1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
16 #include "X86InstrBuilder.h"
17 #include "X86ISelLowering.h"
18 #include "X86TargetMachine.h"
19 #include "X86TargetObjectFile.h"
20 #include "llvm/CallingConv.h"
21 #include "llvm/Constants.h"
22 #include "llvm/DerivedTypes.h"
23 #include "llvm/GlobalAlias.h"
24 #include "llvm/GlobalVariable.h"
25 #include "llvm/Function.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/Intrinsics.h"
28 #include "llvm/LLVMContext.h"
29 #include "llvm/ADT/BitVector.h"
30 #include "llvm/ADT/VectorExtras.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunction.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachineModuleInfo.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/PseudoSourceValue.h"
37 #include "llvm/Support/MathExtras.h"
38 #include "llvm/Support/Debug.h"
39 #include "llvm/Support/ErrorHandling.h"
40 #include "llvm/Target/TargetOptions.h"
41 #include "llvm/ADT/SmallSet.h"
42 #include "llvm/ADT/StringExtras.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/raw_ostream.h"
48 DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
50 // Disable16Bit - 16-bit operations typically have a larger encoding than
51 // corresponding 32-bit instructions, and 16-bit code is slow on some
52 // processors. This is an experimental flag to disable 16-bit operations
53 // (which forces them to be Legalized to 32-bit operations).
55 Disable16Bit("disable-16bit", cl::Hidden,
56 cl::desc("Disable use of 16-bit instructions"));
58 // Forward declarations.
59 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
62 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
63 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
64 default: llvm_unreachable("unknown subtarget type");
65 case X86Subtarget::isDarwin:
66 if (TM.getSubtarget<X86Subtarget>().is64Bit())
67 return new X8664_MachoTargetObjectFile();
68 return new X8632_MachoTargetObjectFile();
69 case X86Subtarget::isELF:
70 return new TargetLoweringObjectFileELF();
71 case X86Subtarget::isMingw:
72 case X86Subtarget::isCygwin:
73 case X86Subtarget::isWindows:
74 return new TargetLoweringObjectFileCOFF();
79 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
80 : TargetLowering(TM, createTLOF(TM)) {
81 Subtarget = &TM.getSubtarget<X86Subtarget>();
82 X86ScalarSSEf64 = Subtarget->hasSSE2();
83 X86ScalarSSEf32 = Subtarget->hasSSE1();
84 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
86 RegInfo = TM.getRegisterInfo();
89 // Set up the TargetLowering object.
91 // X86 is weird, it always uses i8 for shift amounts and setcc results.
92 setShiftAmountType(MVT::i8);
93 setBooleanContents(ZeroOrOneBooleanContent);
94 setSchedulingPreference(SchedulingForRegPressure);
95 setStackPointerRegisterToSaveRestore(X86StackPtr);
97 if (Subtarget->isTargetDarwin()) {
98 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
99 setUseUnderscoreSetJmp(false);
100 setUseUnderscoreLongJmp(false);
101 } else if (Subtarget->isTargetMingw()) {
102 // MS runtime is weird: it exports _setjmp, but longjmp!
103 setUseUnderscoreSetJmp(true);
104 setUseUnderscoreLongJmp(false);
106 setUseUnderscoreSetJmp(true);
107 setUseUnderscoreLongJmp(true);
110 // Set up the register classes.
111 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
113 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
114 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
115 if (Subtarget->is64Bit())
116 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
118 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
120 // We don't accept any truncstore of integer registers.
121 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
123 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
124 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
126 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
127 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
128 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
130 // SETOEQ and SETUNE require checking two conditions.
131 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
132 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
133 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
134 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
135 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
136 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
138 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
140 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
141 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
142 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
144 if (Subtarget->is64Bit()) {
145 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
146 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
147 } else if (!UseSoftFloat) {
148 if (X86ScalarSSEf64) {
149 // We have an impenetrably clever algorithm for ui64->double only.
150 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
152 // We have an algorithm for SSE2, and we turn this into a 64-bit
153 // FILD for other targets.
154 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
157 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
159 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
160 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
163 // SSE has no i16 to fp conversion, only i32
164 if (X86ScalarSSEf32) {
165 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
166 // f32 and f64 cases are Legal, f80 case is not
167 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
169 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
170 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
173 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
174 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
177 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
178 // are Legal, f80 is custom lowered.
179 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
180 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
182 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
184 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
185 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
187 if (X86ScalarSSEf32) {
188 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
189 // f32 and f64 cases are Legal, f80 case is not
190 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
192 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
193 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
196 // Handle FP_TO_UINT by promoting the destination to a larger signed
198 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
199 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
200 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
202 if (Subtarget->is64Bit()) {
203 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
204 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
205 } else if (!UseSoftFloat) {
206 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
207 // Expand FP_TO_UINT into a select.
208 // FIXME: We would like to use a Custom expander here eventually to do
209 // the optimal thing for SSE vs. the default expansion in the legalizer.
210 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
212 // With SSE3 we can use fisttpll to convert to a signed i64; without
213 // SSE, we're stuck with a fistpll.
214 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
217 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
218 if (!X86ScalarSSEf64) {
219 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
220 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
223 // Scalar integer divide and remainder are lowered to use operations that
224 // produce two results, to match the available instructions. This exposes
225 // the two-result form to trivial CSE, which is able to combine x/y and x%y
226 // into a single instruction.
228 // Scalar integer multiply-high is also lowered to use two-result
229 // operations, to match the available instructions. However, plain multiply
230 // (low) operations are left as Legal, as there are single-result
231 // instructions for this in x86. Using the two-result multiply instructions
232 // when both high and low results are needed must be arranged by dagcombine.
233 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
234 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
235 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
236 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
237 setOperationAction(ISD::SREM , MVT::i8 , Expand);
238 setOperationAction(ISD::UREM , MVT::i8 , Expand);
239 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
240 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
241 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
242 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
243 setOperationAction(ISD::SREM , MVT::i16 , Expand);
244 setOperationAction(ISD::UREM , MVT::i16 , Expand);
245 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
246 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
247 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
248 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
249 setOperationAction(ISD::SREM , MVT::i32 , Expand);
250 setOperationAction(ISD::UREM , MVT::i32 , Expand);
251 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
252 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
253 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
254 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
255 setOperationAction(ISD::SREM , MVT::i64 , Expand);
256 setOperationAction(ISD::UREM , MVT::i64 , Expand);
258 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
259 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
260 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
261 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
262 if (Subtarget->is64Bit())
263 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
264 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
265 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
266 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
267 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
268 setOperationAction(ISD::FREM , MVT::f32 , Expand);
269 setOperationAction(ISD::FREM , MVT::f64 , Expand);
270 setOperationAction(ISD::FREM , MVT::f80 , Expand);
271 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
273 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
274 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
275 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
276 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
278 setOperationAction(ISD::CTTZ , MVT::i16 , Expand);
279 setOperationAction(ISD::CTLZ , MVT::i16 , Expand);
281 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
282 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
284 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
285 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
286 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
287 if (Subtarget->is64Bit()) {
288 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
289 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
290 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
293 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
294 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
296 // These should be promoted to a larger select which is supported.
297 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
298 // X86 wants to expand cmov itself.
299 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
301 setOperationAction(ISD::SELECT , MVT::i16 , Expand);
303 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
304 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
305 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
306 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
307 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
308 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
310 setOperationAction(ISD::SETCC , MVT::i16 , Expand);
312 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
313 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
314 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
315 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
316 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
317 if (Subtarget->is64Bit()) {
318 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
319 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
321 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
324 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
325 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
326 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
327 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
328 if (Subtarget->is64Bit())
329 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
330 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
331 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
332 if (Subtarget->is64Bit()) {
333 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
334 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
335 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
336 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
337 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
339 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
340 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
341 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
342 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
343 if (Subtarget->is64Bit()) {
344 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
345 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
346 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
349 if (Subtarget->hasSSE1())
350 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
352 if (!Subtarget->hasSSE2())
353 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
355 // Expand certain atomics
356 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
357 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
358 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
359 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
361 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
362 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
363 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
364 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
366 if (!Subtarget->is64Bit()) {
367 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
372 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
373 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
376 // Use the default ISD::DBG_STOPPOINT.
377 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
378 // FIXME - use subtarget debug flags
379 if (!Subtarget->isTargetDarwin() &&
380 !Subtarget->isTargetELF() &&
381 !Subtarget->isTargetCygMing()) {
382 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
383 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
386 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
387 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
388 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
389 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
390 if (Subtarget->is64Bit()) {
391 setExceptionPointerRegister(X86::RAX);
392 setExceptionSelectorRegister(X86::RDX);
394 setExceptionPointerRegister(X86::EAX);
395 setExceptionSelectorRegister(X86::EDX);
397 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
398 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
400 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
402 setOperationAction(ISD::TRAP, MVT::Other, Legal);
404 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
405 setOperationAction(ISD::VASTART , MVT::Other, Custom);
406 setOperationAction(ISD::VAEND , MVT::Other, Expand);
407 if (Subtarget->is64Bit()) {
408 setOperationAction(ISD::VAARG , MVT::Other, Custom);
409 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
411 setOperationAction(ISD::VAARG , MVT::Other, Expand);
412 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
415 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
416 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
417 if (Subtarget->is64Bit())
418 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
419 if (Subtarget->isTargetCygMing())
420 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
422 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
424 if (!UseSoftFloat && X86ScalarSSEf64) {
425 // f32 and f64 use SSE.
426 // Set up the FP register classes.
427 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
428 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
430 // Use ANDPD to simulate FABS.
431 setOperationAction(ISD::FABS , MVT::f64, Custom);
432 setOperationAction(ISD::FABS , MVT::f32, Custom);
434 // Use XORP to simulate FNEG.
435 setOperationAction(ISD::FNEG , MVT::f64, Custom);
436 setOperationAction(ISD::FNEG , MVT::f32, Custom);
438 // Use ANDPD and ORPD to simulate FCOPYSIGN.
439 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
440 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
442 // We don't support sin/cos/fmod
443 setOperationAction(ISD::FSIN , MVT::f64, Expand);
444 setOperationAction(ISD::FCOS , MVT::f64, Expand);
445 setOperationAction(ISD::FSIN , MVT::f32, Expand);
446 setOperationAction(ISD::FCOS , MVT::f32, Expand);
448 // Expand FP immediates into loads from the stack, except for the special
450 addLegalFPImmediate(APFloat(+0.0)); // xorpd
451 addLegalFPImmediate(APFloat(+0.0f)); // xorps
452 } else if (!UseSoftFloat && X86ScalarSSEf32) {
453 // Use SSE for f32, x87 for f64.
454 // Set up the FP register classes.
455 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
456 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
458 // Use ANDPS to simulate FABS.
459 setOperationAction(ISD::FABS , MVT::f32, Custom);
461 // Use XORP to simulate FNEG.
462 setOperationAction(ISD::FNEG , MVT::f32, Custom);
464 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
466 // Use ANDPS and ORPS to simulate FCOPYSIGN.
467 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
468 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
470 // We don't support sin/cos/fmod
471 setOperationAction(ISD::FSIN , MVT::f32, Expand);
472 setOperationAction(ISD::FCOS , MVT::f32, Expand);
474 // Special cases we handle for FP constants.
475 addLegalFPImmediate(APFloat(+0.0f)); // xorps
476 addLegalFPImmediate(APFloat(+0.0)); // FLD0
477 addLegalFPImmediate(APFloat(+1.0)); // FLD1
478 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
479 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
482 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
483 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
485 } else if (!UseSoftFloat) {
486 // f32 and f64 in x87.
487 // Set up the FP register classes.
488 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
489 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
491 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
492 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
493 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
494 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
497 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
498 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
500 addLegalFPImmediate(APFloat(+0.0)); // FLD0
501 addLegalFPImmediate(APFloat(+1.0)); // FLD1
502 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
503 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
504 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
505 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
506 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
507 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
510 // Long double always uses X87.
512 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
513 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
514 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
517 APFloat TmpFlt(+0.0);
518 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
520 addLegalFPImmediate(TmpFlt); // FLD0
522 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
523 APFloat TmpFlt2(+1.0);
524 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
526 addLegalFPImmediate(TmpFlt2); // FLD1
527 TmpFlt2.changeSign();
528 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
532 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
533 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
537 // Always use a library call for pow.
538 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
539 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
540 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
542 setOperationAction(ISD::FLOG, MVT::f80, Expand);
543 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
544 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
545 setOperationAction(ISD::FEXP, MVT::f80, Expand);
546 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
548 // First set operation action for all vector types to either promote
549 // (for widening) or expand (for scalarization). Then we will selectively
550 // turn on ones that can be effectively codegen'd.
551 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
552 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
553 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
568 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
569 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
599 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
600 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
603 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
604 // with -msoft-float, disable use of MMX as well.
605 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
606 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
607 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
608 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
609 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
610 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
612 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
613 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
614 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
615 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
617 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
618 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
619 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
620 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
622 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
623 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
625 setOperationAction(ISD::AND, MVT::v8i8, Promote);
626 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
627 setOperationAction(ISD::AND, MVT::v4i16, Promote);
628 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
629 setOperationAction(ISD::AND, MVT::v2i32, Promote);
630 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
631 setOperationAction(ISD::AND, MVT::v1i64, Legal);
633 setOperationAction(ISD::OR, MVT::v8i8, Promote);
634 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
635 setOperationAction(ISD::OR, MVT::v4i16, Promote);
636 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
637 setOperationAction(ISD::OR, MVT::v2i32, Promote);
638 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
639 setOperationAction(ISD::OR, MVT::v1i64, Legal);
641 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
642 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
643 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
644 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
645 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
646 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
647 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
649 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
650 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
651 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
652 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
653 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
654 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
655 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
656 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
657 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
659 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
660 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
661 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
662 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
663 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
665 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
666 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
667 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
668 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
670 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
671 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
672 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
673 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
675 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
677 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
678 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
679 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
680 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
681 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
682 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
683 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
684 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
685 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
688 if (!UseSoftFloat && Subtarget->hasSSE1()) {
689 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
691 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
692 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
693 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
694 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
695 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
696 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
697 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
698 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
699 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
700 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
701 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
702 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
705 if (!UseSoftFloat && Subtarget->hasSSE2()) {
706 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
708 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
709 // registers cannot be used even for integer operations.
710 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
711 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
712 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
713 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
715 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
716 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
717 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
718 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
719 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
720 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
721 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
722 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
723 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
724 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
725 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
726 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
727 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
728 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
729 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
730 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
732 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
733 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
734 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
735 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
737 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
738 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
739 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
740 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
741 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
743 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
744 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
745 EVT VT = (MVT::SimpleValueType)i;
746 // Do not attempt to custom lower non-power-of-2 vectors
747 if (!isPowerOf2_32(VT.getVectorNumElements()))
749 // Do not attempt to custom lower non-128-bit vectors
750 if (!VT.is128BitVector())
752 setOperationAction(ISD::BUILD_VECTOR,
753 VT.getSimpleVT().SimpleTy, Custom);
754 setOperationAction(ISD::VECTOR_SHUFFLE,
755 VT.getSimpleVT().SimpleTy, Custom);
756 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
757 VT.getSimpleVT().SimpleTy, Custom);
760 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
761 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
762 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
763 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
764 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
765 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
767 if (Subtarget->is64Bit()) {
768 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
769 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
772 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
773 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
774 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
777 // Do not attempt to promote non-128-bit vectors
778 if (!VT.is128BitVector()) {
781 setOperationAction(ISD::AND, SVT, Promote);
782 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
783 setOperationAction(ISD::OR, SVT, Promote);
784 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
785 setOperationAction(ISD::XOR, SVT, Promote);
786 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
787 setOperationAction(ISD::LOAD, SVT, Promote);
788 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
789 setOperationAction(ISD::SELECT, SVT, Promote);
790 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
793 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
795 // Custom lower v2i64 and v2f64 selects.
796 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
797 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
798 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
799 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
801 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
802 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
803 if (!DisableMMX && Subtarget->hasMMX()) {
804 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
805 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
809 if (Subtarget->hasSSE41()) {
810 // FIXME: Do we need to handle scalar-to-vector here?
811 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
813 // i8 and i16 vectors are custom , because the source register and source
814 // source memory operand types are not the same width. f32 vectors are
815 // custom since the immediate controlling the insert encodes additional
817 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
818 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
819 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
820 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
822 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
823 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
824 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
825 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
827 if (Subtarget->is64Bit()) {
828 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
829 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
833 if (Subtarget->hasSSE42()) {
834 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
837 if (!UseSoftFloat && Subtarget->hasAVX()) {
838 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
839 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
840 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
841 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
843 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
844 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
845 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
846 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
847 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
848 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
849 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
850 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
851 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
852 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
853 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
854 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
855 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
856 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
857 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
859 // Operations to consider commented out -v16i16 v32i8
860 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
861 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
862 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
863 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
864 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
865 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
866 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
867 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
868 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
869 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
870 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
871 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
872 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
873 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
875 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
876 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
877 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
878 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
880 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
881 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
882 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
883 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
884 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
886 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
887 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
888 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
889 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
890 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
891 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
894 // Not sure we want to do this since there are no 256-bit integer
897 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
898 // This includes 256-bit vectors
899 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
900 EVT VT = (MVT::SimpleValueType)i;
902 // Do not attempt to custom lower non-power-of-2 vectors
903 if (!isPowerOf2_32(VT.getVectorNumElements()))
906 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
907 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
908 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
911 if (Subtarget->is64Bit()) {
912 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
913 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
918 // Not sure we want to do this since there are no 256-bit integer
921 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
922 // Including 256-bit vectors
923 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
924 EVT VT = (MVT::SimpleValueType)i;
926 if (!VT.is256BitVector()) {
929 setOperationAction(ISD::AND, VT, Promote);
930 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
931 setOperationAction(ISD::OR, VT, Promote);
932 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
933 setOperationAction(ISD::XOR, VT, Promote);
934 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
935 setOperationAction(ISD::LOAD, VT, Promote);
936 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
937 setOperationAction(ISD::SELECT, VT, Promote);
938 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
941 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
945 // We want to custom lower some of our intrinsics.
946 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
948 // Add/Sub/Mul with overflow operations are custom lowered.
949 setOperationAction(ISD::SADDO, MVT::i32, Custom);
950 setOperationAction(ISD::SADDO, MVT::i64, Custom);
951 setOperationAction(ISD::UADDO, MVT::i32, Custom);
952 setOperationAction(ISD::UADDO, MVT::i64, Custom);
953 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
954 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
955 setOperationAction(ISD::USUBO, MVT::i32, Custom);
956 setOperationAction(ISD::USUBO, MVT::i64, Custom);
957 setOperationAction(ISD::SMULO, MVT::i32, Custom);
958 setOperationAction(ISD::SMULO, MVT::i64, Custom);
960 if (!Subtarget->is64Bit()) {
961 // These libcalls are not available in 32-bit.
962 setLibcallName(RTLIB::SHL_I128, 0);
963 setLibcallName(RTLIB::SRL_I128, 0);
964 setLibcallName(RTLIB::SRA_I128, 0);
967 // We have target-specific dag combine patterns for the following nodes:
968 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
969 setTargetDAGCombine(ISD::BUILD_VECTOR);
970 setTargetDAGCombine(ISD::SELECT);
971 setTargetDAGCombine(ISD::SHL);
972 setTargetDAGCombine(ISD::SRA);
973 setTargetDAGCombine(ISD::SRL);
974 setTargetDAGCombine(ISD::STORE);
975 setTargetDAGCombine(ISD::MEMBARRIER);
976 if (Subtarget->is64Bit())
977 setTargetDAGCombine(ISD::MUL);
979 computeRegisterProperties();
981 // FIXME: These should be based on subtarget info. Plus, the values should
982 // be smaller when we are in optimizing for size mode.
983 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
984 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
985 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
986 setPrefLoopAlignment(16);
987 benefitFromCodePlacementOpt = true;
991 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
996 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
997 /// the desired ByVal argument alignment.
998 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
1001 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1002 if (VTy->getBitWidth() == 128)
1004 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1005 unsigned EltAlign = 0;
1006 getMaxByValAlign(ATy->getElementType(), EltAlign);
1007 if (EltAlign > MaxAlign)
1008 MaxAlign = EltAlign;
1009 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1010 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1011 unsigned EltAlign = 0;
1012 getMaxByValAlign(STy->getElementType(i), EltAlign);
1013 if (EltAlign > MaxAlign)
1014 MaxAlign = EltAlign;
1022 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1023 /// function arguments in the caller parameter area. For X86, aggregates
1024 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1025 /// are at 4-byte boundaries.
1026 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1027 if (Subtarget->is64Bit()) {
1028 // Max of 8 and alignment of type.
1029 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1036 if (Subtarget->hasSSE1())
1037 getMaxByValAlign(Ty, Align);
1041 /// getOptimalMemOpType - Returns the target specific optimal type for load
1042 /// and store operations as a result of memset, memcpy, and memmove
1043 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1046 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1047 bool isSrcConst, bool isSrcStr,
1048 SelectionDAG &DAG) const {
1049 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1050 // linux. This is because the stack realignment code can't handle certain
1051 // cases like PR2962. This should be removed when PR2962 is fixed.
1052 const Function *F = DAG.getMachineFunction().getFunction();
1053 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1054 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1055 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1057 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1060 if (Subtarget->is64Bit() && Size >= 8)
1065 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1067 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1068 SelectionDAG &DAG) const {
1069 if (usesGlobalOffsetTable())
1070 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1071 if (!Subtarget->is64Bit())
1072 // This doesn't have DebugLoc associated with it, but is not really the
1073 // same as a Register.
1074 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1079 /// getFunctionAlignment - Return the Log2 alignment of this function.
1080 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1081 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1084 //===----------------------------------------------------------------------===//
1085 // Return Value Calling Convention Implementation
1086 //===----------------------------------------------------------------------===//
1088 #include "X86GenCallingConv.inc"
1091 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
1092 const SmallVectorImpl<EVT> &OutTys,
1093 const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
1094 SelectionDAG &DAG) {
1095 SmallVector<CCValAssign, 16> RVLocs;
1096 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1097 RVLocs, *DAG.getContext());
1098 return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
1102 X86TargetLowering::LowerReturn(SDValue Chain,
1103 CallingConv::ID CallConv, bool isVarArg,
1104 const SmallVectorImpl<ISD::OutputArg> &Outs,
1105 DebugLoc dl, SelectionDAG &DAG) {
1107 SmallVector<CCValAssign, 16> RVLocs;
1108 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1109 RVLocs, *DAG.getContext());
1110 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1112 // If this is the first return lowered for this function, add the regs to the
1113 // liveout set for the function.
