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 "X86MCTargetExpr.h"
20 #include "X86TargetMachine.h"
21 #include "X86TargetObjectFile.h"
22 #include "llvm/CallingConv.h"
23 #include "llvm/Constants.h"
24 #include "llvm/DerivedTypes.h"
25 #include "llvm/GlobalAlias.h"
26 #include "llvm/GlobalVariable.h"
27 #include "llvm/Function.h"
28 #include "llvm/Instructions.h"
29 #include "llvm/Intrinsics.h"
30 #include "llvm/LLVMContext.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineJumpTableInfo.h"
35 #include "llvm/CodeGen/MachineModuleInfo.h"
36 #include "llvm/CodeGen/MachineRegisterInfo.h"
37 #include "llvm/CodeGen/PseudoSourceValue.h"
38 #include "llvm/MC/MCAsmInfo.h"
39 #include "llvm/MC/MCContext.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 // FIXME: These should be based on subtarget info. Plus, the values should
1005 // be smaller when we are in optimizing for size mode.
1006 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1007 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
1008 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
1009 setPrefLoopAlignment(16);
1010 benefitFromCodePlacementOpt = true;
1014 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
1019 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
1020 /// the desired ByVal argument alignment.
1021 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1024 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1025 if (VTy->getBitWidth() == 128)
1027 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1028 unsigned EltAlign = 0;
1029 getMaxByValAlign(ATy->getElementType(), EltAlign);
1030 if (EltAlign > MaxAlign)
1031 MaxAlign = EltAlign;
1032 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1033 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1034 unsigned EltAlign = 0;
1035 getMaxByValAlign(STy->getElementType(i), EltAlign);
1036 if (EltAlign > MaxAlign)
1037 MaxAlign = EltAlign;
1045 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1046 /// function arguments in the caller parameter area. For X86, aggregates
1047 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1048 /// are at 4-byte boundaries.
1049 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1050 if (Subtarget->is64Bit()) {
1051 // Max of 8 and alignment of type.
1052 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1059 if (Subtarget->hasSSE1())
1060 getMaxByValAlign(Ty, Align);
1064 /// getOptimalMemOpType - Returns the target specific optimal type for load
1065 /// and store operations as a result of memset, memcpy, and memmove
1066 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1069 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1070 bool isSrcConst, bool isSrcStr,
1071 SelectionDAG &DAG) const {
1072 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1073 // linux. This is because the stack realignment code can't handle certain
1074 // cases like PR2962. This should be removed when PR2962 is fixed.
1075 const Function *F = DAG.getMachineFunction().getFunction();
1076 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1077 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1078 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1080 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1083 if (Subtarget->is64Bit() && Size >= 8)
1088 /// getJumpTableEncoding - Return the entry encoding for a jump table in the
1089 /// current function. The returned value is a member of the
1090 /// MachineJumpTableInfo::JTEntryKind enum.
1091 unsigned X86TargetLowering::getJumpTableEncoding() const {
1092 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1094 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1095 Subtarget->isPICStyleGOT())
1096 return MachineJumpTableInfo::EK_Custom32;
1098 // Otherwise, use the normal jump table encoding heuristics.
1099 return TargetLowering::getJumpTableEncoding();
1102 /// getPICBaseSymbol - Return the X86-32 PIC base.
1104 X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
1105 MCContext &Ctx) const {
1106 const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
1107 return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
1108 Twine(MF->getFunctionNumber())+"$pb");
1113 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1114 const MachineBasicBlock *MBB,
1115 unsigned uid,MCContext &Ctx) const{
1116 assert(getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1117 Subtarget->isPICStyleGOT());
1118 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1120 return X86MCTargetExpr::Create(MBB->getSymbol(Ctx),
1121 X86MCTargetExpr::GOTOFF, Ctx);
1124 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1126 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1127 SelectionDAG &DAG) const {
1128 if (!Subtarget->is64Bit())
1129 // This doesn't have DebugLoc associated with it, but is not really the
1130 // same as a Register.
1131 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1136 /// getPICJumpTableRelocBaseExpr - This returns the relocation base for the
1137 /// given PIC jumptable, the same as getPICJumpTableRelocBase, but as an
1139 const MCExpr *X86TargetLowering::
1140 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1141 MCContext &Ctx) const {
1142 // X86-64 uses RIP relative addressing based on the jump table label.
1143 if (Subtarget->isPICStyleRIPRel())
1144 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1146 // Otherwise, the reference is relative to the PIC base.
1147 return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
1150 /// getFunctionAlignment - Return the Log2 alignment of this function.
1151 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1152 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1155 //===----------------------------------------------------------------------===//
1156 // Return Value Calling Convention Implementation
1157 //===----------------------------------------------------------------------===//
1159 #include "X86GenCallingConv.inc"
1162 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1163 const SmallVectorImpl<EVT> &OutTys,
1164 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1165 SelectionDAG &DAG) {
1166 SmallVector<CCValAssign, 16> RVLocs;
1167 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1168 RVLocs, *DAG.getContext());
1169 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1173 X86TargetLowering::LowerReturn(SDValue Chain,
1174 CallingConv::ID CallConv, bool isVarArg,
1175 const SmallVectorImpl<ISD::OutputArg> &Outs,
1176 DebugLoc dl, SelectionDAG &DAG) {
1178 SmallVector<CCValAssign, 16> RVLocs;
1179 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1180 RVLocs, *DAG.getContext());
1181 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1183 // Add the regs to the liveout set for the function.
1184 MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
1185 for (unsigned i = 0; i != RVLocs.size(); ++i)
1186 if (RVLocs[i].isRegLoc() && !MRI.isLiveOut(RVLocs[i].getLocReg()))
1187 MRI.addLiveOut(RVLocs[i].getLocReg());
1191 SmallVector<SDValue, 6> RetOps;
1192 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1193 // Operand #1 = Bytes To Pop
1194 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1196 // Copy the result values into the output registers.
1197 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1198 CCValAssign &VA = RVLocs[i];
1199 assert(VA.isRegLoc() && "Can only return in registers!");
1200 SDValue ValToCopy = Outs[i].Val;
1202 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1203 // the RET instruction and handled by the FP Stackifier.
1204 if (VA.getLocReg() == X86::ST0 ||
1205 VA.getLocReg() == X86::ST1) {
1206 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1207 // change the value to the FP stack register class.
1208 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1209 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1210 RetOps.push_back(ValToCopy);
1211 // Don't emit a copytoreg.
1215 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1216 // which is returned in RAX / RDX.
1217 if (Subtarget->is64Bit()) {
1218 EVT ValVT = ValToCopy.getValueType();
1219 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1220 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1221 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1222 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1226 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1227 Flag = Chain.getValue(1);
1230 // The x86-64 ABI for returning structs by value requires that we copy
1231 // the sret argument into %rax for the return. We saved the argument into
1232 // a virtual register in the entry block, so now we copy the value out
1234 if (Subtarget->is64Bit() &&
1235 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1236 MachineFunction &MF = DAG.getMachineFunction();
1237 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1238 unsigned Reg = FuncInfo->getSRetReturnReg();
1240 Reg = MRI.createVirtualRegister(getRegClassFor(MVT::i64));
1241 FuncInfo->setSRetReturnReg(Reg);
1243 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1245 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1246 Flag = Chain.getValue(1);
1248 // RAX now acts like a return value.
1249 MRI.addLiveOut(X86::RAX);
1252 RetOps[0] = Chain; // Update chain.
1254 // Add the flag if we have it.
1256 RetOps.push_back(Flag);
1258 return DAG.getNode(X86ISD::RET_FLAG, dl,
1259 MVT::Other, &RetOps[0], RetOps.size());
1262 /// LowerCallResult - Lower the result values of a call into the
1263 /// appropriate copies out of appropriate physical registers.
1266 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1267 CallingConv::ID CallConv, bool isVarArg,
1268 const SmallVectorImpl<ISD::InputArg> &Ins,
1269 DebugLoc dl, SelectionDAG &DAG,
1270 SmallVectorImpl<SDValue> &InVals) {
1272 // Assign locations to each value returned by this call.
1273 SmallVector<CCValAssign, 16> RVLocs;
1274 bool Is64Bit = Subtarget->is64Bit();
1275 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1276 RVLocs, *DAG.getContext());
1277 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1279 // Copy all of the result registers out of their specified physreg.
1280 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1281 CCValAssign &VA = RVLocs[i];
1282 EVT CopyVT = VA.getValVT();
1284 // If this is x86-64, and we disabled SSE, we can't return FP values
1285 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1286 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1287 llvm_report_error("SSE register return with SSE disabled");
1290 // If this is a call to a function that returns an fp value on the floating
1291 // point stack, but where we prefer to use the value in xmm registers, copy
1292 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1293 if ((VA.getLocReg() == X86::ST0 ||
1294 VA.getLocReg() == X86::ST1) &&
1295 isScalarFPTypeInSSEReg(VA.getValVT())) {
1300 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1301 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1302 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1303 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1304 MVT::v2i64, InFlag).getValue(1);
1305 Val = Chain.getValue(0);
1306 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1307 Val, DAG.getConstant(0, MVT::i64));
1309 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1310 MVT::i64, InFlag).getValue(1);
1311 Val = Chain.getValue(0);
1313 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1315 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1316 CopyVT, InFlag).getValue(1);
1317 Val = Chain.getValue(0);
1319 InFlag = Chain.getValue(2);
1321 if (CopyVT != VA.getValVT()) {
1322 // Round the F80 the right size, which also moves to the appropriate xmm
1324 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1325 // This truncation won't change the value.
1326 DAG.getIntPtrConstant(1));
1329 InVals.push_back(Val);
1336 //===----------------------------------------------------------------------===//
1337 // C & StdCall & Fast Calling Convention implementation
1338 //===----------------------------------------------------------------------===//
1339 // StdCall calling convention seems to be standard for many Windows' API
1340 // routines and around. It differs from C calling convention just a little:
1341 // callee should clean up the stack, not caller. Symbols should be also
1342 // decorated in some fancy way :) It doesn't support any vector arguments.
1343 // For info on fast calling convention see Fast Calling Convention (tail call)
1344 // implementation LowerX86_32FastCCCallTo.
1346 /// CallIsStructReturn - Determines whether a call uses struct return
1348 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1352 return Outs[0].Flags.isSRet();
1355 /// ArgsAreStructReturn - Determines whether a function uses struct
1356 /// return semantics.
1358 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1362 return Ins[0].Flags.isSRet();
1365 /// IsCalleePop - Determines whether the callee is required to pop its
1366 /// own arguments. Callee pop is necessary to support tail calls.
1367 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1371 switch (CallingConv) {
1374 case CallingConv::X86_StdCall:
1375 return !Subtarget->is64Bit();
1376 case CallingConv::X86_FastCall:
1377 return !Subtarget->is64Bit();
1378 case CallingConv::Fast:
1379 return GuaranteedTailCallOpt;
1383 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1384 /// given CallingConvention value.
1385 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1386 if (Subtarget->is64Bit()) {
1387 if (Subtarget->isTargetWin64())
1388 return CC_X86_Win64_C;
1393 if (CC == CallingConv::X86_FastCall)
1394 return CC_X86_32_FastCall;
1395 else if (CC == CallingConv::Fast)
1396 return CC_X86_32_FastCC;
1401 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1402 /// by "Src" to address "Dst" with size and alignment information specified by
1403 /// the specific parameter attribute. The copy will be passed as a byval
1404 /// function parameter.
1406 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1407 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1409 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1410 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1411 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1414 /// FuncIsMadeTailCallSafe - Return true if the function is being made into
1415 /// a tailcall target by changing its ABI.
1416 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC) {
1417 return GuaranteedTailCallOpt && CC == CallingConv::Fast;
1421 X86TargetLowering::LowerMemArgument(SDValue Chain,
1422 CallingConv::ID CallConv,
1423 const SmallVectorImpl<ISD::InputArg> &Ins,
1424 DebugLoc dl, SelectionDAG &DAG,
1425 const CCValAssign &VA,
1426 MachineFrameInfo *MFI,
1428 // Create the nodes corresponding to a load from this parameter slot.
1429 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1430 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(CallConv);
1431 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1434 // If value is passed by pointer we have address passed instead of the value
1436 if (VA.getLocInfo() == CCValAssign::Indirect)
1437 ValVT = VA.getLocVT();
1439 ValVT = VA.getValVT();
1441 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1442 // changed with more analysis.
1443 // In case of tail call optimization mark all arguments mutable. Since they
1444 // could be overwritten by lowering of arguments in case of a tail call.
1445 if (Flags.isByVal()) {
1446 int FI = MFI->CreateFixedObject(Flags.getByValSize(),
1447 VA.getLocMemOffset(), isImmutable, false);
1448 return DAG.getFrameIndex(FI, getPointerTy());
1450 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1451 VA.getLocMemOffset(), isImmutable, false);
1452 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1453 return DAG.getLoad(ValVT, dl, Chain, FIN,
1454 PseudoSourceValue::getFixedStack(FI), 0,
1460 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1461 CallingConv::ID CallConv,
1463 const SmallVectorImpl<ISD::InputArg> &Ins,
1466 SmallVectorImpl<SDValue> &InVals) {
1468 MachineFunction &MF = DAG.getMachineFunction();
1469 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1471 const Function* Fn = MF.getFunction();
1472 if (Fn->hasExternalLinkage() &&
1473 Subtarget->isTargetCygMing() &&
1474 Fn->getName() == "main")
1475 FuncInfo->setForceFramePointer(true);
1477 MachineFrameInfo *MFI = MF.getFrameInfo();
1478 bool Is64Bit = Subtarget->is64Bit();
1479 bool IsWin64 = Subtarget->isTargetWin64();
1481 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1482 "Var args not supported with calling convention fastcc");
1484 // Assign locations to all of the incoming arguments.
1485 SmallVector<CCValAssign, 16> ArgLocs;
1486 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1487 ArgLocs, *DAG.getContext());
1488 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1490 unsigned LastVal = ~0U;
1492 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1493 CCValAssign &VA = ArgLocs[i];
1494 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1496 assert(VA.getValNo() != LastVal &&
1497 "Don't support value assigned to multiple locs yet");
1498 LastVal = VA.getValNo();
1500 if (VA.isRegLoc()) {
1501 EVT RegVT = VA.getLocVT();
1502 TargetRegisterClass *RC = NULL;
1503 if (RegVT == MVT::i32)
1504 RC = X86::GR32RegisterClass;
1505 else if (Is64Bit && RegVT == MVT::i64)
1506 RC = X86::GR64RegisterClass;
1507 else if (RegVT == MVT::f32)
1508 RC = X86::FR32RegisterClass;
1509 else if (RegVT == MVT::f64)
1510 RC = X86::FR64RegisterClass;
1511 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1512 RC = X86::VR128RegisterClass;
1513 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1514 RC = X86::VR64RegisterClass;
1516 llvm_unreachable("Unknown argument type!");
1518 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1519 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1521 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1522 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1524 if (VA.getLocInfo() == CCValAssign::SExt)
1525 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1526 DAG.getValueType(VA.getValVT()));
1527 else if (VA.getLocInfo() == CCValAssign::ZExt)
1528 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1529 DAG.getValueType(VA.getValVT()));
1530 else if (VA.getLocInfo() == CCValAssign::BCvt)
1531 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1533 if (VA.isExtInLoc()) {
1534 // Handle MMX values passed in XMM regs.
1535 if (RegVT.isVector()) {
1536 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1537 ArgValue, DAG.getConstant(0, MVT::i64));
1538 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1540 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1543 assert(VA.isMemLoc());
1544 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1547 // If value is passed via pointer - do a load.
1548 if (VA.getLocInfo() == CCValAssign::Indirect)
1549 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0,
1552 InVals.push_back(ArgValue);
1555 // The x86-64 ABI for returning structs by value requires that we copy
1556 // the sret argument into %rax for the return. Save the argument into
1557 // a virtual register so that we can access it from the return points.
1558 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1559 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1560 unsigned Reg = FuncInfo->getSRetReturnReg();
1562 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1563 FuncInfo->setSRetReturnReg(Reg);
1565 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1566 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1569 unsigned StackSize = CCInfo.getNextStackOffset();
1570 // Align stack specially for tail calls.
1571 if (FuncIsMadeTailCallSafe(CallConv))
1572 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1574 // If the function takes variable number of arguments, make a frame index for
1575 // the start of the first vararg value... for expansion of llvm.va_start.
1577 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1578 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1581 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1583 // FIXME: We should really autogenerate these arrays
1584 static const unsigned GPR64ArgRegsWin64[] = {
1585 X86::RCX, X86::RDX, X86::R8, X86::R9
1587 static const unsigned XMMArgRegsWin64[] = {
1588 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1590 static const unsigned GPR64ArgRegs64Bit[] = {
1591 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1593 static const unsigned XMMArgRegs64Bit[] = {
1594 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1595 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1597 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1600 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1601 GPR64ArgRegs = GPR64ArgRegsWin64;
1602 XMMArgRegs = XMMArgRegsWin64;
1604 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1605 GPR64ArgRegs = GPR64ArgRegs64Bit;
1606 XMMArgRegs = XMMArgRegs64Bit;
1608 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1610 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1613 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1614 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1615 "SSE register cannot be used when SSE is disabled!");
1616 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1617 "SSE register cannot be used when SSE is disabled!");
1618 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1619 // Kernel mode asks for SSE to be disabled, so don't push them
1621 TotalNumXMMRegs = 0;
1623 // For X86-64, if there are vararg parameters that are passed via
1624 // registers, then we must store them to their spots on the stack so they
1625 // may be loaded by deferencing the result of va_next.
1626 VarArgsGPOffset = NumIntRegs * 8;
1627 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1628 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1629 TotalNumXMMRegs * 16, 16,
1632 // Store the integer parameter registers.
1633 SmallVector<SDValue, 8> MemOps;
1634 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1635 unsigned Offset = VarArgsGPOffset;
1636 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1637 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1638 DAG.getIntPtrConstant(Offset));
1639 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1640 X86::GR64RegisterClass);
1641 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1643 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1644 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1645 Offset, false, false, 0);
1646 MemOps.push_back(Store);
1650 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1651 // Now store the XMM (fp + vector) parameter registers.
1652 SmallVector<SDValue, 11> SaveXMMOps;
1653 SaveXMMOps.push_back(Chain);
1655 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1656 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1657 SaveXMMOps.push_back(ALVal);
1659 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1660 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1662 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1663 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1664 X86::VR128RegisterClass);
1665 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1666 SaveXMMOps.push_back(Val);
1668 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1670 &SaveXMMOps[0], SaveXMMOps.size()));
1673 if (!MemOps.empty())
1674 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1675 &MemOps[0], MemOps.size());
1679 // Some CCs need callee pop.
1680 if (IsCalleePop(isVarArg, CallConv)) {
1681 BytesToPopOnReturn = StackSize; // Callee pops everything.
1683 BytesToPopOnReturn = 0; // Callee pops nothing.
1684 // If this is an sret function, the return should pop the hidden pointer.
1685 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1686 BytesToPopOnReturn = 4;
1690 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1691 if (CallConv == CallingConv::X86_FastCall)
1692 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1695 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1701 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1702 SDValue StackPtr, SDValue Arg,
1703 DebugLoc dl, SelectionDAG &DAG,
1704 const CCValAssign &VA,
1705 ISD::ArgFlagsTy Flags) {
1706 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1707 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1708 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1709 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1710 if (Flags.isByVal()) {
1711 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1713 return DAG.getStore(Chain, dl, Arg, PtrOff,
1714 PseudoSourceValue::getStack(), LocMemOffset,
1718 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1719 /// optimization is performed and it is required.
1721 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1722 SDValue &OutRetAddr, SDValue Chain,
1723 bool IsTailCall, bool Is64Bit,
1724 int FPDiff, DebugLoc dl) {
1725 // Adjust the Return address stack slot.
1726 EVT VT = getPointerTy();
1727 OutRetAddr = getReturnAddressFrameIndex(DAG);
1729 // Load the "old" Return address.
1730 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0, false, false, 0);
1731 return SDValue(OutRetAddr.getNode(), 1);
1734 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1735 /// optimization is performed and it is required (FPDiff!=0).
1737 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1738 SDValue Chain, SDValue RetAddrFrIdx,
1739 bool Is64Bit, int FPDiff, DebugLoc dl) {
1740 // Store the return address to the appropriate stack slot.
1741 if (!FPDiff) return Chain;
1742 // Calculate the new stack slot for the return address.
1743 int SlotSize = Is64Bit ? 8 : 4;
1744 int NewReturnAddrFI =
1745 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize, true,false);
1746 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1747 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1748 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1749 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0,
1755 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1756 CallingConv::ID CallConv, bool isVarArg,
1758 const SmallVectorImpl<ISD::OutputArg> &Outs,
1759 const SmallVectorImpl<ISD::InputArg> &Ins,
1760 DebugLoc dl, SelectionDAG &DAG,
1761 SmallVectorImpl<SDValue> &InVals) {
1762 MachineFunction &MF = DAG.getMachineFunction();
1763 bool Is64Bit = Subtarget->is64Bit();
1764 bool IsStructRet = CallIsStructReturn(Outs);
1765 bool IsSibcall = false;
1768 // Check if it's really possible to do a tail call.
1769 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv, isVarArg,
1772 // Sibcalls are automatically detected tailcalls which do not require
1774 if (!GuaranteedTailCallOpt && isTailCall)
1781 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1782 "Var args not supported with calling convention fastcc");
1784 // Analyze operands of the call, assigning locations to each operand.
1785 SmallVector<CCValAssign, 16> ArgLocs;
1786 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1787 ArgLocs, *DAG.getContext());
1788 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1790 // Get a count of how many bytes are to be pushed on the stack.
1791 unsigned NumBytes = CCInfo.getNextStackOffset();
1793 // This is a sibcall. The memory operands are available in caller's
1794 // own caller's stack.
1796 else if (GuaranteedTailCallOpt && CallConv == CallingConv::Fast)
1797 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1800 if (isTailCall && !IsSibcall) {
1801 // Lower arguments at fp - stackoffset + fpdiff.
1802 unsigned NumBytesCallerPushed =
1803 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1804 FPDiff = NumBytesCallerPushed - NumBytes;
1806 // Set the delta of movement of the returnaddr stackslot.
1807 // But only set if delta is greater than previous delta.
1808 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1809 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1813 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1815 SDValue RetAddrFrIdx;
1816 // Load return adress for tail calls.
1817 if (isTailCall && FPDiff)
1818 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
1819 Is64Bit, FPDiff, dl);
1821 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1822 SmallVector<SDValue, 8> MemOpChains;
1825 // Walk the register/memloc assignments, inserting copies/loads. In the case
1826 // of tail call optimization arguments are handle later.
1827 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1828 CCValAssign &VA = ArgLocs[i];
1829 EVT RegVT = VA.getLocVT();
1830 SDValue Arg = Outs[i].Val;
1831 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1832 bool isByVal = Flags.isByVal();
1834 // Promote the value if needed.
1835 switch (VA.getLocInfo()) {
1836 default: llvm_unreachable("Unknown loc info!");
1837 case CCValAssign::Full: break;
1838 case CCValAssign::SExt:
1839 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1841 case CCValAssign::ZExt:
1842 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1844 case CCValAssign::AExt:
1845 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1846 // Special case: passing MMX values in XMM registers.
1847 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1848 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1849 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1851 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1853 case CCValAssign::BCvt:
1854 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1856 case CCValAssign::Indirect: {
1857 // Store the argument.
1858 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1859 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1860 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1861 PseudoSourceValue::getFixedStack(FI), 0,
1868 if (VA.isRegLoc()) {
1869 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1870 } else if (!IsSibcall && (!isTailCall || isByVal)) {
1871 assert(VA.isMemLoc());
1872 if (StackPtr.getNode() == 0)
1873 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1874 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1875 dl, DAG, VA, Flags));
1879 if (!MemOpChains.empty())
1880 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1881 &MemOpChains[0], MemOpChains.size());
1883 // Build a sequence of copy-to-reg nodes chained together with token chain
1884 // and flag operands which copy the outgoing args into registers.
1886 // Tail call byval lowering might overwrite argument registers so in case of
1887 // tail call optimization the copies to registers are lowered later.
1889 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1890 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1891 RegsToPass[i].second, InFlag);
1892 InFlag = Chain.getValue(1);
1895 if (Subtarget->isPICStyleGOT()) {
1896 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1899 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1900 DAG.getNode(X86ISD::GlobalBaseReg,
1901 DebugLoc::getUnknownLoc(),
1904 InFlag = Chain.getValue(1);
1906 // If we are tail calling and generating PIC/GOT style code load the
1907 // address of the callee into ECX. The value in ecx is used as target of
1908 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1909 // for tail calls on PIC/GOT architectures. Normally we would just put the
1910 // address of GOT into ebx and then call target@PLT. But for tail calls
1911 // ebx would be restored (since ebx is callee saved) before jumping to the
1914 // Note: The actual moving to ECX is done further down.
1915 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1916 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1917 !G->getGlobal()->hasProtectedVisibility())
1918 Callee = LowerGlobalAddress(Callee, DAG);
1919 else if (isa<ExternalSymbolSDNode>(Callee))
1920 Callee = LowerExternalSymbol(Callee, DAG);
1924 if (Is64Bit && isVarArg) {
1925 // From AMD64 ABI document:
1926 // For calls that may call functions that use varargs or stdargs
1927 // (prototype-less calls or calls to functions containing ellipsis (...) in
1928 // the declaration) %al is used as hidden argument to specify the number
1929 // of SSE registers used. The contents of %al do not need to match exactly
1930 // the number of registers, but must be an ubound on the number of SSE
1931 // registers used and is in the range 0 - 8 inclusive.
1933 // FIXME: Verify this on Win64
1934 // Count the number of XMM registers allocated.
1935 static const unsigned XMMArgRegs[] = {
1936 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1937 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1939 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1940 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1941 && "SSE registers cannot be used when SSE is disabled");
1943 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1944 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1945 InFlag = Chain.getValue(1);
1949 // For tail calls lower the arguments to the 'real' stack slot.
1951 // Force all the incoming stack arguments to be loaded from the stack
1952 // before any new outgoing arguments are stored to the stack, because the
1953 // outgoing stack slots may alias the incoming argument stack slots, and
1954 // the alias isn't otherwise explicit. This is slightly more conservative
1955 // than necessary, because it means that each store effectively depends
1956 // on every argument instead of just those arguments it would clobber.
1957 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1959 SmallVector<SDValue, 8> MemOpChains2;
1962 // Do not flag preceeding copytoreg stuff together with the following stuff.
1964 if (GuaranteedTailCallOpt) {
1965 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1966 CCValAssign &VA = ArgLocs[i];
1969 assert(VA.isMemLoc());
1970 SDValue Arg = Outs[i].Val;
1971 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1972 // Create frame index.
1973 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1974 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1975 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1976 FIN = DAG.getFrameIndex(FI, getPointerTy());
1978 if (Flags.isByVal()) {
1979 // Copy relative to framepointer.
1980 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1981 if (StackPtr.getNode() == 0)
1982 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1984 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1986 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1990 // Store relative to framepointer.
1991 MemOpChains2.push_back(
1992 DAG.getStore(ArgChain, dl, Arg, FIN,
1993 PseudoSourceValue::getFixedStack(FI), 0,
1999 if (!MemOpChains2.empty())
2000 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
2001 &MemOpChains2[0], MemOpChains2.size());
2003 // Copy arguments to their registers.
2004 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
2005 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
2006 RegsToPass[i].second, InFlag);
2007 InFlag = Chain.getValue(1);
2011 // Store the return address to the appropriate stack slot.
2012 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
2016 bool WasGlobalOrExternal = false;
2017 if (getTargetMachine().getCodeModel() == CodeModel::Large) {
2018 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
2019 // In the 64-bit large code model, we have to make all calls
2020 // through a register, since the call instruction's 32-bit
2021 // pc-relative offset may not be large enough to hold the whole
2023 } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
2024 WasGlobalOrExternal = true;
2025 // If the callee is a GlobalAddress node (quite common, every direct call
2026 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
2029 // We should use extra load for direct calls to dllimported functions in
2031 GlobalValue *GV = G->getGlobal();
2032 if (!GV->hasDLLImportLinkage()) {
2033 unsigned char OpFlags = 0;
2035 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
2036 // external symbols most go through the PLT in PIC mode. If the symbol
2037 // has hidden or protected visibility, or if it is static or local, then
2038 // we don't need to use the PLT - we can directly call it.
2039 if (Subtarget->isTargetELF() &&
2040 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2041 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
2042 OpFlags = X86II::MO_PLT;
2043 } else if (Subtarget->isPICStyleStubAny() &&
2044 (GV->isDeclaration() || GV->isWeakForLinker()) &&
2045 Subtarget->getDarwinVers() < 9) {
2046 // PC-relative references to external symbols should go through $stub,
2047 // unless we're building with the leopard linker or later, which
2048 // automatically synthesizes these stubs.
2049 OpFlags = X86II::MO_DARWIN_STUB;
2052 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
2053 G->getOffset(), OpFlags);
2055 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
2056 WasGlobalOrExternal = true;
2057 unsigned char OpFlags = 0;
2059 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
2060 // symbols should go through the PLT.
