_psrli_q:
6888 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6890 case Intrinsic::x86_mmx_psrai_w:
6891 NewIntNo = Intrinsic::x86_mmx_psra_w;
6893 case Intrinsic::x86_mmx_psrai_d:
6894 NewIntNo = Intrinsic::x86_mmx_psra_d;
6896 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6902 // The vector shift intrinsics with scalars uses 32b shift amounts but
6903 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6907 ShOps[1] = DAG.getConstant(0, MVT::i32);
6908 if (ShAmtVT == MVT::v4i32) {
6909 ShOps[2] = DAG.getUNDEF(MVT::i32);
6910 ShOps[3] = DAG.getUNDEF(MVT::i32);
6911 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6913 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6916 EVT VT = Op.getValueType();
6917 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6918 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6919 DAG.getConstant(NewIntNo, MVT::i32),
6920 Op.getOperand(1), ShAmt);
6925 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6926 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6927 DebugLoc dl = Op.getDebugLoc();
6930 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6932 DAG.getConstant(TD->getPointerSize(),
6933 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6934 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6935 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6940 // Just load the return address.
6941 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6942 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6943 RetAddrFI, NULL, 0);
6946 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6947 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6948 MFI->setFrameAddressIsTaken(true);
6949 EVT VT = Op.getValueType();
6950 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6951 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6952 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6953 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6955 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6959 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6960 SelectionDAG &DAG) {
6961 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6964 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6966 MachineFunction &MF = DAG.getMachineFunction();
6967 SDValue Chain = Op.getOperand(0);
6968 SDValue Offset = Op.getOperand(1);
6969 SDValue Handler = Op.getOperand(2);
6970 DebugLoc dl = Op.getDebugLoc();
6972 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6974 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6976 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6977 DAG.getIntPtrConstant(-TD->getPointerSize()));
6978 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6979 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6980 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6981 MF.getRegInfo().addLiveOut(StoreAddrReg);
6983 return DAG.getNode(X86ISD::EH_RETURN, dl,
6985 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6988 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6989 SelectionDAG &DAG) {
6990 SDValue Root = Op.getOperand(0);
6991 SDValue Trmp = Op.getOperand(1); // trampoline
6992 SDValue FPtr = Op.getOperand(2); // nested function
6993 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6994 DebugLoc dl = Op.getDebugLoc();
6996 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6998 const X86InstrInfo *TII =
6999 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
7001 if (Subtarget->is64Bit()) {
7002 SDValue OutChains[6];
7004 // Large code-model.
7006 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
7007 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
7009 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7010 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7012 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7014 // Load the pointer to the nested function into R11.
7015 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7016 SDValue Addr = Trmp;
7017 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7020 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7021 DAG.getConstant(2, MVT::i64));
7022 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
7024 // Load the 'nest' parameter value into R10.
7025 // R10 is specified in X86CallingConv.td
7026 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7027 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7028 DAG.getConstant(10, MVT::i64));
7029 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7030 Addr, TrmpAddr, 10);
7032 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7033 DAG.getConstant(12, MVT::i64));
7034 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
7036 // Jump to the nested function.
7037 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7038 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7039 DAG.getConstant(20, MVT::i64));
7040 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7041 Addr, TrmpAddr, 20);
7043 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7044 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7045 DAG.getConstant(22, MVT::i64));
7046 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7050 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7051 return DAG.getMergeValues(Ops, 2, dl);
7053 const Function *Func =
7054 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7055 CallingConv::ID CC = Func->getCallingConv();
7060 llvm_unreachable("Unsupported calling convention");
7061 case CallingConv::C:
7062 case CallingConv::X86_StdCall: {
7063 // Pass 'nest' parameter in ECX.
7064 // Must be kept in sync with X86CallingConv.td
7067 // Check that ECX wasn't needed by an 'inreg' parameter.
7068 const FunctionType *FTy = Func->getFunctionType();
7069 const AttrListPtr &Attrs = Func->getAttributes();
7071 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7072 unsigned InRegCount = 0;
7075 for (FunctionType::param_iterator I = FTy->param_begin(),
7076 E = FTy->param_end(); I != E; ++I, ++Idx)
7077 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7078 // FIXME: should only count parameters that are lowered to integers.
7079 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7081 if (InRegCount > 2) {
7082 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7087 case CallingConv::X86_FastCall:
7088 case CallingConv::Fast:
7089 // Pass 'nest' parameter in EAX.
7090 // Must be kept in sync with X86CallingConv.td
7095 SDValue OutChains[4];
7098 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7099 DAG.getConstant(10, MVT::i32));
7100 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7102 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
7103 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7104 OutChains[0] = DAG.getStore(Root, dl,
7105 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7108 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7109 DAG.getConstant(1, MVT::i32));
7110 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
7112 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
7113 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7114 DAG.getConstant(5, MVT::i32));
7115 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7116 TrmpAddr, 5, false, 1);
7118 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7119 DAG.getConstant(6, MVT::i32));
7120 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
7123 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7124 return DAG.getMergeValues(Ops, 2, dl);
7128 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7130 The rounding mode is in bits 11:10 of FPSR, and has the following
7137 FLT_ROUNDS, on the other hand, expects the following:
7144 To perform the conversion, we do:
7145 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7148 MachineFunction &MF = DAG.getMachineFunction();
7149 const TargetMachine &TM = MF.getTarget();
7150 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7151 unsigned StackAlignment = TFI.getStackAlignment();
7152 EVT VT = Op.getValueType();
7153 DebugLoc dl = Op.getDebugLoc();
7155 // Save FP Control Word to stack slot
7156 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7157 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7159 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7160 DAG.getEntryNode(), StackSlot);
7162 // Load FP Control Word from stack slot
7163 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
7165 // Transform as necessary
7167 DAG.getNode(ISD::SRL, dl, MVT::i16,
7168 DAG.getNode(ISD::AND, dl, MVT::i16,
7169 CWD, DAG.getConstant(0x800, MVT::i16)),
7170 DAG.getConstant(11, MVT::i8));
7172 DAG.getNode(ISD::SRL, dl, MVT::i16,
7173 DAG.getNode(ISD::AND, dl, MVT::i16,
7174 CWD, DAG.getConstant(0x400, MVT::i16)),
7175 DAG.getConstant(9, MVT::i8));
7178 DAG.getNode(ISD::AND, dl, MVT::i16,
7179 DAG.getNode(ISD::ADD, dl, MVT::i16,
7180 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7181 DAG.getConstant(1, MVT::i16)),
7182 DAG.getConstant(3, MVT::i16));
7185 return DAG.getNode((VT.getSizeInBits() < 16 ?
7186 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7189 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7190 EVT VT = Op.getValueType();
7192 unsigned NumBits = VT.getSizeInBits();
7193 DebugLoc dl = Op.getDebugLoc();
7195 Op = Op.getOperand(0);
7196 if (VT == MVT::i8) {
7197 // Zero extend to i32 since there is not an i8 bsr.
7199 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7202 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7203 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7204 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7206 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7209 DAG.getConstant(NumBits+NumBits-1, OpVT),
7210 DAG.getConstant(X86::COND_E, MVT::i8),
7213 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7215 // Finally xor with NumBits-1.
7216 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7219 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7223 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7224 EVT VT = Op.getValueType();
7226 unsigned NumBits = VT.getSizeInBits();
7227 DebugLoc dl = Op.getDebugLoc();
7229 Op = Op.getOperand(0);
7230 if (VT == MVT::i8) {
7232 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7235 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7236 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7237 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7239 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7242 DAG.getConstant(NumBits, OpVT),
7243 DAG.getConstant(X86::COND_E, MVT::i8),
7246 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7249 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7253 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7254 EVT VT = Op.getValueType();
7255 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7256 DebugLoc dl = Op.getDebugLoc();
7258 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7259 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7260 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7261 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7262 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7264 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7265 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7266 // return AloBlo + AloBhi + AhiBlo;
7268 SDValue A = Op.getOperand(0);
7269 SDValue B = Op.getOperand(1);
7271 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7272 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7273 A, DAG.getConstant(32, MVT::i32));
7274 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7275 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7276 B, DAG.getConstant(32, MVT::i32));
7277 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7278 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7280 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7281 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7283 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7284 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7286 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7287 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7288 AloBhi, DAG.getConstant(32, MVT::i32));
7289 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7290 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7291 AhiBlo, DAG.getConstant(32, MVT::i32));
7292 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7293 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7298 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7299 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7300 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7301 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7302 // has only one use.
7303 SDNode *N = Op.getNode();
7304 SDValue LHS = N->getOperand(0);
7305 SDValue RHS = N->getOperand(1);
7306 unsigned BaseOp = 0;
7308 DebugLoc dl = Op.getDebugLoc();
7310 switch (Op.getOpcode()) {
7311 default: llvm_unreachable("Unknown ovf instruction!");
7313 // A subtract of one will be selected as a INC. Note that INC doesn't
7314 // set CF, so we can't do this for UADDO.
7315 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7316 if (C->getAPIntValue() == 1) {
7317 BaseOp = X86ISD::INC;
7321 BaseOp = X86ISD::ADD;
7325 BaseOp = X86ISD::ADD;
7329 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7330 // set CF, so we can't do this for USUBO.
7331 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7332 if (C->getAPIntValue() == 1) {
7333 BaseOp = X86ISD::DEC;
7337 BaseOp = X86ISD::SUB;
7341 BaseOp = X86ISD::SUB;
7345 BaseOp = X86ISD::SMUL;
7349 BaseOp = X86ISD::UMUL;
7354 // Also sets EFLAGS.
7355 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7356 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7359 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7360 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7362 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7366 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7367 EVT T = Op.getValueType();
7368 DebugLoc dl = Op.getDebugLoc();
7371 switch(T.getSimpleVT().SimpleTy) {
7373 assert(false && "Invalid value type!");
7374 case MVT::i8: Reg = X86::AL; size = 1; break;
7375 case MVT::i16: Reg = X86::AX; size = 2; break;
7376 case MVT::i32: Reg = X86::EAX; size = 4; break;
7378 assert(Subtarget->is64Bit() && "Node not type legal!");
7379 Reg = X86::RAX; size = 8;
7382 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7383 Op.getOperand(2), SDValue());
7384 SDValue Ops[] = { cpIn.getValue(0),
7387 DAG.getTargetConstant(size, MVT::i8),
7389 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7390 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7392 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7396 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7397 SelectionDAG &DAG) {
7398 assert(Subtarget->is64Bit() && "Result not type legalized?");
7399 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7400 SDValue TheChain = Op.getOperand(0);
7401 DebugLoc dl = Op.getDebugLoc();
7402 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7403 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7404 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7406 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7407 DAG.getConstant(32, MVT::i8));
7409 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7412 return DAG.getMergeValues(Ops, 2, dl);
7415 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7416 SDNode *Node = Op.getNode();
7417 DebugLoc dl = Node->getDebugLoc();
7418 EVT T = Node->getValueType(0);
7419 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7420 DAG.getConstant(0, T), Node->getOperand(2));
7421 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7422 cast<AtomicSDNode>(Node)->getMemoryVT(),
7423 Node->getOperand(0),
7424 Node->getOperand(1), negOp,
7425 cast<AtomicSDNode>(Node)->getSrcValue(),
7426 cast<AtomicSDNode>(Node)->getAlignment());
7429 /// LowerOperation - Provide custom lowering hooks for some operations.