1114 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1115 for (unsigned i = 0; i != RVLocs.size(); ++i)
1116 if (RVLocs[i].isRegLoc())
1117 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1122 SmallVector<SDValue, 6> RetOps;
1123 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1124 // Operand #1 = Bytes To Pop
1125 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1127 // Copy the result values into the output registers.
1128 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1129 CCValAssign &VA = RVLocs[i];
1130 assert(VA.isRegLoc() && "Can only return in registers!");
1131 SDValue ValToCopy = Outs[i].Val;
1133 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1134 // the RET instruction and handled by the FP Stackifier.
1135 if (VA.getLocReg() == X86::ST0 ||
1136 VA.getLocReg() == X86::ST1) {
1137 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1138 // change the value to the FP stack register class.
1139 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1140 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1141 RetOps.push_back(ValToCopy);
1142 // Don't emit a copytoreg.
1146 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1147 // which is returned in RAX / RDX.
1148 if (Subtarget->is64Bit()) {
1149 EVT ValVT = ValToCopy.getValueType();
1150 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1151 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1152 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1153 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1157 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1158 Flag = Chain.getValue(1);
1161 // The x86-64 ABI for returning structs by value requires that we copy
1162 // the sret argument into %rax for the return. We saved the argument into
1163 // a virtual register in the entry block, so now we copy the value out
1165 if (Subtarget->is64Bit() &&
1166 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1167 MachineFunction &MF = DAG.getMachineFunction();
1168 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1169 unsigned Reg = FuncInfo->getSRetReturnReg();
1171 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1172 FuncInfo->setSRetReturnReg(Reg);
1174 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1176 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1177 Flag = Chain.getValue(1);
1179 // RAX now acts like a return value.
1180 MF.getRegInfo().addLiveOut(X86::RAX);
1183 RetOps[0] = Chain; // Update chain.
1185 // Add the flag if we have it.
1187 RetOps.push_back(Flag);
1189 return DAG.getNode(X86ISD::RET_FLAG, dl,
1190 MVT::Other, &RetOps[0], RetOps.size());
1193 /// LowerCallResult - Lower the result values of a call into the
1194 /// appropriate copies out of appropriate physical registers.
1197 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1198 CallingConv::ID CallConv, bool isVarArg,
1199 const SmallVectorImpl<ISD::InputArg> &Ins,
1200 DebugLoc dl, SelectionDAG &DAG,
1201 SmallVectorImpl<SDValue> &InVals) {
1203 // Assign locations to each value returned by this call.
1204 SmallVector<CCValAssign, 16> RVLocs;
1205 bool Is64Bit = Subtarget->is64Bit();
1206 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1207 RVLocs, *DAG.getContext());
1208 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1210 // Copy all of the result registers out of their specified physreg.
1211 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1212 CCValAssign &VA = RVLocs[i];
1213 EVT CopyVT = VA.getValVT();
1215 // If this is x86-64, and we disabled SSE, we can't return FP values
1216 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1217 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1218 llvm_report_error("SSE register return with SSE disabled");
1221 // If this is a call to a function that returns an fp value on the floating
1222 // point stack, but where we prefer to use the value in xmm registers, copy
1223 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1224 if ((VA.getLocReg() == X86::ST0 ||
1225 VA.getLocReg() == X86::ST1) &&
1226 isScalarFPTypeInSSEReg(VA.getValVT())) {
1231 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1232 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1233 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1234 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1235 MVT::v2i64, InFlag).getValue(1);
1236 Val = Chain.getValue(0);
1237 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1238 Val, DAG.getConstant(0, MVT::i64));
1240 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1241 MVT::i64, InFlag).getValue(1);
1242 Val = Chain.getValue(0);
1244 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1246 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1247 CopyVT, InFlag).getValue(1);
1248 Val = Chain.getValue(0);
1250 InFlag = Chain.getValue(2);
1252 if (CopyVT != VA.getValVT()) {
1253 // Round the F80 the right size, which also moves to the appropriate xmm
1255 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1256 // This truncation won't change the value.
1257 DAG.getIntPtrConstant(1));
1260 InVals.push_back(Val);
1267 //===----------------------------------------------------------------------===//
1268 // C & StdCall & Fast Calling Convention implementation
1269 //===----------------------------------------------------------------------===//
1270 // StdCall calling convention seems to be standard for many Windows' API
1271 // routines and around. It differs from C calling convention just a little:
1272 // callee should clean up the stack, not caller. Symbols should be also
1273 // decorated in some fancy way :) It doesn't support any vector arguments.
1274 // For info on fast calling convention see Fast Calling Convention (tail call)
1275 // implementation LowerX86_32FastCCCallTo.
1277 /// CallIsStructReturn - Determines whether a call uses struct return
1279 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1283 return Outs[0].Flags.isSRet();
1286 /// ArgsAreStructReturn - Determines whether a function uses struct
1287 /// return semantics.
1289 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1293 return Ins[0].Flags.isSRet();
1296 /// IsCalleePop - Determines whether the callee is required to pop its
1297 /// own arguments. Callee pop is necessary to support tail calls.
1298 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1302 switch (CallingConv) {
1305 case CallingConv::X86_StdCall:
1306 return !Subtarget->is64Bit();
1307 case CallingConv::X86_FastCall:
1308 return !Subtarget->is64Bit();
1309 case CallingConv::Fast:
1310 return PerformTailCallOpt;
1314 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1315 /// given CallingConvention value.
1316 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1317 if (Subtarget->is64Bit()) {
1318 if (Subtarget->isTargetWin64())
1319 return CC_X86_Win64_C;
1324 if (CC == CallingConv::X86_FastCall)
1325 return CC_X86_32_FastCall;
1326 else if (CC == CallingConv::Fast)
1327 return CC_X86_32_FastCC;
1332 /// NameDecorationForCallConv - Selects the appropriate decoration to
1333 /// apply to a MachineFunction containing a given calling convention.
1335 X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
1336 if (CallConv == CallingConv::X86_FastCall)
1338 else if (CallConv == CallingConv::X86_StdCall)
1344 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1345 /// by "Src" to address "Dst" with size and alignment information specified by
1346 /// the specific parameter attribute. The copy will be passed as a byval
1347 /// function parameter.
1349 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1350 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1352 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1353 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1354 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1358 X86TargetLowering::LowerMemArgument(SDValue Chain,
1359 CallingConv::ID CallConv,
1360 const SmallVectorImpl<ISD::InputArg> &Ins,
1361 DebugLoc dl, SelectionDAG &DAG,
1362 const CCValAssign &VA,
1363 MachineFrameInfo *MFI,
1366 // Create the nodes corresponding to a load from this parameter slot.
1367 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1368 bool AlwaysUseMutable = (CallConv==CallingConv::Fast) && PerformTailCallOpt;
1369 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1372 // If value is passed by pointer we have address passed instead of the value
1374 if (VA.getLocInfo() == CCValAssign::Indirect)
1375 ValVT = VA.getLocVT();
1377 ValVT = VA.getValVT();
1379 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1380 // changed with more analysis.
1381 // In case of tail call optimization mark all arguments mutable. Since they
1382 // could be overwritten by lowering of arguments in case of a tail call.
1383 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1384 VA.getLocMemOffset(), isImmutable, false);
1385 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1386 if (Flags.isByVal())
1388 return DAG.getLoad(ValVT, dl, Chain, FIN,
1389 PseudoSourceValue::getFixedStack(FI), 0);
1393 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1394 CallingConv::ID CallConv,
1396 const SmallVectorImpl<ISD::InputArg> &Ins,
1399 SmallVectorImpl<SDValue> &InVals) {
1401 MachineFunction &MF = DAG.getMachineFunction();
1402 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1404 const Function* Fn = MF.getFunction();
1405 if (Fn->hasExternalLinkage() &&
1406 Subtarget->isTargetCygMing() &&
1407 Fn->getName() == "main")
1408 FuncInfo->setForceFramePointer(true);
1410 // Decorate the function name.
1411 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1413 MachineFrameInfo *MFI = MF.getFrameInfo();
1414 bool Is64Bit = Subtarget->is64Bit();
1415 bool IsWin64 = Subtarget->isTargetWin64();
1417 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1418 "Var args not supported with calling convention fastcc");
1420 // Assign locations to all of the incoming arguments.
1421 SmallVector<CCValAssign, 16> ArgLocs;
1422 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1423 ArgLocs, *DAG.getContext());
1424 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1426 unsigned LastVal = ~0U;
1428 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1429 CCValAssign &VA = ArgLocs[i];
1430 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1432 assert(VA.getValNo() != LastVal &&
1433 "Don't support value assigned to multiple locs yet");
1434 LastVal = VA.getValNo();
1436 if (VA.isRegLoc()) {
1437 EVT RegVT = VA.getLocVT();
1438 TargetRegisterClass *RC = NULL;
1439 if (RegVT == MVT::i32)
1440 RC = X86::GR32RegisterClass;
1441 else if (Is64Bit && RegVT == MVT::i64)
1442 RC = X86::GR64RegisterClass;
1443 else if (RegVT == MVT::f32)
1444 RC = X86::FR32RegisterClass;
1445 else if (RegVT == MVT::f64)
1446 RC = X86::FR64RegisterClass;
1447 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1448 RC = X86::VR128RegisterClass;
1449 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1450 RC = X86::VR64RegisterClass;
1452 llvm_unreachable("Unknown argument type!");
1454 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1455 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1457 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1458 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1460 if (VA.getLocInfo() == CCValAssign::SExt)
1461 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1462 DAG.getValueType(VA.getValVT()));
1463 else if (VA.getLocInfo() == CCValAssign::ZExt)
1464 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1465 DAG.getValueType(VA.getValVT()));
1466 else if (VA.getLocInfo() == CCValAssign::BCvt)
1467 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1469 if (VA.isExtInLoc()) {
1470 // Handle MMX values passed in XMM regs.
1471 if (RegVT.isVector()) {
1472 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1473 ArgValue, DAG.getConstant(0, MVT::i64));
1474 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1476 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1479 assert(VA.isMemLoc());
1480 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1483 // If value is passed via pointer - do a load.
1484 if (VA.getLocInfo() == CCValAssign::Indirect)
1485 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1487 InVals.push_back(ArgValue);
1490 // The x86-64 ABI for returning structs by value requires that we copy
1491 // the sret argument into %rax for the return. Save the argument into
1492 // a virtual register so that we can access it from the return points.
1493 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1494 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1495 unsigned Reg = FuncInfo->getSRetReturnReg();
1497 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1498 FuncInfo->setSRetReturnReg(Reg);
1500 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1501 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1504 unsigned StackSize = CCInfo.getNextStackOffset();
1505 // align stack specially for tail calls
1506 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1507 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1509 // If the function takes variable number of arguments, make a frame index for
1510 // the start of the first vararg value... for expansion of llvm.va_start.
1512 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1513 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize, true, false);
1516 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1518 // FIXME: We should really autogenerate these arrays
1519 static const unsigned GPR64ArgRegsWin64[] = {
1520 X86::RCX, X86::RDX, X86::R8, X86::R9
1522 static const unsigned XMMArgRegsWin64[] = {
1523 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1525 static const unsigned GPR64ArgRegs64Bit[] = {
1526 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1528 static const unsigned XMMArgRegs64Bit[] = {
1529 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1530 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1532 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1535 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1536 GPR64ArgRegs = GPR64ArgRegsWin64;
1537 XMMArgRegs = XMMArgRegsWin64;
1539 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1540 GPR64ArgRegs = GPR64ArgRegs64Bit;
1541 XMMArgRegs = XMMArgRegs64Bit;
1543 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1545 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1548 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1549 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1550 "SSE register cannot be used when SSE is disabled!");
1551 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1552 "SSE register cannot be used when SSE is disabled!");
1553 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1554 // Kernel mode asks for SSE to be disabled, so don't push them
1556 TotalNumXMMRegs = 0;
1558 // For X86-64, if there are vararg parameters that are passed via
1559 // registers, then we must store them to their spots on the stack so they
1560 // may be loaded by deferencing the result of va_next.
1561 VarArgsGPOffset = NumIntRegs * 8;
1562 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1563 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1564 TotalNumXMMRegs * 16, 16,
1567 // Store the integer parameter registers.
1568 SmallVector<SDValue, 8> MemOps;
1569 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1570 unsigned Offset = VarArgsGPOffset;
1571 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1572 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1573 DAG.getIntPtrConstant(Offset));
1574 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1575 X86::GR64RegisterClass);
1576 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1578 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1579 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1581 MemOps.push_back(Store);
1585 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1586 // Now store the XMM (fp + vector) parameter registers.
1587 SmallVector<SDValue, 11> SaveXMMOps;
1588 SaveXMMOps.push_back(Chain);
1590 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1591 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1592 SaveXMMOps.push_back(ALVal);
1594 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1595 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1597 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1598 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1599 X86::VR128RegisterClass);
1600 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1601 SaveXMMOps.push_back(Val);
1603 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1605 &SaveXMMOps[0], SaveXMMOps.size()));
1608 if (!MemOps.empty())
1609 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1610 &MemOps[0], MemOps.size());
1614 // Some CCs need callee pop.
1615 if (IsCalleePop(isVarArg, CallConv)) {
1616 BytesToPopOnReturn = StackSize; // Callee pops everything.
1617 BytesCallerReserves = 0;
1619 BytesToPopOnReturn = 0; // Callee pops nothing.
1620 // If this is an sret function, the return should pop the hidden pointer.
1621 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1622 BytesToPopOnReturn = 4;
1623 BytesCallerReserves = StackSize;
1627 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1628 if (CallConv == CallingConv::X86_FastCall)
1629 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1632 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1638 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1639 SDValue StackPtr, SDValue Arg,
1640 DebugLoc dl, SelectionDAG &DAG,
1641 const CCValAssign &VA,
1642 ISD::ArgFlagsTy Flags) {
1643 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1644 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1645 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1646 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1647 if (Flags.isByVal()) {
1648 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1650 return DAG.getStore(Chain, dl, Arg, PtrOff,
1651 PseudoSourceValue::getStack(), LocMemOffset);
1654 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1655 /// optimization is performed and it is required.
1657 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1658 SDValue &OutRetAddr,
1664 if (!IsTailCall || FPDiff==0) return Chain;
1666 // Adjust the Return address stack slot.
1667 EVT VT = getPointerTy();
1668 OutRetAddr = getReturnAddressFrameIndex(DAG);
1670 // Load the "old" Return address.
1671 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1672 return SDValue(OutRetAddr.getNode(), 1);
1675 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1676 /// optimization is performed and it is required (FPDiff!=0).
1678 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1679 SDValue Chain, SDValue RetAddrFrIdx,
1680 bool Is64Bit, int FPDiff, DebugLoc dl) {
1681 // Store the return address to the appropriate stack slot.
1682 if (!FPDiff) return Chain;
1683 // Calculate the new stack slot for the return address.
1684 int SlotSize = Is64Bit ? 8 : 4;
1685 int NewReturnAddrFI =
1686 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize,
1688 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1689 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1690 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1691 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1696 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1697 CallingConv::ID CallConv, bool isVarArg,
1699 const SmallVectorImpl<ISD::OutputArg> &Outs,
1700 const SmallVectorImpl<ISD::InputArg> &Ins,
1701 DebugLoc dl, SelectionDAG &DAG,
1702 SmallVectorImpl<SDValue> &InVals) {
1704 MachineFunction &MF = DAG.getMachineFunction();
1705 bool Is64Bit = Subtarget->is64Bit();
1706 bool IsStructRet = CallIsStructReturn(Outs);
1708 assert((!isTailCall ||
1709 (CallConv == CallingConv::Fast && PerformTailCallOpt)) &&
1710 "IsEligibleForTailCallOptimization missed a case!");
1711 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1712 "Var args not supported with calling convention fastcc");
1714 // Analyze operands of the call, assigning locations to each operand.
1715 SmallVector<CCValAssign, 16> ArgLocs;
1716 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1717 ArgLocs, *DAG.getContext());
1718 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1720 // Get a count of how many bytes are to be pushed on the stack.
1721 unsigned NumBytes = CCInfo.getNextStackOffset();
1722 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1723 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1727 // Lower arguments at fp - stackoffset + fpdiff.
1728 unsigned NumBytesCallerPushed =
1729 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1730 FPDiff = NumBytesCallerPushed - NumBytes;
1732 // Set the delta of movement of the returnaddr stackslot.
1733 // But only set if delta is greater than previous delta.
1734 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1735 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1738 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1740 SDValue RetAddrFrIdx;
1741 // Load return adress for tail calls.
1742 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
1745 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1746 SmallVector<SDValue, 8> MemOpChains;
1749 // Walk the register/memloc assignments, inserting copies/loads. In the case
1750 // of tail call optimization arguments are handle later.
1751 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1752 CCValAssign &VA = ArgLocs[i];
1753 EVT RegVT = VA.getLocVT();
1754 SDValue Arg = Outs[i].Val;
1755 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1756 bool isByVal = Flags.isByVal();
1758 // Promote the value if needed.
1759 switch (VA.getLocInfo()) {
1760 default: llvm_unreachable("Unknown loc info!");
1761 case CCValAssign::Full: break;
1762 case CCValAssign::SExt:
1763 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1765 case CCValAssign::ZExt:
1766 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1768 case CCValAssign::AExt:
1769 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1770 // Special case: passing MMX values in XMM registers.
1771 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1772 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1773 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1775 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1777 case CCValAssign::BCvt:
1778 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1780 case CCValAssign::Indirect: {
1781 // Store the argument.
1782 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1783 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1784 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1785 PseudoSourceValue::getFixedStack(FI), 0);
1791 if (VA.isRegLoc()) {
1792 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1794 if (!isTailCall || (isTailCall && isByVal)) {
1795 assert(VA.isMemLoc());
1796 if (StackPtr.getNode() == 0)
1797 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1799 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1800 dl, DAG, VA, Flags));
1805 if (!MemOpChains.empty())
1806 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1807 &MemOpChains[0], MemOpChains.size());
1809 // Build a sequence of copy-to-reg nodes chained together with token chain
1810 // and flag operands which copy the outgoing args into registers.
1812 // Tail call byval lowering might overwrite argument registers so in case of
1813 // tail call optimization the copies to registers are lowered later.
1815 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1816 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1817 RegsToPass[i].second, InFlag);
1818 InFlag = Chain.getValue(1);
1822 if (Subtarget->isPICStyleGOT()) {
1823 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1826 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1827 DAG.getNode(X86ISD::GlobalBaseReg,
1828 DebugLoc::getUnknownLoc(),
1831 InFlag = Chain.getValue(1);
1833 // If we are tail calling and generating PIC/GOT style code load the
1834 // address of the callee into ECX. The value in ecx is used as target of
1835 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1836 // for tail calls on PIC/GOT architectures. Normally we would just put the
1837 // address of GOT into ebx and then call target@PLT. But for tail calls
1838 // ebx would be restored (since ebx is callee saved) before jumping to the
1841 // Note: The actual moving to ECX is done further down.
1842 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1843 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1844 !G->getGlobal()->hasProtectedVisibility())
1845 Callee = LowerGlobalAddress(Callee, DAG);
1846 else if (isa<ExternalSymbolSDNode>(Callee))
1847 Callee = LowerExternalSymbol(Callee, DAG);
1851 if (Is64Bit && isVarArg) {
1852 // From AMD64 ABI document:
1853 // For calls that may call functions that use varargs or stdargs
1854 // (prototype-less calls or calls to functions containing ellipsis (...) in
1855 // the declaration) %al is used as hidden argument to specify the number
1856 // of SSE registers used. The contents of %al do not need to match exactly
1857 // the number of registers, but must be an ubound on the number of SSE
1858 // registers used and is in the range 0 - 8 inclusive.
1860 // FIXME: Verify this on Win64
1861 // Count the number of XMM registers allocated.
1862 static const unsigned XMMArgRegs[] = {
1863 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1864 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1866 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1867 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1868 && "SSE registers cannot be used when SSE is disabled");
1870 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1871 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1872 InFlag = Chain.getValue(1);
1876 // For tail calls lower the arguments to the 'real' stack slot.
1878 // Force all the incoming stack arguments to be loaded from the stack
1879 // before any new outgoing arguments are stored to the stack, because the
1880 // outgoing stack slots may alias the incoming argument stack slots, and
1881 // the alias isn't otherwise explicit. This is slightly more conservative
1882 // than necessary, because it means that each store effectively depends
1883 // on every argument instead of just those arguments it would clobber.
1884 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1886 SmallVector<SDValue, 8> MemOpChains2;
1889 // Do not flag preceeding copytoreg stuff together with the following stuff.
1891 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1892 CCValAssign &VA = ArgLocs[i];
1893 if (!VA.isRegLoc()) {
1894 assert(VA.isMemLoc());
1895 SDValue Arg = Outs[i].Val;
1896 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1897 // Create frame index.
1898 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1899 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1900 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true, false);
1901 FIN = DAG.getFrameIndex(FI, getPointerTy());
1903 if (Flags.isByVal()) {
1904 // Copy relative to framepointer.
1905 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1906 if (StackPtr.getNode() == 0)
1907 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1909 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1911 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1915 // Store relative to framepointer.
1916 MemOpChains2.push_back(
1917 DAG.getStore(ArgChain, dl, Arg, FIN,
1918 PseudoSourceValue::getFixedStack(FI), 0));
1923 if (!MemOpChains2.empty())
1924 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1925 &MemOpChains2[0], MemOpChains2.size());
1927 // Copy arguments to their registers.
1928 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1929 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1930 RegsToPass[i].second, InFlag);
1931 InFlag = Chain.getValue(1);
1935 // Store the return address to the appropriate stack slot.
1936 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1940 // If the callee is a GlobalAddress node (quite common, every direct call is)
1941 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1942 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1943 // We should use extra load for direct calls to dllimported functions in
1945 GlobalValue *GV = G->getGlobal();
1946 if (!GV->hasDLLImportLinkage()) {
1947 unsigned char OpFlags = 0;
1949 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1950 // external symbols most go through the PLT in PIC mode. If the symbol
1951 // has hidden or protected visibility, or if it is static or local, then
1952 // we don't need to use the PLT - we can directly call it.
1953 if (Subtarget->isTargetELF() &&
1954 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1955 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1956 OpFlags = X86II::MO_PLT;
1957 } else if (Subtarget->isPICStyleStubAny() &&
1958 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1959 Subtarget->getDarwinVers() < 9) {
1960 // PC-relative references to external symbols should go through $stub,
1961 // unless we're building with the leopard linker or later, which
1962 // automatically synthesizes these stubs.
1963 OpFlags = X86II::MO_DARWIN_STUB;
1966 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
1967 G->getOffset(), OpFlags);
1969 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1970 unsigned char OpFlags = 0;
1972 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
1973 // symbols should go through the PLT.