2061 if (Subtarget->isTargetELF() &&
2062 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
2063 OpFlags = X86II::MO_PLT;
2064 } else if (Subtarget->isPICStyleStubAny() &&
2065 Subtarget->getDarwinVers() < 9) {
2066 // PC-relative references to external symbols should go through $stub,
2067 // unless we're building with the leopard linker or later, which
2068 // automatically synthesizes these stubs.
2069 OpFlags = X86II::MO_DARWIN_STUB;
2072 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
2076 if (isTailCall && !WasGlobalOrExternal) {
2077 // Force the address into a (call preserved) caller-saved register since
2078 // tailcall must happen after callee-saved registers are poped.
2079 // FIXME: Give it a special register class that contains caller-saved
2080 // register instead?
2081 unsigned TCReg = Is64Bit ? X86::R11 : X86::EAX;
2082 Chain = DAG.getCopyToReg(Chain, dl,
2083 DAG.getRegister(TCReg, getPointerTy()),
2085 Callee = DAG.getRegister(TCReg, getPointerTy());
2088 // Returns a chain & a flag for retval copy to use.
2089 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2090 SmallVector<SDValue, 8> Ops;
2092 if (!IsSibcall && isTailCall) {
2093 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2094 DAG.getIntPtrConstant(0, true), InFlag);
2095 InFlag = Chain.getValue(1);
2098 Ops.push_back(Chain);
2099 Ops.push_back(Callee);
2102 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2104 // Add argument registers to the end of the list so that they are known live
2106 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2107 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2108 RegsToPass[i].second.getValueType()));
2110 // Add an implicit use GOT pointer in EBX.
2111 if (!isTailCall && Subtarget->isPICStyleGOT())
2112 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2114 // Add an implicit use of AL for x86 vararg functions.
2115 if (Is64Bit && isVarArg)
2116 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2118 if (InFlag.getNode())
2119 Ops.push_back(InFlag);
2122 // If this is the first return lowered for this function, add the regs
2123 // to the liveout set for the function.
2124 if (MF.getRegInfo().liveout_empty()) {
2125 SmallVector<CCValAssign, 16> RVLocs;
2126 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2128 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2129 for (unsigned i = 0; i != RVLocs.size(); ++i)
2130 if (RVLocs[i].isRegLoc())
2131 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2134 assert(((Callee.getOpcode() == ISD::Register &&
2135 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2136 cast<RegisterSDNode>(Callee)->getReg() == X86::R11)) ||
2137 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2138 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2139 "Expecting a global address, external symbol, or scratch register");
2141 return DAG.getNode(X86ISD::TC_RETURN, dl,
2142 NodeTys, &Ops[0], Ops.size());
2145 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2146 InFlag = Chain.getValue(1);
2148 // Create the CALLSEQ_END node.
2149 unsigned NumBytesForCalleeToPush;
2150 if (IsCalleePop(isVarArg, CallConv))
2151 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2152 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2153 // If this is a call to a struct-return function, the callee
2154 // pops the hidden struct pointer, so we have to push it back.
2155 // This is common for Darwin/X86, Linux & Mingw32 targets.
2156 NumBytesForCalleeToPush = 4;
2158 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2160 // Returns a flag for retval copy to use.
2162 Chain = DAG.getCALLSEQ_END(Chain,
2163 DAG.getIntPtrConstant(NumBytes, true),
2164 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2167 InFlag = Chain.getValue(1);
2170 // Handle result values, copying them out of physregs into vregs that we
2172 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2173 Ins, dl, DAG, InVals);
2177 //===----------------------------------------------------------------------===//
2178 // Fast Calling Convention (tail call) implementation
2179 //===----------------------------------------------------------------------===//
2181 // Like std call, callee cleans arguments, convention except that ECX is
2182 // reserved for storing the tail called function address. Only 2 registers are
2183 // free for argument passing (inreg). Tail call optimization is performed
2185 // * tailcallopt is enabled
2186 // * caller/callee are fastcc
2187 // On X86_64 architecture with GOT-style position independent code only local
2188 // (within module) calls are supported at the moment.
2189 // To keep the stack aligned according to platform abi the function
2190 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2191 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2192 // If a tail called function callee has more arguments than the caller the
2193 // caller needs to make sure that there is room to move the RETADDR to. This is
2194 // achieved by reserving an area the size of the argument delta right after the
2195 // original REtADDR, but before the saved framepointer or the spilled registers
2196 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2208 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2209 /// for a 16 byte align requirement.
2210 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2211 SelectionDAG& DAG) {
2212 MachineFunction &MF = DAG.getMachineFunction();
2213 const TargetMachine &TM = MF.getTarget();
2214 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2215 unsigned StackAlignment = TFI.getStackAlignment();
2216 uint64_t AlignMask = StackAlignment - 1;
2217 int64_t Offset = StackSize;
2218 uint64_t SlotSize = TD->getPointerSize();
2219 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2220 // Number smaller than 12 so just add the difference.
2221 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2223 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2224 Offset = ((~AlignMask) & Offset) + StackAlignment +
2225 (StackAlignment-SlotSize);
2230 /// MatchingStackOffset - Return true if the given stack call argument is
2231 /// already available in the same position (relatively) of the caller's
2232 /// incoming argument stack.
2234 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
2235 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
2236 const X86InstrInfo *TII) {
2238 if (Arg.getOpcode() == ISD::CopyFromReg) {
2239 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
2240 if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
2242 MachineInstr *Def = MRI->getVRegDef(VR);
2245 if (!Flags.isByVal()) {
2246 if (!TII->isLoadFromStackSlot(Def, FI))
2249 unsigned Opcode = Def->getOpcode();
2250 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r) &&
2251 Def->getOperand(1).isFI()) {
2252 FI = Def->getOperand(1).getIndex();
2253 if (MFI->getObjectSize(FI) != Flags.getByValSize())
2259 LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg);
2262 SDValue Ptr = Ld->getBasePtr();
2263 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
2266 FI = FINode->getIndex();
2269 if (!MFI->isFixedObjectIndex(FI))
2271 return Offset == MFI->getObjectOffset(FI);
2274 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2275 /// for tail call optimization. Targets which want to do tail call
2276 /// optimization should implement this function.
2278 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2279 CallingConv::ID CalleeCC,
2281 const SmallVectorImpl<ISD::OutputArg> &Outs,
2282 const SmallVectorImpl<ISD::InputArg> &Ins,
2283 SelectionDAG& DAG) const {
2284 if (CalleeCC != CallingConv::Fast &&
2285 CalleeCC != CallingConv::C)
2288 // If -tailcallopt is specified, make fastcc functions tail-callable.
2289 const Function *CallerF = DAG.getMachineFunction().getFunction();
2290 if (GuaranteedTailCallOpt) {
2291 if (CalleeCC == CallingConv::Fast &&
2292 CallerF->getCallingConv() == CalleeCC)
2297 // Look for obvious safe cases to perform tail call optimization that does not
2298 // requite ABI changes. This is what gcc calls sibcall.
2300 // Do not tail call optimize vararg calls for now.
2304 // If the callee takes no arguments then go on to check the results of the
2306 if (!Outs.empty()) {
2307 // Check if stack adjustment is needed. For now, do not do this if any
2308 // argument is passed on the stack.
2309 SmallVector<CCValAssign, 16> ArgLocs;
2310 CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
2311 ArgLocs, *DAG.getContext());
2312 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
2313 if (CCInfo.getNextStackOffset()) {
2314 MachineFunction &MF = DAG.getMachineFunction();
2315 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
2317 if (Subtarget->isTargetWin64())
2318 // Win64 ABI has additional complications.
2321 // Check if the arguments are already laid out in the right way as
2322 // the caller's fixed stack objects.
2323 MachineFrameInfo *MFI = MF.getFrameInfo();
2324 const MachineRegisterInfo *MRI = &MF.getRegInfo();
2325 const X86InstrInfo *TII =
2326 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
2327 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2328 CCValAssign &VA = ArgLocs[i];
2329 EVT RegVT = VA.getLocVT();
2330 SDValue Arg = Outs[i].Val;
2331 ISD::ArgFlagsTy Flags = Outs[i].Flags;
2332 if (VA.getLocInfo() == CCValAssign::Indirect)
2334 if (!VA.isRegLoc()) {
2335 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
2347 X86TargetLowering::createFastISel(MachineFunction &mf, MachineModuleInfo *mmo,
2349 DenseMap<const Value *, unsigned> &vm,
2350 DenseMap<const BasicBlock*, MachineBasicBlock*> &bm,
2351 DenseMap<const AllocaInst *, int> &am
2353 , SmallSet<Instruction*, 8> &cil
2356 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2364 //===----------------------------------------------------------------------===//
2365 // Other Lowering Hooks
2366 //===----------------------------------------------------------------------===//
2369 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2370 MachineFunction &MF = DAG.getMachineFunction();
2371 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2372 int ReturnAddrIndex = FuncInfo->getRAIndex();
2374 if (ReturnAddrIndex == 0) {
2375 // Set up a frame object for the return address.
2376 uint64_t SlotSize = TD->getPointerSize();
2377 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2379 FuncInfo->setRAIndex(ReturnAddrIndex);
2382 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2386 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2387 bool hasSymbolicDisplacement) {
2388 // Offset should fit into 32 bit immediate field.
2389 if (!isInt32(Offset))
2392 // If we don't have a symbolic displacement - we don't have any extra
2394 if (!hasSymbolicDisplacement)
2397 // FIXME: Some tweaks might be needed for medium code model.
2398 if (M != CodeModel::Small && M != CodeModel::Kernel)
2401 // For small code model we assume that latest object is 16MB before end of 31
2402 // bits boundary. We may also accept pretty large negative constants knowing
2403 // that all objects are in the positive half of address space.
2404 if (M == CodeModel::Small && Offset < 16*1024*1024)
2407 // For kernel code model we know that all object resist in the negative half
2408 // of 32bits address space. We may not accept negative offsets, since they may
2409 // be just off and we may accept pretty large positive ones.
2410 if (M == CodeModel::Kernel && Offset > 0)
2416 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2417 /// specific condition code, returning the condition code and the LHS/RHS of the
2418 /// comparison to make.
2419 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2420 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2422 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2423 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2424 // X > -1 -> X == 0, jump !sign.
2425 RHS = DAG.getConstant(0, RHS.getValueType());
2426 return X86::COND_NS;
2427 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2428 // X < 0 -> X == 0, jump on sign.
2430 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2432 RHS = DAG.getConstant(0, RHS.getValueType());
2433 return X86::COND_LE;
2437 switch (SetCCOpcode) {
2438 default: llvm_unreachable("Invalid integer condition!");
2439 case ISD::SETEQ: return X86::COND_E;
2440 case ISD::SETGT: return X86::COND_G;
2441 case ISD::SETGE: return X86::COND_GE;
2442 case ISD::SETLT: return X86::COND_L;
2443 case ISD::SETLE: return X86::COND_LE;
2444 case ISD::SETNE: return X86::COND_NE;
2445 case ISD::SETULT: return X86::COND_B;
2446 case ISD::SETUGT: return X86::COND_A;
2447 case ISD::SETULE: return X86::COND_BE;
2448 case ISD::SETUGE: return X86::COND_AE;
2452 // First determine if it is required or is profitable to flip the operands.
2454 // If LHS is a foldable load, but RHS is not, flip the condition.
2455 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2456 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2457 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2458 std::swap(LHS, RHS);
2461 switch (SetCCOpcode) {
2467 std::swap(LHS, RHS);
2471 // On a floating point condition, the flags are set as follows:
2473 // 0 | 0 | 0 | X > Y
2474 // 0 | 0 | 1 | X < Y
2475 // 1 | 0 | 0 | X == Y
2476 // 1 | 1 | 1 | unordered
2477 switch (SetCCOpcode) {
2478 default: llvm_unreachable("Condcode should be pre-legalized away");
2480 case ISD::SETEQ: return X86::COND_E;
2481 case ISD::SETOLT: // flipped
2483 case ISD::SETGT: return X86::COND_A;
2484 case ISD::SETOLE: // flipped
2486 case ISD::SETGE: return X86::COND_AE;
2487 case ISD::SETUGT: // flipped
2489 case ISD::SETLT: return X86::COND_B;
2490 case ISD::SETUGE: // flipped
2492 case ISD::SETLE: return X86::COND_BE;
2494 case ISD::SETNE: return X86::COND_NE;
2495 case ISD::SETUO: return X86::COND_P;
2496 case ISD::SETO: return X86::COND_NP;
2498 case ISD::SETUNE: return X86::COND_INVALID;
2502 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2503 /// code. Current x86 isa includes the following FP cmov instructions:
2504 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2505 static bool hasFPCMov(unsigned X86CC) {
2521 /// isFPImmLegal - Returns true if the target can instruction select the
2522 /// specified FP immediate natively. If false, the legalizer will
2523 /// materialize the FP immediate as a load from a constant pool.
2524 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2525 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2526 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2532 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2533 /// the specified range (L, H].
2534 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2535 return (Val < 0) || (Val >= Low && Val < Hi);
2538 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2539 /// specified value.
2540 static bool isUndefOrEqual(int Val, int CmpVal) {
2541 if (Val < 0 || Val == CmpVal)
2546 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2547 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2548 /// the second operand.
2549 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2550 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2551 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2552 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2553 return (Mask[0] < 2 && Mask[1] < 2);
2557 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2558 SmallVector<int, 8> M;
2560 return ::isPSHUFDMask(M, N->getValueType(0));
2563 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2564 /// is suitable for input to PSHUFHW.
2565 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2566 if (VT != MVT::v8i16)
2569 // Lower quadword copied in order or undef.
2570 for (int i = 0; i != 4; ++i)
2571 if (Mask[i] >= 0 && Mask[i] != i)
2574 // Upper quadword shuffled.
2575 for (int i = 4; i != 8; ++i)
2576 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2582 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2583 SmallVector<int, 8> M;
2585 return ::isPSHUFHWMask(M, N->getValueType(0));
2588 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2589 /// is suitable for input to PSHUFLW.
2590 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2591 if (VT != MVT::v8i16)
2594 // Upper quadword copied in order.
2595 for (int i = 4; i != 8; ++i)
2596 if (Mask[i] >= 0 && Mask[i] != i)
2599 // Lower quadword shuffled.
2600 for (int i = 0; i != 4; ++i)
2607 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2608 SmallVector<int, 8> M;
2610 return ::isPSHUFLWMask(M, N->getValueType(0));
2613 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2614 /// is suitable for input to PALIGNR.
2615 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2617 int i, e = VT.getVectorNumElements();
2619 // Do not handle v2i64 / v2f64 shuffles with palignr.
2620 if (e < 4 || !hasSSSE3)
2623 for (i = 0; i != e; ++i)
2627 // All undef, not a palignr.
2631 // Determine if it's ok to perform a palignr with only the LHS, since we
2632 // don't have access to the actual shuffle elements to see if RHS is undef.
2633 bool Unary = Mask[i] < (int)e;
2634 bool NeedsUnary = false;
2636 int s = Mask[i] - i;
2638 // Check the rest of the elements to see if they are consecutive.
2639 for (++i; i != e; ++i) {
2644 Unary = Unary && (m < (int)e);
2645 NeedsUnary = NeedsUnary || (m < s);
2647 if (NeedsUnary && !Unary)
2649 if (Unary && m != ((s+i) & (e-1)))
2651 if (!Unary && m != (s+i))
2657 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2658 SmallVector<int, 8> M;
2660 return ::isPALIGNRMask(M, N->getValueType(0), true);
2663 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2664 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2665 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2666 int NumElems = VT.getVectorNumElements();
2667 if (NumElems != 2 && NumElems != 4)
2670 int Half = NumElems / 2;
2671 for (int i = 0; i < Half; ++i)
2672 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2674 for (int i = Half; i < NumElems; ++i)
2675 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2681 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2682 SmallVector<int, 8> M;
2684 return ::isSHUFPMask(M, N->getValueType(0));
2687 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2688 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2689 /// half elements to come from vector 1 (which would equal the dest.) and
2690 /// the upper half to come from vector 2.
2691 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2692 int NumElems = VT.getVectorNumElements();
2694 if (NumElems != 2 && NumElems != 4)
2697 int Half = NumElems / 2;
2698 for (int i = 0; i < Half; ++i)
2699 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2701 for (int i = Half; i < NumElems; ++i)
2702 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2707 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2708 SmallVector<int, 8> M;
2710 return isCommutedSHUFPMask(M, N->getValueType(0));
2713 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2714 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2715 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2716 if (N->getValueType(0).getVectorNumElements() != 4)
2719 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2720 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2721 isUndefOrEqual(N->getMaskElt(1), 7) &&
2722 isUndefOrEqual(N->getMaskElt(2), 2) &&
2723 isUndefOrEqual(N->getMaskElt(3), 3);
2726 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2727 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2729 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2730 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2735 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2736 isUndefOrEqual(N->getMaskElt(1), 3) &&
2737 isUndefOrEqual(N->getMaskElt(2), 2) &&
2738 isUndefOrEqual(N->getMaskElt(3), 3);
2741 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2742 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2743 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2744 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2746 if (NumElems != 2 && NumElems != 4)
2749 for (unsigned i = 0; i < NumElems/2; ++i)
2750 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2753 for (unsigned i = NumElems/2; i < NumElems; ++i)
2754 if (!isUndefOrEqual(N->getMaskElt(i), i))
2760 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2761 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2762 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2763 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2765 if (NumElems != 2 && NumElems != 4)
2768 for (unsigned i = 0; i < NumElems/2; ++i)
2769 if (!isUndefOrEqual(N->getMaskElt(i), i))
2772 for (unsigned i = 0; i < NumElems/2; ++i)
2773 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2779 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2780 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2781 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2782 bool V2IsSplat = false) {
2783 int NumElts = VT.getVectorNumElements();
2784 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2787 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2789 int BitI1 = Mask[i+1];
2790 if (!isUndefOrEqual(BitI, j))
2793 if (!isUndefOrEqual(BitI1, NumElts))
2796 if (!isUndefOrEqual(BitI1, j + NumElts))
2803 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2804 SmallVector<int, 8> M;
2806 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2809 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2810 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2811 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2812 bool V2IsSplat = false) {
2813 int NumElts = VT.getVectorNumElements();
2814 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2817 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2819 int BitI1 = Mask[i+1];
2820 if (!isUndefOrEqual(BitI, j + NumElts/2))
2823 if (isUndefOrEqual(BitI1, NumElts))
2826 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2833 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2834 SmallVector<int, 8> M;
2836 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2839 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2840 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2842 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2843 int NumElems = VT.getVectorNumElements();
2844 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2847 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2849 int BitI1 = Mask[i+1];
2850 if (!isUndefOrEqual(BitI, j))
2852 if (!isUndefOrEqual(BitI1, j))
2858 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2859 SmallVector<int, 8> M;
2861 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2864 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2865 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2867 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2868 int NumElems = VT.getVectorNumElements();
2869 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2872 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2874 int BitI1 = Mask[i+1];
2875 if (!isUndefOrEqual(BitI, j))
2877 if (!isUndefOrEqual(BitI1, j))
2883 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2884 SmallVector<int, 8> M;
2886 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2889 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2890 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2891 /// MOVSD, and MOVD, i.e. setting the lowest element.
2892 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2893 if (VT.getVectorElementType().getSizeInBits() < 32)
2896 int NumElts = VT.getVectorNumElements();
2898 if (!isUndefOrEqual(Mask[0], NumElts))
2901 for (int i = 1; i < NumElts; ++i)
2902 if (!isUndefOrEqual(Mask[i], i))
2908 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2909 SmallVector<int, 8> M;
2911 return ::isMOVLMask(M, N->getValueType(0));
2914 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2915 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2916 /// element of vector 2 and the other elements to come from vector 1 in order.
2917 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2918 bool V2IsSplat = false, bool V2IsUndef = false) {
2919 int NumOps = VT.getVectorNumElements();
2920 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2923 if (!isUndefOrEqual(Mask[0], 0))
2926 for (int i = 1; i < NumOps; ++i)
2927 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2928 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2929 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2935 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2936 bool V2IsUndef = false) {
2937 SmallVector<int, 8> M;
2939 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2942 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2943 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2944 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2945 if (N->getValueType(0).getVectorNumElements() != 4)
2948 // Expect 1, 1, 3, 3
2949 for (unsigned i = 0; i < 2; ++i) {
2950 int Elt = N->getMaskElt(i);
2951 if (Elt >= 0 && Elt != 1)
2956 for (unsigned i = 2; i < 4; ++i) {
2957 int Elt = N->getMaskElt(i);
2958 if (Elt >= 0 && Elt != 3)
2963 // Don't use movshdup if it can be done with a shufps.
2964 // FIXME: verify that matching u, u, 3, 3 is what we want.
2968 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2969 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2970 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2971 if (N->getValueType(0).getVectorNumElements() != 4)
2974 // Expect 0, 0, 2, 2
2975 for (unsigned i = 0; i < 2; ++i)
2976 if (N->getMaskElt(i) > 0)
2980 for (unsigned i = 2; i < 4; ++i) {
2981 int Elt = N->getMaskElt(i);
2982 if (Elt >= 0 && Elt != 2)
2987 // Don't use movsldup if it can be done with a shufps.
2991 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2992 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2993 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2994 int e = N->getValueType(0).getVectorNumElements() / 2;
2996 for (int i = 0; i < e; ++i)
2997 if (!isUndefOrEqual(N->getMaskElt(i), i))
2999 for (int i = 0; i < e; ++i)
3000 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
3005 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
3006 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
3007 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
3008 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3009 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
3011 unsigned Shift = (NumOperands == 4) ? 2 : 1;
3013 for (int i = 0; i < NumOperands; ++i) {
3014 int Val = SVOp->getMaskElt(NumOperands-i-1);
3015 if (Val < 0) Val = 0;
3016 if (Val >= NumOperands) Val -= NumOperands;
3018 if (i != NumOperands - 1)
3024 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
3025 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
3026 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
3027 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3029 // 8 nodes, but we only care about the last 4.
3030 for (unsigned i = 7; i >= 4; --i) {
3031 int Val = SVOp->getMaskElt(i);
3040 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
3041 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
3042 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
3043 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3045 // 8 nodes, but we only care about the first 4.
3046 for (int i = 3; i >= 0; --i) {
3047 int Val = SVOp->getMaskElt(i);
3056 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
3057 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
3058 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
3059 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
3060 EVT VVT = N->getValueType(0);
3061 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
3065 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
3066 Val = SVOp->getMaskElt(i);
3070 return (Val - i) * EltSize;
3073 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
3075 bool X86::isZeroNode(SDValue Elt) {
3076 return ((isa<ConstantSDNode>(Elt) &&
3077 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
3078 (isa<ConstantFPSDNode>(Elt) &&
3079 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
3082 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
3083 /// their permute mask.
3084 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
3085 SelectionDAG &DAG) {
3086 EVT VT = SVOp->getValueType(0);
3087 unsigned NumElems = VT.getVectorNumElements();
3088 SmallVector<int, 8> MaskVec;
3090 for (unsigned i = 0; i != NumElems; ++i) {
3091 int idx = SVOp->getMaskElt(i);
3093 MaskVec.push_back(idx);
3094 else if (idx < (int)NumElems)
3095 MaskVec.push_back(idx + NumElems);
3097 MaskVec.push_back(idx - NumElems);
3099 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
3100 SVOp->getOperand(0), &MaskVec[0]);
3103 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
3104 /// the two vector operands have swapped position.
3105 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
3106 unsigned NumElems = VT.getVectorNumElements();
3107 for (unsigned i = 0; i != NumElems; ++i) {
3111 else if (idx < (int)NumElems)
3112 Mask[i] = idx + NumElems;
3114 Mask[i] = idx - NumElems;
3118 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
3119 /// match movhlps. The lower half elements should come from upper half of
3120 /// V1 (and in order), and the upper half elements should come from the upper
3121 /// half of V2 (and in order).
3122 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
3123 if (Op->getValueType(0).getVectorNumElements() != 4)
3125 for (unsigned i = 0, e = 2; i != e; ++i)
3126 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
3128 for (unsigned i = 2; i != 4; ++i)
3129 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
3134 /// isScalarLoadToVector - Returns true if the node is a scalar load that
3135 /// is promoted to a vector. It also returns the LoadSDNode by reference if
3137 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
3138 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
3140 N = N->getOperand(0).getNode();
3141 if (!ISD::isNON_EXTLoad(N))
3144 *LD = cast<LoadSDNode>(N);
3148 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
3149 /// match movlp{s|d}. The lower half elements should come from lower half of
3150 /// V1 (and in order), and the upper half elements should come from the upper
3151 /// half of V2 (and in order). And since V1 will become the source of the
3152 /// MOVLP, it must be either a vector load or a scalar load to vector.
3153 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
3154 ShuffleVectorSDNode *Op) {
3155 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
3157 // Is V2 is a vector load, don't do this transformation. We will try to use
3158 // load folding shufps op.
3159 if (ISD::isNON_EXTLoad(V2))
3162 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
3164 if (NumElems != 2 && NumElems != 4)
3166 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
3167 if (!isUndefOrEqual(Op->getMaskElt(i), i))
3169 for (unsigned i = NumElems/2; i != NumElems; ++i)
3170 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
3175 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
3177 static bool isSplatVector(SDNode *N) {
3178 if (N->getOpcode() != ISD::BUILD_VECTOR)
3181 SDValue SplatValue = N->getOperand(0);
3182 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
3183 if (N->getOperand(i) != SplatValue)
3188 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
3189 /// to an zero vector.
3190 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
3191 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3192 SDValue V1 = N->getOperand(0);
3193 SDValue V2 = N->getOperand(1);
3194 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3195 for (unsigned i = 0; i != NumElems; ++i) {
3196 int Idx = N->getMaskElt(i);
3197 if (Idx >= (int)NumElems) {
3198 unsigned Opc = V2.getOpcode();
3199 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3201 if (Opc != ISD::BUILD_VECTOR ||
3202 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3204 } else if (Idx >= 0) {
3205 unsigned Opc = V1.getOpcode();
3206 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3208 if (Opc != ISD::BUILD_VECTOR ||
3209 !X86::isZeroNode(V1.getOperand(Idx)))
3216 /// getZeroVector - Returns a vector of specified type with all zero elements.
3218 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3220 assert(VT.isVector() && "Expected a vector type");
3222 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3223 // type. This ensures they get CSE'd.
3225 if (VT.getSizeInBits() == 64) { // MMX
3226 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3227 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3228 } else if (HasSSE2) { // SSE2
3229 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3230 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3232 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3233 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3235 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3238 /// getOnesVector - Returns a vector of specified type with all bits set.
3240 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3241 assert(VT.isVector() && "Expected a vector type");
3243 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3244 // type. This ensures they get CSE'd.
3245 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3247 if (VT.getSizeInBits() == 64) // MMX
3248 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3250 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3251 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3255 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3256 /// that point to V2 points to its first element.
3257 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3258 EVT VT = SVOp->getValueType(0);
3259 unsigned NumElems = VT.getVectorNumElements();
3261 bool Changed = false;
3262 SmallVector<int, 8> MaskVec;
3263 SVOp->getMask(MaskVec);
3265 for (unsigned i = 0; i != NumElems; ++i) {
3266 if (MaskVec[i] > (int)NumElems) {
3267 MaskVec[i] = NumElems;
3272 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3273 SVOp->getOperand(1), &MaskVec[0]);
3274 return SDValue(SVOp, 0);
3277 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3278 /// operation of specified width.
3279 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3281 unsigned NumElems = VT.getVectorNumElements();
3282 SmallVector<int, 8> Mask;
3283 Mask.push_back(NumElems);
3284 for (unsigned i = 1; i != NumElems; ++i)
3286 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3289 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3290 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3292 unsigned NumElems = VT.getVectorNumElements();
3293 SmallVector<int, 8> Mask;
3294 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3296 Mask.push_back(i + NumElems);
3298 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3301 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3302 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3304 unsigned NumElems = VT.getVectorNumElements();
3305 unsigned Half = NumElems/2;
3306 SmallVector<int, 8> Mask;
3307 for (unsigned i = 0; i != Half; ++i) {
3308 Mask.push_back(i + Half);
3309 Mask.push_back(i + NumElems + Half);
3311 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3314 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3315 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3317 if (SV->getValueType(0).getVectorNumElements() <= 4)
3318 return SDValue(SV, 0);
3320 EVT PVT = MVT::v4f32;
3321 EVT VT = SV->getValueType(0);
3322 DebugLoc dl = SV->getDebugLoc();
3323 SDValue V1 = SV->getOperand(0);
3324 int NumElems = VT.getVectorNumElements();
3325 int EltNo = SV->getSplatIndex();
3327 // unpack elements to the correct location
3328 while (NumElems > 4) {
3329 if (EltNo < NumElems/2) {
3330 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3332 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3333 EltNo -= NumElems/2;
3338 // Perform the splat.
3339 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3340 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3341 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3342 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3345 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3346 /// vector of zero or undef vector. This produces a shuffle where the low
3347 /// element of V2 is swizzled into the zero/undef vector, landing at element
3348 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3349 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3350 bool isZero, bool HasSSE2,
3351 SelectionDAG &DAG) {
3352 EVT VT = V2.getValueType();
3354 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3355 unsigned NumElems = VT.getVectorNumElements();
3356 SmallVector<int, 16> MaskVec;
3357 for (unsigned i = 0; i != NumElems; ++i)
3358 // If this is the insertion idx, put the low elt of V2 here.