7431 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7432 switch (Op.getOpcode()) {
7433 default: llvm_unreachable("Should not custom lower this!");
7434 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7435 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7436 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7437 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7438 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7439 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7440 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7441 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7442 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7443 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7444 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7445 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7446 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7447 case ISD::SHL_PARTS:
7448 case ISD::SRA_PARTS:
7449 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7450 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7451 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7452 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7453 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7454 case ISD::FABS: return LowerFABS(Op, DAG);
7455 case ISD::FNEG: return LowerFNEG(Op, DAG);
7456 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7457 case ISD::SETCC: return LowerSETCC(Op, DAG);
7458 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7459 case ISD::SELECT: return LowerSELECT(Op, DAG);
7460 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7461 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7462 case ISD::VASTART: return LowerVASTART(Op, DAG);
7463 case ISD::VAARG: return LowerVAARG(Op, DAG);
7464 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7465 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7466 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7467 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7468 case ISD::FRAME_TO_ARGS_OFFSET:
7469 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7470 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7471 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7472 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7473 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7474 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7475 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7476 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7482 case ISD::UMULO: return LowerXALUO(Op, DAG);
7483 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7487 void X86TargetLowering::
7488 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7489 SelectionDAG &DAG, unsigned NewOp) {
7490 EVT T = Node->getValueType(0);
7491 DebugLoc dl = Node->getDebugLoc();
7492 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7494 SDValue Chain = Node->getOperand(0);
7495 SDValue In1 = Node->getOperand(1);
7496 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7497 Node->getOperand(2), DAG.getIntPtrConstant(0));
7498 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7499 Node->getOperand(2), DAG.getIntPtrConstant(1));
7500 SDValue Ops[] = { Chain, In1, In2L, In2H };
7501 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7503 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7504 cast<MemSDNode>(Node)->getMemOperand());
7505 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7506 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7507 Results.push_back(Result.getValue(2));
7510 /// ReplaceNodeResults - Replace a node with an illegal result type
7511 /// with a new node built out of custom code.
7512 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7513 SmallVectorImpl<SDValue>&Results,
7514 SelectionDAG &DAG) {
7515 DebugLoc dl = N->getDebugLoc();
7516 switch (N->getOpcode()) {
7518 assert(false && "Do not know how to custom type legalize this operation!");
7520 case ISD::FP_TO_SINT: {
7521 std::pair<SDValue,SDValue> Vals =
7522 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7523 SDValue FIST = Vals.first, StackSlot = Vals.second;
7524 if (FIST.getNode() != 0) {
7525 EVT VT = N->getValueType(0);
7526 // Return a load from the stack slot.
7527 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7531 case ISD::READCYCLECOUNTER: {
7532 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7533 SDValue TheChain = N->getOperand(0);
7534 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7535 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7537 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7539 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7540 SDValue Ops[] = { eax, edx };
7541 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7542 Results.push_back(edx.getValue(1));
7549 EVT WidenVT = getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
7550 Results.push_back(DAG.UnrollVectorOp(N, WidenVT.getVectorNumElements()));
7553 case ISD::ATOMIC_CMP_SWAP: {
7554 EVT T = N->getValueType(0);
7555 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7556 SDValue cpInL, cpInH;
7557 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7558 DAG.getConstant(0, MVT::i32));
7559 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7560 DAG.getConstant(1, MVT::i32));
7561 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7562 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7564 SDValue swapInL, swapInH;
7565 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7566 DAG.getConstant(0, MVT::i32));
7567 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7568 DAG.getConstant(1, MVT::i32));
7569 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7571 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7572 swapInL.getValue(1));
7573 SDValue Ops[] = { swapInH.getValue(0),
7575 swapInH.getValue(1) };
7576 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7577 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7578 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7579 MVT::i32, Result.getValue(1));
7580 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7581 MVT::i32, cpOutL.getValue(2));
7582 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7583 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7584 Results.push_back(cpOutH.getValue(1));
7587 case ISD::ATOMIC_LOAD_ADD:
7588 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7590 case ISD::ATOMIC_LOAD_AND:
7591 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7593 case ISD::ATOMIC_LOAD_NAND:
7594 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7596 case ISD::ATOMIC_LOAD_OR:
7597 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7599 case ISD::ATOMIC_LOAD_SUB:
7600 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7602 case ISD::ATOMIC_LOAD_XOR:
7603 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7605 case ISD::ATOMIC_SWAP:
7606 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7611 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7613 default: return NULL;
7614 case X86ISD::BSF: return "X86ISD::BSF";
7615 case X86ISD::BSR: return "X86ISD::BSR";
7616 case X86ISD::SHLD: return "X86ISD::SHLD";
7617 case X86ISD::SHRD: return "X86ISD::SHRD";
7618 case X86ISD::FAND: return "X86ISD::FAND";
7619 case X86ISD::FOR: return "X86ISD::FOR";
7620 case X86ISD::FXOR: return "X86ISD::FXOR";
7621 case X86ISD::FSRL: return "X86ISD::FSRL";
7622 case X86ISD::FILD: return "X86ISD::FILD";
7623 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7624 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7625 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7626 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7627 case X86ISD::FLD: return "X86ISD::FLD";
7628 case X86ISD::FST: return "X86ISD::FST";
7629 case X86ISD::CALL: return "X86ISD::CALL";
7630 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7631 case X86ISD::BT: return "X86ISD::BT";
7632 case X86ISD::CMP: return "X86ISD::CMP";
7633 case X86ISD::COMI: return "X86ISD::COMI";
7634 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7635 case X86ISD::SETCC: return "X86ISD::SETCC";
7636 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7637 case X86ISD::CMOV: return "X86ISD::CMOV";
7638 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7639 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7640 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7641 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7642 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7643 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7644 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7645 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7646 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7647 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7648 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7649 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7650 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7651 case X86ISD::FMAX: return "X86ISD::FMAX";
7652 case X86ISD::FMIN: return "X86ISD::FMIN";
7653 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7654 case X86ISD::FRCP: return "X86ISD::FRCP";
7655 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7656 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7657 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7658 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7659 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7660 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7661 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7662 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7663 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7664 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7665 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7666 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7667 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7668 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7669 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7670 case X86ISD::VSHL: return "X86ISD::VSHL";
7671 case X86ISD::VSRL: return "X86ISD::VSRL";
7672 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7673 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7674 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7675 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7676 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7677 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7678 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7679 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7680 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7681 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7682 case X86ISD::ADD: return "X86ISD::ADD";
7683 case X86ISD::SUB: return "X86ISD::SUB";
7684 case X86ISD::SMUL: return "X86ISD::SMUL";
7685 case X86ISD::UMUL: return "X86ISD::UMUL";
7686 case X86ISD::INC: return "X86ISD::INC";
7687 case X86ISD::DEC: return "X86ISD::DEC";
7688 case X86ISD::OR: return "X86ISD::OR";
7689 case X86ISD::XOR: return "X86ISD::XOR";
7690 case X86ISD::AND: return "X86ISD::AND";
7691 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7692 case X86ISD::PTEST: return "X86ISD::PTEST";
7693 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7697 // isLegalAddressingMode - Return true if the addressing mode represented
7698 // by AM is legal for this target, for a load/store of the specified type.
7699 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7700 const Type *Ty) const {
7701 // X86 supports extremely general addressing modes.
7702 CodeModel::Model M = getTargetMachine().getCodeModel();
7704 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7705 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7710 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7712 // If a reference to this global requires an extra load, we can't fold it.
7713 if (isGlobalStubReference(GVFlags))
7716 // If BaseGV requires a register for the PIC base, we cannot also have a
7717 // BaseReg specified.
7718 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7721 // If lower 4G is not available, then we must use rip-relative addressing.
7722 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7732 // These scales always work.
7737 // These scales are formed with basereg+scalereg. Only accept if there is
7742 default: // Other stuff never works.
7750 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7751 if (!Ty1->isInteger() || !Ty2->isInteger())
7753 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7754 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7755 if (NumBits1 <= NumBits2)
7757 return Subtarget->is64Bit() || NumBits1 < 64;
7760 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7761 if (!VT1.isInteger() || !VT2.isInteger())
7763 unsigned NumBits1 = VT1.getSizeInBits();
7764 unsigned NumBits2 = VT2.getSizeInBits();
7765 if (NumBits1 <= NumBits2)
7767 return Subtarget->is64Bit() || NumBits1 < 64;
7770 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7771 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7772 return Ty1->isInteger(32) && Ty2->isInteger(64) && Subtarget->is64Bit();
7775 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7776 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7777 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7780 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7781 // i16 instructions are longer (0x66 prefix) and potentially slower.
7782 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7785 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7786 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7787 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7788 /// are assumed to be legal.
7790 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7792 // Only do shuffles on 128-bit vector types for now.
7793 if (VT.getSizeInBits() == 64)
7796 // FIXME: pshufb, blends, shifts.
7797 return (VT.getVectorNumElements() == 2 ||
7798 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7799 isMOVLMask(M, VT) ||
7800 isSHUFPMask(M, VT) ||
7801 isPSHUFDMask(M, VT) ||
7802 isPSHUFHWMask(M, VT) ||
7803 isPSHUFLWMask(M, VT) ||
7804 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7805 isUNPCKLMask(M, VT) ||
7806 isUNPCKHMask(M, VT) ||
7807 isUNPCKL_v_undef_Mask(M, VT) ||
7808 isUNPCKH_v_undef_Mask(M, VT));
7812 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7814 unsigned NumElts = VT.getVectorNumElements();
7815 // FIXME: This collection of masks seems suspect.