1974 if (Subtarget->isTargetELF() &&
1975 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1976 OpFlags = X86II::MO_PLT;
1977 } else if (Subtarget->isPICStyleStubAny() &&
1978 Subtarget->getDarwinVers() < 9) {
1979 // PC-relative references to external symbols should go through $stub,
1980 // unless we're building with the leopard linker or later, which
1981 // automatically synthesizes these stubs.
1982 OpFlags = X86II::MO_DARWIN_STUB;
1985 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
1987 } else if (isTailCall) {
1988 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
1990 Chain = DAG.getCopyToReg(Chain, dl,
1991 DAG.getRegister(Opc, getPointerTy()),
1993 Callee = DAG.getRegister(Opc, getPointerTy());
1994 // Add register as live out.
1995 MF.getRegInfo().addLiveOut(Opc);
1998 // Returns a chain & a flag for retval copy to use.
1999 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
2000 SmallVector<SDValue, 8> Ops;
2003 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
2004 DAG.getIntPtrConstant(0, true), InFlag);
2005 InFlag = Chain.getValue(1);
2008 Ops.push_back(Chain);
2009 Ops.push_back(Callee);
2012 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
2014 // Add argument registers to the end of the list so that they are known live
2016 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2017 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2018 RegsToPass[i].second.getValueType()));
2020 // Add an implicit use GOT pointer in EBX.
2021 if (!isTailCall && Subtarget->isPICStyleGOT())
2022 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2024 // Add an implicit use of AL for x86 vararg functions.
2025 if (Is64Bit && isVarArg)
2026 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2028 if (InFlag.getNode())
2029 Ops.push_back(InFlag);
2032 // If this is the first return lowered for this function, add the regs
2033 // to the liveout set for the function.
2034 if (MF.getRegInfo().liveout_empty()) {
2035 SmallVector<CCValAssign, 16> RVLocs;
2036 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2038 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2039 for (unsigned i = 0; i != RVLocs.size(); ++i)
2040 if (RVLocs[i].isRegLoc())
2041 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2044 assert(((Callee.getOpcode() == ISD::Register &&
2045 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2046 cast<RegisterSDNode>(Callee)->getReg() == X86::R9)) ||
2047 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2048 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2049 "Expecting an global address, external symbol, or register");
2051 return DAG.getNode(X86ISD::TC_RETURN, dl,
2052 NodeTys, &Ops[0], Ops.size());
2055 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2056 InFlag = Chain.getValue(1);
2058 // Create the CALLSEQ_END node.
2059 unsigned NumBytesForCalleeToPush;
2060 if (IsCalleePop(isVarArg, CallConv))
2061 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2062 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2063 // If this is is a call to a struct-return function, the callee
2064 // pops the hidden struct pointer, so we have to push it back.
2065 // This is common for Darwin/X86, Linux & Mingw32 targets.
2066 NumBytesForCalleeToPush = 4;
2068 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2070 // Returns a flag for retval copy to use.
2071 Chain = DAG.getCALLSEQ_END(Chain,
2072 DAG.getIntPtrConstant(NumBytes, true),
2073 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2076 InFlag = Chain.getValue(1);
2078 // Handle result values, copying them out of physregs into vregs that we
2080 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2081 Ins, dl, DAG, InVals);
2085 //===----------------------------------------------------------------------===//
2086 // Fast Calling Convention (tail call) implementation
2087 //===----------------------------------------------------------------------===//
2089 // Like std call, callee cleans arguments, convention except that ECX is
2090 // reserved for storing the tail called function address. Only 2 registers are
2091 // free for argument passing (inreg). Tail call optimization is performed
2093 // * tailcallopt is enabled
2094 // * caller/callee are fastcc
2095 // On X86_64 architecture with GOT-style position independent code only local
2096 // (within module) calls are supported at the moment.
2097 // To keep the stack aligned according to platform abi the function
2098 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2099 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2100 // If a tail called function callee has more arguments than the caller the
2101 // caller needs to make sure that there is room to move the RETADDR to. This is
2102 // achieved by reserving an area the size of the argument delta right after the
2103 // original REtADDR, but before the saved framepointer or the spilled registers
2104 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2116 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2117 /// for a 16 byte align requirement.
2118 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2119 SelectionDAG& DAG) {
2120 MachineFunction &MF = DAG.getMachineFunction();
2121 const TargetMachine &TM = MF.getTarget();
2122 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2123 unsigned StackAlignment = TFI.getStackAlignment();
2124 uint64_t AlignMask = StackAlignment - 1;
2125 int64_t Offset = StackSize;
2126 uint64_t SlotSize = TD->getPointerSize();
2127 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2128 // Number smaller than 12 so just add the difference.
2129 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2131 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2132 Offset = ((~AlignMask) & Offset) + StackAlignment +
2133 (StackAlignment-SlotSize);
2138 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2139 /// for tail call optimization. Targets which want to do tail call
2140 /// optimization should implement this function.
2142 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2143 CallingConv::ID CalleeCC,
2145 const SmallVectorImpl<ISD::InputArg> &Ins,
2146 SelectionDAG& DAG) const {
2147 MachineFunction &MF = DAG.getMachineFunction();
2148 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2149 return CalleeCC == CallingConv::Fast && CallerCC == CalleeCC;
2153 X86TargetLowering::createFastISel(MachineFunction &mf,
2154 MachineModuleInfo *mmo,
2156 DenseMap<const Value *, unsigned> &vm,
2157 DenseMap<const BasicBlock *,
2158 MachineBasicBlock *> &bm,
2159 DenseMap<const AllocaInst *, int> &am
2161 , SmallSet<Instruction*, 8> &cil
2164 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2172 //===----------------------------------------------------------------------===//
2173 // Other Lowering Hooks
2174 //===----------------------------------------------------------------------===//
2177 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2178 MachineFunction &MF = DAG.getMachineFunction();
2179 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2180 int ReturnAddrIndex = FuncInfo->getRAIndex();
2182 if (ReturnAddrIndex == 0) {
2183 // Set up a frame object for the return address.
2184 uint64_t SlotSize = TD->getPointerSize();
2185 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize,
2187 FuncInfo->setRAIndex(ReturnAddrIndex);
2190 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2194 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2195 bool hasSymbolicDisplacement) {
2196 // Offset should fit into 32 bit immediate field.
2197 if (!isInt32(Offset))
2200 // If we don't have a symbolic displacement - we don't have any extra
2202 if (!hasSymbolicDisplacement)
2205 // FIXME: Some tweaks might be needed for medium code model.
2206 if (M != CodeModel::Small && M != CodeModel::Kernel)
2209 // For small code model we assume that latest object is 16MB before end of 31
2210 // bits boundary. We may also accept pretty large negative constants knowing
2211 // that all objects are in the positive half of address space.
2212 if (M == CodeModel::Small && Offset < 16*1024*1024)
2215 // For kernel code model we know that all object resist in the negative half
2216 // of 32bits address space. We may not accept negative offsets, since they may
2217 // be just off and we may accept pretty large positive ones.
2218 if (M == CodeModel::Kernel && Offset > 0)
2224 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2225 /// specific condition code, returning the condition code and the LHS/RHS of the
2226 /// comparison to make.
2227 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2228 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2230 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2231 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2232 // X > -1 -> X == 0, jump !sign.
2233 RHS = DAG.getConstant(0, RHS.getValueType());
2234 return X86::COND_NS;
2235 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2236 // X < 0 -> X == 0, jump on sign.
2238 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2240 RHS = DAG.getConstant(0, RHS.getValueType());
2241 return X86::COND_LE;
2245 switch (SetCCOpcode) {
2246 default: llvm_unreachable("Invalid integer condition!");
2247 case ISD::SETEQ: return X86::COND_E;
2248 case ISD::SETGT: return X86::COND_G;
2249 case ISD::SETGE: return X86::COND_GE;
2250 case ISD::SETLT: return X86::COND_L;
2251 case ISD::SETLE: return X86::COND_LE;
2252 case ISD::SETNE: return X86::COND_NE;
2253 case ISD::SETULT: return X86::COND_B;
2254 case ISD::SETUGT: return X86::COND_A;
2255 case ISD::SETULE: return X86::COND_BE;
2256 case ISD::SETUGE: return X86::COND_AE;
2260 // First determine if it is required or is profitable to flip the operands.
2262 // If LHS is a foldable load, but RHS is not, flip the condition.
2263 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2264 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2265 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2266 std::swap(LHS, RHS);
2269 switch (SetCCOpcode) {
2275 std::swap(LHS, RHS);
2279 // On a floating point condition, the flags are set as follows:
2281 // 0 | 0 | 0 | X > Y
2282 // 0 | 0 | 1 | X < Y
2283 // 1 | 0 | 0 | X == Y
2284 // 1 | 1 | 1 | unordered
2285 switch (SetCCOpcode) {
2286 default: llvm_unreachable("Condcode should be pre-legalized away");
2288 case ISD::SETEQ: return X86::COND_E;
2289 case ISD::SETOLT: // flipped
2291 case ISD::SETGT: return X86::COND_A;
2292 case ISD::SETOLE: // flipped
2294 case ISD::SETGE: return X86::COND_AE;
2295 case ISD::SETUGT: // flipped
2297 case ISD::SETLT: return X86::COND_B;
2298 case ISD::SETUGE: // flipped
2300 case ISD::SETLE: return X86::COND_BE;
2302 case ISD::SETNE: return X86::COND_NE;
2303 case ISD::SETUO: return X86::COND_P;
2304 case ISD::SETO: return X86::COND_NP;
2306 case ISD::SETUNE: return X86::COND_INVALID;
2310 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2311 /// code. Current x86 isa includes the following FP cmov instructions:
2312 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2313 static bool hasFPCMov(unsigned X86CC) {
2329 /// isFPImmLegal - Returns true if the target can instruction select the
2330 /// specified FP immediate natively. If false, the legalizer will
2331 /// materialize the FP immediate as a load from a constant pool.
2332 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
2333 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
2334 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
2340 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2341 /// the specified range (L, H].
2342 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2343 return (Val < 0) || (Val >= Low && Val < Hi);
2346 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2347 /// specified value.
2348 static bool isUndefOrEqual(int Val, int CmpVal) {
2349 if (Val < 0 || Val == CmpVal)
2354 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2355 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2356 /// the second operand.
2357 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2358 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2359 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2360 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2361 return (Mask[0] < 2 && Mask[1] < 2);
2365 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2366 SmallVector<int, 8> M;
2368 return ::isPSHUFDMask(M, N->getValueType(0));
2371 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2372 /// is suitable for input to PSHUFHW.
2373 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2374 if (VT != MVT::v8i16)
2377 // Lower quadword copied in order or undef.
2378 for (int i = 0; i != 4; ++i)
2379 if (Mask[i] >= 0 && Mask[i] != i)
2382 // Upper quadword shuffled.
2383 for (int i = 4; i != 8; ++i)
2384 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2390 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2391 SmallVector<int, 8> M;
2393 return ::isPSHUFHWMask(M, N->getValueType(0));
2396 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2397 /// is suitable for input to PSHUFLW.
2398 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2399 if (VT != MVT::v8i16)
2402 // Upper quadword copied in order.
2403 for (int i = 4; i != 8; ++i)
2404 if (Mask[i] >= 0 && Mask[i] != i)
2407 // Lower quadword shuffled.
2408 for (int i = 0; i != 4; ++i)
2415 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2416 SmallVector<int, 8> M;
2418 return ::isPSHUFLWMask(M, N->getValueType(0));
2421 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2422 /// is suitable for input to PALIGNR.
2423 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2425 int i, e = VT.getVectorNumElements();
2427 // Do not handle v2i64 / v2f64 shuffles with palignr.
2428 if (e < 4 || !hasSSSE3)
2431 for (i = 0; i != e; ++i)
2435 // All undef, not a palignr.
2439 // Determine if it's ok to perform a palignr with only the LHS, since we
2440 // don't have access to the actual shuffle elements to see if RHS is undef.
2441 bool Unary = Mask[i] < (int)e;
2442 bool NeedsUnary = false;
2444 int s = Mask[i] - i;
2446 // Check the rest of the elements to see if they are consecutive.
2447 for (++i; i != e; ++i) {
2452 Unary = Unary && (m < (int)e);
2453 NeedsUnary = NeedsUnary || (m < s);
2455 if (NeedsUnary && !Unary)
2457 if (Unary && m != ((s+i) & (e-1)))
2459 if (!Unary && m != (s+i))
2465 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2466 SmallVector<int, 8> M;
2468 return ::isPALIGNRMask(M, N->getValueType(0), true);
2471 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2472 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2473 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2474 int NumElems = VT.getVectorNumElements();
2475 if (NumElems != 2 && NumElems != 4)
2478 int Half = NumElems / 2;
2479 for (int i = 0; i < Half; ++i)
2480 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2482 for (int i = Half; i < NumElems; ++i)
2483 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2489 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2490 SmallVector<int, 8> M;
2492 return ::isSHUFPMask(M, N->getValueType(0));
2495 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2496 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2497 /// half elements to come from vector 1 (which would equal the dest.) and
2498 /// the upper half to come from vector 2.
2499 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2500 int NumElems = VT.getVectorNumElements();
2502 if (NumElems != 2 && NumElems != 4)
2505 int Half = NumElems / 2;
2506 for (int i = 0; i < Half; ++i)
2507 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2509 for (int i = Half; i < NumElems; ++i)
2510 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2515 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2516 SmallVector<int, 8> M;
2518 return isCommutedSHUFPMask(M, N->getValueType(0));
2521 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2522 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2523 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2524 if (N->getValueType(0).getVectorNumElements() != 4)
2527 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2528 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2529 isUndefOrEqual(N->getMaskElt(1), 7) &&
2530 isUndefOrEqual(N->getMaskElt(2), 2) &&
2531 isUndefOrEqual(N->getMaskElt(3), 3);
2534 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2535 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2537 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2538 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2543 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2544 isUndefOrEqual(N->getMaskElt(1), 3) &&
2545 isUndefOrEqual(N->getMaskElt(2), 2) &&
2546 isUndefOrEqual(N->getMaskElt(3), 3);
2549 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2550 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2551 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2552 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2554 if (NumElems != 2 && NumElems != 4)
2557 for (unsigned i = 0; i < NumElems/2; ++i)
2558 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2561 for (unsigned i = NumElems/2; i < NumElems; ++i)
2562 if (!isUndefOrEqual(N->getMaskElt(i), i))
2568 /// isMOVLHPSMask - Return true if the specified VECTOR_SHUFFLE operand
2569 /// specifies a shuffle of elements that is suitable for input to MOVLHPS.
2570 bool X86::isMOVLHPSMask(ShuffleVectorSDNode *N) {
2571 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2573 if (NumElems != 2 && NumElems != 4)
2576 for (unsigned i = 0; i < NumElems/2; ++i)
2577 if (!isUndefOrEqual(N->getMaskElt(i), i))
2580 for (unsigned i = 0; i < NumElems/2; ++i)
2581 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2587 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2588 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2589 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2590 bool V2IsSplat = false) {
2591 int NumElts = VT.getVectorNumElements();
2592 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2595 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2597 int BitI1 = Mask[i+1];
2598 if (!isUndefOrEqual(BitI, j))
2601 if (!isUndefOrEqual(BitI1, NumElts))
2604 if (!isUndefOrEqual(BitI1, j + NumElts))
2611 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2612 SmallVector<int, 8> M;
2614 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2617 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2618 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2619 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2620 bool V2IsSplat = false) {
2621 int NumElts = VT.getVectorNumElements();
2622 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2625 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2627 int BitI1 = Mask[i+1];
2628 if (!isUndefOrEqual(BitI, j + NumElts/2))
2631 if (isUndefOrEqual(BitI1, NumElts))
2634 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2641 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2642 SmallVector<int, 8> M;
2644 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2647 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2648 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2650 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2651 int NumElems = VT.getVectorNumElements();
2652 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2655 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2657 int BitI1 = Mask[i+1];
2658 if (!isUndefOrEqual(BitI, j))
2660 if (!isUndefOrEqual(BitI1, j))
2666 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2667 SmallVector<int, 8> M;
2669 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2672 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2673 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2675 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2676 int NumElems = VT.getVectorNumElements();
2677 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2680 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2682 int BitI1 = Mask[i+1];
2683 if (!isUndefOrEqual(BitI, j))
2685 if (!isUndefOrEqual(BitI1, j))
2691 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2692 SmallVector<int, 8> M;
2694 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2697 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2698 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2699 /// MOVSD, and MOVD, i.e. setting the lowest element.
2700 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2701 if (VT.getVectorElementType().getSizeInBits() < 32)
2704 int NumElts = VT.getVectorNumElements();
2706 if (!isUndefOrEqual(Mask[0], NumElts))
2709 for (int i = 1; i < NumElts; ++i)
2710 if (!isUndefOrEqual(Mask[i], i))
2716 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2717 SmallVector<int, 8> M;
2719 return ::isMOVLMask(M, N->getValueType(0));
2722 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2723 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2724 /// element of vector 2 and the other elements to come from vector 1 in order.
2725 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2726 bool V2IsSplat = false, bool V2IsUndef = false) {
2727 int NumOps = VT.getVectorNumElements();
2728 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2731 if (!isUndefOrEqual(Mask[0], 0))
2734 for (int i = 1; i < NumOps; ++i)
2735 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2736 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2737 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2743 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2744 bool V2IsUndef = false) {
2745 SmallVector<int, 8> M;
2747 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2750 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2751 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2752 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2753 if (N->getValueType(0).getVectorNumElements() != 4)
2756 // Expect 1, 1, 3, 3
2757 for (unsigned i = 0; i < 2; ++i) {
2758 int Elt = N->getMaskElt(i);
2759 if (Elt >= 0 && Elt != 1)
2764 for (unsigned i = 2; i < 4; ++i) {
2765 int Elt = N->getMaskElt(i);
2766 if (Elt >= 0 && Elt != 3)
2771 // Don't use movshdup if it can be done with a shufps.
2772 // FIXME: verify that matching u, u, 3, 3 is what we want.
2776 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2777 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2778 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2779 if (N->getValueType(0).getVectorNumElements() != 4)
2782 // Expect 0, 0, 2, 2
2783 for (unsigned i = 0; i < 2; ++i)
2784 if (N->getMaskElt(i) > 0)
2788 for (unsigned i = 2; i < 4; ++i) {
2789 int Elt = N->getMaskElt(i);
2790 if (Elt >= 0 && Elt != 2)
2795 // Don't use movsldup if it can be done with a shufps.
2799 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2800 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2801 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2802 int e = N->getValueType(0).getVectorNumElements() / 2;
2804 for (int i = 0; i < e; ++i)
2805 if (!isUndefOrEqual(N->getMaskElt(i), i))
2807 for (int i = 0; i < e; ++i)
2808 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2813 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2814 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
2815 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2816 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2817 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2819 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2821 for (int i = 0; i < NumOperands; ++i) {
2822 int Val = SVOp->getMaskElt(NumOperands-i-1);
2823 if (Val < 0) Val = 0;
2824 if (Val >= NumOperands) Val -= NumOperands;
2826 if (i != NumOperands - 1)
2832 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2833 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
2834 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2835 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2837 // 8 nodes, but we only care about the last 4.
2838 for (unsigned i = 7; i >= 4; --i) {
2839 int Val = SVOp->getMaskElt(i);
2848 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2849 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
2850 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2851 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2853 // 8 nodes, but we only care about the first 4.
2854 for (int i = 3; i >= 0; --i) {
2855 int Val = SVOp->getMaskElt(i);
2864 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
2865 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
2866 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
2867 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2868 EVT VVT = N->getValueType(0);
2869 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
2873 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
2874 Val = SVOp->getMaskElt(i);
2878 return (Val - i) * EltSize;
2881 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2883 bool X86::isZeroNode(SDValue Elt) {
2884 return ((isa<ConstantSDNode>(Elt) &&
2885 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2886 (isa<ConstantFPSDNode>(Elt) &&
2887 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2890 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2891 /// their permute mask.
2892 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2893 SelectionDAG &DAG) {
2894 EVT VT = SVOp->getValueType(0);
2895 unsigned NumElems = VT.getVectorNumElements();
2896 SmallVector<int, 8> MaskVec;
2898 for (unsigned i = 0; i != NumElems; ++i) {
2899 int idx = SVOp->getMaskElt(i);
2901 MaskVec.push_back(idx);
2902 else if (idx < (int)NumElems)
2903 MaskVec.push_back(idx + NumElems);
2905 MaskVec.push_back(idx - NumElems);
2907 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2908 SVOp->getOperand(0), &MaskVec[0]);
2911 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2912 /// the two vector operands have swapped position.
2913 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
2914 unsigned NumElems = VT.getVectorNumElements();
2915 for (unsigned i = 0; i != NumElems; ++i) {
2919 else if (idx < (int)NumElems)
2920 Mask[i] = idx + NumElems;
2922 Mask[i] = idx - NumElems;
2926 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2927 /// match movhlps. The lower half elements should come from upper half of
2928 /// V1 (and in order), and the upper half elements should come from the upper
2929 /// half of V2 (and in order).
2930 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2931 if (Op->getValueType(0).getVectorNumElements() != 4)
2933 for (unsigned i = 0, e = 2; i != e; ++i)
2934 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2936 for (unsigned i = 2; i != 4; ++i)
2937 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2942 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2943 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2945 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2946 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2948 N = N->getOperand(0).getNode();
2949 if (!ISD::isNON_EXTLoad(N))
2952 *LD = cast<LoadSDNode>(N);
2956 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2957 /// match movlp{s|d}. The lower half elements should come from lower half of
2958 /// V1 (and in order), and the upper half elements should come from the upper
2959 /// half of V2 (and in order). And since V1 will become the source of the
2960 /// MOVLP, it must be either a vector load or a scalar load to vector.
2961 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2962 ShuffleVectorSDNode *Op) {
2963 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2965 // Is V2 is a vector load, don't do this transformation. We will try to use
2966 // load folding shufps op.
2967 if (ISD::isNON_EXTLoad(V2))
2970 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2972 if (NumElems != 2 && NumElems != 4)
2974 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2975 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2977 for (unsigned i = NumElems/2; i != NumElems; ++i)
2978 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2983 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2985 static bool isSplatVector(SDNode *N) {
2986 if (N->getOpcode() != ISD::BUILD_VECTOR)
2989 SDValue SplatValue = N->getOperand(0);
2990 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2991 if (N->getOperand(i) != SplatValue)
2996 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2997 /// to an zero vector.
2998 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
2999 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
3000 SDValue V1 = N->getOperand(0);
3001 SDValue V2 = N->getOperand(1);
3002 unsigned NumElems = N->getValueType(0).getVectorNumElements();
3003 for (unsigned i = 0; i != NumElems; ++i) {
3004 int Idx = N->getMaskElt(i);
3005 if (Idx >= (int)NumElems) {
3006 unsigned Opc = V2.getOpcode();
3007 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
3009 if (Opc != ISD::BUILD_VECTOR ||
3010 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
3012 } else if (Idx >= 0) {
3013 unsigned Opc = V1.getOpcode();
3014 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
3016 if (Opc != ISD::BUILD_VECTOR ||
3017 !X86::isZeroNode(V1.getOperand(Idx)))
3024 /// getZeroVector - Returns a vector of specified type with all zero elements.