3359 MaskVec.push_back(i == Idx ? NumElems : i);
3360 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3363 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3364 /// a shuffle that is zero.
3366 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3367 bool Low, SelectionDAG &DAG) {
3368 unsigned NumZeros = 0;
3369 for (int i = 0; i < NumElems; ++i) {
3370 unsigned Index = Low ? i : NumElems-i-1;
3371 int Idx = SVOp->getMaskElt(Index);
3376 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3377 if (Elt.getNode() && X86::isZeroNode(Elt))
3385 /// isVectorShift - Returns true if the shuffle can be implemented as a
3386 /// logical left or right shift of a vector.
3387 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3388 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3389 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3390 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3393 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3396 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3400 bool SeenV1 = false;
3401 bool SeenV2 = false;
3402 for (int i = NumZeros; i < NumElems; ++i) {
3403 int Val = isLeft ? (i - NumZeros) : i;
3404 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3416 if (SeenV1 && SeenV2)
3419 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3425 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3427 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3428 unsigned NumNonZero, unsigned NumZero,
3429 SelectionDAG &DAG, TargetLowering &TLI) {
3433 DebugLoc dl = Op.getDebugLoc();
3436 for (unsigned i = 0; i < 16; ++i) {
3437 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3438 if (ThisIsNonZero && First) {
3440 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3442 V = DAG.getUNDEF(MVT::v8i16);
3447 SDValue ThisElt(0, 0), LastElt(0, 0);
3448 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3449 if (LastIsNonZero) {
3450 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3451 MVT::i16, Op.getOperand(i-1));
3453 if (ThisIsNonZero) {
3454 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3455 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3456 ThisElt, DAG.getConstant(8, MVT::i8));
3458 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3462 if (ThisElt.getNode())
3463 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3464 DAG.getIntPtrConstant(i/2));
3468 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3471 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3473 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3474 unsigned NumNonZero, unsigned NumZero,
3475 SelectionDAG &DAG, TargetLowering &TLI) {
3479 DebugLoc dl = Op.getDebugLoc();
3482 for (unsigned i = 0; i < 8; ++i) {
3483 bool isNonZero = (NonZeros & (1 << i)) != 0;
3487 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3489 V = DAG.getUNDEF(MVT::v8i16);
3492 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3493 MVT::v8i16, V, Op.getOperand(i),
3494 DAG.getIntPtrConstant(i));
3501 /// getVShift - Return a vector logical shift node.
3503 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3504 unsigned NumBits, SelectionDAG &DAG,
3505 const TargetLowering &TLI, DebugLoc dl) {
3506 bool isMMX = VT.getSizeInBits() == 64;
3507 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3508 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3509 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3510 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3511 DAG.getNode(Opc, dl, ShVT, SrcOp,
3512 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3516 X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
3517 SelectionDAG &DAG) {
3519 // Check if the scalar load can be widened into a vector load. And if
3520 // the address is "base + cst" see if the cst can be "absorbed" into
3521 // the shuffle mask.
3522 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
3523 SDValue Ptr = LD->getBasePtr();
3524 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
3526 EVT PVT = LD->getValueType(0);
3527 if (PVT != MVT::i32 && PVT != MVT::f32)
3532 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
3533 FI = FINode->getIndex();
3535 } else if (Ptr.getOpcode() == ISD::ADD &&
3536 isa<ConstantSDNode>(Ptr.getOperand(1)) &&
3537 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
3538 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
3539 Offset = Ptr.getConstantOperandVal(1);
3540 Ptr = Ptr.getOperand(0);
3545 SDValue Chain = LD->getChain();
3546 // Make sure the stack object alignment is at least 16.
3547 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
3548 if (DAG.InferPtrAlignment(Ptr) < 16) {
3549 if (MFI->isFixedObjectIndex(FI)) {
3550 // Can't change the alignment. FIXME: It's possible to compute
3551 // the exact stack offset and reference FI + adjust offset instead.
3552 // If someone *really* cares about this. That's the way to implement it.
3555 MFI->setObjectAlignment(FI, 16);
3559 // (Offset % 16) must be multiple of 4. Then address is then
3560 // Ptr + (Offset & ~15).
3563 if ((Offset % 16) & 3)
3565 int64_t StartOffset = Offset & ~15;
3567 Ptr = DAG.getNode(ISD::ADD, Ptr.getDebugLoc(), Ptr.getValueType(),
3568 Ptr,DAG.getConstant(StartOffset, Ptr.getValueType()));
3570 int EltNo = (Offset - StartOffset) >> 2;
3571 int Mask[4] = { EltNo, EltNo, EltNo, EltNo };
3572 EVT VT = (PVT == MVT::i32) ? MVT::v4i32 : MVT::v4f32;
3573 SDValue V1 = DAG.getLoad(VT, dl, Chain, Ptr,LD->getSrcValue(),0,
3575 // Canonicalize it to a v4i32 shuffle.
3576 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
3577 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3578 DAG.getVectorShuffle(MVT::v4i32, dl, V1,
3579 DAG.getUNDEF(MVT::v4i32), &Mask[0]));
3586 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3587 DebugLoc dl = Op.getDebugLoc();
3588 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3589 if (ISD::isBuildVectorAllZeros(Op.getNode())
3590 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3591 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3592 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3593 // eliminated on x86-32 hosts.
3594 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3597 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3598 return getOnesVector(Op.getValueType(), DAG, dl);
3599 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3602 EVT VT = Op.getValueType();
3603 EVT ExtVT = VT.getVectorElementType();
3604 unsigned EVTBits = ExtVT.getSizeInBits();
3606 unsigned NumElems = Op.getNumOperands();
3607 unsigned NumZero = 0;
3608 unsigned NumNonZero = 0;
3609 unsigned NonZeros = 0;
3610 bool IsAllConstants = true;
3611 SmallSet<SDValue, 8> Values;
3612 for (unsigned i = 0; i < NumElems; ++i) {
3613 SDValue Elt = Op.getOperand(i);
3614 if (Elt.getOpcode() == ISD::UNDEF)
3617 if (Elt.getOpcode() != ISD::Constant &&
3618 Elt.getOpcode() != ISD::ConstantFP)
3619 IsAllConstants = false;
3620 if (X86::isZeroNode(Elt))
3623 NonZeros |= (1 << i);
3628 if (NumNonZero == 0) {
3629 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3630 return DAG.getUNDEF(VT);
3633 // Special case for single non-zero, non-undef, element.
3634 if (NumNonZero == 1) {
3635 unsigned Idx = CountTrailingZeros_32(NonZeros);
3636 SDValue Item = Op.getOperand(Idx);
3638 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3639 // the value are obviously zero, truncate the value to i32 and do the
3640 // insertion that way. Only do this if the value is non-constant or if the
3641 // value is a constant being inserted into element 0. It is cheaper to do
3642 // a constant pool load than it is to do a movd + shuffle.
3643 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3644 (!IsAllConstants || Idx == 0)) {
3645 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3646 // Handle MMX and SSE both.
3647 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3648 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3650 // Truncate the value (which may itself be a constant) to i32, and
3651 // convert it to a vector with movd (S2V+shuffle to zero extend).
3652 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3653 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3654 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3655 Subtarget->hasSSE2(), DAG);
3657 // Now we have our 32-bit value zero extended in the low element of
3658 // a vector. If Idx != 0, swizzle it into place.
3660 SmallVector<int, 4> Mask;
3661 Mask.push_back(Idx);
3662 for (unsigned i = 1; i != VecElts; ++i)
3664 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3665 DAG.getUNDEF(Item.getValueType()),
3668 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3672 // If we have a constant or non-constant insertion into the low element of
3673 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3674 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3675 // depending on what the source datatype is.
3678 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3679 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3680 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3681 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3682 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3683 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3685 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3686 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3687 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3688 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3689 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3690 Subtarget->hasSSE2(), DAG);
3691 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3695 // Is it a vector logical left shift?
3696 if (NumElems == 2 && Idx == 1 &&
3697 X86::isZeroNode(Op.getOperand(0)) &&
3698 !X86::isZeroNode(Op.getOperand(1))) {
3699 unsigned NumBits = VT.getSizeInBits();
3700 return getVShift(true, VT,
3701 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3702 VT, Op.getOperand(1)),
3703 NumBits/2, DAG, *this, dl);
3706 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3709 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3710 // is a non-constant being inserted into an element other than the low one,
3711 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3712 // movd/movss) to move this into the low element, then shuffle it into
3714 if (EVTBits == 32) {
3715 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3717 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3718 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3719 Subtarget->hasSSE2(), DAG);
3720 SmallVector<int, 8> MaskVec;
3721 for (unsigned i = 0; i < NumElems; i++)
3722 MaskVec.push_back(i == Idx ? 0 : 1);
3723 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3727 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3728 if (Values.size() == 1) {
3729 if (EVTBits == 32) {
3730 // Instead of a shuffle like this:
3731 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
3732 // Check if it's possible to issue this instead.
3733 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
3734 unsigned Idx = CountTrailingZeros_32(NonZeros);
3735 SDValue Item = Op.getOperand(Idx);
3736 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
3737 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
3742 // A vector full of immediates; various special cases are already
3743 // handled, so this is best done with a single constant-pool load.
3747 // Let legalizer expand 2-wide build_vectors.
3748 if (EVTBits == 64) {
3749 if (NumNonZero == 1) {
3750 // One half is zero or undef.
3751 unsigned Idx = CountTrailingZeros_32(NonZeros);
3752 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3753 Op.getOperand(Idx));
3754 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3755 Subtarget->hasSSE2(), DAG);
3760 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3761 if (EVTBits == 8 && NumElems == 16) {
3762 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3764 if (V.getNode()) return V;
3767 if (EVTBits == 16 && NumElems == 8) {
3768 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3770 if (V.getNode()) return V;
3773 // If element VT is == 32 bits, turn it into a number of shuffles.
3774 SmallVector<SDValue, 8> V;
3776 if (NumElems == 4 && NumZero > 0) {
3777 for (unsigned i = 0; i < 4; ++i) {
3778 bool isZero = !(NonZeros & (1 << i));
3780 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3782 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3785 for (unsigned i = 0; i < 2; ++i) {
3786 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3789 V[i] = V[i*2]; // Must be a zero vector.
3792 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3795 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3798 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3803 SmallVector<int, 8> MaskVec;
3804 bool Reverse = (NonZeros & 0x3) == 2;
3805 for (unsigned i = 0; i < 2; ++i)
3806 MaskVec.push_back(Reverse ? 1-i : i);
3807 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3808 for (unsigned i = 0; i < 2; ++i)
3809 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3810 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3813 if (Values.size() > 2) {
3814 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3815 // values to be inserted is equal to the number of elements, in which case
3816 // use the unpack code below in the hopes of matching the consecutive elts
3817 // load merge pattern for shuffles.
3818 // FIXME: We could probably just check that here directly.
3819 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3820 getSubtarget()->hasSSE41()) {
3821 V[0] = DAG.getUNDEF(VT);
3822 for (unsigned i = 0; i < NumElems; ++i)
3823 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3824 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3825 Op.getOperand(i), DAG.getIntPtrConstant(i));
3828 // Expand into a number of unpckl*.
3830 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3831 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3832 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3833 for (unsigned i = 0; i < NumElems; ++i)
3834 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3836 while (NumElems != 0) {
3837 for (unsigned i = 0; i < NumElems; ++i)
3838 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3848 X86TargetLowering::LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
3849 // We support concatenate two MMX registers and place them in a MMX
3850 // register. This is better than doing a stack convert.
3851 DebugLoc dl = Op.getDebugLoc();
3852 EVT ResVT = Op.getValueType();
3853 assert(Op.getNumOperands() == 2);
3854 assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
3855 ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
3857 SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
3858 SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3859 InVec = Op.getOperand(1);
3860 if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
3861 unsigned NumElts = ResVT.getVectorNumElements();
3862 VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3863 VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
3864 InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
3866 InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
3867 SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
3868 Mask[0] = 0; Mask[1] = 2;
3869 VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
3871 return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
3874 // v8i16 shuffles - Prefer shuffles in the following order:
3875 // 1. [all] pshuflw, pshufhw, optional move
3876 // 2. [ssse3] 1 x pshufb
3877 // 3. [ssse3] 2 x pshufb + 1 x por
3878 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3880 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3881 SelectionDAG &DAG, X86TargetLowering &TLI) {
3882 SDValue V1 = SVOp->getOperand(0);
3883 SDValue V2 = SVOp->getOperand(1);
3884 DebugLoc dl = SVOp->getDebugLoc();
3885 SmallVector<int, 8> MaskVals;
3887 // Determine if more than 1 of the words in each of the low and high quadwords
3888 // of the result come from the same quadword of one of the two inputs. Undef
3889 // mask values count as coming from any quadword, for better codegen.
3890 SmallVector<unsigned, 4> LoQuad(4);
3891 SmallVector<unsigned, 4> HiQuad(4);
3892 BitVector InputQuads(4);
3893 for (unsigned i = 0; i < 8; ++i) {
3894 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3895 int EltIdx = SVOp->getMaskElt(i);
3896 MaskVals.push_back(EltIdx);
3905 InputQuads.set(EltIdx / 4);
3908 int BestLoQuad = -1;
3909 unsigned MaxQuad = 1;
3910 for (unsigned i = 0; i < 4; ++i) {
3911 if (LoQuad[i] > MaxQuad) {
3913 MaxQuad = LoQuad[i];
3917 int BestHiQuad = -1;
3919 for (unsigned i = 0; i < 4; ++i) {
3920 if (HiQuad[i] > MaxQuad) {
3922 MaxQuad = HiQuad[i];
3926 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3927 // of the two input vectors, shuffle them into one input vector so only a
3928 // single pshufb instruction is necessary. If There are more than 2 input
3929 // quads, disable the next transformation since it does not help SSSE3.
3930 bool V1Used = InputQuads[0] || InputQuads[1];
3931 bool V2Used = InputQuads[2] || InputQuads[3];
3932 if (TLI.getSubtarget()->hasSSSE3()) {
3933 if (InputQuads.count() == 2 && V1Used && V2Used) {
3934 BestLoQuad = InputQuads.find_first();
3935 BestHiQuad = InputQuads.find_next(BestLoQuad);
3937 if (InputQuads.count() > 2) {
3943 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3944 // the shuffle mask. If a quad is scored as -1, that means that it contains
3945 // words from all 4 input quadwords.
3947 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3948 SmallVector<int, 8> MaskV;
3949 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3950 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3951 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3952 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3953 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3954 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3956 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3957 // source words for the shuffle, to aid later transformations.
3958 bool AllWordsInNewV = true;
3959 bool InOrder[2] = { true, true };
3960 for (unsigned i = 0; i != 8; ++i) {
3961 int idx = MaskVals[i];
3963 InOrder[i/4] = false;
3964 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3966 AllWordsInNewV = false;
3970 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3971 if (AllWordsInNewV) {
3972 for (int i = 0; i != 8; ++i) {
3973 int idx = MaskVals[i];
3976 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3977 if ((idx != i) && idx < 4)
3979 if ((idx != i) && idx > 3)
3988 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3989 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3990 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3991 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3992 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3996 // If we have SSSE3, and all words of the result are from 1 input vector,
3997 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3998 // is present, fall back to case 4.
3999 if (TLI.getSubtarget()->hasSSSE3()) {
4000 SmallVector<SDValue,16> pshufbMask;
4002 // If we have elements from both input vectors, set the high bit of the
4003 // shuffle mask element to zero out elements that come from V2 in the V1
4004 // mask, and elements that come from V1 in the V2 mask, so that the two
4005 // results can be OR'd together.
4006 bool TwoInputs = V1Used && V2Used;
4007 for (unsigned i = 0; i != 8; ++i) {
4008 int EltIdx = MaskVals[i] * 2;
4009 if (TwoInputs && (EltIdx >= 16)) {
4010 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4011 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4014 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4015 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
4017 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
4018 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4019 DAG.getNode(ISD::BUILD_VECTOR, dl,
4020 MVT::v16i8, &pshufbMask[0], 16));
4022 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4024 // Calculate the shuffle mask for the second input, shuffle it, and
4025 // OR it with the first shuffled input.
4027 for (unsigned i = 0; i != 8; ++i) {
4028 int EltIdx = MaskVals[i] * 2;
4030 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4031 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4034 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4035 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
4037 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
4038 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4039 DAG.getNode(ISD::BUILD_VECTOR, dl,
4040 MVT::v16i8, &pshufbMask[0], 16));
4041 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4042 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4045 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
4046 // and update MaskVals with new element order.
4047 BitVector InOrder(8);
4048 if (BestLoQuad >= 0) {
4049 SmallVector<int, 8> MaskV;
4050 for (int i = 0; i != 4; ++i) {
4051 int idx = MaskVals[i];
4053 MaskV.push_back(-1);
4055 } else if ((idx / 4) == BestLoQuad) {
4056 MaskV.push_back(idx & 3);
4059 MaskV.push_back(-1);
4062 for (unsigned i = 4; i != 8; ++i)
4064 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4068 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
4069 // and update MaskVals with the new element order.
4070 if (BestHiQuad >= 0) {
4071 SmallVector<int, 8> MaskV;
4072 for (unsigned i = 0; i != 4; ++i)
4074 for (unsigned i = 4; i != 8; ++i) {
4075 int idx = MaskVals[i];
4077 MaskV.push_back(-1);
4079 } else if ((idx / 4) == BestHiQuad) {
4080 MaskV.push_back((idx & 3) + 4);
4083 MaskV.push_back(-1);
4086 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
4090 // In case BestHi & BestLo were both -1, which means each quadword has a word
4091 // from each of the four input quadwords, calculate the InOrder bitvector now
4092 // before falling through to the insert/extract cleanup.
4093 if (BestLoQuad == -1 && BestHiQuad == -1) {
4095 for (int i = 0; i != 8; ++i)
4096 if (MaskVals[i] < 0 || MaskVals[i] == i)
4100 // The other elements are put in the right place using pextrw and pinsrw.
4101 for (unsigned i = 0; i != 8; ++i) {
4104 int EltIdx = MaskVals[i];
4107 SDValue ExtOp = (EltIdx < 8)
4108 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
4109 DAG.getIntPtrConstant(EltIdx))
4110 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
4111 DAG.getIntPtrConstant(EltIdx - 8));
4112 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
4113 DAG.getIntPtrConstant(i));
4118 // v16i8 shuffles - Prefer shuffles in the following order:
4119 // 1. [ssse3] 1 x pshufb
4120 // 2. [ssse3] 2 x pshufb + 1 x por
4121 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
4123 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
4124 SelectionDAG &DAG, X86TargetLowering &TLI) {
4125 SDValue V1 = SVOp->getOperand(0);
4126 SDValue V2 = SVOp->getOperand(1);
4127 DebugLoc dl = SVOp->getDebugLoc();
4128 SmallVector<int, 16> MaskVals;
4129 SVOp->getMask(MaskVals);
4131 // If we have SSSE3, case 1 is generated when all result bytes come from
4132 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
4133 // present, fall back to case 3.
4134 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
4137 for (unsigned i = 0; i < 16; ++i) {
4138 int EltIdx = MaskVals[i];
4147 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
4148 if (TLI.getSubtarget()->hasSSSE3()) {
4149 SmallVector<SDValue,16> pshufbMask;
4151 // If all result elements are from one input vector, then only translate
4152 // undef mask values to 0x80 (zero out result) in the pshufb mask.
4154 // Otherwise, we have elements from both input vectors, and must zero out
4155 // elements that come from V2 in the first mask, and V1 in the second mask
4156 // so that we can OR them together.
4157 bool TwoInputs = !(V1Only || V2Only);
4158 for (unsigned i = 0; i != 16; ++i) {
4159 int EltIdx = MaskVals[i];
4160 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
4161 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4164 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
4166 // If all the elements are from V2, assign it to V1 and return after
4167 // building the first pshufb.
4170 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
4171 DAG.getNode(ISD::BUILD_VECTOR, dl,
4172 MVT::v16i8, &pshufbMask[0], 16));
4176 // Calculate the shuffle mask for the second input, shuffle it, and
4177 // OR it with the first shuffled input.
4179 for (unsigned i = 0; i != 16; ++i) {
4180 int EltIdx = MaskVals[i];
4182 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
4185 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
4187 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
4188 DAG.getNode(ISD::BUILD_VECTOR, dl,
4189 MVT::v16i8, &pshufbMask[0], 16));
4190 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
4193 // No SSSE3 - Calculate in place words and then fix all out of place words
4194 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
4195 // the 16 different words that comprise the two doublequadword input vectors.
4196 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
4197 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
4198 SDValue NewV = V2Only ? V2 : V1;
4199 for (int i = 0; i != 8; ++i) {
4200 int Elt0 = MaskVals[i*2];
4201 int Elt1 = MaskVals[i*2+1];
4203 // This word of the result is all undef, skip it.
4204 if (Elt0 < 0 && Elt1 < 0)
4207 // This word of the result is already in the correct place, skip it.
4208 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
4210 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
4213 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
4214 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
4217 // If Elt0 and Elt1 are defined, are consecutive, and can be load
4218 // using a single extract together, load it and store it.
4219 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
4220 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4221 DAG.getIntPtrConstant(Elt1 / 2));
4222 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4223 DAG.getIntPtrConstant(i));
4227 // If Elt1 is defined, extract it from the appropriate source. If the
4228 // source byte is not also odd, shift the extracted word left 8 bits
4229 // otherwise clear the bottom 8 bits if we need to do an or.
4231 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
4232 DAG.getIntPtrConstant(Elt1 / 2));
4233 if ((Elt1 & 1) == 0)
4234 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
4235 DAG.getConstant(8, TLI.getShiftAmountTy()));
4237 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
4238 DAG.getConstant(0xFF00, MVT::i16));
4240 // If Elt0 is defined, extract it from the appropriate source. If the
4241 // source byte is not also even, shift the extracted word right 8 bits. If
4242 // Elt1 was also defined, OR the extracted values together before
4243 // inserting them in the result.
4245 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
4246 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
4247 if ((Elt0 & 1) != 0)
4248 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
4249 DAG.getConstant(8, TLI.getShiftAmountTy()));
4251 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
4252 DAG.getConstant(0x00FF, MVT::i16));
4253 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
4256 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
4257 DAG.getIntPtrConstant(i));
4259 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
4262 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
4263 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
4264 /// done when every pair / quad of shuffle mask elements point to elements in
4265 /// the right sequence. e.g.
4266 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
4268 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
4270 TargetLowering &TLI, DebugLoc dl) {
4271 EVT VT = SVOp->getValueType(0);
4272 SDValue V1 = SVOp->getOperand(0);
4273 SDValue V2 = SVOp->getOperand(1);
4274 unsigned NumElems = VT.getVectorNumElements();
4275 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
4276 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
4277 EVT MaskEltVT = MaskVT.getVectorElementType();
4279 switch (VT.getSimpleVT().SimpleTy) {
4280 default: assert(false && "Unexpected!");
4281 case MVT::v4f32: NewVT = MVT::v2f64; break;
4282 case MVT::v4i32: NewVT = MVT::v2i64; break;
4283 case MVT::v8i16: NewVT = MVT::v4i32; break;
4284 case MVT::v16i8: NewVT = MVT::v4i32; break;
4287 if (NewWidth == 2) {
4293 int Scale = NumElems / NewWidth;
4294 SmallVector<int, 8> MaskVec;
4295 for (unsigned i = 0; i < NumElems; i += Scale) {
4297 for (int j = 0; j < Scale; ++j) {
4298 int EltIdx = SVOp->getMaskElt(i+j);
4302 StartIdx = EltIdx - (EltIdx % Scale);
4303 if (EltIdx != StartIdx + j)
4307 MaskVec.push_back(-1);
4309 MaskVec.push_back(StartIdx / Scale);
4312 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4313 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4314 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4317 /// getVZextMovL - Return a zero-extending vector move low node.
4319 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4320 SDValue SrcOp, SelectionDAG &DAG,
4321 const X86Subtarget *Subtarget, DebugLoc dl) {
4322 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4323 LoadSDNode *LD = NULL;
4324 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4325 LD = dyn_cast<LoadSDNode>(SrcOp);
4327 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4329 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4330 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4331 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4332 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4333 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4335 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4336 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4337 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4338 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4346 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4347 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4348 DAG.getNode(ISD::BIT_CONVERT, dl,
4352 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4355 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4356 SDValue V1 = SVOp->getOperand(0);
4357 SDValue V2 = SVOp->getOperand(1);
4358 DebugLoc dl = SVOp->getDebugLoc();
4359 EVT VT = SVOp->getValueType(0);
4361 SmallVector<std::pair<int, int>, 8> Locs;
4363 SmallVector<int, 8> Mask1(4U, -1);
4364 SmallVector<int, 8> PermMask;
4365 SVOp->getMask(PermMask);
4369 for (unsigned i = 0; i != 4; ++i) {
4370 int Idx = PermMask[i];
4372 Locs[i] = std::make_pair(-1, -1);
4374 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4376 Locs[i] = std::make_pair(0, NumLo);
4380 Locs[i] = std::make_pair(1, NumHi);
4382 Mask1[2+NumHi] = Idx;
4388 if (NumLo <= 2 && NumHi <= 2) {
4389 // If no more than two elements come from either vector. This can be
4390 // implemented with two shuffles. First shuffle gather the elements.
4391 // The second shuffle, which takes the first shuffle as both of its
4392 // vector operands, put the elements into the right order.
4393 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4395 SmallVector<int, 8> Mask2(4U, -1);
4397 for (unsigned i = 0; i != 4; ++i) {
4398 if (Locs[i].first == -1)
4401 unsigned Idx = (i < 2) ? 0 : 4;
4402 Idx += Locs[i].first * 2 + Locs[i].second;
4407 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4408 } else if (NumLo == 3 || NumHi == 3) {
4409 // Otherwise, we must have three elements from one vector, call it X, and
4410 // one element from the other, call it Y. First, use a shufps to build an
4411 // intermediate vector with the one element from Y and the element from X
4412 // that will be in the same half in the final destination (the indexes don't
4413 // matter). Then, use a shufps to build the final vector, taking the half
4414 // containing the element from Y from the intermediate, and the other half
4417 // Normalize it so the 3 elements come from V1.
4418 CommuteVectorShuffleMask(PermMask, VT);
4422 // Find the element from V2.
4424 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4425 int Val = PermMask[HiIndex];
4432 Mask1[0] = PermMask[HiIndex];
4434 Mask1[2] = PermMask[HiIndex^1];
4436 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4439 Mask1[0] = PermMask[0];
4440 Mask1[1] = PermMask[1];
4441 Mask1[2] = HiIndex & 1 ? 6 : 4;
4442 Mask1[3] = HiIndex & 1 ? 4 : 6;
4443 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4445 Mask1[0] = HiIndex & 1 ? 2 : 0;
4446 Mask1[1] = HiIndex & 1 ? 0 : 2;
4447 Mask1[2] = PermMask[2];
4448 Mask1[3] = PermMask[3];
4453 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4457 // Break it into (shuffle shuffle_hi, shuffle_lo).
4459 SmallVector<int,8> LoMask(4U, -1);
4460 SmallVector<int,8> HiMask(4U, -1);
4462 SmallVector<int,8> *MaskPtr = &LoMask;
4463 unsigned MaskIdx = 0;
4466 for (unsigned i = 0; i != 4; ++i) {
4473 int Idx = PermMask[i];
4475 Locs[i] = std::make_pair(-1, -1);
4476 } else if (Idx < 4) {
4477 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4478 (*MaskPtr)[LoIdx] = Idx;
4481 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4482 (*MaskPtr)[HiIdx] = Idx;
4487 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4488 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4489 SmallVector<int, 8> MaskOps;
4490 for (unsigned i = 0; i != 4; ++i) {
4491 if (Locs[i].first == -1) {
4492 MaskOps.push_back(-1);
4494 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4495 MaskOps.push_back(Idx);
4498 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4502 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4503 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4504 SDValue V1 = Op.getOperand(0);
4505 SDValue V2 = Op.getOperand(1);
4506 EVT VT = Op.getValueType();
4507 DebugLoc dl = Op.getDebugLoc();
4508 unsigned NumElems = VT.getVectorNumElements();
4509 bool isMMX = VT.getSizeInBits() == 64;
4510 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4511 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4512 bool V1IsSplat = false;
4513 bool V2IsSplat = false;
4515 if (isZeroShuffle(SVOp))
4516 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4518 // Promote splats to v4f32.
4519 if (SVOp->isSplat()) {
4520 if (isMMX || NumElems < 4)
4522 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4525 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4527 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4528 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4529 if (NewOp.getNode())
4530 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4531 LowerVECTOR_SHUFFLE(NewOp, DAG));
4532 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4533 // FIXME: Figure out a cleaner way to do this.
4534 // Try to make use of movq to zero out the top part.
4535 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4536 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4537 if (NewOp.getNode()) {
4538 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4539 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4540 DAG, Subtarget, dl);
4542 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4543 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4544 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4545 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4546 DAG, Subtarget, dl);
4550 if (X86::isPSHUFDMask(SVOp))
4553 // Check if this can be converted into a logical shift.
4554 bool isLeft = false;
4557 bool isShift = getSubtarget()->hasSSE2() &&
4558 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4559 if (isShift && ShVal.hasOneUse()) {
4560 // If the shifted value has multiple uses, it may be cheaper to use
4561 // v_set0 + movlhps or movhlps, etc.