7818 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7819 return (isMOVLMask(Mask, VT) ||
7820 isCommutedMOVLMask(Mask, VT, true) ||
7821 isSHUFPMask(Mask, VT) ||
7822 isCommutedSHUFPMask(Mask, VT));
7827 //===----------------------------------------------------------------------===//
7828 // X86 Scheduler Hooks
7829 //===----------------------------------------------------------------------===//
7831 // private utility function
7833 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7834 MachineBasicBlock *MBB,
7842 TargetRegisterClass *RC,
7843 bool invSrc) const {
7844 // For the atomic bitwise operator, we generate
7847 // ld t1 = [bitinstr.addr]
7848 // op t2 = t1, [bitinstr.val]
7850 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7852 // fallthrough -->nextMBB
7853 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7854 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7855 MachineFunction::iterator MBBIter = MBB;
7858 /// First build the CFG
7859 MachineFunction *F = MBB->getParent();
7860 MachineBasicBlock *thisMBB = MBB;
7861 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7862 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7863 F->insert(MBBIter, newMBB);
7864 F->insert(MBBIter, nextMBB);
7866 // Move all successors to thisMBB to nextMBB
7867 nextMBB->transferSuccessors(thisMBB);
7869 // Update thisMBB to fall through to newMBB
7870 thisMBB->addSuccessor(newMBB);
7872 // newMBB jumps to itself and fall through to nextMBB
7873 newMBB->addSuccessor(nextMBB);
7874 newMBB->addSuccessor(newMBB);
7876 // Insert instructions into newMBB based on incoming instruction
7877 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7878 "unexpected number of operands");
7879 DebugLoc dl = bInstr->getDebugLoc();
7880 MachineOperand& destOper = bInstr->getOperand(0);
7881 MachineOperand* argOpers[2 + X86AddrNumOperands];
7882 int numArgs = bInstr->getNumOperands() - 1;
7883 for (int i=0; i < numArgs; ++i)
7884 argOpers[i] = &bInstr->getOperand(i+1);
7886 // x86 address has 4 operands: base, index, scale, and displacement
7887 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7888 int valArgIndx = lastAddrIndx + 1;
7890 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7891 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7892 for (int i=0; i <= lastAddrIndx; ++i)
7893 (*MIB).addOperand(*argOpers[i]);
7895 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7897 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7902 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7903 assert((argOpers[valArgIndx]->isReg() ||
7904 argOpers[valArgIndx]->isImm()) &&
7906 if (argOpers[valArgIndx]->isReg())
7907 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7909 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7911 (*MIB).addOperand(*argOpers[valArgIndx]);
7913 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7916 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7917 for (int i=0; i <= lastAddrIndx; ++i)
7918 (*MIB).addOperand(*argOpers[i]);
7920 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7921 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7922 bInstr->memoperands_end());
7924 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7928 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7930 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7934 // private utility function: 64 bit atomics on 32 bit host.
7936 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7937 MachineBasicBlock *MBB,
7942 bool invSrc) const {
7943 // For the atomic bitwise operator, we generate
7944 // thisMBB (instructions are in pairs, except cmpxchg8b)
7945 // ld t1,t2 = [bitinstr.addr]
7947 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7948 // op t5, t6 <- out1, out2, [bitinstr.val]
7949 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7950 // mov ECX, EBX <- t5, t6
7951 // mov EAX, EDX <- t1, t2
7952 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7953 // mov t3, t4 <- EAX, EDX
7955 // result in out1, out2
7956 // fallthrough -->nextMBB
7958 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7959 const unsigned LoadOpc = X86::MOV32rm;
7960 const unsigned copyOpc = X86::MOV32rr;
7961 const unsigned NotOpc = X86::NOT32r;
7962 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7963 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7964 MachineFunction::iterator MBBIter = MBB;
7967 /// First build the CFG
7968 MachineFunction *F = MBB->getParent();
7969 MachineBasicBlock *thisMBB = MBB;
7970 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7971 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7972 F->insert(MBBIter, newMBB);
7973 F->insert(MBBIter, nextMBB);
7975 // Move all successors to thisMBB to nextMBB
7976 nextMBB->transferSuccessors(thisMBB);
7978 // Update thisMBB to fall through to newMBB
7979 thisMBB->addSuccessor(newMBB);
7981 // newMBB jumps to itself and fall through to nextMBB
7982 newMBB->addSuccessor(nextMBB);
7983 newMBB->addSuccessor(newMBB);
7985 DebugLoc dl = bInstr->getDebugLoc();
7986 // Insert instructions into newMBB based on incoming instruction
7987 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7988 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7989 "unexpected number of operands");
7990 MachineOperand& dest1Oper = bInstr->getOperand(0);
7991 MachineOperand& dest2Oper = bInstr->getOperand(1);
7992 MachineOperand* argOpers[2 + X86AddrNumOperands];
7993 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7994 argOpers[i] = &bInstr->getOperand(i+2);
7996 // x86 address has 5 operands: base, index, scale, displacement, and segment.
7997 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7999 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8000 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8001 for (int i=0; i <= lastAddrIndx; ++i)
8002 (*MIB).addOperand(*argOpers[i]);
8003 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8004 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8005 // add 4 to displacement.
8006 for (int i=0; i <= lastAddrIndx-2; ++i)
8007 (*MIB).addOperand(*argOpers[i]);
8008 MachineOperand newOp3 = *(argOpers[3]);
8010 newOp3.setImm(newOp3.getImm()+4);
8012 newOp3.setOffset(newOp3.getOffset()+4);
8013 (*MIB).addOperand(newOp3);
8014 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8016 // t3/4 are defined later, at the bottom of the loop
8017 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8018 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8019 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8020 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8021 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8022 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8024 // The subsequent operations should be using the destination registers of
8025 //the PHI instructions.
8027 t1 = F->getRegInfo().createVirtualRegister(RC);
8028 t2 = F->getRegInfo().createVirtualRegister(RC);
8029 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8030 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8032 t1 = dest1Oper.getReg();
8033 t2 = dest2Oper.getReg();
8036 int valArgIndx = lastAddrIndx + 1;
8037 assert((argOpers[valArgIndx]->isReg() ||
8038 argOpers[valArgIndx]->isImm()) &&
8040 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8041 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8042 if (argOpers[valArgIndx]->isReg())
8043 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8045 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8046 if (regOpcL != X86::MOV32rr)
8048 (*MIB).addOperand(*argOpers[valArgIndx]);
8049 assert(argOpers[valArgIndx + 1]->isReg() ==
8050 argOpers[valArgIndx]->isReg());
8051 assert(argOpers[valArgIndx + 1]->isImm() ==
8052 argOpers[valArgIndx]->isImm());
8053 if (argOpers[valArgIndx + 1]->isReg())
8054 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8056 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8057 if (regOpcH != X86::MOV32rr)
8059 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8061 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8063 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8066 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8068 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8071 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8072 for (int i=0; i <= lastAddrIndx; ++i)
8073 (*MIB).addOperand(*argOpers[i]);
8075 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8076 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8077 bInstr->memoperands_end());
8079 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8080 MIB.addReg(X86::EAX);
8081 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8082 MIB.addReg(X86::EDX);
8085 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8087 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8091 // private utility function
8093 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8094 MachineBasicBlock *MBB,
8095 unsigned cmovOpc) const {
8096 // For the atomic min/max operator, we generate
8099 // ld t1 = [min/max.addr]
8100 // mov t2 = [min/max.val]
8102 // cmov[cond] t2 = t1
8104 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8106 // fallthrough -->nextMBB
8108 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8109 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8110 MachineFunction::iterator MBBIter = MBB;
8113 /// First build the CFG
8114 MachineFunction *F = MBB->getParent();
8115 MachineBasicBlock *thisMBB = MBB;
8116 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8117 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8118 F->insert(MBBIter, newMBB);
8119 F->insert(MBBIter, nextMBB);
8121 // Move all successors of thisMBB to nextMBB
8122 nextMBB->transferSuccessors(thisMBB);
8124 // Update thisMBB to fall through to newMBB
8125 thisMBB->addSuccessor(newMBB);
8127 // newMBB jumps to newMBB and fall through to nextMBB
8128 newMBB->addSuccessor(nextMBB);
8129 newMBB->addSuccessor(newMBB);
8131 DebugLoc dl = mInstr->getDebugLoc();
8132 // Insert instructions into newMBB based on incoming instruction
8133 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8134 "unexpected number of operands");
8135 MachineOperand& destOper = mInstr->getOperand(0);
8136 MachineOperand* argOpers[2 + X86AddrNumOperands];
8137 int numArgs = mInstr->getNumOperands() - 1;
8138 for (int i=0; i < numArgs; ++i)
8139 argOpers[i] = &mInstr->getOperand(i+1);
8141 // x86 address has 4 operands: base, index, scale, and displacement
8142 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8143 int valArgIndx = lastAddrIndx + 1;
8145 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8146 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8147 for (int i=0; i <= lastAddrIndx; ++i)
8148 (*MIB).addOperand(*argOpers[i]);
8150 // We only support register and immediate values
8151 assert((argOpers[valArgIndx]->isReg() ||
8152 argOpers[valArgIndx]->isImm()) &&
8155 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8156 if (argOpers[valArgIndx]->isReg())
8157 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8159 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8160 (*MIB).addOperand(*argOpers[valArgIndx]);
8162 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8165 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8170 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8171 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8175 // Cmp and exchange if none has modified the memory location
8176 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8177 for (int i=0; i <= lastAddrIndx; ++i)
8178 (*MIB).addOperand(*argOpers[i]);
8180 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8181 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8182 mInstr->memoperands_end());
8184 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8185 MIB.addReg(X86::EAX);
8188 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
8190 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8194 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8195 // all of this code can be replaced with that in the .td file.
8197 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8198 unsigned numArgs, bool memArg) const {
8200 MachineFunction *F = BB->getParent();
8201 DebugLoc dl = MI->getDebugLoc();
8202 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8206 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8208 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8210 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8212 for (unsigned i = 0; i < numArgs; ++i) {
8213 MachineOperand &Op = MI->getOperand(i+1);
8215 if (!(Op.isReg() && Op.isImplicit()))
8219 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8222 F->DeleteMachineInstr(MI);
8228 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8230 MachineBasicBlock *MBB) const {
8231 // Emit code to save XMM registers to the stack. The ABI says that the
8232 // number of registers to save is given in %al, so it's theoretically
8233 // possible to do an indirect jump trick to avoid saving all of them,
8234 // however this code takes a simpler approach and just executes all
8235 // of the stores if %al is non-zero. It's less code, and it's probably
8236 // easier on the hardware branch predictor, and stores aren't all that
8237 // expensive anyway.
8239 // Create the new basic blocks. One block contains all the XMM stores,
8240 // and one block is the final destination regardless of whether any
8241 // stores were performed.
8242 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8243 MachineFunction *F = MBB->getParent();
8244 MachineFunction::iterator MBBIter = MBB;
8246 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8247 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8248 F->insert(MBBIter, XMMSaveMBB);
8249 F->insert(MBBIter, EndMBB);
8252 // Move any original successors of MBB to the end block.
8253 EndMBB->transferSuccessors(MBB);
8254 // The original block will now fall through to the XMM save block.
8255 MBB->addSuccessor(XMMSaveMBB);
8256 // The XMMSaveMBB will fall through to the end block.
8257 XMMSaveMBB->addSuccessor(EndMBB);
8259 // Now add the instructions.