3026 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3028 assert(VT.isVector() && "Expected a vector type");
3030 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3031 // type. This ensures they get CSE'd.
3033 if (VT.getSizeInBits() == 64) { // MMX
3034 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3035 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3036 } else if (HasSSE2) { // SSE2
3037 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3038 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3040 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3041 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3043 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3046 /// getOnesVector - Returns a vector of specified type with all bits set.
3048 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3049 assert(VT.isVector() && "Expected a vector type");
3051 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3052 // type. This ensures they get CSE'd.
3053 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3055 if (VT.getSizeInBits() == 64) // MMX
3056 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3058 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3059 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3063 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3064 /// that point to V2 points to its first element.
3065 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3066 EVT VT = SVOp->getValueType(0);
3067 unsigned NumElems = VT.getVectorNumElements();
3069 bool Changed = false;
3070 SmallVector<int, 8> MaskVec;
3071 SVOp->getMask(MaskVec);
3073 for (unsigned i = 0; i != NumElems; ++i) {
3074 if (MaskVec[i] > (int)NumElems) {
3075 MaskVec[i] = NumElems;
3080 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3081 SVOp->getOperand(1), &MaskVec[0]);
3082 return SDValue(SVOp, 0);
3085 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3086 /// operation of specified width.
3087 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3089 unsigned NumElems = VT.getVectorNumElements();
3090 SmallVector<int, 8> Mask;
3091 Mask.push_back(NumElems);
3092 for (unsigned i = 1; i != NumElems; ++i)
3094 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3097 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3098 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3100 unsigned NumElems = VT.getVectorNumElements();
3101 SmallVector<int, 8> Mask;
3102 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3104 Mask.push_back(i + NumElems);
3106 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3109 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3110 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3112 unsigned NumElems = VT.getVectorNumElements();
3113 unsigned Half = NumElems/2;
3114 SmallVector<int, 8> Mask;
3115 for (unsigned i = 0; i != Half; ++i) {
3116 Mask.push_back(i + Half);
3117 Mask.push_back(i + NumElems + Half);
3119 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3122 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3123 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3125 if (SV->getValueType(0).getVectorNumElements() <= 4)
3126 return SDValue(SV, 0);
3128 EVT PVT = MVT::v4f32;
3129 EVT VT = SV->getValueType(0);
3130 DebugLoc dl = SV->getDebugLoc();
3131 SDValue V1 = SV->getOperand(0);
3132 int NumElems = VT.getVectorNumElements();
3133 int EltNo = SV->getSplatIndex();
3135 // unpack elements to the correct location
3136 while (NumElems > 4) {
3137 if (EltNo < NumElems/2) {
3138 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3140 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3141 EltNo -= NumElems/2;
3146 // Perform the splat.
3147 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3148 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3149 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3150 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3153 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3154 /// vector of zero or undef vector. This produces a shuffle where the low
3155 /// element of V2 is swizzled into the zero/undef vector, landing at element
3156 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3157 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3158 bool isZero, bool HasSSE2,
3159 SelectionDAG &DAG) {
3160 EVT VT = V2.getValueType();
3162 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3163 unsigned NumElems = VT.getVectorNumElements();
3164 SmallVector<int, 16> MaskVec;
3165 for (unsigned i = 0; i != NumElems; ++i)
3166 // If this is the insertion idx, put the low elt of V2 here.
3167 MaskVec.push_back(i == Idx ? NumElems : i);
3168 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3171 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3172 /// a shuffle that is zero.
3174 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3175 bool Low, SelectionDAG &DAG) {
3176 unsigned NumZeros = 0;
3177 for (int i = 0; i < NumElems; ++i) {
3178 unsigned Index = Low ? i : NumElems-i-1;
3179 int Idx = SVOp->getMaskElt(Index);
3184 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3185 if (Elt.getNode() && X86::isZeroNode(Elt))
3193 /// isVectorShift - Returns true if the shuffle can be implemented as a
3194 /// logical left or right shift of a vector.
3195 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3196 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3197 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3198 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3201 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3204 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3208 bool SeenV1 = false;
3209 bool SeenV2 = false;
3210 for (int i = NumZeros; i < NumElems; ++i) {
3211 int Val = isLeft ? (i - NumZeros) : i;
3212 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3224 if (SeenV1 && SeenV2)
3227 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3233 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3235 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3236 unsigned NumNonZero, unsigned NumZero,
3237 SelectionDAG &DAG, TargetLowering &TLI) {
3241 DebugLoc dl = Op.getDebugLoc();
3244 for (unsigned i = 0; i < 16; ++i) {
3245 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3246 if (ThisIsNonZero && First) {
3248 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3250 V = DAG.getUNDEF(MVT::v8i16);
3255 SDValue ThisElt(0, 0), LastElt(0, 0);
3256 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3257 if (LastIsNonZero) {
3258 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3259 MVT::i16, Op.getOperand(i-1));
3261 if (ThisIsNonZero) {
3262 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3263 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3264 ThisElt, DAG.getConstant(8, MVT::i8));
3266 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3270 if (ThisElt.getNode())
3271 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3272 DAG.getIntPtrConstant(i/2));
3276 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3279 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3281 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3282 unsigned NumNonZero, unsigned NumZero,
3283 SelectionDAG &DAG, TargetLowering &TLI) {
3287 DebugLoc dl = Op.getDebugLoc();
3290 for (unsigned i = 0; i < 8; ++i) {
3291 bool isNonZero = (NonZeros & (1 << i)) != 0;
3295 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3297 V = DAG.getUNDEF(MVT::v8i16);
3300 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3301 MVT::v8i16, V, Op.getOperand(i),
3302 DAG.getIntPtrConstant(i));
3309 /// getVShift - Return a vector logical shift node.
3311 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3312 unsigned NumBits, SelectionDAG &DAG,
3313 const TargetLowering &TLI, DebugLoc dl) {
3314 bool isMMX = VT.getSizeInBits() == 64;
3315 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3316 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3317 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3318 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3319 DAG.getNode(Opc, dl, ShVT, SrcOp,
3320 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3324 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3325 DebugLoc dl = Op.getDebugLoc();
3326 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3327 if (ISD::isBuildVectorAllZeros(Op.getNode())
3328 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3329 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3330 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3331 // eliminated on x86-32 hosts.
3332 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3335 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3336 return getOnesVector(Op.getValueType(), DAG, dl);
3337 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3340 EVT VT = Op.getValueType();
3341 EVT ExtVT = VT.getVectorElementType();
3342 unsigned EVTBits = ExtVT.getSizeInBits();
3344 unsigned NumElems = Op.getNumOperands();
3345 unsigned NumZero = 0;
3346 unsigned NumNonZero = 0;
3347 unsigned NonZeros = 0;
3348 bool IsAllConstants = true;
3349 SmallSet<SDValue, 8> Values;
3350 for (unsigned i = 0; i < NumElems; ++i) {
3351 SDValue Elt = Op.getOperand(i);
3352 if (Elt.getOpcode() == ISD::UNDEF)
3355 if (Elt.getOpcode() != ISD::Constant &&
3356 Elt.getOpcode() != ISD::ConstantFP)
3357 IsAllConstants = false;
3358 if (X86::isZeroNode(Elt))
3361 NonZeros |= (1 << i);
3366 if (NumNonZero == 0) {
3367 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3368 return DAG.getUNDEF(VT);
3371 // Special case for single non-zero, non-undef, element.
3372 if (NumNonZero == 1) {
3373 unsigned Idx = CountTrailingZeros_32(NonZeros);
3374 SDValue Item = Op.getOperand(Idx);
3376 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3377 // the value are obviously zero, truncate the value to i32 and do the
3378 // insertion that way. Only do this if the value is non-constant or if the
3379 // value is a constant being inserted into element 0. It is cheaper to do
3380 // a constant pool load than it is to do a movd + shuffle.
3381 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3382 (!IsAllConstants || Idx == 0)) {
3383 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3384 // Handle MMX and SSE both.
3385 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3386 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3388 // Truncate the value (which may itself be a constant) to i32, and
3389 // convert it to a vector with movd (S2V+shuffle to zero extend).
3390 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3391 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3392 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3393 Subtarget->hasSSE2(), DAG);
3395 // Now we have our 32-bit value zero extended in the low element of
3396 // a vector. If Idx != 0, swizzle it into place.
3398 SmallVector<int, 4> Mask;
3399 Mask.push_back(Idx);
3400 for (unsigned i = 1; i != VecElts; ++i)
3402 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3403 DAG.getUNDEF(Item.getValueType()),
3406 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3410 // If we have a constant or non-constant insertion into the low element of
3411 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3412 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3413 // depending on what the source datatype is.
3416 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3417 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3418 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3419 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3420 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3421 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3423 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3424 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3425 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3426 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3427 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3428 Subtarget->hasSSE2(), DAG);
3429 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3433 // Is it a vector logical left shift?
3434 if (NumElems == 2 && Idx == 1 &&
3435 X86::isZeroNode(Op.getOperand(0)) &&
3436 !X86::isZeroNode(Op.getOperand(1))) {
3437 unsigned NumBits = VT.getSizeInBits();
3438 return getVShift(true, VT,
3439 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3440 VT, Op.getOperand(1)),
3441 NumBits/2, DAG, *this, dl);
3444 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3447 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3448 // is a non-constant being inserted into an element other than the low one,
3449 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3450 // movd/movss) to move this into the low element, then shuffle it into
3452 if (EVTBits == 32) {
3453 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3455 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3456 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3457 Subtarget->hasSSE2(), DAG);
3458 SmallVector<int, 8> MaskVec;
3459 for (unsigned i = 0; i < NumElems; i++)
3460 MaskVec.push_back(i == Idx ? 0 : 1);
3461 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3465 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3466 if (Values.size() == 1)
3469 // A vector full of immediates; various special cases are already
3470 // handled, so this is best done with a single constant-pool load.
3474 // Let legalizer expand 2-wide build_vectors.
3475 if (EVTBits == 64) {
3476 if (NumNonZero == 1) {
3477 // One half is zero or undef.
3478 unsigned Idx = CountTrailingZeros_32(NonZeros);
3479 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3480 Op.getOperand(Idx));
3481 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3482 Subtarget->hasSSE2(), DAG);
3487 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3488 if (EVTBits == 8 && NumElems == 16) {
3489 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3491 if (V.getNode()) return V;
3494 if (EVTBits == 16 && NumElems == 8) {
3495 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3497 if (V.getNode()) return V;
3500 // If element VT is == 32 bits, turn it into a number of shuffles.
3501 SmallVector<SDValue, 8> V;
3503 if (NumElems == 4 && NumZero > 0) {
3504 for (unsigned i = 0; i < 4; ++i) {
3505 bool isZero = !(NonZeros & (1 << i));
3507 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3509 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3512 for (unsigned i = 0; i < 2; ++i) {
3513 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3516 V[i] = V[i*2]; // Must be a zero vector.
3519 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3522 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3525 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3530 SmallVector<int, 8> MaskVec;
3531 bool Reverse = (NonZeros & 0x3) == 2;
3532 for (unsigned i = 0; i < 2; ++i)
3533 MaskVec.push_back(Reverse ? 1-i : i);
3534 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3535 for (unsigned i = 0; i < 2; ++i)
3536 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3537 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3540 if (Values.size() > 2) {
3541 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3542 // values to be inserted is equal to the number of elements, in which case
3543 // use the unpack code below in the hopes of matching the consecutive elts
3544 // load merge pattern for shuffles.
3545 // FIXME: We could probably just check that here directly.
3546 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3547 getSubtarget()->hasSSE41()) {
3548 V[0] = DAG.getUNDEF(VT);
3549 for (unsigned i = 0; i < NumElems; ++i)
3550 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3551 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3552 Op.getOperand(i), DAG.getIntPtrConstant(i));
3555 // Expand into a number of unpckl*.
3557 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3558 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3559 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3560 for (unsigned i = 0; i < NumElems; ++i)
3561 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3563 while (NumElems != 0) {
3564 for (unsigned i = 0; i < NumElems; ++i)
3565 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3574 // v8i16 shuffles - Prefer shuffles in the following order:
3575 // 1. [all] pshuflw, pshufhw, optional move
3576 // 2. [ssse3] 1 x pshufb
3577 // 3. [ssse3] 2 x pshufb + 1 x por
3578 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3580 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3581 SelectionDAG &DAG, X86TargetLowering &TLI) {
3582 SDValue V1 = SVOp->getOperand(0);
3583 SDValue V2 = SVOp->getOperand(1);
3584 DebugLoc dl = SVOp->getDebugLoc();
3585 SmallVector<int, 8> MaskVals;
3587 // Determine if more than 1 of the words in each of the low and high quadwords
3588 // of the result come from the same quadword of one of the two inputs. Undef
3589 // mask values count as coming from any quadword, for better codegen.
3590 SmallVector<unsigned, 4> LoQuad(4);
3591 SmallVector<unsigned, 4> HiQuad(4);
3592 BitVector InputQuads(4);
3593 for (unsigned i = 0; i < 8; ++i) {
3594 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3595 int EltIdx = SVOp->getMaskElt(i);
3596 MaskVals.push_back(EltIdx);
3605 InputQuads.set(EltIdx / 4);
3608 int BestLoQuad = -1;
3609 unsigned MaxQuad = 1;
3610 for (unsigned i = 0; i < 4; ++i) {
3611 if (LoQuad[i] > MaxQuad) {
3613 MaxQuad = LoQuad[i];
3617 int BestHiQuad = -1;
3619 for (unsigned i = 0; i < 4; ++i) {
3620 if (HiQuad[i] > MaxQuad) {
3622 MaxQuad = HiQuad[i];
3626 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3627 // of the two input vectors, shuffle them into one input vector so only a
3628 // single pshufb instruction is necessary. If There are more than 2 input
3629 // quads, disable the next transformation since it does not help SSSE3.
3630 bool V1Used = InputQuads[0] || InputQuads[1];
3631 bool V2Used = InputQuads[2] || InputQuads[3];
3632 if (TLI.getSubtarget()->hasSSSE3()) {
3633 if (InputQuads.count() == 2 && V1Used && V2Used) {
3634 BestLoQuad = InputQuads.find_first();
3635 BestHiQuad = InputQuads.find_next(BestLoQuad);
3637 if (InputQuads.count() > 2) {
3643 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3644 // the shuffle mask. If a quad is scored as -1, that means that it contains
3645 // words from all 4 input quadwords.
3647 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3648 SmallVector<int, 8> MaskV;
3649 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3650 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3651 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3652 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3653 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3654 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3656 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3657 // source words for the shuffle, to aid later transformations.
3658 bool AllWordsInNewV = true;
3659 bool InOrder[2] = { true, true };
3660 for (unsigned i = 0; i != 8; ++i) {
3661 int idx = MaskVals[i];
3663 InOrder[i/4] = false;
3664 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3666 AllWordsInNewV = false;
3670 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3671 if (AllWordsInNewV) {
3672 for (int i = 0; i != 8; ++i) {
3673 int idx = MaskVals[i];
3676 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3677 if ((idx != i) && idx < 4)
3679 if ((idx != i) && idx > 3)
3688 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3689 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3690 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3691 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3692 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3696 // If we have SSSE3, and all words of the result are from 1 input vector,
3697 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3698 // is present, fall back to case 4.
3699 if (TLI.getSubtarget()->hasSSSE3()) {
3700 SmallVector<SDValue,16> pshufbMask;
3702 // If we have elements from both input vectors, set the high bit of the
3703 // shuffle mask element to zero out elements that come from V2 in the V1
3704 // mask, and elements that come from V1 in the V2 mask, so that the two
3705 // results can be OR'd together.
3706 bool TwoInputs = V1Used && V2Used;
3707 for (unsigned i = 0; i != 8; ++i) {
3708 int EltIdx = MaskVals[i] * 2;
3709 if (TwoInputs && (EltIdx >= 16)) {
3710 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3711 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3714 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3715 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3717 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3718 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3719 DAG.getNode(ISD::BUILD_VECTOR, dl,
3720 MVT::v16i8, &pshufbMask[0], 16));
3722 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3724 // Calculate the shuffle mask for the second input, shuffle it, and
3725 // OR it with the first shuffled input.
3727 for (unsigned i = 0; i != 8; ++i) {
3728 int EltIdx = MaskVals[i] * 2;
3730 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3731 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3734 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3735 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3737 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3738 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3739 DAG.getNode(ISD::BUILD_VECTOR, dl,
3740 MVT::v16i8, &pshufbMask[0], 16));
3741 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3742 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3745 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3746 // and update MaskVals with new element order.
3747 BitVector InOrder(8);
3748 if (BestLoQuad >= 0) {
3749 SmallVector<int, 8> MaskV;
3750 for (int i = 0; i != 4; ++i) {
3751 int idx = MaskVals[i];
3753 MaskV.push_back(-1);
3755 } else if ((idx / 4) == BestLoQuad) {
3756 MaskV.push_back(idx & 3);
3759 MaskV.push_back(-1);
3762 for (unsigned i = 4; i != 8; ++i)
3764 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3768 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3769 // and update MaskVals with the new element order.
3770 if (BestHiQuad >= 0) {
3771 SmallVector<int, 8> MaskV;
3772 for (unsigned i = 0; i != 4; ++i)
3774 for (unsigned i = 4; i != 8; ++i) {
3775 int idx = MaskVals[i];
3777 MaskV.push_back(-1);
3779 } else if ((idx / 4) == BestHiQuad) {
3780 MaskV.push_back((idx & 3) + 4);
3783 MaskV.push_back(-1);
3786 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3790 // In case BestHi & BestLo were both -1, which means each quadword has a word
3791 // from each of the four input quadwords, calculate the InOrder bitvector now
3792 // before falling through to the insert/extract cleanup.
3793 if (BestLoQuad == -1 && BestHiQuad == -1) {
3795 for (int i = 0; i != 8; ++i)
3796 if (MaskVals[i] < 0 || MaskVals[i] == i)
3800 // The other elements are put in the right place using pextrw and pinsrw.
3801 for (unsigned i = 0; i != 8; ++i) {
3804 int EltIdx = MaskVals[i];
3807 SDValue ExtOp = (EltIdx < 8)
3808 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3809 DAG.getIntPtrConstant(EltIdx))
3810 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3811 DAG.getIntPtrConstant(EltIdx - 8));
3812 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3813 DAG.getIntPtrConstant(i));
3818 // v16i8 shuffles - Prefer shuffles in the following order:
3819 // 1. [ssse3] 1 x pshufb
3820 // 2. [ssse3] 2 x pshufb + 1 x por
3821 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3823 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3824 SelectionDAG &DAG, X86TargetLowering &TLI) {
3825 SDValue V1 = SVOp->getOperand(0);
3826 SDValue V2 = SVOp->getOperand(1);
3827 DebugLoc dl = SVOp->getDebugLoc();
3828 SmallVector<int, 16> MaskVals;
3829 SVOp->getMask(MaskVals);
3831 // If we have SSSE3, case 1 is generated when all result bytes come from
3832 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3833 // present, fall back to case 3.
3834 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3837 for (unsigned i = 0; i < 16; ++i) {
3838 int EltIdx = MaskVals[i];
3847 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3848 if (TLI.getSubtarget()->hasSSSE3()) {
3849 SmallVector<SDValue,16> pshufbMask;
3851 // If all result elements are from one input vector, then only translate
3852 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3854 // Otherwise, we have elements from both input vectors, and must zero out
3855 // elements that come from V2 in the first mask, and V1 in the second mask
3856 // so that we can OR them together.
3857 bool TwoInputs = !(V1Only || V2Only);
3858 for (unsigned i = 0; i != 16; ++i) {
3859 int EltIdx = MaskVals[i];
3860 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3861 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3864 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3866 // If all the elements are from V2, assign it to V1 and return after
3867 // building the first pshufb.
3870 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3871 DAG.getNode(ISD::BUILD_VECTOR, dl,
3872 MVT::v16i8, &pshufbMask[0], 16));
3876 // Calculate the shuffle mask for the second input, shuffle it, and
3877 // OR it with the first shuffled input.
3879 for (unsigned i = 0; i != 16; ++i) {
3880 int EltIdx = MaskVals[i];
3882 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3885 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3887 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3888 DAG.getNode(ISD::BUILD_VECTOR, dl,
3889 MVT::v16i8, &pshufbMask[0], 16));
3890 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3893 // No SSSE3 - Calculate in place words and then fix all out of place words
3894 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3895 // the 16 different words that comprise the two doublequadword input vectors.
3896 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3897 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3898 SDValue NewV = V2Only ? V2 : V1;
3899 for (int i = 0; i != 8; ++i) {
3900 int Elt0 = MaskVals[i*2];
3901 int Elt1 = MaskVals[i*2+1];
3903 // This word of the result is all undef, skip it.
3904 if (Elt0 < 0 && Elt1 < 0)
3907 // This word of the result is already in the correct place, skip it.
3908 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3910 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3913 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3914 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3917 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3918 // using a single extract together, load it and store it.
3919 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3920 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3921 DAG.getIntPtrConstant(Elt1 / 2));
3922 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3923 DAG.getIntPtrConstant(i));
3927 // If Elt1 is defined, extract it from the appropriate source. If the
3928 // source byte is not also odd, shift the extracted word left 8 bits
3929 // otherwise clear the bottom 8 bits if we need to do an or.
3931 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3932 DAG.getIntPtrConstant(Elt1 / 2));
3933 if ((Elt1 & 1) == 0)
3934 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3935 DAG.getConstant(8, TLI.getShiftAmountTy()));
3937 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3938 DAG.getConstant(0xFF00, MVT::i16));
3940 // If Elt0 is defined, extract it from the appropriate source. If the
3941 // source byte is not also even, shift the extracted word right 8 bits. If
3942 // Elt1 was also defined, OR the extracted values together before
3943 // inserting them in the result.
3945 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3946 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3947 if ((Elt0 & 1) != 0)
3948 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3949 DAG.getConstant(8, TLI.getShiftAmountTy()));
3951 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3952 DAG.getConstant(0x00FF, MVT::i16));
3953 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3956 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3957 DAG.getIntPtrConstant(i));
3959 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3962 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3963 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3964 /// done when every pair / quad of shuffle mask elements point to elements in
3965 /// the right sequence. e.g.