4562 EVT EltVT = VT.getVectorElementType();
4563 ShAmt *= EltVT.getSizeInBits();
4564 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4567 if (X86::isMOVLMask(SVOp)) {
4570 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4571 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4576 // FIXME: fold these into legal mask.
4577 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4578 X86::isMOVSLDUPMask(SVOp) ||
4579 X86::isMOVHLPSMask(SVOp) ||
4580 X86::isMOVLHPSMask(SVOp) ||
4581 X86::isMOVLPMask(SVOp)))
4584 if (ShouldXformToMOVHLPS(SVOp) ||
4585 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4586 return CommuteVectorShuffle(SVOp, DAG);
4589 // No better options. Use a vshl / vsrl.
4590 EVT EltVT = VT.getVectorElementType();
4591 ShAmt *= EltVT.getSizeInBits();
4592 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4595 bool Commuted = false;
4596 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4597 // 1,1,1,1 -> v8i16 though.
4598 V1IsSplat = isSplatVector(V1.getNode());
4599 V2IsSplat = isSplatVector(V2.getNode());
4601 // Canonicalize the splat or undef, if present, to be on the RHS.
4602 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4603 Op = CommuteVectorShuffle(SVOp, DAG);
4604 SVOp = cast<ShuffleVectorSDNode>(Op);
4605 V1 = SVOp->getOperand(0);
4606 V2 = SVOp->getOperand(1);
4607 std::swap(V1IsSplat, V2IsSplat);
4608 std::swap(V1IsUndef, V2IsUndef);
4612 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4613 // Shuffling low element of v1 into undef, just return v1.
4616 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4617 // the instruction selector will not match, so get a canonical MOVL with
4618 // swapped operands to undo the commute.
4619 return getMOVL(DAG, dl, VT, V2, V1);
4622 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4623 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4624 X86::isUNPCKLMask(SVOp) ||
4625 X86::isUNPCKHMask(SVOp))
4629 // Normalize mask so all entries that point to V2 points to its first
4630 // element then try to match unpck{h|l} again. If match, return a
4631 // new vector_shuffle with the corrected mask.
4632 SDValue NewMask = NormalizeMask(SVOp, DAG);
4633 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4634 if (NSVOp != SVOp) {
4635 if (X86::isUNPCKLMask(NSVOp, true)) {
4637 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4644 // Commute is back and try unpck* again.
4645 // FIXME: this seems wrong.
4646 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4647 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4648 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4649 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4650 X86::isUNPCKLMask(NewSVOp) ||
4651 X86::isUNPCKHMask(NewSVOp))
4655 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4657 // Normalize the node to match x86 shuffle ops if needed
4658 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4659 return CommuteVectorShuffle(SVOp, DAG);
4661 // Check for legal shuffle and return?
4662 SmallVector<int, 16> PermMask;
4663 SVOp->getMask(PermMask);
4664 if (isShuffleMaskLegal(PermMask, VT))
4667 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4668 if (VT == MVT::v8i16) {
4669 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4670 if (NewOp.getNode())
4674 if (VT == MVT::v16i8) {
4675 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4676 if (NewOp.getNode())
4680 // Handle all 4 wide cases with a number of shuffles except for MMX.
4681 if (NumElems == 4 && !isMMX)
4682 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4688 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4689 SelectionDAG &DAG) {
4690 EVT VT = Op.getValueType();
4691 DebugLoc dl = Op.getDebugLoc();
4692 if (VT.getSizeInBits() == 8) {
4693 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4694 Op.getOperand(0), Op.getOperand(1));
4695 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4696 DAG.getValueType(VT));
4697 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4698 } else if (VT.getSizeInBits() == 16) {
4699 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4700 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4702 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4703 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4704 DAG.getNode(ISD::BIT_CONVERT, dl,
4708 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4709 Op.getOperand(0), Op.getOperand(1));
4710 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4711 DAG.getValueType(VT));
4712 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4713 } else if (VT == MVT::f32) {
4714 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4715 // the result back to FR32 register. It's only worth matching if the
4716 // result has a single use which is a store or a bitcast to i32. And in
4717 // the case of a store, it's not worth it if the index is a constant 0,
4718 // because a MOVSSmr can be used instead, which is smaller and faster.
4719 if (!Op.hasOneUse())
4721 SDNode *User = *Op.getNode()->use_begin();
4722 if ((User->getOpcode() != ISD::STORE ||
4723 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4724 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4725 (User->getOpcode() != ISD::BIT_CONVERT ||
4726 User->getValueType(0) != MVT::i32))
4728 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4729 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4732 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4733 } else if (VT == MVT::i32) {
4734 // ExtractPS works with constant index.
4735 if (isa<ConstantSDNode>(Op.getOperand(1)))
4743 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4744 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4747 if (Subtarget->hasSSE41()) {
4748 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4753 EVT VT = Op.getValueType();
4754 DebugLoc dl = Op.getDebugLoc();
4755 // TODO: handle v16i8.
4756 if (VT.getSizeInBits() == 16) {
4757 SDValue Vec = Op.getOperand(0);
4758 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4760 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4761 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4762 DAG.getNode(ISD::BIT_CONVERT, dl,
4765 // Transform it so it match pextrw which produces a 32-bit result.
4766 EVT EltVT = MVT::i32;
4767 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4768 Op.getOperand(0), Op.getOperand(1));
4769 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4770 DAG.getValueType(VT));
4771 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4772 } else if (VT.getSizeInBits() == 32) {
4773 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4777 // SHUFPS the element to the lowest double word, then movss.
4778 int Mask[4] = { Idx, -1, -1, -1 };
4779 EVT VVT = Op.getOperand(0).getValueType();
4780 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4781 DAG.getUNDEF(VVT), Mask);
4782 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4783 DAG.getIntPtrConstant(0));
4784 } else if (VT.getSizeInBits() == 64) {
4785 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4786 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4787 // to match extract_elt for f64.
4788 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4792 // UNPCKHPD the element to the lowest double word, then movsd.
4793 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4794 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4795 int Mask[2] = { 1, -1 };
4796 EVT VVT = Op.getOperand(0).getValueType();
4797 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4798 DAG.getUNDEF(VVT), Mask);
4799 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4800 DAG.getIntPtrConstant(0));
4807 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4808 EVT VT = Op.getValueType();
4809 EVT EltVT = VT.getVectorElementType();
4810 DebugLoc dl = Op.getDebugLoc();
4812 SDValue N0 = Op.getOperand(0);
4813 SDValue N1 = Op.getOperand(1);
4814 SDValue N2 = Op.getOperand(2);
4816 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4817 isa<ConstantSDNode>(N2)) {
4818 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4820 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4822 if (N1.getValueType() != MVT::i32)
4823 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4824 if (N2.getValueType() != MVT::i32)
4825 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4826 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4827 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4828 // Bits [7:6] of the constant are the source select. This will always be
4829 // zero here. The DAG Combiner may combine an extract_elt index into these
4830 // bits. For example (insert (extract, 3), 2) could be matched by putting
4831 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4832 // Bits [5:4] of the constant are the destination select. This is the
4833 // value of the incoming immediate.
4834 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4835 // combine either bitwise AND or insert of float 0.0 to set these bits.
4836 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4837 // Create this as a scalar to vector..
4838 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4839 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4840 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4841 // PINSR* works with constant index.
4848 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4849 EVT VT = Op.getValueType();
4850 EVT EltVT = VT.getVectorElementType();
4852 if (Subtarget->hasSSE41())
4853 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4855 if (EltVT == MVT::i8)
4858 DebugLoc dl = Op.getDebugLoc();
4859 SDValue N0 = Op.getOperand(0);
4860 SDValue N1 = Op.getOperand(1);
4861 SDValue N2 = Op.getOperand(2);
4863 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4864 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4865 // as its second argument.
4866 if (N1.getValueType() != MVT::i32)
4867 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4868 if (N2.getValueType() != MVT::i32)
4869 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4870 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4876 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4877 DebugLoc dl = Op.getDebugLoc();
4878 if (Op.getValueType() == MVT::v2f32)
4879 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4880 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4881 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4882 Op.getOperand(0))));
4884 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4885 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4887 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4888 EVT VT = MVT::v2i32;
4889 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4896 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4897 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4900 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4901 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4902 // one of the above mentioned nodes. It has to be wrapped because otherwise
4903 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4904 // be used to form addressing mode. These wrapped nodes will be selected
4907 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4908 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4910 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4912 unsigned char OpFlag = 0;
4913 unsigned WrapperKind = X86ISD::Wrapper;
4914 CodeModel::Model M = getTargetMachine().getCodeModel();
4916 if (Subtarget->isPICStyleRIPRel() &&
4917 (M == CodeModel::Small || M == CodeModel::Kernel))
4918 WrapperKind = X86ISD::WrapperRIP;
4919 else if (Subtarget->isPICStyleGOT())
4920 OpFlag = X86II::MO_GOTOFF;
4921 else if (Subtarget->isPICStyleStubPIC())
4922 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4924 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4926 CP->getOffset(), OpFlag);
4927 DebugLoc DL = CP->getDebugLoc();
4928 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4929 // With PIC, the address is actually $g + Offset.
4931 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4932 DAG.getNode(X86ISD::GlobalBaseReg,
4933 DebugLoc::getUnknownLoc(), getPointerTy()),
4940 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4941 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4943 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4945 unsigned char OpFlag = 0;
4946 unsigned WrapperKind = X86ISD::Wrapper;
4947 CodeModel::Model M = getTargetMachine().getCodeModel();
4949 if (Subtarget->isPICStyleRIPRel() &&
4950 (M == CodeModel::Small || M == CodeModel::Kernel))
4951 WrapperKind = X86ISD::WrapperRIP;
4952 else if (Subtarget->isPICStyleGOT())
4953 OpFlag = X86II::MO_GOTOFF;
4954 else if (Subtarget->isPICStyleStubPIC())
4955 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4957 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4959 DebugLoc DL = JT->getDebugLoc();
4960 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4962 // With PIC, the address is actually $g + Offset.
4964 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4965 DAG.getNode(X86ISD::GlobalBaseReg,
4966 DebugLoc::getUnknownLoc(), getPointerTy()),
4974 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4975 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4977 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4979 unsigned char OpFlag = 0;
4980 unsigned WrapperKind = X86ISD::Wrapper;
4981 CodeModel::Model M = getTargetMachine().getCodeModel();
4983 if (Subtarget->isPICStyleRIPRel() &&
4984 (M == CodeModel::Small || M == CodeModel::Kernel))
4985 WrapperKind = X86ISD::WrapperRIP;
4986 else if (Subtarget->isPICStyleGOT())
4987 OpFlag = X86II::MO_GOTOFF;
4988 else if (Subtarget->isPICStyleStubPIC())
4989 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4991 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4993 DebugLoc DL = Op.getDebugLoc();
4994 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4997 // With PIC, the address is actually $g + Offset.
4998 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4999 !Subtarget->is64Bit()) {
5000 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
5001 DAG.getNode(X86ISD::GlobalBaseReg,
5002 DebugLoc::getUnknownLoc(),
5011 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
5012 // Create the TargetBlockAddressAddress node.
5013 unsigned char OpFlags =
5014 Subtarget->ClassifyBlockAddressReference();
5015 CodeModel::Model M = getTargetMachine().getCodeModel();
5016 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
5017 DebugLoc dl = Op.getDebugLoc();
5018 SDValue Result = DAG.getBlockAddress(BA, getPointerTy(),
5019 /*isTarget=*/true, OpFlags);
5021 if (Subtarget->isPICStyleRIPRel() &&
5022 (M == CodeModel::Small || M == CodeModel::Kernel))
5023 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5025 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5027 // With PIC, the address is actually $g + Offset.
5028 if (isGlobalRelativeToPICBase(OpFlags)) {
5029 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5030 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5038 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
5040 SelectionDAG &DAG) const {
5041 // Create the TargetGlobalAddress node, folding in the constant
5042 // offset if it is legal.
5043 unsigned char OpFlags =
5044 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
5045 CodeModel::Model M = getTargetMachine().getCodeModel();
5047 if (OpFlags == X86II::MO_NO_FLAG &&
5048 X86::isOffsetSuitableForCodeModel(Offset, M)) {
5049 // A direct static reference to a global.
5050 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
5053 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
5056 if (Subtarget->isPICStyleRIPRel() &&
5057 (M == CodeModel::Small || M == CodeModel::Kernel))
5058 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
5060 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
5062 // With PIC, the address is actually $g + Offset.
5063 if (isGlobalRelativeToPICBase(OpFlags)) {
5064 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
5065 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
5069 // For globals that require a load from a stub to get the address, emit the
5071 if (isGlobalStubReference(OpFlags))
5072 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
5073 PseudoSourceValue::getGOT(), 0, false, false, 0);
5075 // If there was a non-zero offset that we didn't fold, create an explicit
5078 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
5079 DAG.getConstant(Offset, getPointerTy()));
5085 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
5086 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
5087 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
5088 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
5092 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
5093 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
5094 unsigned char OperandFlags) {
5095 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5096 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5097 DebugLoc dl = GA->getDebugLoc();
5098 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
5099 GA->getValueType(0),
5103 SDValue Ops[] = { Chain, TGA, *InFlag };
5104 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
5106 SDValue Ops[] = { Chain, TGA };
5107 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
5110 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
5111 MFI->setHasCalls(true);
5113 SDValue Flag = Chain.getValue(1);
5114 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
5117 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
5119 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5122 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
5123 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
5124 DAG.getNode(X86ISD::GlobalBaseReg,
5125 DebugLoc::getUnknownLoc(),
5127 InFlag = Chain.getValue(1);
5129 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
5132 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
5134 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5136 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
5137 X86::RAX, X86II::MO_TLSGD);
5140 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
5141 // "local exec" model.
5142 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
5143 const EVT PtrVT, TLSModel::Model model,
5145 DebugLoc dl = GA->getDebugLoc();
5146 // Get the Thread Pointer
5147 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
5148 DebugLoc::getUnknownLoc(), PtrVT,
5149 DAG.getRegister(is64Bit? X86::FS : X86::GS,
5152 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
5153 NULL, 0, false, false, 0);
5155 unsigned char OperandFlags = 0;
5156 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
5158 unsigned WrapperKind = X86ISD::Wrapper;
5159 if (model == TLSModel::LocalExec) {
5160 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
5161 } else if (is64Bit) {
5162 assert(model == TLSModel::InitialExec);
5163 OperandFlags = X86II::MO_GOTTPOFF;
5164 WrapperKind = X86ISD::WrapperRIP;
5166 assert(model == TLSModel::InitialExec);
5167 OperandFlags = X86II::MO_INDNTPOFF;
5170 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
5172 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
5173 GA->getOffset(), OperandFlags);
5174 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
5176 if (model == TLSModel::InitialExec)
5177 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
5178 PseudoSourceValue::getGOT(), 0, false, false, 0);
5180 // The address of the thread local variable is the add of the thread
5181 // pointer with the offset of the variable.
5182 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
5186 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
5187 // TODO: implement the "local dynamic" model
5188 // TODO: implement the "initial exec"model for pic executables
5189 assert(Subtarget->isTargetELF() &&
5190 "TLS not implemented for non-ELF targets");
5191 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
5192 const GlobalValue *GV = GA->getGlobal();
5194 // If GV is an alias then use the aliasee for determining
5195 // thread-localness.
5196 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
5197 GV = GA->resolveAliasedGlobal(false);
5199 TLSModel::Model model = getTLSModel(GV,
5200 getTargetMachine().getRelocationModel());
5203 case TLSModel::GeneralDynamic:
5204 case TLSModel::LocalDynamic: // not implemented
5205 if (Subtarget->is64Bit())
5206 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
5207 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
5209 case TLSModel::InitialExec:
5210 case TLSModel::LocalExec:
5211 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
5212 Subtarget->is64Bit());
5215 llvm_unreachable("Unreachable");
5220 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
5221 /// take a 2 x i32 value to shift plus a shift amount.
5222 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
5223 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
5224 EVT VT = Op.getValueType();
5225 unsigned VTBits = VT.getSizeInBits();
5226 DebugLoc dl = Op.getDebugLoc();
5227 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
5228 SDValue ShOpLo = Op.getOperand(0);
5229 SDValue ShOpHi = Op.getOperand(1);
5230 SDValue ShAmt = Op.getOperand(2);
5231 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
5232 DAG.getConstant(VTBits - 1, MVT::i8))
5233 : DAG.getConstant(0, VT);
5236 if (Op.getOpcode() == ISD::SHL_PARTS) {
5237 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
5238 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
5240 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
5241 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
5244 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
5245 DAG.getConstant(VTBits, MVT::i8));
5246 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
5247 AndNode, DAG.getConstant(0, MVT::i8));
5250 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5251 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
5252 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
5254 if (Op.getOpcode() == ISD::SHL_PARTS) {
5255 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5256 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5258 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
5259 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
5262 SDValue Ops[2] = { Lo, Hi };
5263 return DAG.getMergeValues(Ops, 2, dl);
5266 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5267 EVT SrcVT = Op.getOperand(0).getValueType();
5269 if (SrcVT.isVector()) {
5270 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
5276 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
5277 "Unknown SINT_TO_FP to lower!");
5279 // These are really Legal; return the operand so the caller accepts it as
5281 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
5283 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
5284 Subtarget->is64Bit()) {
5288 DebugLoc dl = Op.getDebugLoc();
5289 unsigned Size = SrcVT.getSizeInBits()/8;
5290 MachineFunction &MF = DAG.getMachineFunction();
5291 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
5292 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5293 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5295 PseudoSourceValue::getFixedStack(SSFI), 0,
5297 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
5300 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
5302 SelectionDAG &DAG) {
5304 DebugLoc dl = Op.getDebugLoc();
5306 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
5308 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
5310 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
5311 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
5312 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5313 Tys, Ops, array_lengthof(Ops));
5316 Chain = Result.getValue(1);
5317 SDValue InFlag = Result.getValue(2);
5319 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5320 // shouldn't be necessary except that RFP cannot be live across
5321 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5322 MachineFunction &MF = DAG.getMachineFunction();
5323 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5324 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5325 Tys = DAG.getVTList(MVT::Other);
5327 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
5329 Chain = DAG.getNode(X86ISD::FST, dl, Tys, Ops, array_lengthof(Ops));
5330 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5331 PseudoSourceValue::getFixedStack(SSFI), 0,
5338 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5339 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5340 // This algorithm is not obvious. Here it is in C code, more or less:
5342 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5343 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5344 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5346 // Copy ints to xmm registers.
5347 __m128i xh = _mm_cvtsi32_si128( hi );
5348 __m128i xl = _mm_cvtsi32_si128( lo );
5350 // Combine into low half of a single xmm register.
5351 __m128i x = _mm_unpacklo_epi32( xh, xl );
5355 // Merge in appropriate exponents to give the integer bits the right
5357 x = _mm_unpacklo_epi32( x, exp );
5359 // Subtract away the biases to deal with the IEEE-754 double precision
5361 d = _mm_sub_pd( (__m128d) x, bias );
5363 // All conversions up to here are exact. The correctly rounded result is
5364 // calculated using the current rounding mode using the following
5366 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5367 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5368 // store doesn't really need to be here (except
5369 // maybe to zero the other double)
5374 DebugLoc dl = Op.getDebugLoc();
5375 LLVMContext *Context = DAG.getContext();
5377 // Build some magic constants.
5378 std::vector<Constant*> CV0;
5379 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5380 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5381 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5382 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5383 Constant *C0 = ConstantVector::get(CV0);
5384 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5386 std::vector<Constant*> CV1;
5388 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5390 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5391 Constant *C1 = ConstantVector::get(CV1);
5392 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5394 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5395 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5397 DAG.getIntPtrConstant(1)));
5398 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5399 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5401 DAG.getIntPtrConstant(0)));
5402 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5403 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5404 PseudoSourceValue::getConstantPool(), 0,
5406 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5407 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5408 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5409 PseudoSourceValue::getConstantPool(), 0,
5411 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5413 // Add the halves; easiest way is to swap them into another reg first.
5414 int ShufMask[2] = { 1, -1 };
5415 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5416 DAG.getUNDEF(MVT::v2f64), ShufMask);
5417 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5418 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5419 DAG.getIntPtrConstant(0));
5422 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5423 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5424 DebugLoc dl = Op.getDebugLoc();
5425 // FP constant to bias correct the final result.
5426 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5429 // Load the 32-bit value into an XMM register.
5430 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5431 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5433 DAG.getIntPtrConstant(0)));
5435 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5436 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5437 DAG.getIntPtrConstant(0));
5439 // Or the load with the bias.
5440 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5441 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5442 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5444 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5445 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5446 MVT::v2f64, Bias)));
5447 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5448 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5449 DAG.getIntPtrConstant(0));
5451 // Subtract the bias.
5452 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5454 // Handle final rounding.
5455 EVT DestVT = Op.getValueType();
5457 if (DestVT.bitsLT(MVT::f64)) {
5458 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5459 DAG.getIntPtrConstant(0));
5460 } else if (DestVT.bitsGT(MVT::f64)) {
5461 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5464 // Handle final rounding.
5468 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5469 SDValue N0 = Op.getOperand(0);
5470 DebugLoc dl = Op.getDebugLoc();
5472 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5473 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5474 // the optimization here.
5475 if (DAG.SignBitIsZero(N0))
5476 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5478 EVT SrcVT = N0.getValueType();
5479 if (SrcVT == MVT::i64) {
5480 // We only handle SSE2 f64 target here; caller can expand the rest.
5481 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5484 return LowerUINT_TO_FP_i64(Op, DAG);
5485 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5486 return LowerUINT_TO_FP_i32(Op, DAG);
5489 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5491 // Make a 64-bit buffer, and use it to build an FILD.
5492 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5493 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5494 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5495 getPointerTy(), StackSlot, WordOff);
5496 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5497 StackSlot, NULL, 0, false, false, 0);
5498 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5499 OffsetSlot, NULL, 0, false, false, 0);
5500 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5503 std::pair<SDValue,SDValue> X86TargetLowering::
5504 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5505 DebugLoc dl = Op.getDebugLoc();
5507 EVT DstTy = Op.getValueType();
5510 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5514 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5515 DstTy.getSimpleVT() >= MVT::i16 &&
5516 "Unknown FP_TO_SINT to lower!");
5518 // These are really Legal.
5519 if (DstTy == MVT::i32 &&
5520 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5521 return std::make_pair(SDValue(), SDValue());
5522 if (Subtarget->is64Bit() &&
5523 DstTy == MVT::i64 &&
5524 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5525 return std::make_pair(SDValue(), SDValue());
5527 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5529 MachineFunction &MF = DAG.getMachineFunction();
5530 unsigned MemSize = DstTy.getSizeInBits()/8;
5531 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5532 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5535 switch (DstTy.getSimpleVT().SimpleTy) {
5536 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5537 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5538 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5539 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5542 SDValue Chain = DAG.getEntryNode();
5543 SDValue Value = Op.getOperand(0);
5544 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5545 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5546 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5547 PseudoSourceValue::getFixedStack(SSFI), 0,
5549 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5551 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5553 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5554 Chain = Value.getValue(1);
5555 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5556 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5559 // Build the FP_TO_INT*_IN_MEM
5560 SDValue Ops[] = { Chain, Value, StackSlot };
5561 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5563 return std::make_pair(FIST, StackSlot);
5566 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5567 if (Op.getValueType().isVector()) {
5568 if (Op.getValueType() == MVT::v2i32 &&
5569 Op.getOperand(0).getValueType() == MVT::v2f64) {
5575 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5576 SDValue FIST = Vals.first, StackSlot = Vals.second;
5577 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5578 if (FIST.getNode() == 0) return Op;
5581 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5582 FIST, StackSlot, NULL, 0, false, false, 0);
5585 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5586 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5587 SDValue FIST = Vals.first, StackSlot = Vals.second;
5588 assert(FIST.getNode() && "Unexpected failure");
5591 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5592 FIST, StackSlot, NULL, 0, false, false, 0);
5595 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5596 LLVMContext *Context = DAG.getContext();
5597 DebugLoc dl = Op.getDebugLoc();
5598 EVT VT = Op.getValueType();
5601 EltVT = VT.getVectorElementType();
5602 std::vector<Constant*> CV;
5603 if (EltVT == MVT::f64) {
5604 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5608 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5614 Constant *C = ConstantVector::get(CV);
5615 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5616 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5617 PseudoSourceValue::getConstantPool(), 0,
5619 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5622 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5623 LLVMContext *Context = DAG.getContext();
5624 DebugLoc dl = Op.getDebugLoc();
5625 EVT VT = Op.getValueType();
5628 EltVT = VT.getVectorElementType();
5629 std::vector<Constant*> CV;
5630 if (EltVT == MVT::f64) {
5631 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5635 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5641 Constant *C = ConstantVector::get(CV);
5642 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5643 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5644 PseudoSourceValue::getConstantPool(), 0,
5646 if (VT.isVector()) {
5647 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5648 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5649 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5651 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5653 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5657 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5658 LLVMContext *Context = DAG.getContext();
5659 SDValue Op0 = Op.getOperand(0);
5660 SDValue Op1 = Op.getOperand(1);
5661 DebugLoc dl = Op.getDebugLoc();
5662 EVT VT = Op.getValueType();
5663 EVT SrcVT = Op1.getValueType();
5665 // If second operand is smaller, extend it first.
5666 if (SrcVT.bitsLT(VT)) {
5667 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5670 // And if it is bigger, shrink it first.
5671 if (SrcVT.bitsGT(VT)) {
5672 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5676 // At this point the operands and the result should have the same
5677 // type, and that won't be f80 since that is not custom lowered.
5679 // First get the sign bit of second operand.
5680 std::vector<Constant*> CV;
5681 if (SrcVT == MVT::f64) {
5682 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5683 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5685 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5686 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5687 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5688 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5690 Constant *C = ConstantVector::get(CV);
5691 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5692 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5693 PseudoSourceValue::getConstantPool(), 0,
5695 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5697 // Shift sign bit right or left if the two operands have different types.
5698 if (SrcVT.bitsGT(VT)) {
5699 // Op0 is MVT::f32, Op1 is MVT::f64.
5700 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5701 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5702 DAG.getConstant(32, MVT::i32));
5703 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5704 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5705 DAG.getIntPtrConstant(0));
5708 // Clear first operand sign bit.
5710 if (VT == MVT::f64) {
5711 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5712 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5714 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5715 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5716 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5717 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5719 C = ConstantVector::get(CV);
5720 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5721 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5722 PseudoSourceValue::getConstantPool(), 0,
5724 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5726 // Or the value with the sign bit.
5727 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5730 /// Emit nodes that will be selected as "test Op0,Op0", or something
5732 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5733 SelectionDAG &DAG) {
5734 DebugLoc dl = Op.getDebugLoc();
5736 // CF and OF aren't always set the way we want. Determine which
5737 // of these we need.
5738 bool NeedCF = false;
5739 bool NeedOF = false;
5741 case X86::COND_A: case X86::COND_AE:
5742 case X86::COND_B: case X86::COND_BE:
5745 case X86::COND_G: case X86::COND_GE:
5746 case X86::COND_L: case X86::COND_LE:
5747 case X86::COND_O: case X86::COND_NO:
5753 // See if we can use the EFLAGS value from the operand instead of
5754 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5755 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5756 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5757 unsigned Opcode = 0;
5758 unsigned NumOperands = 0;
5759 switch (Op.getNode()->getOpcode()) {
5761 // Due to an isel shortcoming, be conservative if this add is likely to
5762 // be selected as part of a load-modify-store instruction. When the root
5763 // node in a match is a store, isel doesn't know how to remap non-chain
5764 // non-flag uses of other nodes in the match, such as the ADD in this
5765 // case. This leads to the ADD being left around and reselected, with
5766 // the result being two adds in the output.
5767 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5768 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5769 if (UI->getOpcode() == ISD::STORE)
5771 if (ConstantSDNode *C =
5772 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5773 // An add of one will be selected as an INC.
5774 if (C->getAPIntValue() == 1) {
5775 Opcode = X86ISD::INC;
5779 // An add of negative one (subtract of one) will be selected as a DEC.
5780 if (C->getAPIntValue().isAllOnesValue()) {
5781 Opcode = X86ISD::DEC;
5786 // Otherwise use a regular EFLAGS-setting add.
5787 Opcode = X86ISD::ADD;
5791 // If the primary and result isn't used, don't bother using X86ISD::AND,
5792 // because a TEST instruction will be better.
5793 bool NonFlagUse = false;
5794 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5795 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
5797 unsigned UOpNo = UI.getOperandNo();
5798 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
5799 // Look pass truncate.
5800 UOpNo = User->use_begin().getOperandNo();
5801 User = *User->use_begin();
5803 if (User->getOpcode() != ISD::BRCOND &&
5804 User->getOpcode() != ISD::SETCC &&
5805 (User->getOpcode() != ISD::SELECT || UOpNo != 0)) {
5817 // Due to the ISEL shortcoming noted above, be conservative if this op is
5818 // likely to be selected as part of a load-modify-store instruction.
5819 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5820 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5821 if (UI->getOpcode() == ISD::STORE)
5823 // Otherwise use a regular EFLAGS-setting instruction.