8260 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8261 DebugLoc DL = MI->getDebugLoc();
8263 unsigned CountReg = MI->getOperand(0).getReg();
8264 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8265 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8267 if (!Subtarget->isTargetWin64()) {
8268 // If %al is 0, branch around the XMM save block.
8269 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8270 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
8271 MBB->addSuccessor(EndMBB);
8274 // In the XMM save block, save all the XMM argument registers.
8275 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8276 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8277 MachineMemOperand *MMO =
8278 F->getMachineMemOperand(
8279 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8280 MachineMemOperand::MOStore, Offset,
8281 /*Size=*/16, /*Align=*/16);
8282 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8283 .addFrameIndex(RegSaveFrameIndex)
8284 .addImm(/*Scale=*/1)
8285 .addReg(/*IndexReg=*/0)
8286 .addImm(/*Disp=*/Offset)
8287 .addReg(/*Segment=*/0)
8288 .addReg(MI->getOperand(i).getReg())
8289 .addMemOperand(MMO);
8292 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8298 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8299 MachineBasicBlock *BB,
8300 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8301 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8302 DebugLoc DL = MI->getDebugLoc();
8304 // To "insert" a SELECT_CC instruction, we actually have to insert the
8305 // diamond control-flow pattern. The incoming instruction knows the
8306 // destination vreg to set, the condition code register to branch on, the
8307 // true/false values to select between, and a branch opcode to use.
8308 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8309 MachineFunction::iterator It = BB;
8315 // cmpTY ccX, r1, r2
8317 // fallthrough --> copy0MBB
8318 MachineBasicBlock *thisMBB = BB;
8319 MachineFunction *F = BB->getParent();
8320 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8321 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8323 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8324 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8325 F->insert(It, copy0MBB);
8326 F->insert(It, sinkMBB);
8327 // Update machine-CFG edges by first adding all successors of the current
8328 // block to the new block which will contain the Phi node for the select.
8329 // Also inform sdisel of the edge changes.
8330 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8331 E = BB->succ_end(); I != E; ++I) {
8332 EM->insert(std::make_pair(*I, sinkMBB));
8333 sinkMBB->addSuccessor(*I);
8335 // Next, remove all successors of the current block, and add the true
8336 // and fallthrough blocks as its successors.
8337 while (!BB->succ_empty())
8338 BB->removeSuccessor(BB->succ_begin());
8339 // Add the true and fallthrough blocks as its successors.
8340 BB->addSuccessor(copy0MBB);
8341 BB->addSuccessor(sinkMBB);
8344 // %FalseValue = ...
8345 // # fallthrough to sinkMBB
8348 // Update machine-CFG edges
8349 BB->addSuccessor(sinkMBB);
8352 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8355 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8356 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8357 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8359 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8365 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8366 MachineBasicBlock *BB,
8367 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8368 switch (MI->getOpcode()) {
8369 default: assert(false && "Unexpected instr type to insert");
8371 case X86::CMOV_V1I64:
8372 case X86::CMOV_FR32:
8373 case X86::CMOV_FR64:
8374 case X86::CMOV_V4F32:
8375 case X86::CMOV_V2F64:
8376 case X86::CMOV_V2I64:
8377 return EmitLoweredSelect(MI, BB, EM);
8379 case X86::FP32_TO_INT16_IN_MEM:
8380 case X86::FP32_TO_INT32_IN_MEM:
8381 case X86::FP32_TO_INT64_IN_MEM:
8382 case X86::FP64_TO_INT16_IN_MEM:
8383 case X86::FP64_TO_INT32_IN_MEM:
8384 case X86::FP64_TO_INT64_IN_MEM:
8385 case X86::FP80_TO_INT16_IN_MEM:
8386 case X86::FP80_TO_INT32_IN_MEM:
8387 case X86::FP80_TO_INT64_IN_MEM: {
8388 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8389 DebugLoc DL = MI->getDebugLoc();
8391 // Change the floating point control register to use "round towards zero"
8392 // mode when truncating to an integer value.
8393 MachineFunction *F = BB->getParent();
8394 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8395 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8397 // Load the old value of the high byte of the control word...
8399 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8400 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8403 // Set the high part to be round to zero...
8404 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8407 // Reload the modified control word now...
8408 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8410 // Restore the memory image of control word to original value
8411 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8414 // Get the X86 opcode to use.
8416 switch (MI->getOpcode()) {
8417 default: llvm_unreachable("illegal opcode!");
8418 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8419 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8420 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8421 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8422 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8423 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8424 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8425 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8426 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8430 MachineOperand &Op = MI->getOperand(0);
8432 AM.BaseType = X86AddressMode::RegBase;
8433 AM.Base.Reg = Op.getReg();
8435 AM.BaseType = X86AddressMode::FrameIndexBase;
8436 AM.Base.FrameIndex = Op.getIndex();
8438 Op = MI->getOperand(1);
8440 AM.Scale = Op.getImm();
8441 Op = MI->getOperand(2);
8443 AM.IndexReg = Op.getImm();
8444 Op = MI->getOperand(3);
8445 if (Op.isGlobal()) {
8446 AM.GV = Op.getGlobal();
8448 AM.Disp = Op.getImm();
8450 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8451 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8453 // Reload the original control word now.
8454 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8456 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8459 // String/text processing lowering.
8460 case X86::PCMPISTRM128REG:
8461 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8462 case X86::PCMPISTRM128MEM:
8463 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8464 case X86::PCMPESTRM128REG:
8465 return EmitPCMP(MI, BB, 5, false /* in mem */);
8466 case X86::PCMPESTRM128MEM:
8467 return EmitPCMP(MI, BB, 5, true /* in mem */);
8470 case X86::ATOMAND32:
8471 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8472 X86::AND32ri, X86::MOV32rm,
8473 X86::LCMPXCHG32, X86::MOV32rr,
8474 X86::NOT32r, X86::EAX,
8475 X86::GR32RegisterClass);
8477 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8478 X86::OR32ri, X86::MOV32rm,
8479 X86::LCMPXCHG32, X86::MOV32rr,
8480 X86::NOT32r, X86::EAX,
8481 X86::GR32RegisterClass);
8482 case X86::ATOMXOR32:
8483 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8484 X86::XOR32ri, X86::MOV32rm,
8485 X86::LCMPXCHG32, X86::MOV32rr,
8486 X86::NOT32r, X86::EAX,
8487 X86::GR32RegisterClass);
8488 case X86::ATOMNAND32:
8489 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8490 X86::AND32ri, X86::MOV32rm,
8491 X86::LCMPXCHG32, X86::MOV32rr,
8492 X86::NOT32r, X86::EAX,
8493 X86::GR32RegisterClass, true);
8494 case X86::ATOMMIN32:
8495 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8496 case X86::ATOMMAX32:
8497 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8498 case X86::ATOMUMIN32:
8499 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8500 case X86::ATOMUMAX32:
8501 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8503 case X86::ATOMAND16:
8504 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8505 X86::AND16ri, X86::MOV16rm,
8506 X86::LCMPXCHG16, X86::MOV16rr,
8507 X86::NOT16r, X86::AX,
8508 X86::GR16RegisterClass);
8510 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8511 X86::OR16ri, X86::MOV16rm,
8512 X86::LCMPXCHG16, X86::MOV16rr,
8513 X86::NOT16r, X86::AX,
8514 X86::GR16RegisterClass);
8515 case X86::ATOMXOR16:
8516 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8517 X86::XOR16ri, X86::MOV16rm,
8518 X86::LCMPXCHG16, X86::MOV16rr,
8519 X86::NOT16r, X86::AX,
8520 X86::GR16RegisterClass);
8521 case X86::ATOMNAND16:
8522 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8523 X86::AND16ri, X86::MOV16rm,
8524 X86::LCMPXCHG16, X86::MOV16rr,
8525 X86::NOT16r, X86::AX,
8526 X86::GR16RegisterClass, true);
8527 case X86::ATOMMIN16:
8528 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8529 case X86::ATOMMAX16:
8530 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8531 case X86::ATOMUMIN16:
8532 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8533 case X86::ATOMUMAX16:
8534 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8537 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8538 X86::AND8ri, X86::MOV8rm,
8539 X86::LCMPXCHG8, X86::MOV8rr,
8540 X86::NOT8r, X86::AL,
8541 X86::GR8RegisterClass);
8543 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8544 X86::OR8ri, X86::MOV8rm,
8545 X86::LCMPXCHG8, X86::MOV8rr,
8546 X86::NOT8r, X86::AL,
8547 X86::GR8RegisterClass);
8549 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8550 X86::XOR8ri, X86::MOV8rm,
8551 X86::LCMPXCHG8, X86::MOV8rr,
8552 X86::NOT8r, X86::AL,
8553 X86::GR8RegisterClass);
8554 case X86::ATOMNAND8:
8555 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8556 X86::AND8ri, X86::MOV8rm,
8557 X86::LCMPXCHG8, X86::MOV8rr,
8558 X86::NOT8r, X86::AL,
8559 X86::GR8RegisterClass, true);
8560 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8561 // This group is for 64-bit host.
8562 case X86::ATOMAND64:
8563 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8564 X86::AND64ri32, X86::MOV64rm,
8565 X86::LCMPXCHG64, X86::MOV64rr,
8566 X86::NOT64r, X86::RAX,
8567 X86::GR64RegisterClass);
8569 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8570 X86::OR64ri32, X86::MOV64rm,
8571 X86::LCMPXCHG64, X86::MOV64rr,
8572 X86::NOT64r, X86::RAX,
8573 X86::GR64RegisterClass);
8574 case X86::ATOMXOR64:
8575 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8576 X86::XOR64ri32, X86::MOV64rm,
8577 X86::LCMPXCHG64, X86::MOV64rr,
8578 X86::NOT64r, X86::RAX,
8579 X86::GR64RegisterClass);
8580 case X86::ATOMNAND64:
8581 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8582 X86::AND64ri32, X86::MOV64rm,
8583 X86::LCMPXCHG64, X86::MOV64rr,
8584 X86::NOT64r, X86::RAX,
8585 X86::GR64RegisterClass, true);
8586 case X86::ATOMMIN64:
8587 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8588 case X86::ATOMMAX64:
8589 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8590 case X86::ATOMUMIN64:
8591 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8592 case X86::ATOMUMAX64:
8593 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8595 // This group does 64-bit operations on a 32-bit host.