3966 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3968 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3970 TargetLowering &TLI, DebugLoc dl) {
3971 EVT VT = SVOp->getValueType(0);
3972 SDValue V1 = SVOp->getOperand(0);
3973 SDValue V2 = SVOp->getOperand(1);
3974 unsigned NumElems = VT.getVectorNumElements();
3975 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3976 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3977 EVT MaskEltVT = MaskVT.getVectorElementType();
3979 switch (VT.getSimpleVT().SimpleTy) {
3980 default: assert(false && "Unexpected!");
3981 case MVT::v4f32: NewVT = MVT::v2f64; break;
3982 case MVT::v4i32: NewVT = MVT::v2i64; break;
3983 case MVT::v8i16: NewVT = MVT::v4i32; break;
3984 case MVT::v16i8: NewVT = MVT::v4i32; break;
3987 if (NewWidth == 2) {
3993 int Scale = NumElems / NewWidth;
3994 SmallVector<int, 8> MaskVec;
3995 for (unsigned i = 0; i < NumElems; i += Scale) {
3997 for (int j = 0; j < Scale; ++j) {
3998 int EltIdx = SVOp->getMaskElt(i+j);
4002 StartIdx = EltIdx - (EltIdx % Scale);
4003 if (EltIdx != StartIdx + j)
4007 MaskVec.push_back(-1);
4009 MaskVec.push_back(StartIdx / Scale);
4012 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
4013 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
4014 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
4017 /// getVZextMovL - Return a zero-extending vector move low node.
4019 static SDValue getVZextMovL(EVT VT, EVT OpVT,
4020 SDValue SrcOp, SelectionDAG &DAG,
4021 const X86Subtarget *Subtarget, DebugLoc dl) {
4022 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
4023 LoadSDNode *LD = NULL;
4024 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
4025 LD = dyn_cast<LoadSDNode>(SrcOp);
4027 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4029 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4030 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4031 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4032 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4033 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4035 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4036 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4037 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4038 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4046 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4047 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4048 DAG.getNode(ISD::BIT_CONVERT, dl,
4052 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4055 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4056 SDValue V1 = SVOp->getOperand(0);
4057 SDValue V2 = SVOp->getOperand(1);
4058 DebugLoc dl = SVOp->getDebugLoc();
4059 EVT VT = SVOp->getValueType(0);
4061 SmallVector<std::pair<int, int>, 8> Locs;
4063 SmallVector<int, 8> Mask1(4U, -1);
4064 SmallVector<int, 8> PermMask;
4065 SVOp->getMask(PermMask);
4069 for (unsigned i = 0; i != 4; ++i) {
4070 int Idx = PermMask[i];
4072 Locs[i] = std::make_pair(-1, -1);
4074 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4076 Locs[i] = std::make_pair(0, NumLo);
4080 Locs[i] = std::make_pair(1, NumHi);
4082 Mask1[2+NumHi] = Idx;
4088 if (NumLo <= 2 && NumHi <= 2) {
4089 // If no more than two elements come from either vector. This can be
4090 // implemented with two shuffles. First shuffle gather the elements.
4091 // The second shuffle, which takes the first shuffle as both of its
4092 // vector operands, put the elements into the right order.
4093 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4095 SmallVector<int, 8> Mask2(4U, -1);
4097 for (unsigned i = 0; i != 4; ++i) {
4098 if (Locs[i].first == -1)
4101 unsigned Idx = (i < 2) ? 0 : 4;
4102 Idx += Locs[i].first * 2 + Locs[i].second;
4107 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4108 } else if (NumLo == 3 || NumHi == 3) {
4109 // Otherwise, we must have three elements from one vector, call it X, and
4110 // one element from the other, call it Y. First, use a shufps to build an
4111 // intermediate vector with the one element from Y and the element from X
4112 // that will be in the same half in the final destination (the indexes don't
4113 // matter). Then, use a shufps to build the final vector, taking the half
4114 // containing the element from Y from the intermediate, and the other half
4117 // Normalize it so the 3 elements come from V1.
4118 CommuteVectorShuffleMask(PermMask, VT);
4122 // Find the element from V2.
4124 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4125 int Val = PermMask[HiIndex];
4132 Mask1[0] = PermMask[HiIndex];
4134 Mask1[2] = PermMask[HiIndex^1];
4136 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4139 Mask1[0] = PermMask[0];
4140 Mask1[1] = PermMask[1];
4141 Mask1[2] = HiIndex & 1 ? 6 : 4;
4142 Mask1[3] = HiIndex & 1 ? 4 : 6;
4143 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4145 Mask1[0] = HiIndex & 1 ? 2 : 0;
4146 Mask1[1] = HiIndex & 1 ? 0 : 2;
4147 Mask1[2] = PermMask[2];
4148 Mask1[3] = PermMask[3];
4153 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4157 // Break it into (shuffle shuffle_hi, shuffle_lo).
4159 SmallVector<int,8> LoMask(4U, -1);
4160 SmallVector<int,8> HiMask(4U, -1);
4162 SmallVector<int,8> *MaskPtr = &LoMask;
4163 unsigned MaskIdx = 0;
4166 for (unsigned i = 0; i != 4; ++i) {
4173 int Idx = PermMask[i];
4175 Locs[i] = std::make_pair(-1, -1);
4176 } else if (Idx < 4) {
4177 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4178 (*MaskPtr)[LoIdx] = Idx;
4181 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4182 (*MaskPtr)[HiIdx] = Idx;
4187 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4188 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4189 SmallVector<int, 8> MaskOps;
4190 for (unsigned i = 0; i != 4; ++i) {
4191 if (Locs[i].first == -1) {
4192 MaskOps.push_back(-1);
4194 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4195 MaskOps.push_back(Idx);
4198 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4202 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4203 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4204 SDValue V1 = Op.getOperand(0);
4205 SDValue V2 = Op.getOperand(1);
4206 EVT VT = Op.getValueType();
4207 DebugLoc dl = Op.getDebugLoc();
4208 unsigned NumElems = VT.getVectorNumElements();
4209 bool isMMX = VT.getSizeInBits() == 64;
4210 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4211 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4212 bool V1IsSplat = false;
4213 bool V2IsSplat = false;
4215 if (isZeroShuffle(SVOp))
4216 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4218 // Promote splats to v4f32.
4219 if (SVOp->isSplat()) {
4220 if (isMMX || NumElems < 4)
4222 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4225 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4227 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4228 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4229 if (NewOp.getNode())
4230 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4231 LowerVECTOR_SHUFFLE(NewOp, DAG));
4232 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4233 // FIXME: Figure out a cleaner way to do this.
4234 // Try to make use of movq to zero out the top part.
4235 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4236 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4237 if (NewOp.getNode()) {
4238 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4239 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4240 DAG, Subtarget, dl);
4242 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4243 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4244 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4245 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4246 DAG, Subtarget, dl);
4250 if (X86::isPSHUFDMask(SVOp))
4253 // Check if this can be converted into a logical shift.
4254 bool isLeft = false;
4257 bool isShift = getSubtarget()->hasSSE2() &&
4258 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4259 if (isShift && ShVal.hasOneUse()) {
4260 // If the shifted value has multiple uses, it may be cheaper to use
4261 // v_set0 + movlhps or movhlps, etc.
4262 EVT EltVT = VT.getVectorElementType();
4263 ShAmt *= EltVT.getSizeInBits();
4264 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4267 if (X86::isMOVLMask(SVOp)) {
4270 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4271 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4276 // FIXME: fold these into legal mask.
4277 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4278 X86::isMOVSLDUPMask(SVOp) ||
4279 X86::isMOVHLPSMask(SVOp) ||
4280 X86::isMOVLHPSMask(SVOp) ||
4281 X86::isMOVLPMask(SVOp)))
4284 if (ShouldXformToMOVHLPS(SVOp) ||
4285 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4286 return CommuteVectorShuffle(SVOp, DAG);
4289 // No better options. Use a vshl / vsrl.
4290 EVT EltVT = VT.getVectorElementType();
4291 ShAmt *= EltVT.getSizeInBits();
4292 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4295 bool Commuted = false;
4296 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4297 // 1,1,1,1 -> v8i16 though.
4298 V1IsSplat = isSplatVector(V1.getNode());
4299 V2IsSplat = isSplatVector(V2.getNode());
4301 // Canonicalize the splat or undef, if present, to be on the RHS.
4302 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4303 Op = CommuteVectorShuffle(SVOp, DAG);
4304 SVOp = cast<ShuffleVectorSDNode>(Op);
4305 V1 = SVOp->getOperand(0);
4306 V2 = SVOp->getOperand(1);
4307 std::swap(V1IsSplat, V2IsSplat);
4308 std::swap(V1IsUndef, V2IsUndef);
4312 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4313 // Shuffling low element of v1 into undef, just return v1.
4316 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4317 // the instruction selector will not match, so get a canonical MOVL with
4318 // swapped operands to undo the commute.
4319 return getMOVL(DAG, dl, VT, V2, V1);
4322 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4323 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4324 X86::isUNPCKLMask(SVOp) ||
4325 X86::isUNPCKHMask(SVOp))
4329 // Normalize mask so all entries that point to V2 points to its first
4330 // element then try to match unpck{h|l} again. If match, return a
4331 // new vector_shuffle with the corrected mask.
4332 SDValue NewMask = NormalizeMask(SVOp, DAG);
4333 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4334 if (NSVOp != SVOp) {
4335 if (X86::isUNPCKLMask(NSVOp, true)) {
4337 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4344 // Commute is back and try unpck* again.
4345 // FIXME: this seems wrong.
4346 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4347 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4348 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4349 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4350 X86::isUNPCKLMask(NewSVOp) ||
4351 X86::isUNPCKHMask(NewSVOp))
4355 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4357 // Normalize the node to match x86 shuffle ops if needed
4358 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4359 return CommuteVectorShuffle(SVOp, DAG);
4361 // Check for legal shuffle and return?
4362 SmallVector<int, 16> PermMask;
4363 SVOp->getMask(PermMask);
4364 if (isShuffleMaskLegal(PermMask, VT))
4367 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4368 if (VT == MVT::v8i16) {
4369 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4370 if (NewOp.getNode())
4374 if (VT == MVT::v16i8) {
4375 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4376 if (NewOp.getNode())
4380 // Handle all 4 wide cases with a number of shuffles except for MMX.
4381 if (NumElems == 4 && !isMMX)
4382 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4388 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4389 SelectionDAG &DAG) {
4390 EVT VT = Op.getValueType();
4391 DebugLoc dl = Op.getDebugLoc();
4392 if (VT.getSizeInBits() == 8) {
4393 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4394 Op.getOperand(0), Op.getOperand(1));
4395 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4396 DAG.getValueType(VT));
4397 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4398 } else if (VT.getSizeInBits() == 16) {
4399 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4400 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4402 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4403 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4404 DAG.getNode(ISD::BIT_CONVERT, dl,
4408 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4409 Op.getOperand(0), Op.getOperand(1));
4410 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4411 DAG.getValueType(VT));
4412 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4413 } else if (VT == MVT::f32) {
4414 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4415 // the result back to FR32 register. It's only worth matching if the
4416 // result has a single use which is a store or a bitcast to i32. And in
4417 // the case of a store, it's not worth it if the index is a constant 0,
4418 // because a MOVSSmr can be used instead, which is smaller and faster.
4419 if (!Op.hasOneUse())
4421 SDNode *User = *Op.getNode()->use_begin();
4422 if ((User->getOpcode() != ISD::STORE ||
4423 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4424 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4425 (User->getOpcode() != ISD::BIT_CONVERT ||
4426 User->getValueType(0) != MVT::i32))
4428 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4429 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4432 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4433 } else if (VT == MVT::i32) {
4434 // ExtractPS works with constant index.
4435 if (isa<ConstantSDNode>(Op.getOperand(1)))
4443 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4444 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4447 if (Subtarget->hasSSE41()) {
4448 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4453 EVT VT = Op.getValueType();
4454 DebugLoc dl = Op.getDebugLoc();
4455 // TODO: handle v16i8.
4456 if (VT.getSizeInBits() == 16) {
4457 SDValue Vec = Op.getOperand(0);
4458 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4460 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4461 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4462 DAG.getNode(ISD::BIT_CONVERT, dl,
4465 // Transform it so it match pextrw which produces a 32-bit result.
4466 EVT EltVT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy+1);
4467 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4468 Op.getOperand(0), Op.getOperand(1));
4469 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4470 DAG.getValueType(VT));
4471 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4472 } else if (VT.getSizeInBits() == 32) {
4473 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4477 // SHUFPS the element to the lowest double word, then movss.
4478 int Mask[4] = { Idx, -1, -1, -1 };
4479 EVT VVT = Op.getOperand(0).getValueType();
4480 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4481 DAG.getUNDEF(VVT), Mask);
4482 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4483 DAG.getIntPtrConstant(0));
4484 } else if (VT.getSizeInBits() == 64) {
4485 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4486 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4487 // to match extract_elt for f64.
4488 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4492 // UNPCKHPD the element to the lowest double word, then movsd.
4493 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4494 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4495 int Mask[2] = { 1, -1 };
4496 EVT VVT = Op.getOperand(0).getValueType();
4497 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4498 DAG.getUNDEF(VVT), Mask);
4499 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4500 DAG.getIntPtrConstant(0));
4507 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4508 EVT VT = Op.getValueType();
4509 EVT EltVT = VT.getVectorElementType();
4510 DebugLoc dl = Op.getDebugLoc();
4512 SDValue N0 = Op.getOperand(0);
4513 SDValue N1 = Op.getOperand(1);
4514 SDValue N2 = Op.getOperand(2);
4516 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4517 isa<ConstantSDNode>(N2)) {
4518 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4520 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4522 if (N1.getValueType() != MVT::i32)
4523 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4524 if (N2.getValueType() != MVT::i32)
4525 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4526 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4527 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4528 // Bits [7:6] of the constant are the source select. This will always be
4529 // zero here. The DAG Combiner may combine an extract_elt index into these
4530 // bits. For example (insert (extract, 3), 2) could be matched by putting
4531 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4532 // Bits [5:4] of the constant are the destination select. This is the
4533 // value of the incoming immediate.
4534 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4535 // combine either bitwise AND or insert of float 0.0 to set these bits.
4536 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4537 // Create this as a scalar to vector..
4538 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4539 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4540 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4541 // PINSR* works with constant index.
4548 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4549 EVT VT = Op.getValueType();
4550 EVT EltVT = VT.getVectorElementType();
4552 if (Subtarget->hasSSE41())
4553 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4555 if (EltVT == MVT::i8)
4558 DebugLoc dl = Op.getDebugLoc();
4559 SDValue N0 = Op.getOperand(0);
4560 SDValue N1 = Op.getOperand(1);
4561 SDValue N2 = Op.getOperand(2);
4563 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4564 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4565 // as its second argument.
4566 if (N1.getValueType() != MVT::i32)
4567 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4568 if (N2.getValueType() != MVT::i32)
4569 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4570 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4576 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4577 DebugLoc dl = Op.getDebugLoc();
4578 if (Op.getValueType() == MVT::v2f32)
4579 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4580 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4581 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4582 Op.getOperand(0))));
4584 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4585 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4587 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4588 EVT VT = MVT::v2i32;
4589 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4596 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4597 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4600 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4601 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4602 // one of the above mentioned nodes. It has to be wrapped because otherwise
4603 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4604 // be used to form addressing mode. These wrapped nodes will be selected
4607 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4608 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4610 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4612 unsigned char OpFlag = 0;
4613 unsigned WrapperKind = X86ISD::Wrapper;
4614 CodeModel::Model M = getTargetMachine().getCodeModel();
4616 if (Subtarget->isPICStyleRIPRel() &&
4617 (M == CodeModel::Small || M == CodeModel::Kernel))
4618 WrapperKind = X86ISD::WrapperRIP;
4619 else if (Subtarget->isPICStyleGOT())
4620 OpFlag = X86II::MO_GOTOFF;
4621 else if (Subtarget->isPICStyleStubPIC())
4622 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4624 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4626 CP->getOffset(), OpFlag);
4627 DebugLoc DL = CP->getDebugLoc();
4628 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4629 // With PIC, the address is actually $g + Offset.
4631 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4632 DAG.getNode(X86ISD::GlobalBaseReg,
4633 DebugLoc::getUnknownLoc(), getPointerTy()),
4640 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4641 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4643 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4645 unsigned char OpFlag = 0;
4646 unsigned WrapperKind = X86ISD::Wrapper;
4647 CodeModel::Model M = getTargetMachine().getCodeModel();
4649 if (Subtarget->isPICStyleRIPRel() &&
4650 (M == CodeModel::Small || M == CodeModel::Kernel))
4651 WrapperKind = X86ISD::WrapperRIP;
4652 else if (Subtarget->isPICStyleGOT())
4653 OpFlag = X86II::MO_GOTOFF;
4654 else if (Subtarget->isPICStyleStubPIC())
4655 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4657 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4659 DebugLoc DL = JT->getDebugLoc();
4660 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4662 // With PIC, the address is actually $g + Offset.
4664 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4665 DAG.getNode(X86ISD::GlobalBaseReg,
4666 DebugLoc::getUnknownLoc(), getPointerTy()),
4674 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4675 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4677 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4679 unsigned char OpFlag = 0;
4680 unsigned WrapperKind = X86ISD::Wrapper;
4681 CodeModel::Model M = getTargetMachine().getCodeModel();
4683 if (Subtarget->isPICStyleRIPRel() &&
4684 (M == CodeModel::Small || M == CodeModel::Kernel))
4685 WrapperKind = X86ISD::WrapperRIP;
4686 else if (Subtarget->isPICStyleGOT())
4687 OpFlag = X86II::MO_GOTOFF;
4688 else if (Subtarget->isPICStyleStubPIC())
4689 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4691 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4693 DebugLoc DL = Op.getDebugLoc();
4694 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4697 // With PIC, the address is actually $g + Offset.
4698 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4699 !Subtarget->is64Bit()) {
4700 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4701 DAG.getNode(X86ISD::GlobalBaseReg,
4702 DebugLoc::getUnknownLoc(),
4711 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
4712 unsigned WrapperKind = X86ISD::Wrapper;
4713 CodeModel::Model M = getTargetMachine().getCodeModel();
4714 if (Subtarget->isPICStyleRIPRel() &&
4715 (M == CodeModel::Small || M == CodeModel::Kernel))
4716 WrapperKind = X86ISD::WrapperRIP;
4718 DebugLoc DL = Op.getDebugLoc();
4720 BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
4721 SDValue Result = DAG.getBlockAddress(BA, DL, /*isTarget=*/true);
4723 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4729 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4731 SelectionDAG &DAG) const {
4732 // Create the TargetGlobalAddress node, folding in the constant
4733 // offset if it is legal.
4734 unsigned char OpFlags =
4735 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4736 CodeModel::Model M = getTargetMachine().getCodeModel();
4738 if (OpFlags == X86II::MO_NO_FLAG &&
4739 X86::isOffsetSuitableForCodeModel(Offset, M)) {
4740 // A direct static reference to a global.
4741 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4744 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4747 if (Subtarget->isPICStyleRIPRel() &&
4748 (M == CodeModel::Small || M == CodeModel::Kernel))
4749 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4751 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4753 // With PIC, the address is actually $g + Offset.
4754 if (isGlobalRelativeToPICBase(OpFlags)) {
4755 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4756 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4760 // For globals that require a load from a stub to get the address, emit the
4762 if (isGlobalStubReference(OpFlags))
4763 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4764 PseudoSourceValue::getGOT(), 0);
4766 // If there was a non-zero offset that we didn't fold, create an explicit
4769 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4770 DAG.getConstant(Offset, getPointerTy()));
4776 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4777 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4778 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4779 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4783 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4784 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
4785 unsigned char OperandFlags) {
4786 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4787 DebugLoc dl = GA->getDebugLoc();
4788 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4789 GA->getValueType(0),
4793 SDValue Ops[] = { Chain, TGA, *InFlag };
4794 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4796 SDValue Ops[] = { Chain, TGA };
4797 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4799 SDValue Flag = Chain.getValue(1);
4800 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4803 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4805 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4808 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4809 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4810 DAG.getNode(X86ISD::GlobalBaseReg,
4811 DebugLoc::getUnknownLoc(),
4813 InFlag = Chain.getValue(1);
4815 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4818 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4820 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4822 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4823 X86::RAX, X86II::MO_TLSGD);
4826 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4827 // "local exec" model.
4828 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4829 const EVT PtrVT, TLSModel::Model model,
4831 DebugLoc dl = GA->getDebugLoc();
4832 // Get the Thread Pointer
4833 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4834 DebugLoc::getUnknownLoc(), PtrVT,
4835 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4838 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4841 unsigned char OperandFlags = 0;
4842 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
4844 unsigned WrapperKind = X86ISD::Wrapper;
4845 if (model == TLSModel::LocalExec) {
4846 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
4847 } else if (is64Bit) {
4848 assert(model == TLSModel::InitialExec);
4849 OperandFlags = X86II::MO_GOTTPOFF;
4850 WrapperKind = X86ISD::WrapperRIP;
4852 assert(model == TLSModel::InitialExec);
4853 OperandFlags = X86II::MO_INDNTPOFF;
4856 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4858 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4859 GA->getOffset(), OperandFlags);
4860 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
4862 if (model == TLSModel::InitialExec)
4863 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4864 PseudoSourceValue::getGOT(), 0);
4866 // The address of the thread local variable is the add of the thread
4867 // pointer with the offset of the variable.
4868 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4872 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4873 // TODO: implement the "local dynamic" model
4874 // TODO: implement the "initial exec"model for pic executables
4875 assert(Subtarget->isTargetELF() &&
4876 "TLS not implemented for non-ELF targets");
4877 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4878 const GlobalValue *GV = GA->getGlobal();
4880 // If GV is an alias then use the aliasee for determining
4881 // thread-localness.
4882 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
4883 GV = GA->resolveAliasedGlobal(false);
4885 TLSModel::Model model = getTLSModel(GV,
4886 getTargetMachine().getRelocationModel());
4889 case TLSModel::GeneralDynamic:
4890 case TLSModel::LocalDynamic: // not implemented
4891 if (Subtarget->is64Bit())
4892 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4893 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4895 case TLSModel::InitialExec:
4896 case TLSModel::LocalExec:
4897 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
4898 Subtarget->is64Bit());
4901 llvm_unreachable("Unreachable");
4906 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4907 /// take a 2 x i32 value to shift plus a shift amount.