5824 switch (Op.getNode()->getOpcode()) {
5825 case ISD::SUB: Opcode = X86ISD::SUB; break;
5826 case ISD::OR: Opcode = X86ISD::OR; break;
5827 case ISD::XOR: Opcode = X86ISD::XOR; break;
5828 case ISD::AND: Opcode = X86ISD::AND; break;
5829 default: llvm_unreachable("unexpected operator!");
5840 return SDValue(Op.getNode(), 1);
5846 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5847 SmallVector<SDValue, 4> Ops;
5848 for (unsigned i = 0; i != NumOperands; ++i)
5849 Ops.push_back(Op.getOperand(i));
5850 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5851 DAG.ReplaceAllUsesWith(Op, New);
5852 return SDValue(New.getNode(), 1);
5856 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5857 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5858 DAG.getConstant(0, Op.getValueType()));
5861 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5863 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5864 SelectionDAG &DAG) {
5865 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5866 if (C->getAPIntValue() == 0)
5867 return EmitTest(Op0, X86CC, DAG);
5869 DebugLoc dl = Op0.getDebugLoc();
5870 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5873 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
5874 /// if it's possible.
5875 static SDValue LowerToBT(SDValue Op0, ISD::CondCode CC,
5876 DebugLoc dl, SelectionDAG &DAG) {
5878 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5879 if (ConstantSDNode *Op010C =
5880 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5881 if (Op010C->getZExtValue() == 1) {
5882 LHS = Op0.getOperand(0);
5883 RHS = Op0.getOperand(1).getOperand(1);
5885 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5886 if (ConstantSDNode *Op000C =
5887 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5888 if (Op000C->getZExtValue() == 1) {
5889 LHS = Op0.getOperand(1);
5890 RHS = Op0.getOperand(0).getOperand(1);
5892 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5893 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5894 SDValue AndLHS = Op0.getOperand(0);
5895 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5896 LHS = AndLHS.getOperand(0);
5897 RHS = AndLHS.getOperand(1);
5901 if (LHS.getNode()) {
5902 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5903 // instruction. Since the shift amount is in-range-or-undefined, we know
5904 // that doing a bittest on the i16 value is ok. We extend to i32 because
5905 // the encoding for the i16 version is larger than the i32 version.
5906 if (LHS.getValueType() == MVT::i8)
5907 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5909 // If the operand types disagree, extend the shift amount to match. Since
5910 // BT ignores high bits (like shifts) we can use anyextend.
5911 if (LHS.getValueType() != RHS.getValueType())
5912 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5914 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5915 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5916 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5917 DAG.getConstant(Cond, MVT::i8), BT);
5923 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5924 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5925 SDValue Op0 = Op.getOperand(0);
5926 SDValue Op1 = Op.getOperand(1);
5927 DebugLoc dl = Op.getDebugLoc();
5928 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5930 // Optimize to BT if possible.
5931 // Lower (X & (1 << N)) == 0 to BT(X, N).
5932 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5933 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5934 if (Op0.getOpcode() == ISD::AND &&
5936 Op1.getOpcode() == ISD::Constant &&
5937 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5938 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5939 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
5940 if (NewSetCC.getNode())
5944 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5945 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5946 if (X86CC == X86::COND_INVALID)
5949 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5951 // Use sbb x, x to materialize carry bit into a GPR.
5952 if (X86CC == X86::COND_B)
5953 return DAG.getNode(ISD::AND, dl, MVT::i8,
5954 DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
5955 DAG.getConstant(X86CC, MVT::i8), Cond),
5956 DAG.getConstant(1, MVT::i8));
5958 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5959 DAG.getConstant(X86CC, MVT::i8), Cond);
5962 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5964 SDValue Op0 = Op.getOperand(0);
5965 SDValue Op1 = Op.getOperand(1);
5966 SDValue CC = Op.getOperand(2);
5967 EVT VT = Op.getValueType();
5968 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5969 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5970 DebugLoc dl = Op.getDebugLoc();
5974 EVT VT0 = Op0.getValueType();
5975 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5976 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5979 switch (SetCCOpcode) {
5982 case ISD::SETEQ: SSECC = 0; break;
5984 case ISD::SETGT: Swap = true; // Fallthrough
5986 case ISD::SETOLT: SSECC = 1; break;
5988 case ISD::SETGE: Swap = true; // Fallthrough
5990 case ISD::SETOLE: SSECC = 2; break;
5991 case ISD::SETUO: SSECC = 3; break;
5993 case ISD::SETNE: SSECC = 4; break;
5994 case ISD::SETULE: Swap = true;
5995 case ISD::SETUGE: SSECC = 5; break;
5996 case ISD::SETULT: Swap = true;
5997 case ISD::SETUGT: SSECC = 6; break;
5998 case ISD::SETO: SSECC = 7; break;
6001 std::swap(Op0, Op1);
6003 // In the two special cases we can't handle, emit two comparisons.
6005 if (SetCCOpcode == ISD::SETUEQ) {
6007 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
6008 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
6009 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
6011 else if (SetCCOpcode == ISD::SETONE) {
6013 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
6014 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
6015 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
6017 llvm_unreachable("Illegal FP comparison");
6019 // Handle all other FP comparisons here.
6020 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
6023 // We are handling one of the integer comparisons here. Since SSE only has
6024 // GT and EQ comparisons for integer, swapping operands and multiple
6025 // operations may be required for some comparisons.
6026 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
6027 bool Swap = false, Invert = false, FlipSigns = false;
6029 switch (VT.getSimpleVT().SimpleTy) {
6032 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
6034 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
6036 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
6037 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
6040 switch (SetCCOpcode) {
6042 case ISD::SETNE: Invert = true;
6043 case ISD::SETEQ: Opc = EQOpc; break;
6044 case ISD::SETLT: Swap = true;
6045 case ISD::SETGT: Opc = GTOpc; break;
6046 case ISD::SETGE: Swap = true;
6047 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
6048 case ISD::SETULT: Swap = true;
6049 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
6050 case ISD::SETUGE: Swap = true;
6051 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
6054 std::swap(Op0, Op1);
6056 // Since SSE has no unsigned integer comparisons, we need to flip the sign
6057 // bits of the inputs before performing those operations.
6059 EVT EltVT = VT.getVectorElementType();
6060 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
6062 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
6063 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
6065 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
6066 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
6069 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
6071 // If the logical-not of the result is required, perform that now.
6073 Result = DAG.getNOT(dl, Result, VT);
6078 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
6079 static bool isX86LogicalCmp(SDValue Op) {
6080 unsigned Opc = Op.getNode()->getOpcode();
6081 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
6083 if (Op.getResNo() == 1 &&
6084 (Opc == X86ISD::ADD ||
6085 Opc == X86ISD::SUB ||
6086 Opc == X86ISD::SMUL ||
6087 Opc == X86ISD::UMUL ||
6088 Opc == X86ISD::INC ||
6089 Opc == X86ISD::DEC ||
6090 Opc == X86ISD::OR ||
6091 Opc == X86ISD::XOR ||
6092 Opc == X86ISD::AND))
6098 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
6099 bool addTest = true;
6100 SDValue Cond = Op.getOperand(0);
6101 DebugLoc dl = Op.getDebugLoc();
6104 if (Cond.getOpcode() == ISD::SETCC) {
6105 SDValue NewCond = LowerSETCC(Cond, DAG);
6106 if (NewCond.getNode())
6110 // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
6111 SDValue Op1 = Op.getOperand(1);
6112 SDValue Op2 = Op.getOperand(2);
6113 if (Cond.getOpcode() == X86ISD::SETCC &&
6114 cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
6115 SDValue Cmp = Cond.getOperand(1);
6116 if (Cmp.getOpcode() == X86ISD::CMP) {
6117 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
6118 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
6119 ConstantSDNode *RHSC =
6120 dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
6121 if (N1C && N1C->isAllOnesValue() &&
6122 N2C && N2C->isNullValue() &&
6123 RHSC && RHSC->isNullValue()) {
6124 SDValue CmpOp0 = Cmp.getOperand(0);
6125 Cmp = DAG.getNode(X86ISD::CMP, dl, CmpOp0.getValueType(),
6126 CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
6127 return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
6128 DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
6133 // Look pass (and (setcc_carry (cmp ...)), 1).
6134 if (Cond.getOpcode() == ISD::AND &&
6135 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6136 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6137 if (C && C->getAPIntValue() == 1)
6138 Cond = Cond.getOperand(0);
6141 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6142 // setting operand in place of the X86ISD::SETCC.
6143 if (Cond.getOpcode() == X86ISD::SETCC ||
6144 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6145 CC = Cond.getOperand(0);
6147 SDValue Cmp = Cond.getOperand(1);
6148 unsigned Opc = Cmp.getOpcode();
6149 EVT VT = Op.getValueType();
6151 bool IllegalFPCMov = false;
6152 if (VT.isFloatingPoint() && !VT.isVector() &&
6153 !isScalarFPTypeInSSEReg(VT)) // FPStack?
6154 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
6156 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
6157 Opc == X86ISD::BT) { // FIXME
6164 // Look pass the truncate.
6165 if (Cond.getOpcode() == ISD::TRUNCATE)
6166 Cond = Cond.getOperand(0);
6168 // We know the result of AND is compared against zero. Try to match
6170 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6171 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6172 if (NewSetCC.getNode()) {
6173 CC = NewSetCC.getOperand(0);
6174 Cond = NewSetCC.getOperand(1);
6181 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6182 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6185 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
6186 // condition is true.
6187 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
6188 SDValue Ops[] = { Op2, Op1, CC, Cond };
6189 return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
6192 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
6193 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
6194 // from the AND / OR.
6195 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
6196 Opc = Op.getOpcode();
6197 if (Opc != ISD::OR && Opc != ISD::AND)
6199 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6200 Op.getOperand(0).hasOneUse() &&
6201 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
6202 Op.getOperand(1).hasOneUse());
6205 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
6206 // 1 and that the SETCC node has a single use.
6207 static bool isXor1OfSetCC(SDValue Op) {
6208 if (Op.getOpcode() != ISD::XOR)
6210 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
6211 if (N1C && N1C->getAPIntValue() == 1) {
6212 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
6213 Op.getOperand(0).hasOneUse();
6218 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
6219 bool addTest = true;
6220 SDValue Chain = Op.getOperand(0);
6221 SDValue Cond = Op.getOperand(1);
6222 SDValue Dest = Op.getOperand(2);
6223 DebugLoc dl = Op.getDebugLoc();
6226 if (Cond.getOpcode() == ISD::SETCC) {
6227 SDValue NewCond = LowerSETCC(Cond, DAG);
6228 if (NewCond.getNode())
6232 // FIXME: LowerXALUO doesn't handle these!!
6233 else if (Cond.getOpcode() == X86ISD::ADD ||
6234 Cond.getOpcode() == X86ISD::SUB ||
6235 Cond.getOpcode() == X86ISD::SMUL ||
6236 Cond.getOpcode() == X86ISD::UMUL)
6237 Cond = LowerXALUO(Cond, DAG);
6240 // Look pass (and (setcc_carry (cmp ...)), 1).
6241 if (Cond.getOpcode() == ISD::AND &&
6242 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
6243 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
6244 if (C && C->getAPIntValue() == 1)
6245 Cond = Cond.getOperand(0);
6248 // If condition flag is set by a X86ISD::CMP, then use it as the condition
6249 // setting operand in place of the X86ISD::SETCC.
6250 if (Cond.getOpcode() == X86ISD::SETCC ||
6251 Cond.getOpcode() == X86ISD::SETCC_CARRY) {
6252 CC = Cond.getOperand(0);
6254 SDValue Cmp = Cond.getOperand(1);
6255 unsigned Opc = Cmp.getOpcode();
6256 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
6257 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
6261 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
6265 // These can only come from an arithmetic instruction with overflow,
6266 // e.g. SADDO, UADDO.
6267 Cond = Cond.getNode()->getOperand(1);
6274 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
6275 SDValue Cmp = Cond.getOperand(0).getOperand(1);
6276 if (CondOpc == ISD::OR) {
6277 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
6278 // two branches instead of an explicit OR instruction with a
6280 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6281 isX86LogicalCmp(Cmp)) {
6282 CC = Cond.getOperand(0).getOperand(0);
6283 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6284 Chain, Dest, CC, Cmp);
6285 CC = Cond.getOperand(1).getOperand(0);
6289 } else { // ISD::AND
6290 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
6291 // two branches instead of an explicit AND instruction with a
6292 // separate test. However, we only do this if this block doesn't
6293 // have a fall-through edge, because this requires an explicit
6294 // jmp when the condition is false.
6295 if (Cmp == Cond.getOperand(1).getOperand(1) &&
6296 isX86LogicalCmp(Cmp) &&
6297 Op.getNode()->hasOneUse()) {
6298 X86::CondCode CCode =
6299 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6300 CCode = X86::GetOppositeBranchCondition(CCode);
6301 CC = DAG.getConstant(CCode, MVT::i8);
6302 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
6303 // Look for an unconditional branch following this conditional branch.
6304 // We need this because we need to reverse the successors in order
6305 // to implement FCMP_OEQ.
6306 if (User.getOpcode() == ISD::BR) {
6307 SDValue FalseBB = User.getOperand(1);
6309 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
6310 assert(NewBR == User);
6313 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6314 Chain, Dest, CC, Cmp);
6315 X86::CondCode CCode =
6316 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
6317 CCode = X86::GetOppositeBranchCondition(CCode);
6318 CC = DAG.getConstant(CCode, MVT::i8);
6324 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
6325 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
6326 // It should be transformed during dag combiner except when the condition
6327 // is set by a arithmetics with overflow node.
6328 X86::CondCode CCode =
6329 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
6330 CCode = X86::GetOppositeBranchCondition(CCode);
6331 CC = DAG.getConstant(CCode, MVT::i8);
6332 Cond = Cond.getOperand(0).getOperand(1);
6338 // Look pass the truncate.
6339 if (Cond.getOpcode() == ISD::TRUNCATE)
6340 Cond = Cond.getOperand(0);
6342 // We know the result of AND is compared against zero. Try to match
6344 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
6345 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
6346 if (NewSetCC.getNode()) {
6347 CC = NewSetCC.getOperand(0);
6348 Cond = NewSetCC.getOperand(1);
6355 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
6356 Cond = EmitTest(Cond, X86::COND_NE, DAG);
6358 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
6359 Chain, Dest, CC, Cond);
6363 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
6364 // Calls to _alloca is needed to probe the stack when allocating more than 4k
6365 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
6366 // that the guard pages used by the OS virtual memory manager are allocated in
6367 // correct sequence.
6369 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
6370 SelectionDAG &DAG) {
6371 assert(Subtarget->isTargetCygMing() &&
6372 "This should be used only on Cygwin/Mingw targets");
6373 DebugLoc dl = Op.getDebugLoc();
6376 SDValue Chain = Op.getOperand(0);
6377 SDValue Size = Op.getOperand(1);
6378 // FIXME: Ensure alignment here
6382 EVT IntPtr = getPointerTy();
6383 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
6385 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
6387 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
6388 Flag = Chain.getValue(1);
6390 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
6391 SDValue Ops[] = { Chain,
6392 DAG.getTargetExternalSymbol("_alloca", IntPtr),
6393 DAG.getRegister(X86::EAX, IntPtr),
6394 DAG.getRegister(X86StackPtr, SPTy),
6396 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
6397 Flag = Chain.getValue(1);
6399 Chain = DAG.getCALLSEQ_END(Chain,
6400 DAG.getIntPtrConstant(0, true),
6401 DAG.getIntPtrConstant(0, true),
6404 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
6406 SDValue Ops1[2] = { Chain.getValue(0), Chain };
6407 return DAG.getMergeValues(Ops1, 2, dl);
6411 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6413 SDValue Dst, SDValue Src,
6414 SDValue Size, unsigned Align,
6416 uint64_t DstSVOff) {
6417 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6419 // If not DWORD aligned or size is more than the threshold, call the library.
6420 // The libc version is likely to be faster for these cases. It can use the
6421 // address value and run time information about the CPU.
6422 if ((Align & 3) != 0 ||
6424 ConstantSize->getZExtValue() >
6425 getSubtarget()->getMaxInlineSizeThreshold()) {
6426 SDValue InFlag(0, 0);
6428 // Check to see if there is a specialized entry-point for memory zeroing.
6429 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6431 if (const char *bzeroEntry = V &&
6432 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6433 EVT IntPtr = getPointerTy();
6434 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6435 TargetLowering::ArgListTy Args;
6436 TargetLowering::ArgListEntry Entry;
6438 Entry.Ty = IntPtrTy;
6439 Args.push_back(Entry);
6441 Args.push_back(Entry);
6442 std::pair<SDValue,SDValue> CallResult =
6443 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6444 false, false, false, false,
6445 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6446 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl,
6447 DAG.GetOrdering(Chain.getNode()));
6448 return CallResult.second;
6451 // Otherwise have the target-independent code call memset.
6455 uint64_t SizeVal = ConstantSize->getZExtValue();
6456 SDValue InFlag(0, 0);
6459 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6460 unsigned BytesLeft = 0;
6461 bool TwoRepStos = false;
6464 uint64_t Val = ValC->getZExtValue() & 255;
6466 // If the value is a constant, then we can potentially use larger sets.
6467 switch (Align & 3) {
6468 case 2: // WORD aligned
6471 Val = (Val << 8) | Val;
6473 case 0: // DWORD aligned
6476 Val = (Val << 8) | Val;
6477 Val = (Val << 16) | Val;
6478 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6481 Val = (Val << 32) | Val;
6484 default: // Byte aligned
6487 Count = DAG.getIntPtrConstant(SizeVal);
6491 if (AVT.bitsGT(MVT::i8)) {
6492 unsigned UBytes = AVT.getSizeInBits() / 8;
6493 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6494 BytesLeft = SizeVal % UBytes;
6497 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6499 InFlag = Chain.getValue(1);
6502 Count = DAG.getIntPtrConstant(SizeVal);
6503 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6504 InFlag = Chain.getValue(1);
6507 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6510 InFlag = Chain.getValue(1);
6511 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6514 InFlag = Chain.getValue(1);
6516 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6517 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6518 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6521 InFlag = Chain.getValue(1);
6523 EVT CVT = Count.getValueType();
6524 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6525 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6526 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6529 InFlag = Chain.getValue(1);
6530 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6531 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
6532 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
6533 } else if (BytesLeft) {
6534 // Handle the last 1 - 7 bytes.
6535 unsigned Offset = SizeVal - BytesLeft;
6536 EVT AddrVT = Dst.getValueType();
6537 EVT SizeVT = Size.getValueType();
6539 Chain = DAG.getMemset(Chain, dl,
6540 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6541 DAG.getConstant(Offset, AddrVT)),
6543 DAG.getConstant(BytesLeft, SizeVT),
6544 Align, DstSV, DstSVOff + Offset);
6547 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6552 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6553 SDValue Chain, SDValue Dst, SDValue Src,
6554 SDValue Size, unsigned Align,
6556 const Value *DstSV, uint64_t DstSVOff,
6557 const Value *SrcSV, uint64_t SrcSVOff) {
6558 // This requires the copy size to be a constant, preferrably
6559 // within a subtarget-specific limit.
6560 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6563 uint64_t SizeVal = ConstantSize->getZExtValue();
6564 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6567 /// If not DWORD aligned, call the library.
6568 if ((Align & 3) != 0)
6573 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6576 unsigned UBytes = AVT.getSizeInBits() / 8;
6577 unsigned CountVal = SizeVal / UBytes;
6578 SDValue Count = DAG.getIntPtrConstant(CountVal);
6579 unsigned BytesLeft = SizeVal % UBytes;
6581 SDValue InFlag(0, 0);
6582 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6585 InFlag = Chain.getValue(1);
6586 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6589 InFlag = Chain.getValue(1);
6590 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6593 InFlag = Chain.getValue(1);
6595 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6596 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
6597 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
6598 array_lengthof(Ops));
6600 SmallVector<SDValue, 4> Results;
6601 Results.push_back(RepMovs);
6603 // Handle the last 1 - 7 bytes.
6604 unsigned Offset = SizeVal - BytesLeft;
6605 EVT DstVT = Dst.getValueType();
6606 EVT SrcVT = Src.getValueType();
6607 EVT SizeVT = Size.getValueType();
6608 Results.push_back(DAG.getMemcpy(Chain, dl,
6609 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6610 DAG.getConstant(Offset, DstVT)),
6611 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6612 DAG.getConstant(Offset, SrcVT)),
6613 DAG.getConstant(BytesLeft, SizeVT),
6614 Align, AlwaysInline,
6615 DstSV, DstSVOff + Offset,
6616 SrcSV, SrcSVOff + Offset));
6619 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6620 &Results[0], Results.size());
6623 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6624 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6625 DebugLoc dl = Op.getDebugLoc();
6627 if (!Subtarget->is64Bit()) {
6628 // vastart just stores the address of the VarArgsFrameIndex slot into the
6629 // memory location argument.
6630 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6631 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0,
6636 // gp_offset (0 - 6 * 8)
6637 // fp_offset (48 - 48 + 8 * 16)
6638 // overflow_arg_area (point to parameters coming in memory).
6640 SmallVector<SDValue, 8> MemOps;
6641 SDValue FIN = Op.getOperand(1);
6643 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6644 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6645 FIN, SV, 0, false, false, 0);
6646 MemOps.push_back(Store);
6649 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6650 FIN, DAG.getIntPtrConstant(4));
6651 Store = DAG.getStore(Op.getOperand(0), dl,
6652 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6653 FIN, SV, 0, false, false, 0);
6654 MemOps.push_back(Store);
6656 // Store ptr to overflow_arg_area
6657 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6658 FIN, DAG.getIntPtrConstant(4));
6659 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6660 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0,
6662 MemOps.push_back(Store);
6664 // Store ptr to reg_save_area.
6665 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6666 FIN, DAG.getIntPtrConstant(8));
6667 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6668 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0,
6670 MemOps.push_back(Store);
6671 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6672 &MemOps[0], MemOps.size());
6675 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6676 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6677 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6678 SDValue Chain = Op.getOperand(0);
6679 SDValue SrcPtr = Op.getOperand(1);
6680 SDValue SrcSV = Op.getOperand(2);
6682 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6686 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6687 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6688 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6689 SDValue Chain = Op.getOperand(0);
6690 SDValue DstPtr = Op.getOperand(1);
6691 SDValue SrcPtr = Op.getOperand(2);
6692 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6693 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6694 DebugLoc dl = Op.getDebugLoc();
6696 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6697 DAG.getIntPtrConstant(24), 8, false,
6698 DstSV, 0, SrcSV, 0);
6702 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6703 DebugLoc dl = Op.getDebugLoc();
6704 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6706 default: return SDValue(); // Don't custom lower most intrinsics.
6707 // Comparison intrinsics.
6708 case Intrinsic::x86_sse_comieq_ss:
6709 case Intrinsic::x86_sse_comilt_ss:
6710 case Intrinsic::x86_sse_comile_ss:
6711 case Intrinsic::x86_sse_comigt_ss:
6712 case Intrinsic::x86_sse_comige_ss:
6713 case Intrinsic::x86_sse_comineq_ss:
6714 case Intrinsic::x86_sse_ucomieq_ss:
6715 case Intrinsic::x86_sse_ucomilt_ss:
6716 case Intrinsic::x86_sse_ucomile_ss:
6717 case Intrinsic::x86_sse_ucomigt_ss:
6718 case Intrinsic::x86_sse_ucomige_ss:
6719 case Intrinsic::x86_sse_ucomineq_ss:
6720 case Intrinsic::x86_sse2_comieq_sd:
6721 case Intrinsic::x86_sse2_comilt_sd:
6722 case Intrinsic::x86_sse2_comile_sd:
6723 case Intrinsic::x86_sse2_comigt_sd:
6724 case Intrinsic::x86_sse2_comige_sd:
6725 case Intrinsic::x86_sse2_comineq_sd:
6726 case Intrinsic::x86_sse2_ucomieq_sd:
6727 case Intrinsic::x86_sse2_ucomilt_sd:
6728 case Intrinsic::x86_sse2_ucomile_sd:
6729 case Intrinsic::x86_sse2_ucomigt_sd:
6730 case Intrinsic::x86_sse2_ucomige_sd:
6731 case Intrinsic::x86_sse2_ucomineq_sd: {
6733 ISD::CondCode CC = ISD::SETCC_INVALID;
6736 case Intrinsic::x86_sse_comieq_ss:
6737 case Intrinsic::x86_sse2_comieq_sd:
6741 case Intrinsic::x86_sse_comilt_ss:
6742 case Intrinsic::x86_sse2_comilt_sd:
6746 case Intrinsic::x86_sse_comile_ss:
6747 case Intrinsic::x86_sse2_comile_sd:
6751 case Intrinsic::x86_sse_comigt_ss:
6752 case Intrinsic::x86_sse2_comigt_sd:
6756 case Intrinsic::x86_sse_comige_ss:
6757 case Intrinsic::x86_sse2_comige_sd:
6761 case Intrinsic::x86_sse_comineq_ss:
6762 case Intrinsic::x86_sse2_comineq_sd:
6766 case Intrinsic::x86_sse_ucomieq_ss:
6767 case Intrinsic::x86_sse2_ucomieq_sd:
6768 Opc = X86ISD::UCOMI;
6771 case Intrinsic::x86_sse_ucomilt_ss:
6772 case Intrinsic::x86_sse2_ucomilt_sd:
6773 Opc = X86ISD::UCOMI;
6776 case Intrinsic::x86_sse_ucomile_ss:
6777 case Intrinsic::x86_sse2_ucomile_sd:
6778 Opc = X86ISD::UCOMI;
6781 case Intrinsic::x86_sse_ucomigt_ss:
6782 case Intrinsic::x86_sse2_ucomigt_sd:
6783 Opc = X86ISD::UCOMI;
6786 case Intrinsic::x86_sse_ucomige_ss:
6787 case Intrinsic::x86_sse2_ucomige_sd:
6788 Opc = X86ISD::UCOMI;
6791 case Intrinsic::x86_sse_ucomineq_ss:
6792 case Intrinsic::x86_sse2_ucomineq_sd:
6793 Opc = X86ISD::UCOMI;
6798 SDValue LHS = Op.getOperand(1);
6799 SDValue RHS = Op.getOperand(2);
6800 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6801 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6802 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6803 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6804 DAG.getConstant(X86CC, MVT::i8), Cond);
6805 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6807 // ptest intrinsics. The intrinsic these come from are designed to return
6808 // an integer value, not just an instruction so lower it to the ptest
6809 // pattern and a setcc for the result.
6810 case Intrinsic::x86_sse41_ptestz:
6811 case Intrinsic::x86_sse41_ptestc:
6812 case Intrinsic::x86_sse41_ptestnzc:{
6815 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6816 case Intrinsic::x86_sse41_ptestz:
6818 X86CC = X86::COND_E;
6820 case Intrinsic::x86_sse41_ptestc:
6822 X86CC = X86::COND_B;
6824 case Intrinsic::x86_sse41_ptestnzc:
6826 X86CC = X86::COND_A;
6830 SDValue LHS = Op.getOperand(1);
6831 SDValue RHS = Op.getOperand(2);
6832 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6833 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6834 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6835 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6838 // Fix vector shift instructions where the last operand is a non-immediate
6840 case Intrinsic::x86_sse2_pslli_w:
6841 case Intrinsic::x86_sse2_pslli_d:
6842 case Intrinsic::x86_sse2_pslli_q:
6843 case Intrinsic::x86_sse2_psrli_w:
6844 case Intrinsic::x86_sse2_psrli_d:
6845 case Intrinsic::x86_sse2_psrli_q:
6846 case Intrinsic::x86_sse2_psrai_w:
6847 case Intrinsic::x86_sse2_psrai_d:
6848 case Intrinsic::x86_mmx_pslli_w:
6849 case Intrinsic::x86_mmx_pslli_d:
6850 case Intrinsic::x86_mmx_pslli_q:
6851 case Intrinsic::x86_mmx_psrli_w:
6852 case Intrinsic::x86_mmx_psrli_d:
6853 case Intrinsic::x86_mmx_psrli_q:
6854 case Intrinsic::x86_mmx_psrai_w:
6855 case Intrinsic::x86_mmx_psrai_d: {
6856 SDValue ShAmt = Op.getOperand(2);
6857 if (isa<ConstantSDNode>(ShAmt))
6860 unsigned NewIntNo = 0;
6861 EVT ShAmtVT = MVT::v4i32;
6863 case Intrinsic::x86_sse2_pslli_w:
6864 NewIntNo = Intrinsic::x86_sse2_psll_w;
6866 case Intrinsic::x86_sse2_pslli_d:
6867 NewIntNo = Intrinsic::x86_sse2_psll_d;
6869 case Intrinsic::x86_sse2_pslli_q:
6870 NewIntNo = Intrinsic::x86_sse2_psll_q;
6872 case Intrinsic::x86_sse2_psrli_w:
6873 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6875 case Intrinsic::x86_sse2_psrli_d:
6876 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6878 case Intrinsic::x86_sse2_psrli_q:
6879 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6881 case Intrinsic::x86_sse2_psrai_w:
6882 NewIntNo = Intrinsic::x86_sse2_psra_w;
6884 case Intrinsic::x86_sse2_psrai_d:
6885 NewIntNo = Intrinsic::x86_sse2_psra_d;
6888 ShAmtVT = MVT::v2i32;
6890 case Intrinsic::x86_mmx_pslli_w:
6891 NewIntNo = Intrinsic::x86_mmx_psll_w;
6893 case Intrinsic::x86_mmx_pslli_d:
6894 NewIntNo = Intrinsic::x86_mmx_psll_d;
6896 case Intrinsic::x86_mmx_pslli_q:
6897 NewIntNo = Intrinsic::x86_mmx_psll_q;
6899 case Intrinsic::x86_mmx_psrli_w:
6900 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6902 case Intrinsic::x86_mmx_psrli_d:
6903 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6905 case Intrinsic::x86_mmx_psrli_q:
6906 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6908 case Intrinsic::x86_mmx_psrai_w:
6909 NewIntNo = Intrinsic::x86_mmx_psra_w;
6911 case Intrinsic::x86_mmx_psrai_d:
6912 NewIntNo = Intrinsic::x86_mmx_psra_d;
6914 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6920 // The vector shift intrinsics with scalars uses 32b shift amounts but
6921 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6925 ShOps[1] = DAG.getConstant(0, MVT::i32);
6926 if (ShAmtVT == MVT::v4i32) {
6927 ShOps[2] = DAG.getUNDEF(MVT::i32);
6928 ShOps[3] = DAG.getUNDEF(MVT::i32);
6929 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6931 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6934 EVT VT = Op.getValueType();
6935 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6936 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6937 DAG.getConstant(NewIntNo, MVT::i32),
6938 Op.getOperand(1), ShAmt);
6943 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6944 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6945 DebugLoc dl = Op.getDebugLoc();
6948 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6950 DAG.getConstant(TD->getPointerSize(),
6951 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6952 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6953 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6955 NULL, 0, false, false, 0);
6958 // Just load the return address.