8596 case X86::ATOMAND6432:
8597 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8598 X86::AND32rr, X86::AND32rr,
8599 X86::AND32ri, X86::AND32ri,
8601 case X86::ATOMOR6432:
8602 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8603 X86::OR32rr, X86::OR32rr,
8604 X86::OR32ri, X86::OR32ri,
8606 case X86::ATOMXOR6432:
8607 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8608 X86::XOR32rr, X86::XOR32rr,
8609 X86::XOR32ri, X86::XOR32ri,
8611 case X86::ATOMNAND6432:
8612 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8613 X86::AND32rr, X86::AND32rr,
8614 X86::AND32ri, X86::AND32ri,
8616 case X86::ATOMADD6432:
8617 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8618 X86::ADD32rr, X86::ADC32rr,
8619 X86::ADD32ri, X86::ADC32ri,
8621 case X86::ATOMSUB6432:
8622 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8623 X86::SUB32rr, X86::SBB32rr,
8624 X86::SUB32ri, X86::SBB32ri,
8626 case X86::ATOMSWAP6432:
8627 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8628 X86::MOV32rr, X86::MOV32rr,
8629 X86::MOV32ri, X86::MOV32ri,
8631 case X86::VASTART_SAVE_XMM_REGS:
8632 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8636 //===----------------------------------------------------------------------===//
8637 // X86 Optimization Hooks
8638 //===----------------------------------------------------------------------===//
8640 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8644 const SelectionDAG &DAG,
8645 unsigned Depth) const {
8646 unsigned Opc = Op.getOpcode();
8647 assert((Opc >= ISD::BUILTIN_OP_END ||
8648 Opc == ISD::INTRINSIC_WO_CHAIN ||
8649 Opc == ISD::INTRINSIC_W_CHAIN ||
8650 Opc == ISD::INTRINSIC_VOID) &&
8651 "Should use MaskedValueIsZero if you don't know whether Op"
8652 " is a target node!");
8654 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8666 // These nodes' second result is a boolean.
8667 if (Op.getResNo() == 0)
8671 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8672 Mask.getBitWidth() - 1);
8677 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8678 /// node is a GlobalAddress + offset.
8679 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8680 GlobalValue* &GA, int64_t &Offset) const{
8681 if (N->getOpcode() == X86ISD::Wrapper) {
8682 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8683 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8684 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8688 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8691 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8692 EVT EltVT, LoadSDNode *&LDBase,
8693 unsigned &LastLoadedElt,
8694 SelectionDAG &DAG, MachineFrameInfo *MFI,
8695 const TargetLowering &TLI) {
8697 LastLoadedElt = -1U;
8698 for (unsigned i = 0; i < NumElems; ++i) {
8699 if (N->getMaskElt(i) < 0) {
8705 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8706 if (!Elt.getNode() ||
8707 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8710 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8712 LDBase = cast<LoadSDNode>(Elt.getNode());
8716 if (Elt.getOpcode() == ISD::UNDEF)
8719 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8720 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8727 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8728 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8729 /// if the load addresses are consecutive, non-overlapping, and in the right
8730 /// order. In the case of v2i64, it will see if it can rewrite the
8731 /// shuffle to be an appropriate build vector so it can take advantage of
8732 // performBuildVectorCombine.
8733 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8734 const TargetLowering &TLI) {
8735 DebugLoc dl = N->getDebugLoc();
8736 EVT VT = N->getValueType(0);
8737 EVT EltVT = VT.getVectorElementType();
8738 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8739 unsigned NumElems = VT.getVectorNumElements();
8741 if (VT.getSizeInBits() != 128)
8744 // Try to combine a vector_shuffle into a 128-bit load.
8745 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8746 LoadSDNode *LD = NULL;
8747 unsigned LastLoadedElt;
8748 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8752 if (LastLoadedElt == NumElems - 1) {
8753 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8754 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8755 LD->getSrcValue(), LD->getSrcValueOffset(),
8757 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8758 LD->getSrcValue(), LD->getSrcValueOffset(),
8759 LD->isVolatile(), LD->getAlignment());
8760 } else if (NumElems == 4 && LastLoadedElt == 1) {
8761 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8762 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8763 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8764 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8769 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8770 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8771 const X86Subtarget *Subtarget) {
8772 DebugLoc DL = N->getDebugLoc();
8773 SDValue Cond = N->getOperand(0);
8774 // Get the LHS/RHS of the select.
8775 SDValue LHS = N->getOperand(1);
8776 SDValue RHS = N->getOperand(2);
8778 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8779 // instructions have the peculiarity that if either operand is a NaN,
8780 // they chose what we call the RHS operand (and as such are not symmetric).
8781 // It happens that this matches the semantics of the common C idiom
8782 // x<y?x:y and related forms, so we can recognize these cases.
8783 if (Subtarget->hasSSE2() &&
8784 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8785 Cond.getOpcode() == ISD::SETCC) {
8786 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8788 unsigned Opcode = 0;
8789 // Check for x CC y ? x : y.
8790 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8794 // This can be a min if we can prove that at least one of the operands
8796 if (!FiniteOnlyFPMath()) {
8797 if (DAG.isKnownNeverNaN(RHS)) {
8798 // Put the potential NaN in the RHS so that SSE will preserve it.
8799 std::swap(LHS, RHS);
8800 } else if (!DAG.isKnownNeverNaN(LHS))
8803 Opcode = X86ISD::FMIN;
8806 // This can be a min if we can prove that at least one of the operands
8808 if (!FiniteOnlyFPMath()) {
8809 if (DAG.isKnownNeverNaN(LHS)) {
8810 // Put the potential NaN in the RHS so that SSE will preserve it.
8811 std::swap(LHS, RHS);
8812 } else if (!DAG.isKnownNeverNaN(RHS))
8815 Opcode = X86ISD::FMIN;
8818 // This can be a min, but if either operand is a NaN we need it to
8819 // preserve the original LHS.
8820 std::swap(LHS, RHS);
8824 Opcode = X86ISD::FMIN;
8828 // This can be a max if we can prove that at least one of the operands
8830 if (!FiniteOnlyFPMath()) {
8831 if (DAG.isKnownNeverNaN(LHS)) {
8832 // Put the potential NaN in the RHS so that SSE will preserve it.
8833 std::swap(LHS, RHS);
8834 } else if (!DAG.isKnownNeverNaN(RHS))
8837 Opcode = X86ISD::FMAX;
8840 // This can be a max if we can prove that at least one of the operands
8842 if (!FiniteOnlyFPMath()) {
8843 if (DAG.isKnownNeverNaN(RHS)) {
8844 // Put the potential NaN in the RHS so that SSE will preserve it.
8845 std::swap(LHS, RHS);
8846 } else if (!DAG.isKnownNeverNaN(LHS))
8849 Opcode = X86ISD::FMAX;
8852 // This can be a max, but if either operand is a NaN we need it to
8853 // preserve the original LHS.
8854 std::swap(LHS, RHS);
8858 Opcode = X86ISD::FMAX;
8861 // Check for x CC y ? y : x -- a min/max with reversed arms.
8862 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8866 // This can be a min if we can prove that at least one of the operands
8868 if (!FiniteOnlyFPMath()) {
8869 if (DAG.isKnownNeverNaN(RHS)) {
8870 // Put the potential NaN in the RHS so that SSE will preserve it.
8871 std::swap(LHS, RHS);
8872 } else if (!DAG.isKnownNeverNaN(LHS))
8875 Opcode = X86ISD::FMIN;
8878 // This can be a min if we can prove that at least one of the operands
8880 if (!FiniteOnlyFPMath()) {
8881 if (DAG.isKnownNeverNaN(LHS)) {
8882 // Put the potential NaN in the RHS so that SSE will preserve it.
8883 std::swap(LHS, RHS);
8884 } else if (!DAG.isKnownNeverNaN(RHS))
8887 Opcode = X86ISD::FMIN;
8890 // This can be a min, but if either operand is a NaN we need it to
8891 // preserve the original LHS.
8892 std::swap(LHS, RHS);
8896 Opcode = X86ISD::FMIN;
8900 // This can be a max if we can prove that at least one of the operands
8902 if (!FiniteOnlyFPMath()) {
8903 if (DAG.isKnownNeverNaN(LHS)) {
8904 // Put the potential NaN in the RHS so that SSE will preserve it.
8905 std::swap(LHS, RHS);
8906 } else if (!DAG.isKnownNeverNaN(RHS))
8909 Opcode = X86ISD::FMAX;
8912 // This can be a max if we can prove that at least one of the operands
8914 if (!FiniteOnlyFPMath()) {
8915 if (DAG.isKnownNeverNaN(RHS)) {
8916 // Put the potential NaN in the RHS so that SSE will preserve it.
8917 std::swap(LHS, RHS);
8918 } else if (!DAG.isKnownNeverNaN(LHS))
8921 Opcode = X86ISD::FMAX;
8924 // This can be a max, but if either operand is a NaN we need it to
8925 // preserve the original LHS.
8926 std::swap(LHS, RHS);
8930 Opcode = X86ISD::FMAX;
8936 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8939 // If this is a select between two integer constants, try to do some
8941 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8942 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8943 // Don't do this for crazy integer types.
8944 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8945 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8946 // so that TrueC (the true value) is larger than FalseC.
8947 bool NeedsCondInvert = false;
8949 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8950 // Efficiently invertible.
8951 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8952 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8953 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8954 NeedsCondInvert = true;
8955 std::swap(TrueC, FalseC);
8958 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8959 if (FalseC->getAPIntValue() == 0 &&
8960 TrueC->getAPIntValue().isPowerOf2()) {
8961 if (NeedsCondInvert) // Invert the condition if needed.
8962 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8963 DAG.getConstant(1, Cond.getValueType()));
8965 // Zero extend the condition if needed.
8966 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8968 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8969 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8970 DAG.getConstant(ShAmt, MVT::i8));
8973 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8974 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8975 if (NeedsCondInvert) // Invert the condition if needed.
8976 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8977 DAG.getConstant(1, Cond.getValueType()));
8979 // Zero extend the condition if needed.
8980 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8981 FalseC->getValueType(0), Cond);
8982 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8983 SDValue(FalseC, 0));
8986 // Optimize cases that will turn into an LEA instruction. This requires
8987 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8988 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8989 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8990 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8992 bool isFastMultiplier = false;
8994 switch ((unsigned char)Diff) {
8996 case 1: // result = add base, cond
8997 case 2: // result = lea base( , cond*2)
8998 case 3: // result = lea base(cond, cond*2)
8999 case 4: // result = lea base( , cond*4)
9000 case 5: // result = lea base(cond, cond*4)
9001 case 8: // result = lea base( , cond*8)
9002 case 9: // result = lea base(cond, cond*8)
9003 isFastMultiplier = true;
9008 if (isFastMultiplier) {
9009 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9010 if (NeedsCondInvert) // Invert the condition if needed.
9011 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9012 DAG.getConstant(1, Cond.getValueType()));
9014 // Zero extend the condition if needed.
9015 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9017 // Scale the condition by the difference.
9019 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9020 DAG.getConstant(Diff, Cond.getValueType()));
9022 // Add the base if non-zero.