4908 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4909 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4910 EVT VT = Op.getValueType();
4911 unsigned VTBits = VT.getSizeInBits();
4912 DebugLoc dl = Op.getDebugLoc();
4913 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4914 SDValue ShOpLo = Op.getOperand(0);
4915 SDValue ShOpHi = Op.getOperand(1);
4916 SDValue ShAmt = Op.getOperand(2);
4917 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4918 DAG.getConstant(VTBits - 1, MVT::i8))
4919 : DAG.getConstant(0, VT);
4922 if (Op.getOpcode() == ISD::SHL_PARTS) {
4923 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4924 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4926 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4927 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4930 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4931 DAG.getConstant(VTBits, MVT::i8));
4932 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4933 AndNode, DAG.getConstant(0, MVT::i8));
4936 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4937 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4938 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4940 if (Op.getOpcode() == ISD::SHL_PARTS) {
4941 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4942 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4944 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4945 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4948 SDValue Ops[2] = { Lo, Hi };
4949 return DAG.getMergeValues(Ops, 2, dl);
4952 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4953 EVT SrcVT = Op.getOperand(0).getValueType();
4955 if (SrcVT.isVector()) {
4956 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
4962 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4963 "Unknown SINT_TO_FP to lower!");
4965 // These are really Legal; return the operand so the caller accepts it as
4967 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4969 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
4970 Subtarget->is64Bit()) {
4974 DebugLoc dl = Op.getDebugLoc();
4975 unsigned Size = SrcVT.getSizeInBits()/8;
4976 MachineFunction &MF = DAG.getMachineFunction();
4977 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
4978 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4979 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4981 PseudoSourceValue::getFixedStack(SSFI), 0);
4982 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
4985 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
4987 SelectionDAG &DAG) {
4989 DebugLoc dl = Op.getDebugLoc();
4991 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4993 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4995 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4996 SmallVector<SDValue, 8> Ops;
4997 Ops.push_back(Chain);
4998 Ops.push_back(StackSlot);
4999 Ops.push_back(DAG.getValueType(SrcVT));
5000 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
5001 Tys, &Ops[0], Ops.size());
5004 Chain = Result.getValue(1);
5005 SDValue InFlag = Result.getValue(2);
5007 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
5008 // shouldn't be necessary except that RFP cannot be live across
5009 // multiple blocks. When stackifier is fixed, they can be uncoupled.
5010 MachineFunction &MF = DAG.getMachineFunction();
5011 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8, false);
5012 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5013 Tys = DAG.getVTList(MVT::Other);
5014 SmallVector<SDValue, 8> Ops;
5015 Ops.push_back(Chain);
5016 Ops.push_back(Result);
5017 Ops.push_back(StackSlot);
5018 Ops.push_back(DAG.getValueType(Op.getValueType()));
5019 Ops.push_back(InFlag);
5020 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
5021 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
5022 PseudoSourceValue::getFixedStack(SSFI), 0);
5028 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
5029 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
5030 // This algorithm is not obvious. Here it is in C code, more or less:
5032 double uint64_to_double( uint32_t hi, uint32_t lo ) {
5033 static const __m128i exp = { 0x4330000045300000ULL, 0 };
5034 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
5036 // Copy ints to xmm registers.
5037 __m128i xh = _mm_cvtsi32_si128( hi );
5038 __m128i xl = _mm_cvtsi32_si128( lo );
5040 // Combine into low half of a single xmm register.
5041 __m128i x = _mm_unpacklo_epi32( xh, xl );
5045 // Merge in appropriate exponents to give the integer bits the right
5047 x = _mm_unpacklo_epi32( x, exp );
5049 // Subtract away the biases to deal with the IEEE-754 double precision
5051 d = _mm_sub_pd( (__m128d) x, bias );
5053 // All conversions up to here are exact. The correctly rounded result is
5054 // calculated using the current rounding mode using the following
5056 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5057 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5058 // store doesn't really need to be here (except
5059 // maybe to zero the other double)
5064 DebugLoc dl = Op.getDebugLoc();
5065 LLVMContext *Context = DAG.getContext();
5067 // Build some magic constants.
5068 std::vector<Constant*> CV0;
5069 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5070 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5071 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5072 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5073 Constant *C0 = ConstantVector::get(CV0);
5074 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5076 std::vector<Constant*> CV1;
5078 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5080 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5081 Constant *C1 = ConstantVector::get(CV1);
5082 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5084 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5085 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5087 DAG.getIntPtrConstant(1)));
5088 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5089 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5091 DAG.getIntPtrConstant(0)));
5092 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5093 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5094 PseudoSourceValue::getConstantPool(), 0,
5096 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5097 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5098 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5099 PseudoSourceValue::getConstantPool(), 0,
5101 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5103 // Add the halves; easiest way is to swap them into another reg first.
5104 int ShufMask[2] = { 1, -1 };
5105 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5106 DAG.getUNDEF(MVT::v2f64), ShufMask);
5107 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5108 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5109 DAG.getIntPtrConstant(0));
5112 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5113 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5114 DebugLoc dl = Op.getDebugLoc();
5115 // FP constant to bias correct the final result.
5116 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5119 // Load the 32-bit value into an XMM register.
5120 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5121 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5123 DAG.getIntPtrConstant(0)));
5125 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5126 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5127 DAG.getIntPtrConstant(0));
5129 // Or the load with the bias.
5130 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5131 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5132 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5134 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5135 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5136 MVT::v2f64, Bias)));
5137 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5138 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5139 DAG.getIntPtrConstant(0));
5141 // Subtract the bias.
5142 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5144 // Handle final rounding.
5145 EVT DestVT = Op.getValueType();
5147 if (DestVT.bitsLT(MVT::f64)) {
5148 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5149 DAG.getIntPtrConstant(0));
5150 } else if (DestVT.bitsGT(MVT::f64)) {
5151 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5154 // Handle final rounding.
5158 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5159 SDValue N0 = Op.getOperand(0);
5160 DebugLoc dl = Op.getDebugLoc();
5162 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5163 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5164 // the optimization here.
5165 if (DAG.SignBitIsZero(N0))
5166 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5168 EVT SrcVT = N0.getValueType();
5169 if (SrcVT == MVT::i64) {
5170 // We only handle SSE2 f64 target here; caller can expand the rest.
5171 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5174 return LowerUINT_TO_FP_i64(Op, DAG);
5175 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5176 return LowerUINT_TO_FP_i32(Op, DAG);
5179 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5181 // Make a 64-bit buffer, and use it to build an FILD.
5182 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5183 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5184 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5185 getPointerTy(), StackSlot, WordOff);
5186 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5187 StackSlot, NULL, 0);
5188 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5189 OffsetSlot, NULL, 0);
5190 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5193 std::pair<SDValue,SDValue> X86TargetLowering::
5194 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5195 DebugLoc dl = Op.getDebugLoc();
5197 EVT DstTy = Op.getValueType();
5200 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5204 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5205 DstTy.getSimpleVT() >= MVT::i16 &&
5206 "Unknown FP_TO_SINT to lower!");
5208 // These are really Legal.
5209 if (DstTy == MVT::i32 &&
5210 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5211 return std::make_pair(SDValue(), SDValue());
5212 if (Subtarget->is64Bit() &&
5213 DstTy == MVT::i64 &&
5214 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5215 return std::make_pair(SDValue(), SDValue());
5217 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5219 MachineFunction &MF = DAG.getMachineFunction();
5220 unsigned MemSize = DstTy.getSizeInBits()/8;
5221 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5222 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5225 switch (DstTy.getSimpleVT().SimpleTy) {
5226 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5227 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5228 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5229 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5232 SDValue Chain = DAG.getEntryNode();
5233 SDValue Value = Op.getOperand(0);
5234 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5235 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5236 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5237 PseudoSourceValue::getFixedStack(SSFI), 0);
5238 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5240 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5242 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5243 Chain = Value.getValue(1);
5244 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
5245 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5248 // Build the FP_TO_INT*_IN_MEM
5249 SDValue Ops[] = { Chain, Value, StackSlot };
5250 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5252 return std::make_pair(FIST, StackSlot);
5255 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5256 if (Op.getValueType().isVector()) {
5257 if (Op.getValueType() == MVT::v2i32 &&
5258 Op.getOperand(0).getValueType() == MVT::v2f64) {
5264 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5265 SDValue FIST = Vals.first, StackSlot = Vals.second;
5266 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5267 if (FIST.getNode() == 0) return Op;
5270 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5271 FIST, StackSlot, NULL, 0);
5274 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5275 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5276 SDValue FIST = Vals.first, StackSlot = Vals.second;
5277 assert(FIST.getNode() && "Unexpected failure");
5280 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5281 FIST, StackSlot, NULL, 0);
5284 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5285 LLVMContext *Context = DAG.getContext();
5286 DebugLoc dl = Op.getDebugLoc();
5287 EVT VT = Op.getValueType();
5290 EltVT = VT.getVectorElementType();
5291 std::vector<Constant*> CV;
5292 if (EltVT == MVT::f64) {
5293 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5297 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5303 Constant *C = ConstantVector::get(CV);
5304 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5305 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5306 PseudoSourceValue::getConstantPool(), 0,
5308 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5311 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5312 LLVMContext *Context = DAG.getContext();
5313 DebugLoc dl = Op.getDebugLoc();
5314 EVT VT = Op.getValueType();
5317 EltVT = VT.getVectorElementType();
5318 std::vector<Constant*> CV;
5319 if (EltVT == MVT::f64) {
5320 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5324 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5330 Constant *C = ConstantVector::get(CV);
5331 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5332 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5333 PseudoSourceValue::getConstantPool(), 0,
5335 if (VT.isVector()) {
5336 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5337 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5338 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5340 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5342 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5346 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5347 LLVMContext *Context = DAG.getContext();
5348 SDValue Op0 = Op.getOperand(0);
5349 SDValue Op1 = Op.getOperand(1);
5350 DebugLoc dl = Op.getDebugLoc();
5351 EVT VT = Op.getValueType();
5352 EVT SrcVT = Op1.getValueType();
5354 // If second operand is smaller, extend it first.
5355 if (SrcVT.bitsLT(VT)) {
5356 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5359 // And if it is bigger, shrink it first.
5360 if (SrcVT.bitsGT(VT)) {
5361 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5365 // At this point the operands and the result should have the same
5366 // type, and that won't be f80 since that is not custom lowered.
5368 // First get the sign bit of second operand.
5369 std::vector<Constant*> CV;
5370 if (SrcVT == MVT::f64) {
5371 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5372 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5374 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5375 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5376 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5377 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5379 Constant *C = ConstantVector::get(CV);
5380 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5381 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5382 PseudoSourceValue::getConstantPool(), 0,
5384 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5386 // Shift sign bit right or left if the two operands have different types.
5387 if (SrcVT.bitsGT(VT)) {
5388 // Op0 is MVT::f32, Op1 is MVT::f64.
5389 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5390 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5391 DAG.getConstant(32, MVT::i32));
5392 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5393 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5394 DAG.getIntPtrConstant(0));
5397 // Clear first operand sign bit.
5399 if (VT == MVT::f64) {
5400 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5401 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5403 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5404 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5405 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5406 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5408 C = ConstantVector::get(CV);
5409 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5410 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5411 PseudoSourceValue::getConstantPool(), 0,
5413 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5415 // Or the value with the sign bit.
5416 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5419 /// Emit nodes that will be selected as "test Op0,Op0", or something
5421 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5422 SelectionDAG &DAG) {
5423 DebugLoc dl = Op.getDebugLoc();
5425 // CF and OF aren't always set the way we want. Determine which
5426 // of these we need.
5427 bool NeedCF = false;
5428 bool NeedOF = false;
5430 case X86::COND_A: case X86::COND_AE:
5431 case X86::COND_B: case X86::COND_BE:
5434 case X86::COND_G: case X86::COND_GE:
5435 case X86::COND_L: case X86::COND_LE:
5436 case X86::COND_O: case X86::COND_NO:
5442 // See if we can use the EFLAGS value from the operand instead of
5443 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5444 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5445 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5446 unsigned Opcode = 0;
5447 unsigned NumOperands = 0;
5448 switch (Op.getNode()->getOpcode()) {
5450 // Due to an isel shortcoming, be conservative if this add is likely to
5451 // be selected as part of a load-modify-store instruction. When the root
5452 // node in a match is a store, isel doesn't know how to remap non-chain
5453 // non-flag uses of other nodes in the match, such as the ADD in this
5454 // case. This leads to the ADD being left around and reselected, with
5455 // the result being two adds in the output.
5456 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5457 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5458 if (UI->getOpcode() == ISD::STORE)
5460 if (ConstantSDNode *C =
5461 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5462 // An add of one will be selected as an INC.
5463 if (C->getAPIntValue() == 1) {
5464 Opcode = X86ISD::INC;
5468 // An add of negative one (subtract of one) will be selected as a DEC.
5469 if (C->getAPIntValue().isAllOnesValue()) {
5470 Opcode = X86ISD::DEC;
5475 // Otherwise use a regular EFLAGS-setting add.
5476 Opcode = X86ISD::ADD;
5480 // If the primary and result isn't used, don't bother using X86ISD::AND,
5481 // because a TEST instruction will be better.
5482 bool NonFlagUse = false;
5483 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5484 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5485 if (UI->getOpcode() != ISD::BRCOND &&
5486 UI->getOpcode() != ISD::SELECT &&
5487 UI->getOpcode() != ISD::SETCC) {
5498 // Due to the ISEL shortcoming noted above, be conservative if this op is
5499 // likely to be selected as part of a load-modify-store instruction.
5500 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5501 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5502 if (UI->getOpcode() == ISD::STORE)
5504 // Otherwise use a regular EFLAGS-setting instruction.
5505 switch (Op.getNode()->getOpcode()) {
5506 case ISD::SUB: Opcode = X86ISD::SUB; break;
5507 case ISD::OR: Opcode = X86ISD::OR; break;
5508 case ISD::XOR: Opcode = X86ISD::XOR; break;
5509 case ISD::AND: Opcode = X86ISD::AND; break;
5510 default: llvm_unreachable("unexpected operator!");
5521 return SDValue(Op.getNode(), 1);
5527 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5528 SmallVector<SDValue, 4> Ops;
5529 for (unsigned i = 0; i != NumOperands; ++i)
5530 Ops.push_back(Op.getOperand(i));
5531 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5532 DAG.ReplaceAllUsesWith(Op, New);
5533 return SDValue(New.getNode(), 1);
5537 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5538 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5539 DAG.getConstant(0, Op.getValueType()));
5542 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5544 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5545 SelectionDAG &DAG) {
5546 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5547 if (C->getAPIntValue() == 0)
5548 return EmitTest(Op0, X86CC, DAG);
5550 DebugLoc dl = Op0.getDebugLoc();
5551 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5554 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5555 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5556 SDValue Op0 = Op.getOperand(0);
5557 SDValue Op1 = Op.getOperand(1);
5558 DebugLoc dl = Op.getDebugLoc();
5559 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5561 // Lower (X & (1 << N)) == 0 to BT(X, N).
5562 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5563 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5564 if (Op0.getOpcode() == ISD::AND &&
5566 Op1.getOpcode() == ISD::Constant &&
5567 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5568 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5570 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5571 if (ConstantSDNode *Op010C =
5572 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5573 if (Op010C->getZExtValue() == 1) {
5574 LHS = Op0.getOperand(0);
5575 RHS = Op0.getOperand(1).getOperand(1);
5577 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5578 if (ConstantSDNode *Op000C =
5579 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5580 if (Op000C->getZExtValue() == 1) {
5581 LHS = Op0.getOperand(1);
5582 RHS = Op0.getOperand(0).getOperand(1);
5584 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5585 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5586 SDValue AndLHS = Op0.getOperand(0);
5587 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5588 LHS = AndLHS.getOperand(0);
5589 RHS = AndLHS.getOperand(1);
5593 if (LHS.getNode()) {
5594 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5595 // instruction. Since the shift amount is in-range-or-undefined, we know
5596 // that doing a bittest on the i16 value is ok. We extend to i32 because
5597 // the encoding for the i16 version is larger than the i32 version.
5598 if (LHS.getValueType() == MVT::i8)
5599 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5601 // If the operand types disagree, extend the shift amount to match. Since
5602 // BT ignores high bits (like shifts) we can use anyextend.
5603 if (LHS.getValueType() != RHS.getValueType())
5604 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5606 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5607 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5608 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5609 DAG.getConstant(Cond, MVT::i8), BT);
5613 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5614 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5615 if (X86CC == X86::COND_INVALID)
5618 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5619 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5620 DAG.getConstant(X86CC, MVT::i8), Cond);
5623 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5625 SDValue Op0 = Op.getOperand(0);
5626 SDValue Op1 = Op.getOperand(1);
5627 SDValue CC = Op.getOperand(2);
5628 EVT VT = Op.getValueType();
5629 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5630 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5631 DebugLoc dl = Op.getDebugLoc();
5635 EVT VT0 = Op0.getValueType();
5636 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5637 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5640 switch (SetCCOpcode) {
5643 case ISD::SETEQ: SSECC = 0; break;
5645 case ISD::SETGT: Swap = true; // Fallthrough
5647 case ISD::SETOLT: SSECC = 1; break;
5649 case ISD::SETGE: Swap = true; // Fallthrough
5651 case ISD::SETOLE: SSECC = 2; break;
5652 case ISD::SETUO: SSECC = 3; break;
5654 case ISD::SETNE: SSECC = 4; break;
5655 case ISD::SETULE: Swap = true;
5656 case ISD::SETUGE: SSECC = 5; break;
5657 case ISD::SETULT: Swap = true;
5658 case ISD::SETUGT: SSECC = 6; break;
5659 case ISD::SETO: SSECC = 7; break;
5662 std::swap(Op0, Op1);
5664 // In the two special cases we can't handle, emit two comparisons.
5666 if (SetCCOpcode == ISD::SETUEQ) {
5668 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5669 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5670 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5672 else if (SetCCOpcode == ISD::SETONE) {
5674 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5675 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5676 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5678 llvm_unreachable("Illegal FP comparison");
5680 // Handle all other FP comparisons here.
5681 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5684 // We are handling one of the integer comparisons here. Since SSE only has
5685 // GT and EQ comparisons for integer, swapping operands and multiple
5686 // operations may be required for some comparisons.
5687 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5688 bool Swap = false, Invert = false, FlipSigns = false;
5690 switch (VT.getSimpleVT().SimpleTy) {
5693 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5695 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5697 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5698 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5701 switch (SetCCOpcode) {
5703 case ISD::SETNE: Invert = true;
5704 case ISD::SETEQ: Opc = EQOpc; break;
5705 case ISD::SETLT: Swap = true;
5706 case ISD::SETGT: Opc = GTOpc; break;
5707 case ISD::SETGE: Swap = true;
5708 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5709 case ISD::SETULT: Swap = true;
5710 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5711 case ISD::SETUGE: Swap = true;
5712 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5715 std::swap(Op0, Op1);
5717 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5718 // bits of the inputs before performing those operations.
5720 EVT EltVT = VT.getVectorElementType();
5721 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5723 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5724 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5726 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5727 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5730 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5732 // If the logical-not of the result is required, perform that now.
5734 Result = DAG.getNOT(dl, Result, VT);
5739 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5740 static bool isX86LogicalCmp(SDValue Op) {
5741 unsigned Opc = Op.getNode()->getOpcode();
5742 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5744 if (Op.getResNo() == 1 &&
5745 (Opc == X86ISD::ADD ||
5746 Opc == X86ISD::SUB ||
5747 Opc == X86ISD::SMUL ||
5748 Opc == X86ISD::UMUL ||
5749 Opc == X86ISD::INC ||
5750 Opc == X86ISD::DEC ||
5751 Opc == X86ISD::OR ||
5752 Opc == X86ISD::XOR ||
5753 Opc == X86ISD::AND))
5759 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5760 bool addTest = true;
5761 SDValue Cond = Op.getOperand(0);
5762 DebugLoc dl = Op.getDebugLoc();
5765 if (Cond.getOpcode() == ISD::SETCC) {
5766 SDValue NewCond = LowerSETCC(Cond, DAG);
5767 if (NewCond.getNode())
5771 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5772 // setting operand in place of the X86ISD::SETCC.
5773 if (Cond.getOpcode() == X86ISD::SETCC) {
5774 CC = Cond.getOperand(0);
5776 SDValue Cmp = Cond.getOperand(1);
5777 unsigned Opc = Cmp.getOpcode();
5778 EVT VT = Op.getValueType();
5780 bool IllegalFPCMov = false;
5781 if (VT.isFloatingPoint() && !VT.isVector() &&
5782 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5783 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5785 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5786 Opc == X86ISD::BT) { // FIXME
5793 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5794 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5797 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5798 SmallVector<SDValue, 4> Ops;
5799 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5800 // condition is true.
5801 Ops.push_back(Op.getOperand(2));
5802 Ops.push_back(Op.getOperand(1));
5804 Ops.push_back(Cond);
5805 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5808 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5809 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5810 // from the AND / OR.
5811 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5812 Opc = Op.getOpcode();
5813 if (Opc != ISD::OR && Opc != ISD::AND)
5815 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5816 Op.getOperand(0).hasOneUse() &&
5817 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5818 Op.getOperand(1).hasOneUse());
5821 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5822 // 1 and that the SETCC node has a single use.
5823 static bool isXor1OfSetCC(SDValue Op) {
5824 if (Op.getOpcode() != ISD::XOR)
5826 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5827 if (N1C && N1C->getAPIntValue() == 1) {
5828 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5829 Op.getOperand(0).hasOneUse();
5834 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5835 bool addTest = true;
5836 SDValue Chain = Op.getOperand(0);
5837 SDValue Cond = Op.getOperand(1);
5838 SDValue Dest = Op.getOperand(2);
5839 DebugLoc dl = Op.getDebugLoc();
5842 if (Cond.getOpcode() == ISD::SETCC) {
5843 SDValue NewCond = LowerSETCC(Cond, DAG);
5844 if (NewCond.getNode())
5848 // FIXME: LowerXALUO doesn't handle these!!
5849 else if (Cond.getOpcode() == X86ISD::ADD ||
5850 Cond.getOpcode() == X86ISD::SUB ||
5851 Cond.getOpcode() == X86ISD::SMUL ||
5852 Cond.getOpcode() == X86ISD::UMUL)
5853 Cond = LowerXALUO(Cond, DAG);
5856 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5857 // setting operand in place of the X86ISD::SETCC.
5858 if (Cond.getOpcode() == X86ISD::SETCC) {
5859 CC = Cond.getOperand(0);
5861 SDValue Cmp = Cond.getOperand(1);
5862 unsigned Opc = Cmp.getOpcode();
5863 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5864 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5868 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5872 // These can only come from an arithmetic instruction with overflow,
5873 // e.g. SADDO, UADDO.
5874 Cond = Cond.getNode()->getOperand(1);
5881 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5882 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5883 if (CondOpc == ISD::OR) {
5884 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5885 // two branches instead of an explicit OR instruction with a
5887 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5888 isX86LogicalCmp(Cmp)) {
5889 CC = Cond.getOperand(0).getOperand(0);
5890 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5891 Chain, Dest, CC, Cmp);
5892 CC = Cond.getOperand(1).getOperand(0);
5896 } else { // ISD::AND
5897 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5898 // two branches instead of an explicit AND instruction with a
5899 // separate test. However, we only do this if this block doesn't
5900 // have a fall-through edge, because this requires an explicit
5901 // jmp when the condition is false.