6959 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6960 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6961 RetAddrFI, NULL, 0, false, false, 0);
6964 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6965 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6966 MFI->setFrameAddressIsTaken(true);
6967 EVT VT = Op.getValueType();
6968 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6969 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6970 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6971 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6973 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0,
6978 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6979 SelectionDAG &DAG) {
6980 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6983 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6985 MachineFunction &MF = DAG.getMachineFunction();
6986 SDValue Chain = Op.getOperand(0);
6987 SDValue Offset = Op.getOperand(1);
6988 SDValue Handler = Op.getOperand(2);
6989 DebugLoc dl = Op.getDebugLoc();
6991 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6993 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6995 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6996 DAG.getIntPtrConstant(-TD->getPointerSize()));
6997 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6998 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0, false, false, 0);
6999 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
7000 MF.getRegInfo().addLiveOut(StoreAddrReg);
7002 return DAG.getNode(X86ISD::EH_RETURN, dl,
7004 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
7007 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
7008 SelectionDAG &DAG) {
7009 SDValue Root = Op.getOperand(0);
7010 SDValue Trmp = Op.getOperand(1); // trampoline
7011 SDValue FPtr = Op.getOperand(2); // nested function
7012 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
7013 DebugLoc dl = Op.getDebugLoc();
7015 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
7017 if (Subtarget->is64Bit()) {
7018 SDValue OutChains[6];
7020 // Large code-model.
7021 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
7022 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
7024 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
7025 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
7027 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
7029 // Load the pointer to the nested function into R11.
7030 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
7031 SDValue Addr = Trmp;
7032 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7033 Addr, TrmpAddr, 0, false, false, 0);
7035 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7036 DAG.getConstant(2, MVT::i64));
7037 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2,
7040 // Load the 'nest' parameter value into R10.
7041 // R10 is specified in X86CallingConv.td
7042 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
7043 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7044 DAG.getConstant(10, MVT::i64));
7045 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7046 Addr, TrmpAddr, 10, false, false, 0);
7048 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7049 DAG.getConstant(12, MVT::i64));
7050 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12,
7053 // Jump to the nested function.
7054 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
7055 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7056 DAG.getConstant(20, MVT::i64));
7057 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
7058 Addr, TrmpAddr, 20, false, false, 0);
7060 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
7061 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
7062 DAG.getConstant(22, MVT::i64));
7063 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
7064 TrmpAddr, 22, false, false, 0);
7067 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
7068 return DAG.getMergeValues(Ops, 2, dl);
7070 const Function *Func =
7071 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
7072 CallingConv::ID CC = Func->getCallingConv();
7077 llvm_unreachable("Unsupported calling convention");
7078 case CallingConv::C:
7079 case CallingConv::X86_StdCall: {
7080 // Pass 'nest' parameter in ECX.
7081 // Must be kept in sync with X86CallingConv.td
7084 // Check that ECX wasn't needed by an 'inreg' parameter.
7085 const FunctionType *FTy = Func->getFunctionType();
7086 const AttrListPtr &Attrs = Func->getAttributes();
7088 if (!Attrs.isEmpty() && !Func->isVarArg()) {
7089 unsigned InRegCount = 0;
7092 for (FunctionType::param_iterator I = FTy->param_begin(),
7093 E = FTy->param_end(); I != E; ++I, ++Idx)
7094 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
7095 // FIXME: should only count parameters that are lowered to integers.
7096 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
7098 if (InRegCount > 2) {
7099 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
7104 case CallingConv::X86_FastCall:
7105 case CallingConv::Fast:
7106 // Pass 'nest' parameter in EAX.
7107 // Must be kept in sync with X86CallingConv.td
7112 SDValue OutChains[4];
7115 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7116 DAG.getConstant(10, MVT::i32));
7117 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
7119 // This is storing the opcode for MOV32ri.
7120 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
7121 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
7122 OutChains[0] = DAG.getStore(Root, dl,
7123 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
7124 Trmp, TrmpAddr, 0, false, false, 0);
7126 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7127 DAG.getConstant(1, MVT::i32));
7128 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1,
7131 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
7132 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7133 DAG.getConstant(5, MVT::i32));
7134 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
7135 TrmpAddr, 5, false, false, 1);
7137 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
7138 DAG.getConstant(6, MVT::i32));
7139 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6,
7143 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
7144 return DAG.getMergeValues(Ops, 2, dl);
7148 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
7150 The rounding mode is in bits 11:10 of FPSR, and has the following
7157 FLT_ROUNDS, on the other hand, expects the following:
7164 To perform the conversion, we do:
7165 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
7168 MachineFunction &MF = DAG.getMachineFunction();
7169 const TargetMachine &TM = MF.getTarget();
7170 const TargetFrameInfo &TFI = *TM.getFrameInfo();
7171 unsigned StackAlignment = TFI.getStackAlignment();
7172 EVT VT = Op.getValueType();
7173 DebugLoc dl = Op.getDebugLoc();
7175 // Save FP Control Word to stack slot
7176 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
7177 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
7179 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
7180 DAG.getEntryNode(), StackSlot);
7182 // Load FP Control Word from stack slot
7183 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0,
7186 // Transform as necessary
7188 DAG.getNode(ISD::SRL, dl, MVT::i16,
7189 DAG.getNode(ISD::AND, dl, MVT::i16,
7190 CWD, DAG.getConstant(0x800, MVT::i16)),
7191 DAG.getConstant(11, MVT::i8));
7193 DAG.getNode(ISD::SRL, dl, MVT::i16,
7194 DAG.getNode(ISD::AND, dl, MVT::i16,
7195 CWD, DAG.getConstant(0x400, MVT::i16)),
7196 DAG.getConstant(9, MVT::i8));
7199 DAG.getNode(ISD::AND, dl, MVT::i16,
7200 DAG.getNode(ISD::ADD, dl, MVT::i16,
7201 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
7202 DAG.getConstant(1, MVT::i16)),
7203 DAG.getConstant(3, MVT::i16));
7206 return DAG.getNode((VT.getSizeInBits() < 16 ?
7207 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
7210 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
7211 EVT VT = Op.getValueType();
7213 unsigned NumBits = VT.getSizeInBits();
7214 DebugLoc dl = Op.getDebugLoc();
7216 Op = Op.getOperand(0);
7217 if (VT == MVT::i8) {
7218 // Zero extend to i32 since there is not an i8 bsr.
7220 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7223 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
7224 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7225 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
7227 // If src is zero (i.e. bsr sets ZF), returns NumBits.
7230 DAG.getConstant(NumBits+NumBits-1, OpVT),
7231 DAG.getConstant(X86::COND_E, MVT::i8),
7234 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7236 // Finally xor with NumBits-1.
7237 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
7240 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7244 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
7245 EVT VT = Op.getValueType();
7247 unsigned NumBits = VT.getSizeInBits();
7248 DebugLoc dl = Op.getDebugLoc();
7250 Op = Op.getOperand(0);
7251 if (VT == MVT::i8) {
7253 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
7256 // Issue a bsf (scan bits forward) which also sets EFLAGS.
7257 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
7258 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
7260 // If src is zero (i.e. bsf sets ZF), returns NumBits.
7263 DAG.getConstant(NumBits, OpVT),
7264 DAG.getConstant(X86::COND_E, MVT::i8),
7267 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops, array_lengthof(Ops));
7270 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
7274 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
7275 EVT VT = Op.getValueType();
7276 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
7277 DebugLoc dl = Op.getDebugLoc();
7279 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
7280 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
7281 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
7282 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
7283 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
7285 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
7286 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
7287 // return AloBlo + AloBhi + AhiBlo;
7289 SDValue A = Op.getOperand(0);
7290 SDValue B = Op.getOperand(1);
7292 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7293 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7294 A, DAG.getConstant(32, MVT::i32));
7295 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7296 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
7297 B, DAG.getConstant(32, MVT::i32));
7298 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7299 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7301 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7302 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7304 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7305 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
7307 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7308 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7309 AloBhi, DAG.getConstant(32, MVT::i32));
7310 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
7311 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
7312 AhiBlo, DAG.getConstant(32, MVT::i32));
7313 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
7314 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
7319 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
7320 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
7321 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
7322 // looks for this combo and may remove the "setcc" instruction if the "setcc"
7323 // has only one use.
7324 SDNode *N = Op.getNode();
7325 SDValue LHS = N->getOperand(0);
7326 SDValue RHS = N->getOperand(1);
7327 unsigned BaseOp = 0;
7329 DebugLoc dl = Op.getDebugLoc();
7331 switch (Op.getOpcode()) {
7332 default: llvm_unreachable("Unknown ovf instruction!");
7334 // A subtract of one will be selected as a INC. Note that INC doesn't
7335 // set CF, so we can't do this for UADDO.
7336 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7337 if (C->getAPIntValue() == 1) {
7338 BaseOp = X86ISD::INC;
7342 BaseOp = X86ISD::ADD;
7346 BaseOp = X86ISD::ADD;
7350 // A subtract of one will be selected as a DEC. Note that DEC doesn't
7351 // set CF, so we can't do this for USUBO.
7352 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
7353 if (C->getAPIntValue() == 1) {
7354 BaseOp = X86ISD::DEC;
7358 BaseOp = X86ISD::SUB;
7362 BaseOp = X86ISD::SUB;
7366 BaseOp = X86ISD::SMUL;
7370 BaseOp = X86ISD::UMUL;
7375 // Also sets EFLAGS.
7376 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
7377 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
7380 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
7381 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
7383 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
7387 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
7388 EVT T = Op.getValueType();
7389 DebugLoc dl = Op.getDebugLoc();
7392 switch(T.getSimpleVT().SimpleTy) {
7394 assert(false && "Invalid value type!");
7395 case MVT::i8: Reg = X86::AL; size = 1; break;
7396 case MVT::i16: Reg = X86::AX; size = 2; break;
7397 case MVT::i32: Reg = X86::EAX; size = 4; break;
7399 assert(Subtarget->is64Bit() && "Node not type legal!");
7400 Reg = X86::RAX; size = 8;
7403 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
7404 Op.getOperand(2), SDValue());
7405 SDValue Ops[] = { cpIn.getValue(0),
7408 DAG.getTargetConstant(size, MVT::i8),
7410 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7411 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7413 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7417 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7418 SelectionDAG &DAG) {
7419 assert(Subtarget->is64Bit() && "Result not type legalized?");
7420 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7421 SDValue TheChain = Op.getOperand(0);
7422 DebugLoc dl = Op.getDebugLoc();
7423 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7424 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7425 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7427 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7428 DAG.getConstant(32, MVT::i8));
7430 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7433 return DAG.getMergeValues(Ops, 2, dl);
7436 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7437 SDNode *Node = Op.getNode();
7438 DebugLoc dl = Node->getDebugLoc();
7439 EVT T = Node->getValueType(0);
7440 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7441 DAG.getConstant(0, T), Node->getOperand(2));
7442 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7443 cast<AtomicSDNode>(Node)->getMemoryVT(),
7444 Node->getOperand(0),
7445 Node->getOperand(1), negOp,
7446 cast<AtomicSDNode>(Node)->getSrcValue(),
7447 cast<AtomicSDNode>(Node)->getAlignment());
7450 /// LowerOperation - Provide custom lowering hooks for some operations.
7452 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7453 switch (Op.getOpcode()) {
7454 default: llvm_unreachable("Should not custom lower this!");
7455 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7456 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7457 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7458 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
7459 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7460 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7461 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7462 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7463 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7464 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7465 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7466 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7467 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7468 case ISD::SHL_PARTS:
7469 case ISD::SRA_PARTS:
7470 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7471 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7472 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7473 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7474 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7475 case ISD::FABS: return LowerFABS(Op, DAG);
7476 case ISD::FNEG: return LowerFNEG(Op, DAG);
7477 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7478 case ISD::SETCC: return LowerSETCC(Op, DAG);
7479 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7480 case ISD::SELECT: return LowerSELECT(Op, DAG);
7481 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7482 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7483 case ISD::VASTART: return LowerVASTART(Op, DAG);
7484 case ISD::VAARG: return LowerVAARG(Op, DAG);
7485 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7486 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7487 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7488 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7489 case ISD::FRAME_TO_ARGS_OFFSET:
7490 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7491 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7492 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7493 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7494 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7495 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7496 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7497 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7503 case ISD::UMULO: return LowerXALUO(Op, DAG);
7504 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7508 void X86TargetLowering::
7509 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7510 SelectionDAG &DAG, unsigned NewOp) {
7511 EVT T = Node->getValueType(0);
7512 DebugLoc dl = Node->getDebugLoc();
7513 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7515 SDValue Chain = Node->getOperand(0);
7516 SDValue In1 = Node->getOperand(1);
7517 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7518 Node->getOperand(2), DAG.getIntPtrConstant(0));
7519 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7520 Node->getOperand(2), DAG.getIntPtrConstant(1));
7521 SDValue Ops[] = { Chain, In1, In2L, In2H };
7522 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7524 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7525 cast<MemSDNode>(Node)->getMemOperand());
7526 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7527 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7528 Results.push_back(Result.getValue(2));
7531 /// ReplaceNodeResults - Replace a node with an illegal result type
7532 /// with a new node built out of custom code.
7533 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7534 SmallVectorImpl<SDValue>&Results,
7535 SelectionDAG &DAG) {
7536 DebugLoc dl = N->getDebugLoc();
7537 switch (N->getOpcode()) {
7539 assert(false && "Do not know how to custom type legalize this operation!");
7541 case ISD::FP_TO_SINT: {
7542 std::pair<SDValue,SDValue> Vals =
7543 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7544 SDValue FIST = Vals.first, StackSlot = Vals.second;
7545 if (FIST.getNode() != 0) {
7546 EVT VT = N->getValueType(0);
7547 // Return a load from the stack slot.
7548 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0,
7553 case ISD::READCYCLECOUNTER: {
7554 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7555 SDValue TheChain = N->getOperand(0);
7556 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7557 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7559 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7561 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7562 SDValue Ops[] = { eax, edx };
7563 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7564 Results.push_back(edx.getValue(1));
7567 case ISD::ATOMIC_CMP_SWAP: {
7568 EVT T = N->getValueType(0);
7569 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7570 SDValue cpInL, cpInH;
7571 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7572 DAG.getConstant(0, MVT::i32));
7573 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7574 DAG.getConstant(1, MVT::i32));
7575 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7576 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7578 SDValue swapInL, swapInH;
7579 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7580 DAG.getConstant(0, MVT::i32));
7581 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7582 DAG.getConstant(1, MVT::i32));
7583 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7585 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7586 swapInL.getValue(1));
7587 SDValue Ops[] = { swapInH.getValue(0),
7589 swapInH.getValue(1) };
7590 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7591 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7592 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7593 MVT::i32, Result.getValue(1));
7594 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7595 MVT::i32, cpOutL.getValue(2));
7596 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7597 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7598 Results.push_back(cpOutH.getValue(1));
7601 case ISD::ATOMIC_LOAD_ADD:
7602 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7604 case ISD::ATOMIC_LOAD_AND:
7605 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7607 case ISD::ATOMIC_LOAD_NAND:
7608 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7610 case ISD::ATOMIC_LOAD_OR:
7611 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7613 case ISD::ATOMIC_LOAD_SUB:
7614 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7616 case ISD::ATOMIC_LOAD_XOR:
7617 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7619 case ISD::ATOMIC_SWAP:
7620 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7625 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7627 default: return NULL;
7628 case X86ISD::BSF: return "X86ISD::BSF";
7629 case X86ISD::BSR: return "X86ISD::BSR";
7630 case X86ISD::SHLD: return "X86ISD::SHLD";
7631 case X86ISD::SHRD: return "X86ISD::SHRD";
7632 case X86ISD::FAND: return "X86ISD::FAND";
7633 case X86ISD::FOR: return "X86ISD::FOR";
7634 case X86ISD::FXOR: return "X86ISD::FXOR";
7635 case X86ISD::FSRL: return "X86ISD::FSRL";
7636 case X86ISD::FILD: return "X86ISD::FILD";
7637 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7638 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7639 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7640 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7641 case X86ISD::FLD: return "X86ISD::FLD";
7642 case X86ISD::FST: return "X86ISD::FST";
7643 case X86ISD::CALL: return "X86ISD::CALL";
7644 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7645 case X86ISD::BT: return "X86ISD::BT";
7646 case X86ISD::CMP: return "X86ISD::CMP";
7647 case X86ISD::COMI: return "X86ISD::COMI";
7648 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7649 case X86ISD::SETCC: return "X86ISD::SETCC";
7650 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
7651 case X86ISD::CMOV: return "X86ISD::CMOV";
7652 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7653 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7654 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7655 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7656 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7657 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7658 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7659 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7660 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7661 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7662 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7663 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7664 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7665 case X86ISD::FMAX: return "X86ISD::FMAX";
7666 case X86ISD::FMIN: return "X86ISD::FMIN";
7667 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7668 case X86ISD::FRCP: return "X86ISD::FRCP";
7669 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7670 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7671 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7672 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7673 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7674 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7675 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7676 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7677 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7678 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7679 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7680 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7681 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7682 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7683 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7684 case X86ISD::VSHL: return "X86ISD::VSHL";
7685 case X86ISD::VSRL: return "X86ISD::VSRL";
7686 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7687 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7688 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7689 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7690 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7691 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7692 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7693 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7694 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7695 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7696 case X86ISD::ADD: return "X86ISD::ADD";
7697 case X86ISD::SUB: return "X86ISD::SUB";
7698 case X86ISD::SMUL: return "X86ISD::SMUL";
7699 case X86ISD::UMUL: return "X86ISD::UMUL";
7700 case X86ISD::INC: return "X86ISD::INC";
7701 case X86ISD::DEC: return "X86ISD::DEC";
7702 case X86ISD::OR: return "X86ISD::OR";
7703 case X86ISD::XOR: return "X86ISD::XOR";
7704 case X86ISD::AND: return "X86ISD::AND";
7705 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7706 case X86ISD::PTEST: return "X86ISD::PTEST";
7707 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7711 // isLegalAddressingMode - Return true if the addressing mode represented
7712 // by AM is legal for this target, for a load/store of the specified type.
7713 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7714 const Type *Ty) const {
7715 // X86 supports extremely general addressing modes.
7716 CodeModel::Model M = getTargetMachine().getCodeModel();
7718 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7719 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7724 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7726 // If a reference to this global requires an extra load, we can't fold it.
7727 if (isGlobalStubReference(GVFlags))
7730 // If BaseGV requires a register for the PIC base, we cannot also have a
7731 // BaseReg specified.
7732 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7735 // If lower 4G is not available, then we must use rip-relative addressing.
7736 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7746 // These scales always work.
7751 // These scales are formed with basereg+scalereg. Only accept if there is
7756 default: // Other stuff never works.
7764 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7765 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
7767 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7768 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7769 if (NumBits1 <= NumBits2)
7771 return Subtarget->is64Bit() || NumBits1 < 64;
7774 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7775 if (!VT1.isInteger() || !VT2.isInteger())
7777 unsigned NumBits1 = VT1.getSizeInBits();
7778 unsigned NumBits2 = VT2.getSizeInBits();
7779 if (NumBits1 <= NumBits2)
7781 return Subtarget->is64Bit() || NumBits1 < 64;
7784 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7785 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7786 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
7789 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7790 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7791 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7794 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7795 // i16 instructions are longer (0x66 prefix) and potentially slower.
7796 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7799 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7800 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7801 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7802 /// are assumed to be legal.
7804 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7806 // Only do shuffles on 128-bit vector types for now.
7807 if (VT.getSizeInBits() == 64)
7810 // FIXME: pshufb, blends, shifts.
7811 return (VT.getVectorNumElements() == 2 ||
7812 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7813 isMOVLMask(M, VT) ||
7814 isSHUFPMask(M, VT) ||
7815 isPSHUFDMask(M, VT) ||
7816 isPSHUFHWMask(M, VT) ||
7817 isPSHUFLWMask(M, VT) ||
7818 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7819 isUNPCKLMask(M, VT) ||
7820 isUNPCKHMask(M, VT) ||
7821 isUNPCKL_v_undef_Mask(M, VT) ||
7822 isUNPCKH_v_undef_Mask(M, VT));
7826 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7828 unsigned NumElts = VT.getVectorNumElements();
7829 // FIXME: This collection of masks seems suspect.
7832 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7833 return (isMOVLMask(Mask, VT) ||
7834 isCommutedMOVLMask(Mask, VT, true) ||
7835 isSHUFPMask(Mask, VT) ||
7836 isCommutedSHUFPMask(Mask, VT));
7841 //===----------------------------------------------------------------------===//
7842 // X86 Scheduler Hooks
7843 //===----------------------------------------------------------------------===//
7845 // private utility function
7847 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7848 MachineBasicBlock *MBB,
7856 TargetRegisterClass *RC,
7857 bool invSrc) const {
7858 // For the atomic bitwise operator, we generate
7861 // ld t1 = [bitinstr.addr]
7862 // op t2 = t1, [bitinstr.val]
7864 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7866 // fallthrough -->nextMBB
7867 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7868 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7869 MachineFunction::iterator MBBIter = MBB;
7872 /// First build the CFG
7873 MachineFunction *F = MBB->getParent();
7874 MachineBasicBlock *thisMBB = MBB;
7875 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7876 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7877 F->insert(MBBIter, newMBB);
7878 F->insert(MBBIter, nextMBB);
7880 // Move all successors to thisMBB to nextMBB
7881 nextMBB->transferSuccessors(thisMBB);
7883 // Update thisMBB to fall through to newMBB
7884 thisMBB->addSuccessor(newMBB);
7886 // newMBB jumps to itself and fall through to nextMBB
7887 newMBB->addSuccessor(nextMBB);
7888 newMBB->addSuccessor(newMBB);
7890 // Insert instructions into newMBB based on incoming instruction
7891 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7892 "unexpected number of operands");
7893 DebugLoc dl = bInstr->getDebugLoc();
7894 MachineOperand& destOper = bInstr->getOperand(0);
7895 MachineOperand* argOpers[2 + X86AddrNumOperands];
7896 int numArgs = bInstr->getNumOperands() - 1;
7897 for (int i=0; i < numArgs; ++i)
7898 argOpers[i] = &bInstr->getOperand(i+1);
7900 // x86 address has 4 operands: base, index, scale, and displacement
7901 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7902 int valArgIndx = lastAddrIndx + 1;
7904 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7905 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7906 for (int i=0; i <= lastAddrIndx; ++i)
7907 (*MIB).addOperand(*argOpers[i]);
7909 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7911 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7916 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7917 assert((argOpers[valArgIndx]->isReg() ||
7918 argOpers[valArgIndx]->isImm()) &&
7920 if (argOpers[valArgIndx]->isReg())
7921 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7923 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7925 (*MIB).addOperand(*argOpers[valArgIndx]);
7927 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7930 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7931 for (int i=0; i <= lastAddrIndx; ++i)
7932 (*MIB).addOperand(*argOpers[i]);
7934 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7935 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7936 bInstr->memoperands_end());
7938 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7942 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
7944 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7948 // private utility function: 64 bit atomics on 32 bit host.
7950 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7951 MachineBasicBlock *MBB,
7956 bool invSrc) const {
7957 // For the atomic bitwise operator, we generate
7958 // thisMBB (instructions are in pairs, except cmpxchg8b)
7959 // ld t1,t2 = [bitinstr.addr]
7961 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7962 // op t5, t6 <- out1, out2, [bitinstr.val]
7963 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7964 // mov ECX, EBX <- t5, t6
7965 // mov EAX, EDX <- t1, t2
7966 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7967 // mov t3, t4 <- EAX, EDX
7969 // result in out1, out2
7970 // fallthrough -->nextMBB
7972 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7973 const unsigned LoadOpc = X86::MOV32rm;
7974 const unsigned copyOpc = X86::MOV32rr;
7975 const unsigned NotOpc = X86::NOT32r;
7976 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7977 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7978 MachineFunction::iterator MBBIter = MBB;
7981 /// First build the CFG
7982 MachineFunction *F = MBB->getParent();
7983 MachineBasicBlock *thisMBB = MBB;
7984 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7985 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7986 F->insert(MBBIter, newMBB);
7987 F->insert(MBBIter, nextMBB);
7989 // Move all successors to thisMBB to nextMBB
7990 nextMBB->transferSuccessors(thisMBB);
7992 // Update thisMBB to fall through to newMBB
7993 thisMBB->addSuccessor(newMBB);
7995 // newMBB jumps to itself and fall through to nextMBB
7996 newMBB->addSuccessor(nextMBB);
7997 newMBB->addSuccessor(newMBB);
7999 DebugLoc dl = bInstr->getDebugLoc();
8000 // Insert instructions into newMBB based on incoming instruction
8001 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
8002 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
8003 "unexpected number of operands");
8004 MachineOperand& dest1Oper = bInstr->getOperand(0);
8005 MachineOperand& dest2Oper = bInstr->getOperand(1);
8006 MachineOperand* argOpers[2 + X86AddrNumOperands];
8007 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
8008 argOpers[i] = &bInstr->getOperand(i+2);
8010 // x86 address has 5 operands: base, index, scale, displacement, and segment.
8011 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8013 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
8014 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
8015 for (int i=0; i <= lastAddrIndx; ++i)
8016 (*MIB).addOperand(*argOpers[i]);
8017 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
8018 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
8019 // add 4 to displacement.
8020 for (int i=0; i <= lastAddrIndx-2; ++i)
8021 (*MIB).addOperand(*argOpers[i]);
8022 MachineOperand newOp3 = *(argOpers[3]);
8024 newOp3.setImm(newOp3.getImm()+4);
8026 newOp3.setOffset(newOp3.getOffset()+4);
8027 (*MIB).addOperand(newOp3);
8028 (*MIB).addOperand(*argOpers[lastAddrIndx]);
8030 // t3/4 are defined later, at the bottom of the loop
8031 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
8032 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
8033 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
8034 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
8035 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
8036 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
8038 // The subsequent operations should be using the destination registers of
8039 //the PHI instructions.