9023 if (FalseC->getAPIntValue() != 0)
9024 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9025 SDValue(FalseC, 0));
9035 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9036 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9037 TargetLowering::DAGCombinerInfo &DCI) {
9038 DebugLoc DL = N->getDebugLoc();
9040 // If the flag operand isn't dead, don't touch this CMOV.
9041 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9044 // If this is a select between two integer constants, try to do some
9045 // optimizations. Note that the operands are ordered the opposite of SELECT
9047 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9048 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9049 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9050 // larger than FalseC (the false value).
9051 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9053 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9054 CC = X86::GetOppositeBranchCondition(CC);
9055 std::swap(TrueC, FalseC);
9058 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9059 // This is efficient for any integer data type (including i8/i16) and
9061 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9062 SDValue Cond = N->getOperand(3);
9063 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9064 DAG.getConstant(CC, MVT::i8), Cond);
9066 // Zero extend the condition if needed.
9067 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9069 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9070 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9071 DAG.getConstant(ShAmt, MVT::i8));
9072 if (N->getNumValues() == 2) // Dead flag value?
9073 return DCI.CombineTo(N, Cond, SDValue());
9077 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9078 // for any integer data type, including i8/i16.
9079 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9080 SDValue Cond = N->getOperand(3);
9081 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9082 DAG.getConstant(CC, MVT::i8), Cond);
9084 // Zero extend the condition if needed.
9085 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9086 FalseC->getValueType(0), Cond);
9087 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9088 SDValue(FalseC, 0));
9090 if (N->getNumValues() == 2) // Dead flag value?
9091 return DCI.CombineTo(N, Cond, SDValue());
9095 // Optimize cases that will turn into an LEA instruction. This requires
9096 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9097 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9098 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9099 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9101 bool isFastMultiplier = false;
9103 switch ((unsigned char)Diff) {
9105 case 1: // result = add base, cond
9106 case 2: // result = lea base( , cond*2)
9107 case 3: // result = lea base(cond, cond*2)
9108 case 4: // result = lea base( , cond*4)
9109 case 5: // result = lea base(cond, cond*4)
9110 case 8: // result = lea base( , cond*8)
9111 case 9: // result = lea base(cond, cond*8)
9112 isFastMultiplier = true;
9117 if (isFastMultiplier) {
9118 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9119 SDValue Cond = N->getOperand(3);
9120 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9121 DAG.getConstant(CC, MVT::i8), Cond);
9122 // Zero extend the condition if needed.
9123 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9125 // Scale the condition by the difference.
9127 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9128 DAG.getConstant(Diff, Cond.getValueType()));
9130 // Add the base if non-zero.
9131 if (FalseC->getAPIntValue() != 0)
9132 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9133 SDValue(FalseC, 0));
9134 if (N->getNumValues() == 2) // Dead flag value?
9135 return DCI.CombineTo(N, Cond, SDValue());
9145 /// PerformMulCombine - Optimize a single multiply with constant into two
9146 /// in order to implement it with two cheaper instructions, e.g.
9147 /// LEA + SHL, LEA + LEA.
9148 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9149 TargetLowering::DAGCombinerInfo &DCI) {
9150 if (DAG.getMachineFunction().
9151 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
9154 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9157 EVT VT = N->getValueType(0);
9161 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9164 uint64_t MulAmt = C->getZExtValue();
9165 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9168 uint64_t MulAmt1 = 0;
9169 uint64_t MulAmt2 = 0;
9170 if ((MulAmt % 9) == 0) {
9172 MulAmt2 = MulAmt / 9;
9173 } else if ((MulAmt % 5) == 0) {
9175 MulAmt2 = MulAmt / 5;
9176 } else if ((MulAmt % 3) == 0) {
9178 MulAmt2 = MulAmt / 3;
9181 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9182 DebugLoc DL = N->getDebugLoc();
9184 if (isPowerOf2_64(MulAmt2) &&
9185 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9186 // If second multiplifer is pow2, issue it first. We want the multiply by
9187 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9189 std::swap(MulAmt1, MulAmt2);
9192 if (isPowerOf2_64(MulAmt1))
9193 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9194 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9196 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9197 DAG.getConstant(MulAmt1, VT));
9199 if (isPowerOf2_64(MulAmt2))
9200 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9201 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9203 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9204 DAG.getConstant(MulAmt2, VT));
9206 // Do not add new nodes to DAG combiner worklist.
9207 DCI.CombineTo(N, NewMul, false);
9212 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9213 SDValue N0 = N->getOperand(0);
9214 SDValue N1 = N->getOperand(1);
9215 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9216 EVT VT = N0.getValueType();
9218 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9219 // since the result of setcc_c is all zero's or all ones.
9220 if (N1C && N0.getOpcode() == ISD::AND &&
9221 N0.getOperand(1).getOpcode() == ISD::Constant) {
9222 SDValue N00 = N0.getOperand(0);
9223 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9224 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9225 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9226 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9227 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9228 APInt ShAmt = N1C->getAPIntValue();
9229 Mask = Mask.shl(ShAmt);
9231 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9232 N00, DAG.getConstant(Mask, VT));
9239 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9241 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9242 const X86Subtarget *Subtarget) {
9243 EVT VT = N->getValueType(0);
9244 if (!VT.isVector() && VT.isInteger() &&
9245 N->getOpcode() == ISD::SHL)
9246 return PerformSHLCombine(N, DAG);
9248 // On X86 with SSE2 support, we can transform this to a vector shift if
9249 // all elements are shifted by the same amount. We can't do this in legalize
9250 // because the a constant vector is typically transformed to a constant pool
9251 // so we have no knowledge of the shift amount.
9252 if (!Subtarget->hasSSE2())
9255 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9258 SDValue ShAmtOp = N->getOperand(1);
9259 EVT EltVT = VT.getVectorElementType();
9260 DebugLoc DL = N->getDebugLoc();
9261 SDValue BaseShAmt = SDValue();
9262 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9263 unsigned NumElts = VT.getVectorNumElements();
9265 for (; i != NumElts; ++i) {
9266 SDValue Arg = ShAmtOp.getOperand(i);
9267 if (Arg.getOpcode() == ISD::UNDEF) continue;
9271 for (; i != NumElts; ++i) {
9272 SDValue Arg = ShAmtOp.getOperand(i);
9273 if (Arg.getOpcode() == ISD::UNDEF) continue;
9274 if (Arg != BaseShAmt) {
9278 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9279 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9280 SDValue InVec = ShAmtOp.getOperand(0);
9281 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9282 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9284 for (; i != NumElts; ++i) {
9285 SDValue Arg = InVec.getOperand(i);
9286 if (Arg.getOpcode() == ISD::UNDEF) continue;
9290 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9291 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9292 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9293 if (C->getZExtValue() == SplatIdx)
9294 BaseShAmt = InVec.getOperand(1);
9297 if (BaseShAmt.getNode() == 0)
9298 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9299 DAG.getIntPtrConstant(0));
9303 // The shift amount is an i32.
9304 if (EltVT.bitsGT(MVT::i32))
9305 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9306 else if (EltVT.bitsLT(MVT::i32))
9307 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9309 // The shift amount is identical so we can do a vector shift.
9310 SDValue ValOp = N->getOperand(0);
9311 switch (N->getOpcode()) {
9313 llvm_unreachable("Unknown shift opcode!");
9316 if (VT == MVT::v2i64)
9317 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9318 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9320 if (VT == MVT::v4i32)
9321 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9322 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9324 if (VT == MVT::v8i16)
9325 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9326 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9330 if (VT == MVT::v4i32)
9331 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9332 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9334 if (VT == MVT::v8i16)
9335 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9336 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9340 if (VT == MVT::v2i64)
9341 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9342 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9344 if (VT == MVT::v4i32)
9345 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9346 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9348 if (VT == MVT::v8i16)
9349 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9350 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9357 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9358 const X86Subtarget *Subtarget) {
9359 EVT VT = N->getValueType(0);
9360 if (VT != MVT::i64 || !Subtarget->is64Bit())
9363 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9364 SDValue N0 = N->getOperand(0);
9365 SDValue N1 = N->getOperand(1);
9366 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9368 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9371 SDValue ShAmt0 = N0.getOperand(1);
9372 if (ShAmt0.getValueType() != MVT::i8)
9374 SDValue ShAmt1 = N1.getOperand(1);
9375 if (ShAmt1.getValueType() != MVT::i8)
9377 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9378 ShAmt0 = ShAmt0.getOperand(0);
9379 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9380 ShAmt1 = ShAmt1.getOperand(0);
9382 DebugLoc DL = N->getDebugLoc();
9383 unsigned Opc = X86ISD::SHLD;
9384 SDValue Op0 = N0.getOperand(0);
9385 SDValue Op1 = N1.getOperand(0);
9386 if (ShAmt0.getOpcode() == ISD::SUB) {
9388 std::swap(Op0, Op1);
9389 std::swap(ShAmt0, ShAmt1);
9392 if (ShAmt1.getOpcode() == ISD::SUB) {
9393 SDValue Sum = ShAmt1.getOperand(0);
9394 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9395 if (SumC->getSExtValue() == 64 &&
9396 ShAmt1.getOperand(1) == ShAmt0)
9397 return DAG.getNode(Opc, DL, VT,
9399 DAG.getNode(ISD::TRUNCATE, DL,
9402 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9403 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9405 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9406 return DAG.getNode(Opc, DL, VT,
9407 N0.getOperand(0), N1.getOperand(0),
9408 DAG.getNode(ISD::TRUNCATE, DL,
9415 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9416 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9417 const X86Subtarget *Subtarget) {
9418 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9419 // the FP state in cases where an emms may be missing.
9420 // A preferable solution to the general problem is to figure out the right
9421 // places to insert EMMS. This qualifies as a quick hack.
9423 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9424 StoreSDNode *St = cast<StoreSDNode>(N);
9425 EVT VT = St->getValue().getValueType();
9426 if (VT.getSizeInBits() != 64)
9429 const Function *F = DAG.getMachineFunction().getFunction();
9430 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9431 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9432 && Subtarget->hasSSE2();
9433 if ((VT.isVector() ||
9434 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9435 isa<LoadSDNode>(St->getValue()) &&
9436 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9437 St->getChain().hasOneUse() && !St->isVolatile()) {
9438 SDNode* LdVal = St->getValue().getNode();
9440 int TokenFactorIndex = -1;
9441 SmallVector<SDValue, 8> Ops;
9442 SDNode* ChainVal = St->getChain().getNode();
9443 // Must be a store of a load. We currently handle two cases: the load
9444 // is a direct child, and it's under an intervening TokenFactor. It is
9445 // possible to dig deeper under nested TokenFactors.