5902 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5903 isX86LogicalCmp(Cmp) &&
5904 Op.getNode()->hasOneUse()) {
5905 X86::CondCode CCode =
5906 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5907 CCode = X86::GetOppositeBranchCondition(CCode);
5908 CC = DAG.getConstant(CCode, MVT::i8);
5909 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5910 // Look for an unconditional branch following this conditional branch.
5911 // We need this because we need to reverse the successors in order
5912 // to implement FCMP_OEQ.
5913 if (User.getOpcode() == ISD::BR) {
5914 SDValue FalseBB = User.getOperand(1);
5916 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5917 assert(NewBR == User);
5920 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5921 Chain, Dest, CC, Cmp);
5922 X86::CondCode CCode =
5923 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5924 CCode = X86::GetOppositeBranchCondition(CCode);
5925 CC = DAG.getConstant(CCode, MVT::i8);
5931 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5932 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5933 // It should be transformed during dag combiner except when the condition
5934 // is set by a arithmetics with overflow node.
5935 X86::CondCode CCode =
5936 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5937 CCode = X86::GetOppositeBranchCondition(CCode);
5938 CC = DAG.getConstant(CCode, MVT::i8);
5939 Cond = Cond.getOperand(0).getOperand(1);
5945 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5946 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5948 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5949 Chain, Dest, CC, Cond);
5953 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5954 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5955 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5956 // that the guard pages used by the OS virtual memory manager are allocated in
5957 // correct sequence.
5959 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5960 SelectionDAG &DAG) {
5961 assert(Subtarget->isTargetCygMing() &&
5962 "This should be used only on Cygwin/Mingw targets");
5963 DebugLoc dl = Op.getDebugLoc();
5966 SDValue Chain = Op.getOperand(0);
5967 SDValue Size = Op.getOperand(1);
5968 // FIXME: Ensure alignment here
5972 EVT IntPtr = getPointerTy();
5973 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5975 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5977 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5978 Flag = Chain.getValue(1);
5980 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5981 SDValue Ops[] = { Chain,
5982 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5983 DAG.getRegister(X86::EAX, IntPtr),
5984 DAG.getRegister(X86StackPtr, SPTy),
5986 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5987 Flag = Chain.getValue(1);
5989 Chain = DAG.getCALLSEQ_END(Chain,
5990 DAG.getIntPtrConstant(0, true),
5991 DAG.getIntPtrConstant(0, true),
5994 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5996 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5997 return DAG.getMergeValues(Ops1, 2, dl);
6001 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
6003 SDValue Dst, SDValue Src,
6004 SDValue Size, unsigned Align,
6006 uint64_t DstSVOff) {
6007 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6009 // If not DWORD aligned or size is more than the threshold, call the library.
6010 // The libc version is likely to be faster for these cases. It can use the
6011 // address value and run time information about the CPU.
6012 if ((Align & 3) != 0 ||
6014 ConstantSize->getZExtValue() >
6015 getSubtarget()->getMaxInlineSizeThreshold()) {
6016 SDValue InFlag(0, 0);
6018 // Check to see if there is a specialized entry-point for memory zeroing.
6019 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
6021 if (const char *bzeroEntry = V &&
6022 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
6023 EVT IntPtr = getPointerTy();
6024 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
6025 TargetLowering::ArgListTy Args;
6026 TargetLowering::ArgListEntry Entry;
6028 Entry.Ty = IntPtrTy;
6029 Args.push_back(Entry);
6031 Args.push_back(Entry);
6032 std::pair<SDValue,SDValue> CallResult =
6033 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
6034 false, false, false, false,
6035 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
6036 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
6037 return CallResult.second;
6040 // Otherwise have the target-independent code call memset.
6044 uint64_t SizeVal = ConstantSize->getZExtValue();
6045 SDValue InFlag(0, 0);
6048 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
6049 unsigned BytesLeft = 0;
6050 bool TwoRepStos = false;
6053 uint64_t Val = ValC->getZExtValue() & 255;
6055 // If the value is a constant, then we can potentially use larger sets.
6056 switch (Align & 3) {
6057 case 2: // WORD aligned
6060 Val = (Val << 8) | Val;
6062 case 0: // DWORD aligned
6065 Val = (Val << 8) | Val;
6066 Val = (Val << 16) | Val;
6067 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6070 Val = (Val << 32) | Val;
6073 default: // Byte aligned
6076 Count = DAG.getIntPtrConstant(SizeVal);
6080 if (AVT.bitsGT(MVT::i8)) {
6081 unsigned UBytes = AVT.getSizeInBits() / 8;
6082 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6083 BytesLeft = SizeVal % UBytes;
6086 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6088 InFlag = Chain.getValue(1);
6091 Count = DAG.getIntPtrConstant(SizeVal);
6092 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6093 InFlag = Chain.getValue(1);
6096 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6099 InFlag = Chain.getValue(1);
6100 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6103 InFlag = Chain.getValue(1);
6105 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6106 SmallVector<SDValue, 8> Ops;
6107 Ops.push_back(Chain);
6108 Ops.push_back(DAG.getValueType(AVT));
6109 Ops.push_back(InFlag);
6110 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6113 InFlag = Chain.getValue(1);
6115 EVT CVT = Count.getValueType();
6116 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6117 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6118 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6121 InFlag = Chain.getValue(1);
6122 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6124 Ops.push_back(Chain);
6125 Ops.push_back(DAG.getValueType(MVT::i8));
6126 Ops.push_back(InFlag);
6127 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6128 } else if (BytesLeft) {
6129 // Handle the last 1 - 7 bytes.
6130 unsigned Offset = SizeVal - BytesLeft;
6131 EVT AddrVT = Dst.getValueType();
6132 EVT SizeVT = Size.getValueType();
6134 Chain = DAG.getMemset(Chain, dl,
6135 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6136 DAG.getConstant(Offset, AddrVT)),
6138 DAG.getConstant(BytesLeft, SizeVT),
6139 Align, DstSV, DstSVOff + Offset);
6142 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6147 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6148 SDValue Chain, SDValue Dst, SDValue Src,
6149 SDValue Size, unsigned Align,
6151 const Value *DstSV, uint64_t DstSVOff,
6152 const Value *SrcSV, uint64_t SrcSVOff) {
6153 // This requires the copy size to be a constant, preferrably
6154 // within a subtarget-specific limit.
6155 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6158 uint64_t SizeVal = ConstantSize->getZExtValue();
6159 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6162 /// If not DWORD aligned, call the library.
6163 if ((Align & 3) != 0)
6168 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6171 unsigned UBytes = AVT.getSizeInBits() / 8;
6172 unsigned CountVal = SizeVal / UBytes;
6173 SDValue Count = DAG.getIntPtrConstant(CountVal);
6174 unsigned BytesLeft = SizeVal % UBytes;
6176 SDValue InFlag(0, 0);
6177 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6180 InFlag = Chain.getValue(1);
6181 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6184 InFlag = Chain.getValue(1);
6185 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6188 InFlag = Chain.getValue(1);
6190 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6191 SmallVector<SDValue, 8> Ops;
6192 Ops.push_back(Chain);
6193 Ops.push_back(DAG.getValueType(AVT));
6194 Ops.push_back(InFlag);
6195 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
6197 SmallVector<SDValue, 4> Results;
6198 Results.push_back(RepMovs);
6200 // Handle the last 1 - 7 bytes.
6201 unsigned Offset = SizeVal - BytesLeft;
6202 EVT DstVT = Dst.getValueType();
6203 EVT SrcVT = Src.getValueType();
6204 EVT SizeVT = Size.getValueType();
6205 Results.push_back(DAG.getMemcpy(Chain, dl,
6206 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6207 DAG.getConstant(Offset, DstVT)),
6208 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6209 DAG.getConstant(Offset, SrcVT)),
6210 DAG.getConstant(BytesLeft, SizeVT),
6211 Align, AlwaysInline,
6212 DstSV, DstSVOff + Offset,
6213 SrcSV, SrcSVOff + Offset));
6216 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6217 &Results[0], Results.size());
6220 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6221 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6222 DebugLoc dl = Op.getDebugLoc();
6224 if (!Subtarget->is64Bit()) {
6225 // vastart just stores the address of the VarArgsFrameIndex slot into the
6226 // memory location argument.
6227 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6228 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6232 // gp_offset (0 - 6 * 8)
6233 // fp_offset (48 - 48 + 8 * 16)
6234 // overflow_arg_area (point to parameters coming in memory).
6236 SmallVector<SDValue, 8> MemOps;
6237 SDValue FIN = Op.getOperand(1);
6239 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6240 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6242 MemOps.push_back(Store);
6245 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6246 FIN, DAG.getIntPtrConstant(4));
6247 Store = DAG.getStore(Op.getOperand(0), dl,
6248 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6250 MemOps.push_back(Store);
6252 // Store ptr to overflow_arg_area
6253 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6254 FIN, DAG.getIntPtrConstant(4));
6255 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6256 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6257 MemOps.push_back(Store);
6259 // Store ptr to reg_save_area.
6260 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6261 FIN, DAG.getIntPtrConstant(8));
6262 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6263 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6264 MemOps.push_back(Store);
6265 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6266 &MemOps[0], MemOps.size());
6269 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6270 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6271 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6272 SDValue Chain = Op.getOperand(0);
6273 SDValue SrcPtr = Op.getOperand(1);
6274 SDValue SrcSV = Op.getOperand(2);
6276 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6280 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6281 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6282 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6283 SDValue Chain = Op.getOperand(0);
6284 SDValue DstPtr = Op.getOperand(1);
6285 SDValue SrcPtr = Op.getOperand(2);
6286 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6287 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6288 DebugLoc dl = Op.getDebugLoc();
6290 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6291 DAG.getIntPtrConstant(24), 8, false,
6292 DstSV, 0, SrcSV, 0);
6296 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6297 DebugLoc dl = Op.getDebugLoc();
6298 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6300 default: return SDValue(); // Don't custom lower most intrinsics.
6301 // Comparison intrinsics.
6302 case Intrinsic::x86_sse_comieq_ss:
6303 case Intrinsic::x86_sse_comilt_ss:
6304 case Intrinsic::x86_sse_comile_ss:
6305 case Intrinsic::x86_sse_comigt_ss:
6306 case Intrinsic::x86_sse_comige_ss:
6307 case Intrinsic::x86_sse_comineq_ss:
6308 case Intrinsic::x86_sse_ucomieq_ss:
6309 case Intrinsic::x86_sse_ucomilt_ss:
6310 case Intrinsic::x86_sse_ucomile_ss:
6311 case Intrinsic::x86_sse_ucomigt_ss:
6312 case Intrinsic::x86_sse_ucomige_ss:
6313 case Intrinsic::x86_sse_ucomineq_ss:
6314 case Intrinsic::x86_sse2_comieq_sd:
6315 case Intrinsic::x86_sse2_comilt_sd:
6316 case Intrinsic::x86_sse2_comile_sd:
6317 case Intrinsic::x86_sse2_comigt_sd:
6318 case Intrinsic::x86_sse2_comige_sd:
6319 case Intrinsic::x86_sse2_comineq_sd:
6320 case Intrinsic::x86_sse2_ucomieq_sd:
6321 case Intrinsic::x86_sse2_ucomilt_sd:
6322 case Intrinsic::x86_sse2_ucomile_sd:
6323 case Intrinsic::x86_sse2_ucomigt_sd:
6324 case Intrinsic::x86_sse2_ucomige_sd:
6325 case Intrinsic::x86_sse2_ucomineq_sd: {
6327 ISD::CondCode CC = ISD::SETCC_INVALID;
6330 case Intrinsic::x86_sse_comieq_ss:
6331 case Intrinsic::x86_sse2_comieq_sd:
6335 case Intrinsic::x86_sse_comilt_ss:
6336 case Intrinsic::x86_sse2_comilt_sd:
6340 case Intrinsic::x86_sse_comile_ss:
6341 case Intrinsic::x86_sse2_comile_sd:
6345 case Intrinsic::x86_sse_comigt_ss:
6346 case Intrinsic::x86_sse2_comigt_sd:
6350 case Intrinsic::x86_sse_comige_ss:
6351 case Intrinsic::x86_sse2_comige_sd:
6355 case Intrinsic::x86_sse_comineq_ss:
6356 case Intrinsic::x86_sse2_comineq_sd:
6360 case Intrinsic::x86_sse_ucomieq_ss:
6361 case Intrinsic::x86_sse2_ucomieq_sd:
6362 Opc = X86ISD::UCOMI;
6365 case Intrinsic::x86_sse_ucomilt_ss:
6366 case Intrinsic::x86_sse2_ucomilt_sd:
6367 Opc = X86ISD::UCOMI;
6370 case Intrinsic::x86_sse_ucomile_ss:
6371 case Intrinsic::x86_sse2_ucomile_sd:
6372 Opc = X86ISD::UCOMI;
6375 case Intrinsic::x86_sse_ucomigt_ss:
6376 case Intrinsic::x86_sse2_ucomigt_sd:
6377 Opc = X86ISD::UCOMI;
6380 case Intrinsic::x86_sse_ucomige_ss:
6381 case Intrinsic::x86_sse2_ucomige_sd:
6382 Opc = X86ISD::UCOMI;
6385 case Intrinsic::x86_sse_ucomineq_ss:
6386 case Intrinsic::x86_sse2_ucomineq_sd:
6387 Opc = X86ISD::UCOMI;
6392 SDValue LHS = Op.getOperand(1);
6393 SDValue RHS = Op.getOperand(2);
6394 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6395 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
6396 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6397 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6398 DAG.getConstant(X86CC, MVT::i8), Cond);
6399 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6401 // ptest intrinsics. The intrinsic these come from are designed to return
6402 // an integer value, not just an instruction so lower it to the ptest
6403 // pattern and a setcc for the result.
6404 case Intrinsic::x86_sse41_ptestz:
6405 case Intrinsic::x86_sse41_ptestc:
6406 case Intrinsic::x86_sse41_ptestnzc:{
6409 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6410 case Intrinsic::x86_sse41_ptestz:
6412 X86CC = X86::COND_E;
6414 case Intrinsic::x86_sse41_ptestc:
6416 X86CC = X86::COND_B;
6418 case Intrinsic::x86_sse41_ptestnzc:
6420 X86CC = X86::COND_A;
6424 SDValue LHS = Op.getOperand(1);
6425 SDValue RHS = Op.getOperand(2);
6426 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6427 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6428 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6429 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6432 // Fix vector shift instructions where the last operand is a non-immediate
6434 case Intrinsic::x86_sse2_pslli_w:
6435 case Intrinsic::x86_sse2_pslli_d:
6436 case Intrinsic::x86_sse2_pslli_q:
6437 case Intrinsic::x86_sse2_psrli_w:
6438 case Intrinsic::x86_sse2_psrli_d:
6439 case Intrinsic::x86_sse2_psrli_q:
6440 case Intrinsic::x86_sse2_psrai_w:
6441 case Intrinsic::x86_sse2_psrai_d:
6442 case Intrinsic::x86_mmx_pslli_w:
6443 case Intrinsic::x86_mmx_pslli_d:
6444 case Intrinsic::x86_mmx_pslli_q:
6445 case Intrinsic::x86_mmx_psrli_w:
6446 case Intrinsic::x86_mmx_psrli_d:
6447 case Intrinsic::x86_mmx_psrli_q:
6448 case Intrinsic::x86_mmx_psrai_w:
6449 case Intrinsic::x86_mmx_psrai_d: {
6450 SDValue ShAmt = Op.getOperand(2);
6451 if (isa<ConstantSDNode>(ShAmt))
6454 unsigned NewIntNo = 0;
6455 EVT ShAmtVT = MVT::v4i32;
6457 case Intrinsic::x86_sse2_pslli_w:
6458 NewIntNo = Intrinsic::x86_sse2_psll_w;
6460 case Intrinsic::x86_sse2_pslli_d:
6461 NewIntNo = Intrinsic::x86_sse2_psll_d;
6463 case Intrinsic::x86_sse2_pslli_q:
6464 NewIntNo = Intrinsic::x86_sse2_psll_q;
6466 case Intrinsic::x86_sse2_psrli_w:
6467 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6469 case Intrinsic::x86_sse2_psrli_d:
6470 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6472 case Intrinsic::x86_sse2_psrli_q:
6473 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6475 case Intrinsic::x86_sse2_psrai_w:
6476 NewIntNo = Intrinsic::x86_sse2_psra_w;
6478 case Intrinsic::x86_sse2_psrai_d:
6479 NewIntNo = Intrinsic::x86_sse2_psra_d;
6482 ShAmtVT = MVT::v2i32;
6484 case Intrinsic::x86_mmx_pslli_w:
6485 NewIntNo = Intrinsic::x86_mmx_psll_w;
6487 case Intrinsic::x86_mmx_pslli_d:
6488 NewIntNo = Intrinsic::x86_mmx_psll_d;
6490 case Intrinsic::x86_mmx_pslli_q:
6491 NewIntNo = Intrinsic::x86_mmx_psll_q;
6493 case Intrinsic::x86_mmx_psrli_w:
6494 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6496 case Intrinsic::x86_mmx_psrli_d:
6497 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6499 case Intrinsic::x86_mmx_psrli_q:
6500 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6502 case Intrinsic::x86_mmx_psrai_w:
6503 NewIntNo = Intrinsic::x86_mmx_psra_w;
6505 case Intrinsic::x86_mmx_psrai_d:
6506 NewIntNo = Intrinsic::x86_mmx_psra_d;
6508 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6514 // The vector shift intrinsics with scalars uses 32b shift amounts but
6515 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6519 ShOps[1] = DAG.getConstant(0, MVT::i32);
6520 if (ShAmtVT == MVT::v4i32) {
6521 ShOps[2] = DAG.getUNDEF(MVT::i32);
6522 ShOps[3] = DAG.getUNDEF(MVT::i32);
6523 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6525 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6528 EVT VT = Op.getValueType();
6529 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6530 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6531 DAG.getConstant(NewIntNo, MVT::i32),
6532 Op.getOperand(1), ShAmt);
6537 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6538 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6539 DebugLoc dl = Op.getDebugLoc();
6542 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6544 DAG.getConstant(TD->getPointerSize(),
6545 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6546 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6547 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6552 // Just load the return address.
6553 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6554 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6555 RetAddrFI, NULL, 0);
6558 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6559 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6560 MFI->setFrameAddressIsTaken(true);
6561 EVT VT = Op.getValueType();
6562 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6563 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6564 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6565 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6567 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6571 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6572 SelectionDAG &DAG) {
6573 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6576 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6578 MachineFunction &MF = DAG.getMachineFunction();
6579 SDValue Chain = Op.getOperand(0);
6580 SDValue Offset = Op.getOperand(1);
6581 SDValue Handler = Op.getOperand(2);
6582 DebugLoc dl = Op.getDebugLoc();
6584 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6586 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6588 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6589 DAG.getIntPtrConstant(-TD->getPointerSize()));
6590 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6591 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6592 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6593 MF.getRegInfo().addLiveOut(StoreAddrReg);
6595 return DAG.getNode(X86ISD::EH_RETURN, dl,
6597 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6600 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6601 SelectionDAG &DAG) {
6602 SDValue Root = Op.getOperand(0);
6603 SDValue Trmp = Op.getOperand(1); // trampoline
6604 SDValue FPtr = Op.getOperand(2); // nested function
6605 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6606 DebugLoc dl = Op.getDebugLoc();
6608 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6610 const X86InstrInfo *TII =
6611 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6613 if (Subtarget->is64Bit()) {
6614 SDValue OutChains[6];
6616 // Large code-model.
6618 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6619 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6621 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6622 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6624 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6626 // Load the pointer to the nested function into R11.
6627 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6628 SDValue Addr = Trmp;
6629 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6632 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6633 DAG.getConstant(2, MVT::i64));
6634 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6636 // Load the 'nest' parameter value into R10.
6637 // R10 is specified in X86CallingConv.td
6638 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6639 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6640 DAG.getConstant(10, MVT::i64));
6641 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6642 Addr, TrmpAddr, 10);
6644 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6645 DAG.getConstant(12, MVT::i64));
6646 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6648 // Jump to the nested function.
6649 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6650 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6651 DAG.getConstant(20, MVT::i64));
6652 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6653 Addr, TrmpAddr, 20);
6655 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6656 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6657 DAG.getConstant(22, MVT::i64));
6658 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6662 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6663 return DAG.getMergeValues(Ops, 2, dl);
6665 const Function *Func =
6666 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6667 CallingConv::ID CC = Func->getCallingConv();
6672 llvm_unreachable("Unsupported calling convention");
6673 case CallingConv::C:
6674 case CallingConv::X86_StdCall: {
6675 // Pass 'nest' parameter in ECX.
6676 // Must be kept in sync with X86CallingConv.td
6679 // Check that ECX wasn't needed by an 'inreg' parameter.
6680 const FunctionType *FTy = Func->getFunctionType();
6681 const AttrListPtr &Attrs = Func->getAttributes();
6683 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6684 unsigned InRegCount = 0;
6687 for (FunctionType::param_iterator I = FTy->param_begin(),
6688 E = FTy->param_end(); I != E; ++I, ++Idx)
6689 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6690 // FIXME: should only count parameters that are lowered to integers.
6691 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6693 if (InRegCount > 2) {
6694 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6699 case CallingConv::X86_FastCall:
6700 case CallingConv::Fast:
6701 // Pass 'nest' parameter in EAX.
6702 // Must be kept in sync with X86CallingConv.td
6707 SDValue OutChains[4];
6710 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6711 DAG.getConstant(10, MVT::i32));
6712 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6714 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6715 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6716 OutChains[0] = DAG.getStore(Root, dl,
6717 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6720 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6721 DAG.getConstant(1, MVT::i32));
6722 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6724 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6725 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6726 DAG.getConstant(5, MVT::i32));
6727 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6728 TrmpAddr, 5, false, 1);
6730 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6731 DAG.getConstant(6, MVT::i32));
6732 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6735 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6736 return DAG.getMergeValues(Ops, 2, dl);
6740 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6742 The rounding mode is in bits 11:10 of FPSR, and has the following
6749 FLT_ROUNDS, on the other hand, expects the following:
6756 To perform the conversion, we do:
6757 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6760 MachineFunction &MF = DAG.getMachineFunction();
6761 const TargetMachine &TM = MF.getTarget();
6762 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6763 unsigned StackAlignment = TFI.getStackAlignment();
6764 EVT VT = Op.getValueType();
6765 DebugLoc dl = Op.getDebugLoc();
6767 // Save FP Control Word to stack slot
6768 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
6769 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6771 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6772 DAG.getEntryNode(), StackSlot);
6774 // Load FP Control Word from stack slot
6775 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6777 // Transform as necessary
6779 DAG.getNode(ISD::SRL, dl, MVT::i16,
6780 DAG.getNode(ISD::AND, dl, MVT::i16,
6781 CWD, DAG.getConstant(0x800, MVT::i16)),
6782 DAG.getConstant(11, MVT::i8));
6784 DAG.getNode(ISD::SRL, dl, MVT::i16,
6785 DAG.getNode(ISD::AND, dl, MVT::i16,
6786 CWD, DAG.getConstant(0x400, MVT::i16)),
6787 DAG.getConstant(9, MVT::i8));
6790 DAG.getNode(ISD::AND, dl, MVT::i16,
6791 DAG.getNode(ISD::ADD, dl, MVT::i16,
6792 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6793 DAG.getConstant(1, MVT::i16)),
6794 DAG.getConstant(3, MVT::i16));
6797 return DAG.getNode((VT.getSizeInBits() < 16 ?