8041 t1 = F->getRegInfo().createVirtualRegister(RC);
8042 t2 = F->getRegInfo().createVirtualRegister(RC);
8043 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t1).addReg(dest1Oper.getReg());
8044 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), t2).addReg(dest2Oper.getReg());
8046 t1 = dest1Oper.getReg();
8047 t2 = dest2Oper.getReg();
8050 int valArgIndx = lastAddrIndx + 1;
8051 assert((argOpers[valArgIndx]->isReg() ||
8052 argOpers[valArgIndx]->isImm()) &&
8054 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
8055 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
8056 if (argOpers[valArgIndx]->isReg())
8057 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
8059 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
8060 if (regOpcL != X86::MOV32rr)
8062 (*MIB).addOperand(*argOpers[valArgIndx]);
8063 assert(argOpers[valArgIndx + 1]->isReg() ==
8064 argOpers[valArgIndx]->isReg());
8065 assert(argOpers[valArgIndx + 1]->isImm() ==
8066 argOpers[valArgIndx]->isImm());
8067 if (argOpers[valArgIndx + 1]->isReg())
8068 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
8070 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
8071 if (regOpcH != X86::MOV32rr)
8073 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
8075 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
8077 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
8080 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
8082 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
8085 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
8086 for (int i=0; i <= lastAddrIndx; ++i)
8087 (*MIB).addOperand(*argOpers[i]);
8089 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8090 (*MIB).setMemRefs(bInstr->memoperands_begin(),
8091 bInstr->memoperands_end());
8093 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
8094 MIB.addReg(X86::EAX);
8095 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
8096 MIB.addReg(X86::EDX);
8099 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8101 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
8105 // private utility function
8107 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
8108 MachineBasicBlock *MBB,
8109 unsigned cmovOpc) const {
8110 // For the atomic min/max operator, we generate
8113 // ld t1 = [min/max.addr]
8114 // mov t2 = [min/max.val]
8116 // cmov[cond] t2 = t1
8118 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
8120 // fallthrough -->nextMBB
8122 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8123 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8124 MachineFunction::iterator MBBIter = MBB;
8127 /// First build the CFG
8128 MachineFunction *F = MBB->getParent();
8129 MachineBasicBlock *thisMBB = MBB;
8130 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
8131 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
8132 F->insert(MBBIter, newMBB);
8133 F->insert(MBBIter, nextMBB);
8135 // Move all successors of thisMBB to nextMBB
8136 nextMBB->transferSuccessors(thisMBB);
8138 // Update thisMBB to fall through to newMBB
8139 thisMBB->addSuccessor(newMBB);
8141 // newMBB jumps to newMBB and fall through to nextMBB
8142 newMBB->addSuccessor(nextMBB);
8143 newMBB->addSuccessor(newMBB);
8145 DebugLoc dl = mInstr->getDebugLoc();
8146 // Insert instructions into newMBB based on incoming instruction
8147 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
8148 "unexpected number of operands");
8149 MachineOperand& destOper = mInstr->getOperand(0);
8150 MachineOperand* argOpers[2 + X86AddrNumOperands];
8151 int numArgs = mInstr->getNumOperands() - 1;
8152 for (int i=0; i < numArgs; ++i)
8153 argOpers[i] = &mInstr->getOperand(i+1);
8155 // x86 address has 4 operands: base, index, scale, and displacement
8156 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
8157 int valArgIndx = lastAddrIndx + 1;
8159 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8160 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
8161 for (int i=0; i <= lastAddrIndx; ++i)
8162 (*MIB).addOperand(*argOpers[i]);
8164 // We only support register and immediate values
8165 assert((argOpers[valArgIndx]->isReg() ||
8166 argOpers[valArgIndx]->isImm()) &&
8169 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8170 if (argOpers[valArgIndx]->isReg())
8171 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8173 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
8174 (*MIB).addOperand(*argOpers[valArgIndx]);
8176 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
8179 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
8184 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
8185 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
8189 // Cmp and exchange if none has modified the memory location
8190 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
8191 for (int i=0; i <= lastAddrIndx; ++i)
8192 (*MIB).addOperand(*argOpers[i]);
8194 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
8195 (*MIB).setMemRefs(mInstr->memoperands_begin(),
8196 mInstr->memoperands_end());
8198 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
8199 MIB.addReg(X86::EAX);
8202 BuildMI(newMBB, dl, TII->get(X86::JNE_4)).addMBB(newMBB);
8204 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
8208 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
8209 // all of this code can be replaced with that in the .td file.
8211 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
8212 unsigned numArgs, bool memArg) const {
8214 MachineFunction *F = BB->getParent();
8215 DebugLoc dl = MI->getDebugLoc();
8216 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8220 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
8222 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
8224 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
8226 for (unsigned i = 0; i < numArgs; ++i) {
8227 MachineOperand &Op = MI->getOperand(i+1);
8229 if (!(Op.isReg() && Op.isImplicit()))
8233 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
8236 F->DeleteMachineInstr(MI);
8242 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
8244 MachineBasicBlock *MBB) const {
8245 // Emit code to save XMM registers to the stack. The ABI says that the
8246 // number of registers to save is given in %al, so it's theoretically
8247 // possible to do an indirect jump trick to avoid saving all of them,
8248 // however this code takes a simpler approach and just executes all
8249 // of the stores if %al is non-zero. It's less code, and it's probably
8250 // easier on the hardware branch predictor, and stores aren't all that
8251 // expensive anyway.
8253 // Create the new basic blocks. One block contains all the XMM stores,
8254 // and one block is the final destination regardless of whether any
8255 // stores were performed.
8256 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
8257 MachineFunction *F = MBB->getParent();
8258 MachineFunction::iterator MBBIter = MBB;
8260 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
8261 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
8262 F->insert(MBBIter, XMMSaveMBB);
8263 F->insert(MBBIter, EndMBB);
8266 // Move any original successors of MBB to the end block.
8267 EndMBB->transferSuccessors(MBB);
8268 // The original block will now fall through to the XMM save block.
8269 MBB->addSuccessor(XMMSaveMBB);
8270 // The XMMSaveMBB will fall through to the end block.
8271 XMMSaveMBB->addSuccessor(EndMBB);
8273 // Now add the instructions.
8274 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8275 DebugLoc DL = MI->getDebugLoc();
8277 unsigned CountReg = MI->getOperand(0).getReg();
8278 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
8279 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
8281 if (!Subtarget->isTargetWin64()) {
8282 // If %al is 0, branch around the XMM save block.
8283 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
8284 BuildMI(MBB, DL, TII->get(X86::JE_4)).addMBB(EndMBB);
8285 MBB->addSuccessor(EndMBB);
8288 // In the XMM save block, save all the XMM argument registers.
8289 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
8290 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
8291 MachineMemOperand *MMO =
8292 F->getMachineMemOperand(
8293 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
8294 MachineMemOperand::MOStore, Offset,
8295 /*Size=*/16, /*Align=*/16);
8296 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
8297 .addFrameIndex(RegSaveFrameIndex)
8298 .addImm(/*Scale=*/1)
8299 .addReg(/*IndexReg=*/0)
8300 .addImm(/*Disp=*/Offset)
8301 .addReg(/*Segment=*/0)
8302 .addReg(MI->getOperand(i).getReg())
8303 .addMemOperand(MMO);
8306 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8312 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
8313 MachineBasicBlock *BB,
8314 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8315 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8316 DebugLoc DL = MI->getDebugLoc();
8318 // To "insert" a SELECT_CC instruction, we actually have to insert the
8319 // diamond control-flow pattern. The incoming instruction knows the
8320 // destination vreg to set, the condition code register to branch on, the
8321 // true/false values to select between, and a branch opcode to use.
8322 const BasicBlock *LLVM_BB = BB->getBasicBlock();
8323 MachineFunction::iterator It = BB;
8329 // cmpTY ccX, r1, r2
8331 // fallthrough --> copy0MBB
8332 MachineBasicBlock *thisMBB = BB;
8333 MachineFunction *F = BB->getParent();
8334 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
8335 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
8337 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
8338 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
8339 F->insert(It, copy0MBB);
8340 F->insert(It, sinkMBB);
8341 // Update machine-CFG edges by first adding all successors of the current
8342 // block to the new block which will contain the Phi node for the select.
8343 // Also inform sdisel of the edge changes.
8344 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
8345 E = BB->succ_end(); I != E; ++I) {
8346 EM->insert(std::make_pair(*I, sinkMBB));
8347 sinkMBB->addSuccessor(*I);
8349 // Next, remove all successors of the current block, and add the true
8350 // and fallthrough blocks as its successors.
8351 while (!BB->succ_empty())
8352 BB->removeSuccessor(BB->succ_begin());
8353 // Add the true and fallthrough blocks as its successors.
8354 BB->addSuccessor(copy0MBB);
8355 BB->addSuccessor(sinkMBB);
8358 // %FalseValue = ...
8359 // # fallthrough to sinkMBB
8362 // Update machine-CFG edges
8363 BB->addSuccessor(sinkMBB);
8366 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
8369 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
8370 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
8371 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
8373 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8379 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
8380 MachineBasicBlock *BB,
8381 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
8382 switch (MI->getOpcode()) {
8383 default: assert(false && "Unexpected instr type to insert");
8385 case X86::CMOV_V1I64:
8386 case X86::CMOV_FR32:
8387 case X86::CMOV_FR64:
8388 case X86::CMOV_V4F32:
8389 case X86::CMOV_V2F64:
8390 case X86::CMOV_V2I64:
8391 return EmitLoweredSelect(MI, BB, EM);
8393 case X86::FP32_TO_INT16_IN_MEM:
8394 case X86::FP32_TO_INT32_IN_MEM:
8395 case X86::FP32_TO_INT64_IN_MEM:
8396 case X86::FP64_TO_INT16_IN_MEM:
8397 case X86::FP64_TO_INT32_IN_MEM:
8398 case X86::FP64_TO_INT64_IN_MEM:
8399 case X86::FP80_TO_INT16_IN_MEM:
8400 case X86::FP80_TO_INT32_IN_MEM:
8401 case X86::FP80_TO_INT64_IN_MEM: {
8402 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
8403 DebugLoc DL = MI->getDebugLoc();
8405 // Change the floating point control register to use "round towards zero"
8406 // mode when truncating to an integer value.
8407 MachineFunction *F = BB->getParent();
8408 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
8409 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
8411 // Load the old value of the high byte of the control word...
8413 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
8414 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8417 // Set the high part to be round to zero...
8418 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8421 // Reload the modified control word now...
8422 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8424 // Restore the memory image of control word to original value
8425 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8428 // Get the X86 opcode to use.
8430 switch (MI->getOpcode()) {
8431 default: llvm_unreachable("illegal opcode!");
8432 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8433 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8434 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8435 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8436 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8437 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8438 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8439 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8440 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8444 MachineOperand &Op = MI->getOperand(0);
8446 AM.BaseType = X86AddressMode::RegBase;
8447 AM.Base.Reg = Op.getReg();
8449 AM.BaseType = X86AddressMode::FrameIndexBase;
8450 AM.Base.FrameIndex = Op.getIndex();
8452 Op = MI->getOperand(1);
8454 AM.Scale = Op.getImm();
8455 Op = MI->getOperand(2);
8457 AM.IndexReg = Op.getImm();
8458 Op = MI->getOperand(3);
8459 if (Op.isGlobal()) {
8460 AM.GV = Op.getGlobal();
8462 AM.Disp = Op.getImm();
8464 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8465 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8467 // Reload the original control word now.
8468 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8470 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8473 // String/text processing lowering.
8474 case X86::PCMPISTRM128REG:
8475 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8476 case X86::PCMPISTRM128MEM:
8477 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8478 case X86::PCMPESTRM128REG:
8479 return EmitPCMP(MI, BB, 5, false /* in mem */);
8480 case X86::PCMPESTRM128MEM:
8481 return EmitPCMP(MI, BB, 5, true /* in mem */);
8484 case X86::ATOMAND32:
8485 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8486 X86::AND32ri, X86::MOV32rm,
8487 X86::LCMPXCHG32, X86::MOV32rr,
8488 X86::NOT32r, X86::EAX,
8489 X86::GR32RegisterClass);
8491 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8492 X86::OR32ri, X86::MOV32rm,
8493 X86::LCMPXCHG32, X86::MOV32rr,
8494 X86::NOT32r, X86::EAX,
8495 X86::GR32RegisterClass);
8496 case X86::ATOMXOR32:
8497 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8498 X86::XOR32ri, X86::MOV32rm,
8499 X86::LCMPXCHG32, X86::MOV32rr,
8500 X86::NOT32r, X86::EAX,
8501 X86::GR32RegisterClass);
8502 case X86::ATOMNAND32:
8503 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8504 X86::AND32ri, X86::MOV32rm,
8505 X86::LCMPXCHG32, X86::MOV32rr,
8506 X86::NOT32r, X86::EAX,
8507 X86::GR32RegisterClass, true);
8508 case X86::ATOMMIN32:
8509 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8510 case X86::ATOMMAX32:
8511 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8512 case X86::ATOMUMIN32:
8513 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8514 case X86::ATOMUMAX32:
8515 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8517 case X86::ATOMAND16:
8518 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8519 X86::AND16ri, X86::MOV16rm,
8520 X86::LCMPXCHG16, X86::MOV16rr,
8521 X86::NOT16r, X86::AX,
8522 X86::GR16RegisterClass);
8524 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8525 X86::OR16ri, X86::MOV16rm,
8526 X86::LCMPXCHG16, X86::MOV16rr,
8527 X86::NOT16r, X86::AX,
8528 X86::GR16RegisterClass);
8529 case X86::ATOMXOR16:
8530 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8531 X86::XOR16ri, X86::MOV16rm,
8532 X86::LCMPXCHG16, X86::MOV16rr,
8533 X86::NOT16r, X86::AX,
8534 X86::GR16RegisterClass);
8535 case X86::ATOMNAND16:
8536 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8537 X86::AND16ri, X86::MOV16rm,
8538 X86::LCMPXCHG16, X86::MOV16rr,
8539 X86::NOT16r, X86::AX,
8540 X86::GR16RegisterClass, true);
8541 case X86::ATOMMIN16:
8542 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8543 case X86::ATOMMAX16:
8544 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8545 case X86::ATOMUMIN16:
8546 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8547 case X86::ATOMUMAX16:
8548 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8551 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8552 X86::AND8ri, X86::MOV8rm,
8553 X86::LCMPXCHG8, X86::MOV8rr,
8554 X86::NOT8r, X86::AL,
8555 X86::GR8RegisterClass);
8557 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8558 X86::OR8ri, X86::MOV8rm,
8559 X86::LCMPXCHG8, X86::MOV8rr,
8560 X86::NOT8r, X86::AL,
8561 X86::GR8RegisterClass);
8563 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8564 X86::XOR8ri, X86::MOV8rm,
8565 X86::LCMPXCHG8, X86::MOV8rr,
8566 X86::NOT8r, X86::AL,
8567 X86::GR8RegisterClass);
8568 case X86::ATOMNAND8:
8569 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8570 X86::AND8ri, X86::MOV8rm,
8571 X86::LCMPXCHG8, X86::MOV8rr,
8572 X86::NOT8r, X86::AL,
8573 X86::GR8RegisterClass, true);
8574 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8575 // This group is for 64-bit host.
8576 case X86::ATOMAND64:
8577 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8578 X86::AND64ri32, X86::MOV64rm,
8579 X86::LCMPXCHG64, X86::MOV64rr,
8580 X86::NOT64r, X86::RAX,
8581 X86::GR64RegisterClass);
8583 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8584 X86::OR64ri32, X86::MOV64rm,
8585 X86::LCMPXCHG64, X86::MOV64rr,
8586 X86::NOT64r, X86::RAX,
8587 X86::GR64RegisterClass);
8588 case X86::ATOMXOR64:
8589 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8590 X86::XOR64ri32, X86::MOV64rm,
8591 X86::LCMPXCHG64, X86::MOV64rr,
8592 X86::NOT64r, X86::RAX,
8593 X86::GR64RegisterClass);
8594 case X86::ATOMNAND64:
8595 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8596 X86::AND64ri32, X86::MOV64rm,
8597 X86::LCMPXCHG64, X86::MOV64rr,
8598 X86::NOT64r, X86::RAX,
8599 X86::GR64RegisterClass, true);
8600 case X86::ATOMMIN64:
8601 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8602 case X86::ATOMMAX64:
8603 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8604 case X86::ATOMUMIN64:
8605 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8606 case X86::ATOMUMAX64:
8607 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8609 // This group does 64-bit operations on a 32-bit host.
8610 case X86::ATOMAND6432:
8611 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8612 X86::AND32rr, X86::AND32rr,
8613 X86::AND32ri, X86::AND32ri,
8615 case X86::ATOMOR6432:
8616 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8617 X86::OR32rr, X86::OR32rr,
8618 X86::OR32ri, X86::OR32ri,
8620 case X86::ATOMXOR6432:
8621 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8622 X86::XOR32rr, X86::XOR32rr,
8623 X86::XOR32ri, X86::XOR32ri,
8625 case X86::ATOMNAND6432:
8626 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8627 X86::AND32rr, X86::AND32rr,
8628 X86::AND32ri, X86::AND32ri,
8630 case X86::ATOMADD6432:
8631 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8632 X86::ADD32rr, X86::ADC32rr,
8633 X86::ADD32ri, X86::ADC32ri,
8635 case X86::ATOMSUB6432:
8636 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8637 X86::SUB32rr, X86::SBB32rr,
8638 X86::SUB32ri, X86::SBB32ri,
8640 case X86::ATOMSWAP6432:
8641 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8642 X86::MOV32rr, X86::MOV32rr,
8643 X86::MOV32ri, X86::MOV32ri,
8645 case X86::VASTART_SAVE_XMM_REGS:
8646 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8650 //===----------------------------------------------------------------------===//
8651 // X86 Optimization Hooks
8652 //===----------------------------------------------------------------------===//
8654 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8658 const SelectionDAG &DAG,
8659 unsigned Depth) const {
8660 unsigned Opc = Op.getOpcode();
8661 assert((Opc >= ISD::BUILTIN_OP_END ||
8662 Opc == ISD::INTRINSIC_WO_CHAIN ||
8663 Opc == ISD::INTRINSIC_W_CHAIN ||
8664 Opc == ISD::INTRINSIC_VOID) &&
8665 "Should use MaskedValueIsZero if you don't know whether Op"
8666 " is a target node!");
8668 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8680 // These nodes' second result is a boolean.
8681 if (Op.getResNo() == 0)
8685 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8686 Mask.getBitWidth() - 1);
8691 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8692 /// node is a GlobalAddress + offset.
8693 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8694 GlobalValue* &GA, int64_t &Offset) const{
8695 if (N->getOpcode() == X86ISD::Wrapper) {
8696 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8697 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8698 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8702 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8705 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8706 EVT EltVT, LoadSDNode *&LDBase,
8707 unsigned &LastLoadedElt,
8708 SelectionDAG &DAG, MachineFrameInfo *MFI,
8709 const TargetLowering &TLI) {
8711 LastLoadedElt = -1U;
8712 for (unsigned i = 0; i < NumElems; ++i) {
8713 if (N->getMaskElt(i) < 0) {
8719 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8720 if (!Elt.getNode() ||
8721 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8724 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8726 LDBase = cast<LoadSDNode>(Elt.getNode());
8730 if (Elt.getOpcode() == ISD::UNDEF)
8733 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8734 if (!DAG.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i))
8741 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8742 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8743 /// if the load addresses are consecutive, non-overlapping, and in the right
8744 /// order. In the case of v2i64, it will see if it can rewrite the
8745 /// shuffle to be an appropriate build vector so it can take advantage of
8746 // performBuildVectorCombine.
8747 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8748 const TargetLowering &TLI) {
8749 DebugLoc dl = N->getDebugLoc();
8750 EVT VT = N->getValueType(0);
8751 EVT EltVT = VT.getVectorElementType();
8752 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8753 unsigned NumElems = VT.getVectorNumElements();
8755 if (VT.getSizeInBits() != 128)
8758 // Try to combine a vector_shuffle into a 128-bit load.
8759 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8760 LoadSDNode *LD = NULL;
8761 unsigned LastLoadedElt;
8762 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8766 if (LastLoadedElt == NumElems - 1) {
8767 if (DAG.InferPtrAlignment(LD->getBasePtr()) >= 16)
8768 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8769 LD->getSrcValue(), LD->getSrcValueOffset(),
8770 LD->isVolatile(), LD->isNonTemporal(), 0);
8771 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8772 LD->getSrcValue(), LD->getSrcValueOffset(),
8773 LD->isVolatile(), LD->isNonTemporal(),
8774 LD->getAlignment());
8775 } else if (NumElems == 4 && LastLoadedElt == 1) {
8776 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8777 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8778 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8779 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8784 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8785 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8786 const X86Subtarget *Subtarget) {
8787 DebugLoc DL = N->getDebugLoc();
8788 SDValue Cond = N->getOperand(0);
8789 // Get the LHS/RHS of the select.
8790 SDValue LHS = N->getOperand(1);
8791 SDValue RHS = N->getOperand(2);
8793 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8794 // instructions have the peculiarity that if either operand is a NaN,
8795 // they chose what we call the RHS operand (and as such are not symmetric).
8796 // It happens that this matches the semantics of the common C idiom
8797 // x<y?x:y and related forms, so we can recognize these cases.
8798 if (Subtarget->hasSSE2() &&
8799 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8800 Cond.getOpcode() == ISD::SETCC) {
8801 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8803 unsigned Opcode = 0;
8804 // Check for x CC y ? x : y.
8805 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8809 // This can be a min if we can prove that at least one of the operands
8811 if (!FiniteOnlyFPMath()) {
8812 if (DAG.isKnownNeverNaN(RHS)) {
8813 // Put the potential NaN in the RHS so that SSE will preserve it.
8814 std::swap(LHS, RHS);
8815 } else if (!DAG.isKnownNeverNaN(LHS))
8818 Opcode = X86ISD::FMIN;
8821 // This can be a min if we can prove that at least one of the operands
8823 if (!FiniteOnlyFPMath()) {
8824 if (DAG.isKnownNeverNaN(LHS)) {
8825 // Put the potential NaN in the RHS so that SSE will preserve it.
8826 std::swap(LHS, RHS);
8827 } else if (!DAG.isKnownNeverNaN(RHS))
8830 Opcode = X86ISD::FMIN;
8833 // This can be a min, but if either operand is a NaN we need it to
8834 // preserve the original LHS.
8835 std::swap(LHS, RHS);
8839 Opcode = X86ISD::FMIN;
8843 // This can be a max if we can prove that at least one of the operands
8845 if (!FiniteOnlyFPMath()) {
8846 if (DAG.isKnownNeverNaN(LHS)) {
8847 // Put the potential NaN in the RHS so that SSE will preserve it.
8848 std::swap(LHS, RHS);
8849 } else if (!DAG.isKnownNeverNaN(RHS))
8852 Opcode = X86ISD::FMAX;
8855 // This can be a max if we can prove that at least one of the operands
8857 if (!FiniteOnlyFPMath()) {
8858 if (DAG.isKnownNeverNaN(RHS)) {
8859 // Put the potential NaN in the RHS so that SSE will preserve it.
8860 std::swap(LHS, RHS);
8861 } else if (!DAG.isKnownNeverNaN(LHS))
8864 Opcode = X86ISD::FMAX;
8867 // This can be a max, but if either operand is a NaN we need it to
8868 // preserve the original LHS.
8869 std::swap(LHS, RHS);
8873 Opcode = X86ISD::FMAX;
8876 // Check for x CC y ? y : x -- a min/max with reversed arms.
8877 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8881 // This can be a min if we can prove that at least one of the operands
8883 if (!FiniteOnlyFPMath()) {
8884 if (DAG.isKnownNeverNaN(RHS)) {
8885 // Put the potential NaN in the RHS so that SSE will preserve it.
8886 std::swap(LHS, RHS);
8887 } else if (!DAG.isKnownNeverNaN(LHS))
8890 Opcode = X86ISD::FMIN;
8893 // This can be a min if we can prove that at least one of the operands
8895 if (!FiniteOnlyFPMath()) {
8896 if (DAG.isKnownNeverNaN(LHS)) {
8897 // Put the potential NaN in the RHS so that SSE will preserve it.
8898 std::swap(LHS, RHS);
8899 } else if (!DAG.isKnownNeverNaN(RHS))
8902 Opcode = X86ISD::FMIN;
8905 // This can be a min, but if either operand is a NaN we need it to
8906 // preserve the original LHS.
8907 std::swap(LHS, RHS);
8911 Opcode = X86ISD::FMIN;
8915 // This can be a max if we can prove that at least one of the operands
8917 if (!FiniteOnlyFPMath()) {
8918 if (DAG.isKnownNeverNaN(LHS)) {
8919 // Put the potential NaN in the RHS so that SSE will preserve it.
8920 std::swap(LHS, RHS);
8921 } else if (!DAG.isKnownNeverNaN(RHS))
8924 Opcode = X86ISD::FMAX;
8927 // This can be a max if we can prove that at least one of the operands
8929 if (!FiniteOnlyFPMath()) {
8930 if (DAG.isKnownNeverNaN(RHS)) {
8931 // Put the potential NaN in the RHS so that SSE will preserve it.
8932 std::swap(LHS, RHS);
8933 } else if (!DAG.isKnownNeverNaN(LHS))
8936 Opcode = X86ISD::FMAX;
8939 // This can be a max, but if either operand is a NaN we need it to
8940 // preserve the original LHS.
8941 std::swap(LHS, RHS);
8945 Opcode = X86ISD::FMAX;
8951 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8954 // If this is a select between two integer constants, try to do some
8956 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8957 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8958 // Don't do this for crazy integer types.
8959 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8960 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8961 // so that TrueC (the true value) is larger than FalseC.
8962 bool NeedsCondInvert = false;
8964 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8965 // Efficiently invertible.
8966 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8967 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8968 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8969 NeedsCondInvert = true;
8970 std::swap(TrueC, FalseC);
8973 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8974 if (FalseC->getAPIntValue() == 0 &&
8975 TrueC->getAPIntValue().isPowerOf2()) {
8976 if (NeedsCondInvert) // Invert the condition if needed.
8977 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8978 DAG.getConstant(1, Cond.getValueType()));
8980 // Zero extend the condition if needed.
8981 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8983 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8984 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8985 DAG.getConstant(ShAmt, MVT::i8));
8988 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8989 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8990 if (NeedsCondInvert) // Invert the condition if needed.
8991 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8992 DAG.getConstant(1, Cond.getValueType()));
8994 // Zero extend the condition if needed.
8995 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8996 FalseC->getValueType(0), Cond);
8997 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8998 SDValue(FalseC, 0));
9001 // Optimize cases that will turn into an LEA instruction. This requires
9002 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9003 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9004 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9005 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9007 bool isFastMultiplier = false;
9009 switch ((unsigned char)Diff) {
9011 case 1: // result = add base, cond
9012 case 2: // result = lea base( , cond*2)
9013 case 3: // result = lea base(cond, cond*2)
9014 case 4: // result = lea base( , cond*4)
9015 case 5: // result = lea base(cond, cond*4)
9016 case 8: // result = lea base( , cond*8)
9017 case 9: // result = lea base(cond, cond*8)
9018 isFastMultiplier = true;
9023 if (isFastMultiplier) {
9024 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9025 if (NeedsCondInvert) // Invert the condition if needed.
9026 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
9027 DAG.getConstant(1, Cond.getValueType()));
9029 // Zero extend the condition if needed.
9030 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9032 // Scale the condition by the difference.
9034 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9035 DAG.getConstant(Diff, Cond.getValueType()));
9037 // Add the base if non-zero.
9038 if (FalseC->getAPIntValue() != 0)
9039 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9040 SDValue(FalseC, 0));
9050 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
9051 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
9052 TargetLowering::DAGCombinerInfo &DCI) {
9053 DebugLoc DL = N->getDebugLoc();
9055 // If the flag operand isn't dead, don't touch this CMOV.
9056 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
9059 // If this is a select between two integer constants, try to do some
9060 // optimizations. Note that the operands are ordered the opposite of SELECT
9062 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
9063 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
9064 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
9065 // larger than FalseC (the false value).
9066 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
9068 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
9069 CC = X86::GetOppositeBranchCondition(CC);
9070 std::swap(TrueC, FalseC);
9073 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
9074 // This is efficient for any integer data type (including i8/i16) and
9076 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
9077 SDValue Cond = N->getOperand(3);
9078 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9079 DAG.getConstant(CC, MVT::i8), Cond);
9081 // Zero extend the condition if needed.
9082 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
9084 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
9085 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
9086 DAG.getConstant(ShAmt, MVT::i8));
9087 if (N->getNumValues() == 2) // Dead flag value?
9088 return DCI.CombineTo(N, Cond, SDValue());
9092 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
9093 // for any integer data type, including i8/i16.
9094 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
9095 SDValue Cond = N->getOperand(3);
9096 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9097 DAG.getConstant(CC, MVT::i8), Cond);
9099 // Zero extend the condition if needed.
9100 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
9101 FalseC->getValueType(0), Cond);
9102 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9103 SDValue(FalseC, 0));
9105 if (N->getNumValues() == 2) // Dead flag value?
9106 return DCI.CombineTo(N, Cond, SDValue());
9110 // Optimize cases that will turn into an LEA instruction. This requires
9111 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
9112 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
9113 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
9114 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
9116 bool isFastMultiplier = false;
9118 switch ((unsigned char)Diff) {
9120 case 1: // result = add base, cond
9121 case 2: // result = lea base( , cond*2)
9122 case 3: // result = lea base(cond, cond*2)
9123 case 4: // result = lea base( , cond*4)
9124 case 5: // result = lea base(cond, cond*4)
9125 case 8: // result = lea base( , cond*8)
9126 case 9: // result = lea base(cond, cond*8)
9127 isFastMultiplier = true;
9132 if (isFastMultiplier) {
9133 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
9134 SDValue Cond = N->getOperand(3);
9135 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
9136 DAG.getConstant(CC, MVT::i8), Cond);
9137 // Zero extend the condition if needed.
9138 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
9140 // Scale the condition by the difference.
9142 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
9143 DAG.getConstant(Diff, Cond.getValueType()));
9145 // Add the base if non-zero.
9146 if (FalseC->getAPIntValue() != 0)
9147 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
9148 SDValue(FalseC, 0));
9149 if (N->getNumValues() == 2) // Dead flag value?
9150 return DCI.CombineTo(N, Cond, SDValue());
9160 /// PerformMulCombine - Optimize a single multiply with constant into two
9161 /// in order to implement it with two cheaper instructions, e.g.
9162 /// LEA + SHL, LEA + LEA.
9163 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
9164 TargetLowering::DAGCombinerInfo &DCI) {
9165 if (DAG.getMachineFunction().