9446 if (ChainVal == LdVal)
9447 Ld = cast<LoadSDNode>(St->getChain());
9448 else if (St->getValue().hasOneUse() &&
9449 ChainVal->getOpcode() == ISD::TokenFactor) {
9450 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9451 if (ChainVal->getOperand(i).getNode() == LdVal) {
9452 TokenFactorIndex = i;
9453 Ld = cast<LoadSDNode>(St->getValue());
9455 Ops.push_back(ChainVal->getOperand(i));
9459 if (!Ld || !ISD::isNormalLoad(Ld))
9462 // If this is not the MMX case, i.e. we are just turning i64 load/store
9463 // into f64 load/store, avoid the transformation if there are multiple
9464 // uses of the loaded value.
9465 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9468 DebugLoc LdDL = Ld->getDebugLoc();
9469 DebugLoc StDL = N->getDebugLoc();
9470 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9471 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9473 if (Subtarget->is64Bit() || F64IsLegal) {
9474 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9475 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9476 Ld->getBasePtr(), Ld->getSrcValue(),
9477 Ld->getSrcValueOffset(), Ld->isVolatile(),
9478 Ld->getAlignment());
9479 SDValue NewChain = NewLd.getValue(1);
9480 if (TokenFactorIndex != -1) {
9481 Ops.push_back(NewChain);
9482 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9485 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9486 St->getSrcValue(), St->getSrcValueOffset(),
9487 St->isVolatile(), St->getAlignment());
9490 // Otherwise, lower to two pairs of 32-bit loads / stores.
9491 SDValue LoAddr = Ld->getBasePtr();
9492 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9493 DAG.getConstant(4, MVT::i32));
9495 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9496 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9497 Ld->isVolatile(), Ld->getAlignment());
9498 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9499 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9501 MinAlign(Ld->getAlignment(), 4));
9503 SDValue NewChain = LoLd.getValue(1);
9504 if (TokenFactorIndex != -1) {
9505 Ops.push_back(LoLd);
9506 Ops.push_back(HiLd);
9507 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9511 LoAddr = St->getBasePtr();
9512 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9513 DAG.getConstant(4, MVT::i32));
9515 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9516 St->getSrcValue(), St->getSrcValueOffset(),
9517 St->isVolatile(), St->getAlignment());
9518 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9520 St->getSrcValueOffset() + 4,
9522 MinAlign(St->getAlignment(), 4));
9523 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9528 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9529 /// X86ISD::FXOR nodes.
9530 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9531 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9532 // F[X]OR(0.0, x) -> x
9533 // F[X]OR(x, 0.0) -> x
9534 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9535 if (C->getValueAPF().isPosZero())
9536 return N->getOperand(1);
9537 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9538 if (C->getValueAPF().isPosZero())
9539 return N->getOperand(0);
9543 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9544 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9545 // FAND(0.0, x) -> 0.0
9546 // FAND(x, 0.0) -> 0.0
9547 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9548 if (C->getValueAPF().isPosZero())
9549 return N->getOperand(0);
9550 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9551 if (C->getValueAPF().isPosZero())
9552 return N->getOperand(1);
9556 static SDValue PerformBTCombine(SDNode *N,
9558 TargetLowering::DAGCombinerInfo &DCI) {
9559 // BT ignores high bits in the bit index operand.
9560 SDValue Op1 = N->getOperand(1);
9561 if (Op1.hasOneUse()) {
9562 unsigned BitWidth = Op1.getValueSizeInBits();
9563 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9564 APInt KnownZero, KnownOne;
9565 TargetLowering::TargetLoweringOpt TLO(DAG);
9566 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9567 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9568 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9569 DCI.CommitTargetLoweringOpt(TLO);
9574 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9575 SDValue Op = N->getOperand(0);
9576 if (Op.getOpcode() == ISD::BIT_CONVERT)
9577 Op = Op.getOperand(0);
9578 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9579 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9580 VT.getVectorElementType().getSizeInBits() ==
9581 OpVT.getVectorElementType().getSizeInBits()) {
9582 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9587 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9588 // Locked instructions, in turn, have implicit fence semantics (all memory
9589 // operations are flushed before issuing the locked instruction, and the
9590 // are not buffered), so we can fold away the common pattern of
9591 // fence-atomic-fence.
9592 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9593 SDValue atomic = N->getOperand(0);
9594 switch (atomic.getOpcode()) {
9595 case ISD::ATOMIC_CMP_SWAP:
9596 case ISD::ATOMIC_SWAP:
9597 case ISD::ATOMIC_LOAD_ADD:
9598 case ISD::ATOMIC_LOAD_SUB:
9599 case ISD::ATOMIC_LOAD_AND:
9600 case ISD::ATOMIC_LOAD_OR:
9601 case ISD::ATOMIC_LOAD_XOR:
9602 case ISD::ATOMIC_LOAD_NAND:
9603 case ISD::ATOMIC_LOAD_MIN:
9604 case ISD::ATOMIC_LOAD_MAX:
9605 case ISD::ATOMIC_LOAD_UMIN:
9606 case ISD::ATOMIC_LOAD_UMAX:
9612 SDValue fence = atomic.getOperand(0);
9613 if (fence.getOpcode() != ISD::MEMBARRIER)
9616 switch (atomic.getOpcode()) {
9617 case ISD::ATOMIC_CMP_SWAP:
9618 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9619 atomic.getOperand(1), atomic.getOperand(2),
9620 atomic.getOperand(3));
9621 case ISD::ATOMIC_SWAP:
9622 case ISD::ATOMIC_LOAD_ADD:
9623 case ISD::ATOMIC_LOAD_SUB:
9624 case ISD::ATOMIC_LOAD_AND:
9625 case ISD::ATOMIC_LOAD_OR:
9626 case ISD::ATOMIC_LOAD_XOR:
9627 case ISD::ATOMIC_LOAD_NAND:
9628 case ISD::ATOMIC_LOAD_MIN:
9629 case ISD::ATOMIC_LOAD_MAX:
9630 case ISD::ATOMIC_LOAD_UMIN:
9631 case ISD::ATOMIC_LOAD_UMAX:
9632 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9633 atomic.getOperand(1), atomic.getOperand(2));
9639 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9640 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9641 // (and (i32 x86isd::setcc_carry), 1)
9642 // This eliminates the zext. This transformation is necessary because
9643 // ISD::SETCC is always legalized to i8.
9644 DebugLoc dl = N->getDebugLoc();
9645 SDValue N0 = N->getOperand(0);
9646 EVT VT = N->getValueType(0);
9647 if (N0.getOpcode() == ISD::AND &&
9649 N0.getOperand(0).hasOneUse()) {
9650 SDValue N00 = N0.getOperand(0);
9651 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9653 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9654 if (!C || C->getZExtValue() != 1)
9656 return DAG.getNode(ISD::AND, dl, VT,
9657 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9658 N00.getOperand(0), N00.getOperand(1)),
9659 DAG.getConstant(1, VT));
9665 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9666 DAGCombinerInfo &DCI) const {
9667 SelectionDAG &DAG = DCI.DAG;
9668 switch (N->getOpcode()) {
9670 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9671 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9672 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9673 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9676 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9677 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9678 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9680 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9681 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9682 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9683 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9684 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9685 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9691 //===----------------------------------------------------------------------===//
9692 // X86 Inline Assembly Support
9693 //===----------------------------------------------------------------------===//
9695 static bool LowerToBSwap(CallInst *CI) {
9696 // FIXME: this should verify that we are targetting a 486 or better. If not,
9697 // we will turn this bswap into something that will be lowered to logical ops
9698 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9699 // so don't worry about this.
9701 // Verify this is a simple bswap.
9702 if (CI->getNumOperands() != 2 ||
9703 CI->getType() != CI->getOperand(1)->getType() ||
9704 !CI->getType()->isInteger())
9707 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9708 if (!Ty || Ty->getBitWidth() % 16 != 0)
9711 // Okay, we can do this xform, do so now.
9712 const Type *Tys[] = { Ty };
9713 Module *M = CI->getParent()->getParent()->getParent();
9714 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9716 Value *Op = CI->getOperand(1);
9717 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9719 CI->replaceAllUsesWith(Op);
9720 CI->eraseFromParent();
9724 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9725 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9726 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9728 std::string AsmStr = IA->getAsmString();
9730 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9731 SmallVector<StringRef, 4> AsmPieces;
9732 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9734 switch (AsmPieces.size()) {
9735 default: return false;
9737 AsmStr = AsmPieces[0];
9739 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9742 if (AsmPieces.size() == 2 &&
9743 (AsmPieces[0] == "bswap" ||
9744 AsmPieces[0] == "bswapq" ||
9745 AsmPieces[0] == "bswapl") &&
9746 (AsmPieces[1] == "$0" ||
9747 AsmPieces[1] == "${0:q}")) {
9748 // No need to check constraints, nothing other than the equivalent of
9749 // "=r,0" would be valid here.
9750 return LowerToBSwap(CI);
9752 // rorw $$8, ${0:w} --> llvm.bswap.i16
9753 if (CI->getType()->isInteger(16) &&
9754 AsmPieces.size() == 3 &&
9755 AsmPieces[0] == "rorw" &&
9756 AsmPieces[1] == "$$8," &&
9757 AsmPieces[2] == "${0:w}" &&
9758 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9759 return LowerToBSwap(CI);
9763 if (CI->getType()->isInteger(64) &&
9764 Constraints.size() >= 2 &&
9765 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9766 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9767 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9768 SmallVector<StringRef, 4> Words;
9769 SplitString(AsmPieces[0], Words, " \t");
9770 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9772 SplitString(AsmPieces[1], Words, " \t");
9773 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9775 SplitString(AsmPieces[2], Words, " \t,");
9776 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9777 Words[2] == "%edx") {
9778 return LowerToBSwap(CI);
9790 /// getConstraintType - Given a constraint letter, return the type of
9791 /// constraint it is for this target.
9792 X86TargetLowering::ConstraintType
9793 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9794 if (Constraint.size() == 1) {
9795 switch (Constraint[0]) {
9807 return C_RegisterClass;
9815 return TargetLowering::getConstraintType(Constraint);
9818 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9819 /// with another that has more specific requirements based on the type of the
9820 /// corresponding operand.
9821 const char *X86TargetLowering::
9822 LowerXConstraint(EVT ConstraintVT) const {
9823 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9824 // 'f' like normal targets.
9825 if (ConstraintVT.isFloatingPoint()) {
9826 if (Subtarget->hasSSE2())
9828 if (Subtarget->hasSSE1())
9832 return TargetLowering::LowerXConstraint(ConstraintVT);
9835 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9836 /// vector. If it is invalid, don't add anything to Ops.
9837 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9840 std::vector<SDValue>&Ops,
9841 SelectionDAG &DAG) const {
9842 SDValue Result(0, 0);
9844 switch (Constraint) {
9847 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9848 if (C->getZExtValue() <= 31) {
9849 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9855 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9856 if (C->getZExtValue() <= 63) {
9857 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9863 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9864 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9865 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9871 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9872 if (C->getZExtValue() <= 255) {
9873 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9879 // 32-bit signed value
9880 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9881 const ConstantInt *CI = C->getConstantIntValue();
9882 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9883 C->getSExtValue())) {
9884 // Widen to 64 bits here to get it sign extended.
9885 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9888 // FIXME gcc accepts some relocatable values here too, but only in certain
9889 // memory models; it's complicated.
9894 // 32-bit unsigned value
9895 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9896 const ConstantInt *CI = C->getConstantIntValue();
9897 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9898 C->getZExtValue())) {
9899 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9903 // FIXME gcc accepts some relocatable values here too, but only in certain
9904 // memory models; it's complicated.
9908 // Literal immediates are always ok.
9909 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9910 // Widen to 64 bits here to get it sign extended.
9911 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9915 // If we are in non-pic codegen mode, we allow the address of a global (with
9916 // an optional displacement) to be used with 'i'.
9917 GlobalAddressSDNode *GA = 0;
9920 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9922 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9923 Offset += GA->getOffset();
9925 } else if (Op.getOpcode() == ISD::ADD) {
9926 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9927 Offset += C->getZExtValue();
9928 Op = Op.getOperand(0);
9931 } else if (Op.getOpcode() == ISD::SUB) {
9932 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9933 Offset += -C->getZExtValue();
9934 Op = Op.getOperand(0);
9939 // Otherwise, this isn't something we can handle, reject it.
9943 GlobalValue *GV = GA->getGlobal();
9944 // If we require an extra load to get this address, as in PIC mode, we
9946 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9947 getTargetMachine())))
9951 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9953 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9959 if (Result.getNode()) {
9960 Ops.push_back(Result);
9963 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9967 std::vector<unsigned> X86TargetLowering::
9968 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9970 if (Constraint.size() == 1) {
9971 // FIXME: not handling fp-stack yet!
9972 switch (Constraint[0]) { // GCC X86 Constraint Letters
9973 default: break; // Unknown constraint letter
9974 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9975 if (Subtarget->is64Bit()) {
9977 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9978 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9979 X86::R10D,X86::R11D,X86::R12D,
9980 X86::R13D,X86::R14D,X86::R15D,
9981 X86::EBP, X86::ESP, 0);
9982 else if (VT == MVT::i16)
9983 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9984 X86::SI, X86::DI, X86::R8W,X86::R9W,
9985 X86::R10W,X86::R11W,X86::R12W,
9986 X86::R13W,X86::R14W,X86::R15W,
9987 X86::BP, X86::SP, 0);
9988 else if (VT == MVT::i8)
9989 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9990 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9991 X86::R10B,X86::R11B,X86::R12B,
9992 X86::R13B,X86::R14B,X86::R15B,
9993 X86::BPL, X86::SPL, 0);
9995 else if (VT == MVT::i64)
9996 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9997 X86::RSI, X86::RDI, X86::R8, X86::R9,
9998 X86::R10, X86::R11, X86::R12,
9999 X86::R13, X86::R14, X86::R15,
10000 X86::RBP, X86::RSP, 0);
10004 // 32-bit fallthrough
10005 case 'Q': // Q_REGS
10006 if (VT == MVT::i32)
10007 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10008 else if (VT == MVT::i16)
10009 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10010 else if (VT == MVT::i8)
10011 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10012 else if (VT == MVT::i64)
10013 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10018 return std::vector<unsigned>();
10021 std::pair<unsigned, const TargetRegisterClass*>
10022 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10024 // First, see if this is a constraint that directly corresponds to an LLVM
10026 if (Constraint.size() == 1) {
10027 // GCC Constraint Letters
10028 switch (Constraint[0]) {
10030 case 'r': // GENERAL_REGS
10031 case 'l': // INDEX_REGS
10033 return std::make_pair(0U, X86::GR8RegisterClass);
10034 if (VT == MVT::i16)
10035 return std::make_pair(0U, X86::GR16RegisterClass);
10036 if (VT == MVT::i32 || !Subtarget->is64Bit())
10037 return std::make_pair(0U, X86::GR32RegisterClass);
10038 return std::make_pair(0U, X86::GR64RegisterClass);
10039 case 'R': // LEGACY_REGS
10041 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10042 if (VT == MVT::i16)
10043 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10044 if (VT == MVT::i32 || !Subtarget->is64Bit())
10045 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10046 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10047 case 'f': // FP Stack registers.
10048 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10049 // value to the correct fpstack register class.
10050 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10051 return std::make_pair(0U, X86::RFP32RegisterClass);
10052 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10053 return std::make_pair(0U, X86::RFP64RegisterClass);
10054 return std::make_pair(0U, X86::RFP80RegisterClass);
10055 case 'y': // MMX_REGS if MMX allowed.
10056 if (!Subtarget->hasMMX()) break;
10057 return std::make_pair(0U, X86::VR64RegisterClass);
10058 case 'Y': // SSE_REGS if SSE2 allowed
10059 if (!Subtarget->hasSSE2()) break;
10061 case 'x': // SSE_REGS if SSE1 allowed
10062 if (!Subtarget->hasSSE1()) break;
10064 switch (VT.getSimpleVT().SimpleTy) {
10066 // Scalar SSE types.
10069 return std::make_pair(0U, X86::FR32RegisterClass);
10072 return std::make_pair(0U, X86::FR64RegisterClass);
10080 return std::make_pair(0U, X86::VR128RegisterClass);
10086 // Use the default implementation in TargetLowering to convert the register
10087 // constraint into a member of a register class.
10088 std::pair<unsigned, const TargetRegisterClass*> Res;
10089 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10091 // Not found as a standard register?
10092 if (Res.second == 0) {
10093 // Map st(0) -> st(7) -> ST0
10094 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10095 tolower(Constraint[1]) == 's' &&
10096 tolower(Constraint[2]) == 't' &&
10097 Constraint[3] == '(' &&
10098 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10099 Constraint[5] == ')' &&
10100 Constraint[6] == '}') {
10102 Res.first = X86::ST0+Constraint[4]-'0';
10103 Res.second = X86::RFP80RegisterClass;
10107 // GCC allows "st(0)" to be called just plain "st".
10108 if (StringRef("{st}").equals_lower(Constraint)) {
10109 Res.first = X86::ST0;
10110 Res.second = X86::RFP80RegisterClass;
10115 if (StringRef("{flags}").equals_lower(Constraint)) {
10116 Res.first = X86::EFLAGS;
10117 Res.second = X86::CCRRegisterClass;
10121 // 'A' means EAX + EDX.
10122 if (Constraint == "A") {
10123 Res.first = X86::EAX;
10124 Res.second = X86::GR32_ADRegisterClass;
10130 // Otherwise, check to see if this is a register class of the wrong value
10131 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10132 // turn into {ax},{dx}.
10133 if (Res.second->hasType(VT))
10134 return Res; // Correct type already, nothing to do.
10136 // All of the single-register GCC register classes map their values onto
10137 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10138 // really want an 8-bit or 32-bit register, map to the appropriate register
10139 // class and return the appropriate register.
10140 if (Res.second == X86::GR16RegisterClass) {
10141 if (VT == MVT::i8) {
10142 unsigned DestReg = 0;
10143 switch (Res.first) {
10145 case X86::AX: DestReg = X86::AL; break;
10146 case X86::DX: DestReg = X86::DL; break;
10147 case X86::CX: DestReg = X86::CL; break;
10148 case X86::BX: DestReg = X86::BL; break;
10151 Res.first = DestReg;
10152 Res.second = X86::GR8RegisterClass;
10154 } else if (VT == MVT::i32) {
10155 unsigned DestReg = 0;
10156 switch (Res.first) {
10158 case X86::AX: DestReg = X86::EAX; break;
10159 case X86::DX: DestReg = X86::EDX; break;
10160 case X86::CX: DestReg = X86::ECX; break;
10161 case X86::BX: DestReg = X86::EBX; break;
10162 case X86::SI: DestReg = X86::ESI; break;
10163 case X86::DI: DestReg = X86::EDI; break;
10164 case X86::BP: DestReg = X86::EBP; break;
10165 case X86::SP: DestReg = X86::ESP; break;
10168 Res.first = DestReg;
10169 Res.second = X86::GR32RegisterClass;
10171 } else if (VT == MVT::i64) {
10172 unsigned DestReg = 0;
10173 switch (Res.first) {
10175 case X86::AX: DestReg = X86::RAX; break;
10176 case X86::DX: DestReg = X86::RDX; break;
10177 case X86::CX: DestReg = X86::RCX; break;
10178 case X86::BX: DestReg = X86::RBX; break;
10179 case X86::SI: DestReg = X86::RSI; break;
10180 case X86::DI: DestReg = X86::RDI; break;
10181 case X86::BP: DestReg = X86::RBP; break;
10182 case X86::SP: DestReg = X86::RSP; break;
10185 Res.first = DestReg;
10186 Res.second = X86::GR64RegisterClass;
10189 } else if (Res.second == X86::FR32RegisterClass ||
10190 Res.second == X86::FR64RegisterClass ||
10191 Res.second == X86::VR128RegisterClass) {
10192 // Handle references to XMM physical registers that got mapped into the
10193 // wrong class. This can happen with constraints like {xmm0} where the
10194 // target independent register mapper will just pick the first match it can
10195 // find, ignoring the required type.
10196 if (VT == MVT::f32)
10197 Res.second = X86::FR32RegisterClass;
10198 else if (VT == MVT::f64)
10199 Res.second = X86::FR64RegisterClass;
10200 else if (X86::VR128RegisterClass->hasType(VT))
10201 Res.second = X86::VR128RegisterClass;
10207 //===----------------------------------------------------------------------===//
10208 // X86 Widen vector type
10209 //===----------------------------------------------------------------------===//
10211 /// getWidenVectorType: given a vector type, returns the type to widen
10212 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10213 /// If there is no vector type that we want to widen to, returns MVT::Other
10214 /// When and where to widen is target dependent based on the cost of
10215 /// scalarizing vs using the wider vector type.
10217 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10218 assert(VT.isVector());
10219 if (isTypeLegal(VT))
10222 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10223 // type based on element type. This would speed up our search (though
10224 // it may not be worth it since the size of the list is relatively
10226 EVT EltVT = VT.getVectorElementType();
10227 unsigned NElts = VT.getVectorNumElements();
10229 // On X86, it make sense to widen any vector wider than 1
10233 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10234 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10235 EVT SVT = (MVT::SimpleValueType)nVT;
10237 if (isTypeLegal(SVT) &&
10238 SVT.getVectorElementType() == EltVT &&
10239 SVT.getVectorNumElements() > NElts)