9166 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
9169 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
9172 EVT VT = N->getValueType(0);
9176 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
9179 uint64_t MulAmt = C->getZExtValue();
9180 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
9183 uint64_t MulAmt1 = 0;
9184 uint64_t MulAmt2 = 0;
9185 if ((MulAmt % 9) == 0) {
9187 MulAmt2 = MulAmt / 9;
9188 } else if ((MulAmt % 5) == 0) {
9190 MulAmt2 = MulAmt / 5;
9191 } else if ((MulAmt % 3) == 0) {
9193 MulAmt2 = MulAmt / 3;
9196 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
9197 DebugLoc DL = N->getDebugLoc();
9199 if (isPowerOf2_64(MulAmt2) &&
9200 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
9201 // If second multiplifer is pow2, issue it first. We want the multiply by
9202 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
9204 std::swap(MulAmt1, MulAmt2);
9207 if (isPowerOf2_64(MulAmt1))
9208 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
9209 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
9211 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
9212 DAG.getConstant(MulAmt1, VT));
9214 if (isPowerOf2_64(MulAmt2))
9215 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
9216 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
9218 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
9219 DAG.getConstant(MulAmt2, VT));
9221 // Do not add new nodes to DAG combiner worklist.
9222 DCI.CombineTo(N, NewMul, false);
9227 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
9228 SDValue N0 = N->getOperand(0);
9229 SDValue N1 = N->getOperand(1);
9230 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
9231 EVT VT = N0.getValueType();
9233 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
9234 // since the result of setcc_c is all zero's or all ones.
9235 if (N1C && N0.getOpcode() == ISD::AND &&
9236 N0.getOperand(1).getOpcode() == ISD::Constant) {
9237 SDValue N00 = N0.getOperand(0);
9238 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
9239 ((N00.getOpcode() == ISD::ANY_EXTEND ||
9240 N00.getOpcode() == ISD::ZERO_EXTEND) &&
9241 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
9242 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
9243 APInt ShAmt = N1C->getAPIntValue();
9244 Mask = Mask.shl(ShAmt);
9246 return DAG.getNode(ISD::AND, N->getDebugLoc(), VT,
9247 N00, DAG.getConstant(Mask, VT));
9254 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
9256 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
9257 const X86Subtarget *Subtarget) {
9258 EVT VT = N->getValueType(0);
9259 if (!VT.isVector() && VT.isInteger() &&
9260 N->getOpcode() == ISD::SHL)
9261 return PerformSHLCombine(N, DAG);
9263 // On X86 with SSE2 support, we can transform this to a vector shift if
9264 // all elements are shifted by the same amount. We can't do this in legalize
9265 // because the a constant vector is typically transformed to a constant pool
9266 // so we have no knowledge of the shift amount.
9267 if (!Subtarget->hasSSE2())
9270 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
9273 SDValue ShAmtOp = N->getOperand(1);
9274 EVT EltVT = VT.getVectorElementType();
9275 DebugLoc DL = N->getDebugLoc();
9276 SDValue BaseShAmt = SDValue();
9277 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
9278 unsigned NumElts = VT.getVectorNumElements();
9280 for (; i != NumElts; ++i) {
9281 SDValue Arg = ShAmtOp.getOperand(i);
9282 if (Arg.getOpcode() == ISD::UNDEF) continue;
9286 for (; i != NumElts; ++i) {
9287 SDValue Arg = ShAmtOp.getOperand(i);
9288 if (Arg.getOpcode() == ISD::UNDEF) continue;
9289 if (Arg != BaseShAmt) {
9293 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
9294 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
9295 SDValue InVec = ShAmtOp.getOperand(0);
9296 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
9297 unsigned NumElts = InVec.getValueType().getVectorNumElements();
9299 for (; i != NumElts; ++i) {
9300 SDValue Arg = InVec.getOperand(i);
9301 if (Arg.getOpcode() == ISD::UNDEF) continue;
9305 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
9306 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
9307 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
9308 if (C->getZExtValue() == SplatIdx)
9309 BaseShAmt = InVec.getOperand(1);
9312 if (BaseShAmt.getNode() == 0)
9313 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
9314 DAG.getIntPtrConstant(0));
9318 // The shift amount is an i32.
9319 if (EltVT.bitsGT(MVT::i32))
9320 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
9321 else if (EltVT.bitsLT(MVT::i32))
9322 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
9324 // The shift amount is identical so we can do a vector shift.
9325 SDValue ValOp = N->getOperand(0);
9326 switch (N->getOpcode()) {
9328 llvm_unreachable("Unknown shift opcode!");
9331 if (VT == MVT::v2i64)
9332 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9333 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
9335 if (VT == MVT::v4i32)
9336 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9337 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
9339 if (VT == MVT::v8i16)
9340 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9341 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
9345 if (VT == MVT::v4i32)
9346 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9347 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
9349 if (VT == MVT::v8i16)
9350 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9351 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
9355 if (VT == MVT::v2i64)
9356 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9357 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
9359 if (VT == MVT::v4i32)
9360 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9361 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
9363 if (VT == MVT::v8i16)
9364 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
9365 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
9372 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
9373 const X86Subtarget *Subtarget) {
9374 EVT VT = N->getValueType(0);
9375 if (VT != MVT::i64 || !Subtarget->is64Bit())
9378 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
9379 SDValue N0 = N->getOperand(0);
9380 SDValue N1 = N->getOperand(1);
9381 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
9383 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
9386 SDValue ShAmt0 = N0.getOperand(1);
9387 if (ShAmt0.getValueType() != MVT::i8)
9389 SDValue ShAmt1 = N1.getOperand(1);
9390 if (ShAmt1.getValueType() != MVT::i8)
9392 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
9393 ShAmt0 = ShAmt0.getOperand(0);
9394 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
9395 ShAmt1 = ShAmt1.getOperand(0);
9397 DebugLoc DL = N->getDebugLoc();
9398 unsigned Opc = X86ISD::SHLD;
9399 SDValue Op0 = N0.getOperand(0);
9400 SDValue Op1 = N1.getOperand(0);
9401 if (ShAmt0.getOpcode() == ISD::SUB) {
9403 std::swap(Op0, Op1);
9404 std::swap(ShAmt0, ShAmt1);
9407 if (ShAmt1.getOpcode() == ISD::SUB) {
9408 SDValue Sum = ShAmt1.getOperand(0);
9409 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
9410 if (SumC->getSExtValue() == 64 &&
9411 ShAmt1.getOperand(1) == ShAmt0)
9412 return DAG.getNode(Opc, DL, VT,
9414 DAG.getNode(ISD::TRUNCATE, DL,
9417 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
9418 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
9420 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == 64)
9421 return DAG.getNode(Opc, DL, VT,
9422 N0.getOperand(0), N1.getOperand(0),
9423 DAG.getNode(ISD::TRUNCATE, DL,
9430 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
9431 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
9432 const X86Subtarget *Subtarget) {
9433 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
9434 // the FP state in cases where an emms may be missing.
9435 // A preferable solution to the general problem is to figure out the right
9436 // places to insert EMMS. This qualifies as a quick hack.
9438 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
9439 StoreSDNode *St = cast<StoreSDNode>(N);
9440 EVT VT = St->getValue().getValueType();
9441 if (VT.getSizeInBits() != 64)
9444 const Function *F = DAG.getMachineFunction().getFunction();
9445 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
9446 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
9447 && Subtarget->hasSSE2();
9448 if ((VT.isVector() ||
9449 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
9450 isa<LoadSDNode>(St->getValue()) &&
9451 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
9452 St->getChain().hasOneUse() && !St->isVolatile()) {
9453 SDNode* LdVal = St->getValue().getNode();
9455 int TokenFactorIndex = -1;
9456 SmallVector<SDValue, 8> Ops;
9457 SDNode* ChainVal = St->getChain().getNode();
9458 // Must be a store of a load. We currently handle two cases: the load
9459 // is a direct child, and it's under an intervening TokenFactor. It is
9460 // possible to dig deeper under nested TokenFactors.
9461 if (ChainVal == LdVal)
9462 Ld = cast<LoadSDNode>(St->getChain());
9463 else if (St->getValue().hasOneUse() &&
9464 ChainVal->getOpcode() == ISD::TokenFactor) {
9465 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
9466 if (ChainVal->getOperand(i).getNode() == LdVal) {
9467 TokenFactorIndex = i;
9468 Ld = cast<LoadSDNode>(St->getValue());
9470 Ops.push_back(ChainVal->getOperand(i));
9474 if (!Ld || !ISD::isNormalLoad(Ld))
9477 // If this is not the MMX case, i.e. we are just turning i64 load/store
9478 // into f64 load/store, avoid the transformation if there are multiple
9479 // uses of the loaded value.
9480 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
9483 DebugLoc LdDL = Ld->getDebugLoc();
9484 DebugLoc StDL = N->getDebugLoc();
9485 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
9486 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
9488 if (Subtarget->is64Bit() || F64IsLegal) {
9489 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
9490 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
9491 Ld->getBasePtr(), Ld->getSrcValue(),
9492 Ld->getSrcValueOffset(), Ld->isVolatile(),
9493 Ld->isNonTemporal(), Ld->getAlignment());
9494 SDValue NewChain = NewLd.getValue(1);
9495 if (TokenFactorIndex != -1) {
9496 Ops.push_back(NewChain);
9497 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9500 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9501 St->getSrcValue(), St->getSrcValueOffset(),
9502 St->isVolatile(), St->isNonTemporal(),
9503 St->getAlignment());
9506 // Otherwise, lower to two pairs of 32-bit loads / stores.
9507 SDValue LoAddr = Ld->getBasePtr();
9508 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9509 DAG.getConstant(4, MVT::i32));
9511 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9512 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9513 Ld->isVolatile(), Ld->isNonTemporal(),
9514 Ld->getAlignment());
9515 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9516 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9517 Ld->isVolatile(), Ld->isNonTemporal(),
9518 MinAlign(Ld->getAlignment(), 4));
9520 SDValue NewChain = LoLd.getValue(1);
9521 if (TokenFactorIndex != -1) {
9522 Ops.push_back(LoLd);
9523 Ops.push_back(HiLd);
9524 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9528 LoAddr = St->getBasePtr();
9529 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9530 DAG.getConstant(4, MVT::i32));
9532 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9533 St->getSrcValue(), St->getSrcValueOffset(),
9534 St->isVolatile(), St->isNonTemporal(),
9535 St->getAlignment());
9536 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9538 St->getSrcValueOffset() + 4,
9540 St->isNonTemporal(),
9541 MinAlign(St->getAlignment(), 4));
9542 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9547 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9548 /// X86ISD::FXOR nodes.
9549 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9550 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9551 // F[X]OR(0.0, x) -> x
9552 // F[X]OR(x, 0.0) -> x
9553 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9554 if (C->getValueAPF().isPosZero())
9555 return N->getOperand(1);
9556 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9557 if (C->getValueAPF().isPosZero())
9558 return N->getOperand(0);
9562 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9563 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9564 // FAND(0.0, x) -> 0.0
9565 // FAND(x, 0.0) -> 0.0
9566 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9567 if (C->getValueAPF().isPosZero())
9568 return N->getOperand(0);
9569 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9570 if (C->getValueAPF().isPosZero())
9571 return N->getOperand(1);
9575 static SDValue PerformBTCombine(SDNode *N,
9577 TargetLowering::DAGCombinerInfo &DCI) {
9578 // BT ignores high bits in the bit index operand.
9579 SDValue Op1 = N->getOperand(1);
9580 if (Op1.hasOneUse()) {
9581 unsigned BitWidth = Op1.getValueSizeInBits();
9582 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9583 APInt KnownZero, KnownOne;
9584 TargetLowering::TargetLoweringOpt TLO(DAG);
9585 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9586 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9587 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9588 DCI.CommitTargetLoweringOpt(TLO);
9593 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9594 SDValue Op = N->getOperand(0);
9595 if (Op.getOpcode() == ISD::BIT_CONVERT)
9596 Op = Op.getOperand(0);
9597 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9598 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9599 VT.getVectorElementType().getSizeInBits() ==
9600 OpVT.getVectorElementType().getSizeInBits()) {
9601 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9606 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9607 // Locked instructions, in turn, have implicit fence semantics (all memory
9608 // operations are flushed before issuing the locked instruction, and the
9609 // are not buffered), so we can fold away the common pattern of
9610 // fence-atomic-fence.
9611 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9612 SDValue atomic = N->getOperand(0);
9613 switch (atomic.getOpcode()) {
9614 case ISD::ATOMIC_CMP_SWAP:
9615 case ISD::ATOMIC_SWAP:
9616 case ISD::ATOMIC_LOAD_ADD:
9617 case ISD::ATOMIC_LOAD_SUB:
9618 case ISD::ATOMIC_LOAD_AND:
9619 case ISD::ATOMIC_LOAD_OR:
9620 case ISD::ATOMIC_LOAD_XOR:
9621 case ISD::ATOMIC_LOAD_NAND:
9622 case ISD::ATOMIC_LOAD_MIN:
9623 case ISD::ATOMIC_LOAD_MAX:
9624 case ISD::ATOMIC_LOAD_UMIN:
9625 case ISD::ATOMIC_LOAD_UMAX:
9631 SDValue fence = atomic.getOperand(0);
9632 if (fence.getOpcode() != ISD::MEMBARRIER)
9635 switch (atomic.getOpcode()) {
9636 case ISD::ATOMIC_CMP_SWAP:
9637 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9638 atomic.getOperand(1), atomic.getOperand(2),
9639 atomic.getOperand(3));
9640 case ISD::ATOMIC_SWAP:
9641 case ISD::ATOMIC_LOAD_ADD:
9642 case ISD::ATOMIC_LOAD_SUB:
9643 case ISD::ATOMIC_LOAD_AND:
9644 case ISD::ATOMIC_LOAD_OR:
9645 case ISD::ATOMIC_LOAD_XOR:
9646 case ISD::ATOMIC_LOAD_NAND:
9647 case ISD::ATOMIC_LOAD_MIN:
9648 case ISD::ATOMIC_LOAD_MAX:
9649 case ISD::ATOMIC_LOAD_UMIN:
9650 case ISD::ATOMIC_LOAD_UMAX:
9651 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9652 atomic.getOperand(1), atomic.getOperand(2));
9658 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG) {
9659 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
9660 // (and (i32 x86isd::setcc_carry), 1)
9661 // This eliminates the zext. This transformation is necessary because
9662 // ISD::SETCC is always legalized to i8.
9663 DebugLoc dl = N->getDebugLoc();
9664 SDValue N0 = N->getOperand(0);
9665 EVT VT = N->getValueType(0);
9666 if (N0.getOpcode() == ISD::AND &&
9668 N0.getOperand(0).hasOneUse()) {
9669 SDValue N00 = N0.getOperand(0);
9670 if (N00.getOpcode() != X86ISD::SETCC_CARRY)
9672 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
9673 if (!C || C->getZExtValue() != 1)
9675 return DAG.getNode(ISD::AND, dl, VT,
9676 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
9677 N00.getOperand(0), N00.getOperand(1)),
9678 DAG.getConstant(1, VT));
9684 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9685 DAGCombinerInfo &DCI) const {
9686 SelectionDAG &DAG = DCI.DAG;
9687 switch (N->getOpcode()) {
9689 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9690 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9691 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9692 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9695 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9696 case ISD::OR: return PerformOrCombine(N, DAG, Subtarget);
9697 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9699 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9700 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9701 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9702 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9703 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9704 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
9710 //===----------------------------------------------------------------------===//
9711 // X86 Inline Assembly Support
9712 //===----------------------------------------------------------------------===//
9714 static bool LowerToBSwap(CallInst *CI) {
9715 // FIXME: this should verify that we are targetting a 486 or better. If not,
9716 // we will turn this bswap into something that will be lowered to logical ops
9717 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9718 // so don't worry about this.
9720 // Verify this is a simple bswap.
9721 if (CI->getNumOperands() != 2 ||
9722 CI->getType() != CI->getOperand(1)->getType() ||
9723 !CI->getType()->isIntegerTy())
9726 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9727 if (!Ty || Ty->getBitWidth() % 16 != 0)
9730 // Okay, we can do this xform, do so now.
9731 const Type *Tys[] = { Ty };
9732 Module *M = CI->getParent()->getParent()->getParent();
9733 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9735 Value *Op = CI->getOperand(1);
9736 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9738 CI->replaceAllUsesWith(Op);
9739 CI->eraseFromParent();
9743 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9744 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9745 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9747 std::string AsmStr = IA->getAsmString();
9749 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9750 SmallVector<StringRef, 4> AsmPieces;
9751 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9753 switch (AsmPieces.size()) {
9754 default: return false;
9756 AsmStr = AsmPieces[0];
9758 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9761 if (AsmPieces.size() == 2 &&
9762 (AsmPieces[0] == "bswap" ||
9763 AsmPieces[0] == "bswapq" ||
9764 AsmPieces[0] == "bswapl") &&
9765 (AsmPieces[1] == "$0" ||
9766 AsmPieces[1] == "${0:q}")) {
9767 // No need to check constraints, nothing other than the equivalent of
9768 // "=r,0" would be valid here.
9769 return LowerToBSwap(CI);
9771 // rorw $$8, ${0:w} --> llvm.bswap.i16
9772 if (CI->getType()->isIntegerTy(16) &&
9773 AsmPieces.size() == 3 &&
9774 AsmPieces[0] == "rorw" &&
9775 AsmPieces[1] == "$$8," &&
9776 AsmPieces[2] == "${0:w}" &&
9777 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9778 return LowerToBSwap(CI);
9782 if (CI->getType()->isIntegerTy(64) &&
9783 Constraints.size() >= 2 &&
9784 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9785 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9786 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9787 SmallVector<StringRef, 4> Words;
9788 SplitString(AsmPieces[0], Words, " \t");
9789 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9791 SplitString(AsmPieces[1], Words, " \t");
9792 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9794 SplitString(AsmPieces[2], Words, " \t,");
9795 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9796 Words[2] == "%edx") {
9797 return LowerToBSwap(CI);
9809 /// getConstraintType - Given a constraint letter, return the type of
9810 /// constraint it is for this target.
9811 X86TargetLowering::ConstraintType
9812 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9813 if (Constraint.size() == 1) {
9814 switch (Constraint[0]) {
9826 return C_RegisterClass;
9834 return TargetLowering::getConstraintType(Constraint);
9837 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9838 /// with another that has more specific requirements based on the type of the
9839 /// corresponding operand.
9840 const char *X86TargetLowering::
9841 LowerXConstraint(EVT ConstraintVT) const {
9842 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9843 // 'f' like normal targets.
9844 if (ConstraintVT.isFloatingPoint()) {
9845 if (Subtarget->hasSSE2())
9847 if (Subtarget->hasSSE1())
9851 return TargetLowering::LowerXConstraint(ConstraintVT);
9854 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9855 /// vector. If it is invalid, don't add anything to Ops.
9856 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9859 std::vector<SDValue>&Ops,
9860 SelectionDAG &DAG) const {
9861 SDValue Result(0, 0);
9863 switch (Constraint) {
9866 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9867 if (C->getZExtValue() <= 31) {
9868 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9874 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9875 if (C->getZExtValue() <= 63) {
9876 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9882 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9883 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9884 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9890 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9891 if (C->getZExtValue() <= 255) {
9892 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9898 // 32-bit signed value
9899 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9900 const ConstantInt *CI = C->getConstantIntValue();
9901 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9902 C->getSExtValue())) {
9903 // Widen to 64 bits here to get it sign extended.
9904 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9907 // FIXME gcc accepts some relocatable values here too, but only in certain
9908 // memory models; it's complicated.
9913 // 32-bit unsigned value
9914 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9915 const ConstantInt *CI = C->getConstantIntValue();
9916 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9917 C->getZExtValue())) {
9918 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9922 // FIXME gcc accepts some relocatable values here too, but only in certain
9923 // memory models; it's complicated.
9927 // Literal immediates are always ok.
9928 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9929 // Widen to 64 bits here to get it sign extended.
9930 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9934 // If we are in non-pic codegen mode, we allow the address of a global (with
9935 // an optional displacement) to be used with 'i'.
9936 GlobalAddressSDNode *GA = 0;
9939 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9941 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9942 Offset += GA->getOffset();
9944 } else if (Op.getOpcode() == ISD::ADD) {
9945 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9946 Offset += C->getZExtValue();
9947 Op = Op.getOperand(0);
9950 } else if (Op.getOpcode() == ISD::SUB) {
9951 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9952 Offset += -C->getZExtValue();
9953 Op = Op.getOperand(0);
9958 // Otherwise, this isn't something we can handle, reject it.
9962 GlobalValue *GV = GA->getGlobal();
9963 // If we require an extra load to get this address, as in PIC mode, we
9965 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9966 getTargetMachine())))
9970 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9972 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9978 if (Result.getNode()) {
9979 Ops.push_back(Result);
9982 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9986 std::vector<unsigned> X86TargetLowering::
9987 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9989 if (Constraint.size() == 1) {
9990 // FIXME: not handling fp-stack yet!
9991 switch (Constraint[0]) { // GCC X86 Constraint Letters
9992 default: break; // Unknown constraint letter
9993 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9994 if (Subtarget->is64Bit()) {
9996 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9997 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9998 X86::R10D,X86::R11D,X86::R12D,
9999 X86::R13D,X86::R14D,X86::R15D,
10000 X86::EBP, X86::ESP, 0);
10001 else if (VT == MVT::i16)
10002 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
10003 X86::SI, X86::DI, X86::R8W,X86::R9W,
10004 X86::R10W,X86::R11W,X86::R12W,
10005 X86::R13W,X86::R14W,X86::R15W,
10006 X86::BP, X86::SP, 0);
10007 else if (VT == MVT::i8)
10008 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
10009 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
10010 X86::R10B,X86::R11B,X86::R12B,
10011 X86::R13B,X86::R14B,X86::R15B,
10012 X86::BPL, X86::SPL, 0);
10014 else if (VT == MVT::i64)
10015 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
10016 X86::RSI, X86::RDI, X86::R8, X86::R9,
10017 X86::R10, X86::R11, X86::R12,
10018 X86::R13, X86::R14, X86::R15,
10019 X86::RBP, X86::RSP, 0);
10023 // 32-bit fallthrough
10024 case 'Q': // Q_REGS
10025 if (VT == MVT::i32)
10026 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
10027 else if (VT == MVT::i16)
10028 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
10029 else if (VT == MVT::i8)
10030 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
10031 else if (VT == MVT::i64)
10032 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
10037 return std::vector<unsigned>();
10040 std::pair<unsigned, const TargetRegisterClass*>
10041 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
10043 // First, see if this is a constraint that directly corresponds to an LLVM
10045 if (Constraint.size() == 1) {
10046 // GCC Constraint Letters
10047 switch (Constraint[0]) {
10049 case 'r': // GENERAL_REGS
10050 case 'l': // INDEX_REGS
10052 return std::make_pair(0U, X86::GR8RegisterClass);
10053 if (VT == MVT::i16)
10054 return std::make_pair(0U, X86::GR16RegisterClass);
10055 if (VT == MVT::i32 || !Subtarget->is64Bit())
10056 return std::make_pair(0U, X86::GR32RegisterClass);
10057 return std::make_pair(0U, X86::GR64RegisterClass);
10058 case 'R': // LEGACY_REGS
10060 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
10061 if (VT == MVT::i16)
10062 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
10063 if (VT == MVT::i32 || !Subtarget->is64Bit())
10064 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
10065 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
10066 case 'f': // FP Stack registers.
10067 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
10068 // value to the correct fpstack register class.
10069 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
10070 return std::make_pair(0U, X86::RFP32RegisterClass);
10071 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
10072 return std::make_pair(0U, X86::RFP64RegisterClass);
10073 return std::make_pair(0U, X86::RFP80RegisterClass);
10074 case 'y': // MMX_REGS if MMX allowed.
10075 if (!Subtarget->hasMMX()) break;
10076 return std::make_pair(0U, X86::VR64RegisterClass);
10077 case 'Y': // SSE_REGS if SSE2 allowed
10078 if (!Subtarget->hasSSE2()) break;
10080 case 'x': // SSE_REGS if SSE1 allowed
10081 if (!Subtarget->hasSSE1()) break;
10083 switch (VT.getSimpleVT().SimpleTy) {
10085 // Scalar SSE types.
10088 return std::make_pair(0U, X86::FR32RegisterClass);
10091 return std::make_pair(0U, X86::FR64RegisterClass);
10099 return std::make_pair(0U, X86::VR128RegisterClass);
10105 // Use the default implementation in TargetLowering to convert the register
10106 // constraint into a member of a register class.
10107 std::pair<unsigned, const TargetRegisterClass*> Res;
10108 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
10110 // Not found as a standard register?
10111 if (Res.second == 0) {
10112 // Map st(0) -> st(7) -> ST0
10113 if (Constraint.size() == 7 && Constraint[0] == '{' &&
10114 tolower(Constraint[1]) == 's' &&
10115 tolower(Constraint[2]) == 't' &&
10116 Constraint[3] == '(' &&
10117 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
10118 Constraint[5] == ')' &&
10119 Constraint[6] == '}') {
10121 Res.first = X86::ST0+Constraint[4]-'0';
10122 Res.second = X86::RFP80RegisterClass;
10126 // GCC allows "st(0)" to be called just plain "st".
10127 if (StringRef("{st}").equals_lower(Constraint)) {
10128 Res.first = X86::ST0;
10129 Res.second = X86::RFP80RegisterClass;
10134 if (StringRef("{flags}").equals_lower(Constraint)) {
10135 Res.first = X86::EFLAGS;
10136 Res.second = X86::CCRRegisterClass;
10140 // 'A' means EAX + EDX.
10141 if (Constraint == "A") {
10142 Res.first = X86::EAX;
10143 Res.second = X86::GR32_ADRegisterClass;
10149 // Otherwise, check to see if this is a register class of the wrong value
10150 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
10151 // turn into {ax},{dx}.
10152 if (Res.second->hasType(VT))
10153 return Res; // Correct type already, nothing to do.
10155 // All of the single-register GCC register classes map their values onto
10156 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
10157 // really want an 8-bit or 32-bit register, map to the appropriate register
10158 // class and return the appropriate register.
10159 if (Res.second == X86::GR16RegisterClass) {
10160 if (VT == MVT::i8) {
10161 unsigned DestReg = 0;
10162 switch (Res.first) {
10164 case X86::AX: DestReg = X86::AL; break;
10165 case X86::DX: DestReg = X86::DL; break;
10166 case X86::CX: DestReg = X86::CL; break;
10167 case X86::BX: DestReg = X86::BL; break;
10170 Res.first = DestReg;
10171 Res.second = X86::GR8RegisterClass;
10173 } else if (VT == MVT::i32) {
10174 unsigned DestReg = 0;
10175 switch (Res.first) {
10177 case X86::AX: DestReg = X86::EAX; break;
10178 case X86::DX: DestReg = X86::EDX; break;
10179 case X86::CX: DestReg = X86::ECX; break;
10180 case X86::BX: DestReg = X86::EBX; break;
10181 case X86::SI: DestReg = X86::ESI; break;
10182 case X86::DI: DestReg = X86::EDI; break;
10183 case X86::BP: DestReg = X86::EBP; break;
10184 case X86::SP: DestReg = X86::ESP; break;
10187 Res.first = DestReg;
10188 Res.second = X86::GR32RegisterClass;
10190 } else if (VT == MVT::i64) {
10191 unsigned DestReg = 0;
10192 switch (Res.first) {
10194 case X86::AX: DestReg = X86::RAX; break;
10195 case X86::DX: DestReg = X86::RDX; break;
10196 case X86::CX: DestReg = X86::RCX; break;
10197 case X86::BX: DestReg = X86::RBX; break;
10198 case X86::SI: DestReg = X86::RSI; break;
10199 case X86::DI: DestReg = X86::RDI; break;
10200 case X86::BP: DestReg = X86::RBP; break;
10201 case X86::SP: DestReg = X86::RSP; break;
10204 Res.first = DestReg;
10205 Res.second = X86::GR64RegisterClass;
10208 } else if (Res.second == X86::FR32RegisterClass ||
10209 Res.second == X86::FR64RegisterClass ||
10210 Res.second == X86::VR128RegisterClass) {
10211 // Handle references to XMM physical registers that got mapped into the
10212 // wrong class. This can happen with constraints like {xmm0} where the
10213 // target independent register mapper will just pick the first match it can
10214 // find, ignoring the required type.
10215 if (VT == MVT::f32)
10216 Res.second = X86::FR32RegisterClass;
10217 else if (VT == MVT::f64)
10218 Res.second = X86::FR64RegisterClass;
10219 else if (X86::VR128RegisterClass->hasType(VT))
10220 Res.second = X86::VR128RegisterClass;
10226 //===----------------------------------------------------------------------===//
10227 // X86 Widen vector type
10228 //===----------------------------------------------------------------------===//
10230 /// getWidenVectorType: given a vector type, returns the type to widen
10231 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
10232 /// If there is no vector type that we want to widen to, returns MVT::Other
10233 /// When and where to widen is target dependent based on the cost of
10234 /// scalarizing vs using the wider vector type.
10236 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
10237 assert(VT.isVector());
10238 if (isTypeLegal(VT))
10241 // TODO: In computeRegisterProperty, we can compute the list of legal vector
10242 // type based on element type. This would speed up our search (though
10243 // it may not be worth it since the size of the list is relatively
10245 EVT EltVT = VT.getVectorElementType();
10246 unsigned NElts = VT.getVectorNumElements();
10248 // On X86, it make sense to widen any vector wider than 1
10252 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
10253 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
10254 EVT SVT = (MVT::SimpleValueType)nVT;
10256 if (isTypeLegal(SVT) &&
10257 SVT.getVectorElementType() == EltVT &&
10258 SVT.getVectorNumElements() > NElts)