6798 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6801 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6802 EVT VT = Op.getValueType();
6804 unsigned NumBits = VT.getSizeInBits();
6805 DebugLoc dl = Op.getDebugLoc();
6807 Op = Op.getOperand(0);
6808 if (VT == MVT::i8) {
6809 // Zero extend to i32 since there is not an i8 bsr.
6811 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6814 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6815 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6816 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6818 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6819 SmallVector<SDValue, 4> Ops;
6821 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6822 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6823 Ops.push_back(Op.getValue(1));
6824 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6826 // Finally xor with NumBits-1.
6827 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6830 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6834 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6835 EVT VT = Op.getValueType();
6837 unsigned NumBits = VT.getSizeInBits();
6838 DebugLoc dl = Op.getDebugLoc();
6840 Op = Op.getOperand(0);
6841 if (VT == MVT::i8) {
6843 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6846 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6847 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6848 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6850 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6851 SmallVector<SDValue, 4> Ops;
6853 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6854 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6855 Ops.push_back(Op.getValue(1));
6856 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6859 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6863 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6864 EVT VT = Op.getValueType();
6865 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6866 DebugLoc dl = Op.getDebugLoc();
6868 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6869 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6870 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6871 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6872 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6874 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6875 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6876 // return AloBlo + AloBhi + AhiBlo;
6878 SDValue A = Op.getOperand(0);
6879 SDValue B = Op.getOperand(1);
6881 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6882 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6883 A, DAG.getConstant(32, MVT::i32));
6884 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6885 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6886 B, DAG.getConstant(32, MVT::i32));
6887 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6888 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6890 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6891 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6893 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6894 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6896 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6897 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6898 AloBhi, DAG.getConstant(32, MVT::i32));
6899 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6900 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6901 AhiBlo, DAG.getConstant(32, MVT::i32));
6902 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6903 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6908 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6909 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6910 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6911 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6912 // has only one use.
6913 SDNode *N = Op.getNode();
6914 SDValue LHS = N->getOperand(0);
6915 SDValue RHS = N->getOperand(1);
6916 unsigned BaseOp = 0;
6918 DebugLoc dl = Op.getDebugLoc();
6920 switch (Op.getOpcode()) {
6921 default: llvm_unreachable("Unknown ovf instruction!");
6923 // A subtract of one will be selected as a INC. Note that INC doesn't
6924 // set CF, so we can't do this for UADDO.
6925 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6926 if (C->getAPIntValue() == 1) {
6927 BaseOp = X86ISD::INC;
6931 BaseOp = X86ISD::ADD;
6935 BaseOp = X86ISD::ADD;
6939 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6940 // set CF, so we can't do this for USUBO.
6941 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6942 if (C->getAPIntValue() == 1) {
6943 BaseOp = X86ISD::DEC;
6947 BaseOp = X86ISD::SUB;
6951 BaseOp = X86ISD::SUB;
6955 BaseOp = X86ISD::SMUL;
6959 BaseOp = X86ISD::UMUL;
6964 // Also sets EFLAGS.
6965 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6966 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6969 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6970 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6972 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6976 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6977 EVT T = Op.getValueType();
6978 DebugLoc dl = Op.getDebugLoc();
6981 switch(T.getSimpleVT().SimpleTy) {
6983 assert(false && "Invalid value type!");
6984 case MVT::i8: Reg = X86::AL; size = 1; break;
6985 case MVT::i16: Reg = X86::AX; size = 2; break;
6986 case MVT::i32: Reg = X86::EAX; size = 4; break;
6988 assert(Subtarget->is64Bit() && "Node not type legal!");
6989 Reg = X86::RAX; size = 8;
6992 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6993 Op.getOperand(2), SDValue());
6994 SDValue Ops[] = { cpIn.getValue(0),
6997 DAG.getTargetConstant(size, MVT::i8),
6999 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7000 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
7002 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
7006 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
7007 SelectionDAG &DAG) {
7008 assert(Subtarget->is64Bit() && "Result not type legalized?");
7009 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7010 SDValue TheChain = Op.getOperand(0);
7011 DebugLoc dl = Op.getDebugLoc();
7012 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7013 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
7014 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
7016 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
7017 DAG.getConstant(32, MVT::i8));
7019 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
7022 return DAG.getMergeValues(Ops, 2, dl);
7025 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
7026 SDNode *Node = Op.getNode();
7027 DebugLoc dl = Node->getDebugLoc();
7028 EVT T = Node->getValueType(0);
7029 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
7030 DAG.getConstant(0, T), Node->getOperand(2));
7031 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
7032 cast<AtomicSDNode>(Node)->getMemoryVT(),
7033 Node->getOperand(0),
7034 Node->getOperand(1), negOp,
7035 cast<AtomicSDNode>(Node)->getSrcValue(),
7036 cast<AtomicSDNode>(Node)->getAlignment());
7039 /// LowerOperation - Provide custom lowering hooks for some operations.
7041 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
7042 switch (Op.getOpcode()) {
7043 default: llvm_unreachable("Should not custom lower this!");
7044 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
7045 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
7046 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
7047 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
7048 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
7049 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
7050 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
7051 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
7052 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
7053 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
7054 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7055 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
7056 case ISD::SHL_PARTS:
7057 case ISD::SRA_PARTS:
7058 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7059 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7060 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7061 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7062 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7063 case ISD::FABS: return LowerFABS(Op, DAG);
7064 case ISD::FNEG: return LowerFNEG(Op, DAG);
7065 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7066 case ISD::SETCC: return LowerSETCC(Op, DAG);
7067 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7068 case ISD::SELECT: return LowerSELECT(Op, DAG);
7069 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7070 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7071 case ISD::VASTART: return LowerVASTART(Op, DAG);
7072 case ISD::VAARG: return LowerVAARG(Op, DAG);
7073 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7074 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7075 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7076 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7077 case ISD::FRAME_TO_ARGS_OFFSET:
7078 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7079 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7080 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7081 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7082 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7083 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7084 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7085 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7091 case ISD::UMULO: return LowerXALUO(Op, DAG);
7092 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7096 void X86TargetLowering::
7097 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7098 SelectionDAG &DAG, unsigned NewOp) {
7099 EVT T = Node->getValueType(0);
7100 DebugLoc dl = Node->getDebugLoc();
7101 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7103 SDValue Chain = Node->getOperand(0);
7104 SDValue In1 = Node->getOperand(1);
7105 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7106 Node->getOperand(2), DAG.getIntPtrConstant(0));
7107 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7108 Node->getOperand(2), DAG.getIntPtrConstant(1));
7109 SDValue Ops[] = { Chain, In1, In2L, In2H };
7110 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7112 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7113 cast<MemSDNode>(Node)->getMemOperand());
7114 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7115 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7116 Results.push_back(Result.getValue(2));
7119 /// ReplaceNodeResults - Replace a node with an illegal result type
7120 /// with a new node built out of custom code.
7121 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7122 SmallVectorImpl<SDValue>&Results,
7123 SelectionDAG &DAG) {
7124 DebugLoc dl = N->getDebugLoc();
7125 switch (N->getOpcode()) {
7127 assert(false && "Do not know how to custom type legalize this operation!");
7129 case ISD::FP_TO_SINT: {
7130 std::pair<SDValue,SDValue> Vals =
7131 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7132 SDValue FIST = Vals.first, StackSlot = Vals.second;
7133 if (FIST.getNode() != 0) {
7134 EVT VT = N->getValueType(0);
7135 // Return a load from the stack slot.
7136 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7140 case ISD::READCYCLECOUNTER: {
7141 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7142 SDValue TheChain = N->getOperand(0);
7143 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7144 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7146 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7148 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7149 SDValue Ops[] = { eax, edx };
7150 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7151 Results.push_back(edx.getValue(1));
7154 case ISD::ATOMIC_CMP_SWAP: {
7155 EVT T = N->getValueType(0);
7156 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7157 SDValue cpInL, cpInH;
7158 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7159 DAG.getConstant(0, MVT::i32));
7160 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7161 DAG.getConstant(1, MVT::i32));
7162 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7163 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7165 SDValue swapInL, swapInH;
7166 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7167 DAG.getConstant(0, MVT::i32));
7168 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7169 DAG.getConstant(1, MVT::i32));
7170 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7172 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7173 swapInL.getValue(1));
7174 SDValue Ops[] = { swapInH.getValue(0),
7176 swapInH.getValue(1) };
7177 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7178 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7179 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7180 MVT::i32, Result.getValue(1));
7181 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7182 MVT::i32, cpOutL.getValue(2));
7183 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7184 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7185 Results.push_back(cpOutH.getValue(1));
7188 case ISD::ATOMIC_LOAD_ADD:
7189 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7191 case ISD::ATOMIC_LOAD_AND:
7192 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7194 case ISD::ATOMIC_LOAD_NAND:
7195 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7197 case ISD::ATOMIC_LOAD_OR:
7198 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7200 case ISD::ATOMIC_LOAD_SUB:
7201 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7203 case ISD::ATOMIC_LOAD_XOR:
7204 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7206 case ISD::ATOMIC_SWAP:
7207 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7212 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7214 default: return NULL;
7215 case X86ISD::BSF: return "X86ISD::BSF";
7216 case X86ISD::BSR: return "X86ISD::BSR";
7217 case X86ISD::SHLD: return "X86ISD::SHLD";
7218 case X86ISD::SHRD: return "X86ISD::SHRD";
7219 case X86ISD::FAND: return "X86ISD::FAND";
7220 case X86ISD::FOR: return "X86ISD::FOR";
7221 case X86ISD::FXOR: return "X86ISD::FXOR";
7222 case X86ISD::FSRL: return "X86ISD::FSRL";
7223 case X86ISD::FILD: return "X86ISD::FILD";
7224 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7225 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7226 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7227 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7228 case X86ISD::FLD: return "X86ISD::FLD";
7229 case X86ISD::FST: return "X86ISD::FST";
7230 case X86ISD::CALL: return "X86ISD::CALL";
7231 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7232 case X86ISD::BT: return "X86ISD::BT";
7233 case X86ISD::CMP: return "X86ISD::CMP";
7234 case X86ISD::COMI: return "X86ISD::COMI";
7235 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7236 case X86ISD::SETCC: return "X86ISD::SETCC";
7237 case X86ISD::CMOV: return "X86ISD::CMOV";
7238 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7239 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7240 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7241 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7242 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7243 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7244 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7245 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7246 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7247 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7248 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7249 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7250 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7251 case X86ISD::FMAX: return "X86ISD::FMAX";
7252 case X86ISD::FMIN: return "X86ISD::FMIN";
7253 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7254 case X86ISD::FRCP: return "X86ISD::FRCP";
7255 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7256 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7257 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7258 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7259 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7260 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7261 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7262 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7263 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7264 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7265 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7266 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7267 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7268 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7269 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7270 case X86ISD::VSHL: return "X86ISD::VSHL";
7271 case X86ISD::VSRL: return "X86ISD::VSRL";
7272 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7273 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7274 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7275 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7276 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7277 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7278 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7279 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7280 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7281 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7282 case X86ISD::ADD: return "X86ISD::ADD";
7283 case X86ISD::SUB: return "X86ISD::SUB";
7284 case X86ISD::SMUL: return "X86ISD::SMUL";
7285 case X86ISD::UMUL: return "X86ISD::UMUL";
7286 case X86ISD::INC: return "X86ISD::INC";
7287 case X86ISD::DEC: return "X86ISD::DEC";
7288 case X86ISD::OR: return "X86ISD::OR";
7289 case X86ISD::XOR: return "X86ISD::XOR";
7290 case X86ISD::AND: return "X86ISD::AND";
7291 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7292 case X86ISD::PTEST: return "X86ISD::PTEST";
7293 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7297 // isLegalAddressingMode - Return true if the addressing mode represented
7298 // by AM is legal for this target, for a load/store of the specified type.
7299 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7300 const Type *Ty) const {
7301 // X86 supports extremely general addressing modes.
7302 CodeModel::Model M = getTargetMachine().getCodeModel();
7304 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7305 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7310 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7312 // If a reference to this global requires an extra load, we can't fold it.
7313 if (isGlobalStubReference(GVFlags))
7316 // If BaseGV requires a register for the PIC base, we cannot also have a
7317 // BaseReg specified.
7318 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7321 // If lower 4G is not available, then we must use rip-relative addressing.
7322 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7332 // These scales always work.
7337 // These scales are formed with basereg+scalereg. Only accept if there is
7342 default: // Other stuff never works.
7350 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7351 if (!Ty1->isInteger() || !Ty2->isInteger())
7353 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7354 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7355 if (NumBits1 <= NumBits2)
7357 return Subtarget->is64Bit() || NumBits1 < 64;
7360 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7361 if (!VT1.isInteger() || !VT2.isInteger())
7363 unsigned NumBits1 = VT1.getSizeInBits();
7364 unsigned NumBits2 = VT2.getSizeInBits();
7365 if (NumBits1 <= NumBits2)
7367 return Subtarget->is64Bit() || NumBits1 < 64;
7370 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7371 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7372 return Ty1 == Type::getInt32Ty(Ty1->getContext()) &&
7373 Ty2 == Type::getInt64Ty(Ty1->getContext()) && Subtarget->is64Bit();
7376 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7377 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7378 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7381 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7382 // i16 instructions are longer (0x66 prefix) and potentially slower.
7383 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7386 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7387 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7388 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7389 /// are assumed to be legal.
7391 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7393 // Only do shuffles on 128-bit vector types for now.
7394 if (VT.getSizeInBits() == 64)
7397 // FIXME: pshufb, blends, shifts.
7398 return (VT.getVectorNumElements() == 2 ||
7399 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7400 isMOVLMask(M, VT) ||
7401 isSHUFPMask(M, VT) ||
7402 isPSHUFDMask(M, VT) ||
7403 isPSHUFHWMask(M, VT) ||
7404 isPSHUFLWMask(M, VT) ||
7405 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7406 isUNPCKLMask(M, VT) ||
7407 isUNPCKHMask(M, VT) ||
7408 isUNPCKL_v_undef_Mask(M, VT) ||
7409 isUNPCKH_v_undef_Mask(M, VT));
7413 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7415 unsigned NumElts = VT.getVectorNumElements();
7416 // FIXME: This collection of masks seems suspect.
7419 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7420 return (isMOVLMask(Mask, VT) ||
7421 isCommutedMOVLMask(Mask, VT, true) ||
7422 isSHUFPMask(Mask, VT) ||
7423 isCommutedSHUFPMask(Mask, VT));
7428 //===----------------------------------------------------------------------===//
7429 // X86 Scheduler Hooks
7430 //===----------------------------------------------------------------------===//
7432 // private utility function
7434 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7435 MachineBasicBlock *MBB,
7443 TargetRegisterClass *RC,
7444 bool invSrc) const {
7445 // For the atomic bitwise operator, we generate
7448 // ld t1 = [bitinstr.addr]
7449 // op t2 = t1, [bitinstr.val]
7451 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7453 // fallthrough -->nextMBB
7454 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7455 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7456 MachineFunction::iterator MBBIter = MBB;
7459 /// First build the CFG
7460 MachineFunction *F = MBB->getParent();
7461 MachineBasicBlock *thisMBB = MBB;
7462 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7463 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7464 F->insert(MBBIter, newMBB);
7465 F->insert(MBBIter, nextMBB);
7467 // Move all successors to thisMBB to nextMBB
7468 nextMBB->transferSuccessors(thisMBB);
7470 // Update thisMBB to fall through to newMBB
7471 thisMBB->addSuccessor(newMBB);
7473 // newMBB jumps to itself and fall through to nextMBB
7474 newMBB->addSuccessor(nextMBB);
7475 newMBB->addSuccessor(newMBB);
7477 // Insert instructions into newMBB based on incoming instruction
7478 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7479 "unexpected number of operands");
7480 DebugLoc dl = bInstr->getDebugLoc();
7481 MachineOperand& destOper = bInstr->getOperand(0);
7482 MachineOperand* argOpers[2 + X86AddrNumOperands];
7483 int numArgs = bInstr->getNumOperands() - 1;
7484 for (int i=0; i < numArgs; ++i)
7485 argOpers[i] = &bInstr->getOperand(i+1);
7487 // x86 address has 4 operands: base, index, scale, and displacement
7488 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7489 int valArgIndx = lastAddrIndx + 1;
7491 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7492 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7493 for (int i=0; i <= lastAddrIndx; ++i)
7494 (*MIB).addOperand(*argOpers[i]);
7496 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7498 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7503 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7504 assert((argOpers[valArgIndx]->isReg() ||
7505 argOpers[valArgIndx]->isImm()) &&
7507 if (argOpers[valArgIndx]->isReg())
7508 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7510 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7512 (*MIB).addOperand(*argOpers[valArgIndx]);
7514 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7517 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7518 for (int i=0; i <= lastAddrIndx; ++i)
7519 (*MIB).addOperand(*argOpers[i]);
7521 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7522 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7523 bInstr->memoperands_end());
7525 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7529 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7531 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7535 // private utility function: 64 bit atomics on 32 bit host.
7537 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7538 MachineBasicBlock *MBB,
7543 bool invSrc) const {
7544 // For the atomic bitwise operator, we generate
7545 // thisMBB (instructions are in pairs, except cmpxchg8b)
7546 // ld t1,t2 = [bitinstr.addr]
7548 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7549 // op t5, t6 <- out1, out2, [bitinstr.val]
7550 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7551 // mov ECX, EBX <- t5, t6
7552 // mov EAX, EDX <- t1, t2
7553 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7554 // mov t3, t4 <- EAX, EDX
7556 // result in out1, out2
7557 // fallthrough -->nextMBB
7559 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7560 const unsigned LoadOpc = X86::MOV32rm;
7561 const unsigned copyOpc = X86::MOV32rr;
7562 const unsigned NotOpc = X86::NOT32r;
7563 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7564 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7565 MachineFunction::iterator MBBIter = MBB;
7568 /// First build the CFG
7569 MachineFunction *F = MBB->getParent();
7570 MachineBasicBlock *thisMBB = MBB;
7571 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7572 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7573 F->insert(MBBIter, newMBB);
7574 F->insert(MBBIter, nextMBB);
7576 // Move all successors to thisMBB to nextMBB
7577 nextMBB->transferSuccessors(thisMBB);
7579 // Update thisMBB to fall through to newMBB
7580 thisMBB->addSuccessor(newMBB);
7582 // newMBB jumps to itself and fall through to nextMBB
7583 newMBB->addSuccessor(nextMBB);
7584 newMBB->addSuccessor(newMBB);
7586 DebugLoc dl = bInstr->getDebugLoc();
7587 // Insert instructions into newMBB based on incoming instruction
7588 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7589 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7590 "unexpected number of operands");
7591 MachineOperand& dest1Oper = bInstr->getOperand(0);
7592 MachineOperand& dest2Oper = bInstr->getOperand(1);
7593 MachineOperand* argOpers[2 + X86AddrNumOperands];
7594 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7595 argOpers[i] = &bInstr->getOperand(i+2);
7597 // x86 address has 4 operands: base, index, scale, and displacement
7598 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7600 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7601 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7602 for (int i=0; i <= lastAddrIndx; ++i)
7603 (*MIB).addOperand(*argOpers[i]);
7604 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7605 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7606 // add 4 to displacement.
7607 for (int i=0; i <= lastAddrIndx-2; ++i)
7608 (*MIB).addOperand(*argOpers[i]);
7609 MachineOperand newOp3 = *(argOpers[3]);
7611 newOp3.setImm(newOp3.getImm()+4);
7613 newOp3.setOffset(newOp3.getOffset()+4);
7614 (*MIB).addOperand(newOp3);
7615 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7617 // t3/4 are defined later, at the bottom of the loop
7618 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7619 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7620 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7621 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7622 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7623 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7625 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7626 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7628 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7629 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7635 int valArgIndx = lastAddrIndx + 1;
7636 assert((argOpers[valArgIndx]->isReg() ||
7637 argOpers[valArgIndx]->isImm()) &&
7639 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7640 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7641 if (argOpers[valArgIndx]->isReg())
7642 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7644 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7645 if (regOpcL != X86::MOV32rr)
7647 (*MIB).addOperand(*argOpers[valArgIndx]);
7648 assert(argOpers[valArgIndx + 1]->isReg() ==
7649 argOpers[valArgIndx]->isReg());
7650 assert(argOpers[valArgIndx + 1]->isImm() ==
7651 argOpers[valArgIndx]->isImm());
7652 if (argOpers[valArgIndx + 1]->isReg())
7653 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7655 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7656 if (regOpcH != X86::MOV32rr)
7658 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7660 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7662 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7665 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7667 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7670 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7671 for (int i=0; i <= lastAddrIndx; ++i)
7672 (*MIB).addOperand(*argOpers[i]);
7674 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7675 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7676 bInstr->memoperands_end());
7678 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7679 MIB.addReg(X86::EAX);
7680 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7681 MIB.addReg(X86::EDX);
7684 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7686 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7690 // private utility function
7692 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7693 MachineBasicBlock *MBB,
7694 unsigned cmovOpc) const {
7695 // For the atomic min/max operator, we generate
7698 // ld t1 = [min/max.addr]
7699 // mov t2 = [min/max.val]
7701 // cmov[cond] t2 = t1
7703 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7705 // fallthrough -->nextMBB
7707 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7708 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7709 MachineFunction::iterator MBBIter = MBB;
7712 /// First build the CFG
7713 MachineFunction *F = MBB->getParent();
7714 MachineBasicBlock *thisMBB = MBB;
7715 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7716 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7717 F->insert(MBBIter, newMBB);
7718 F->insert(MBBIter, nextMBB);
7720 // Move all successors of thisMBB to nextMBB
7721 nextMBB->transferSuccessors(thisMBB);
7723 // Update thisMBB to fall through to newMBB
7724 thisMBB->addSuccessor(newMBB);
7726 // newMBB jumps to newMBB and fall through to nextMBB
7727 newMBB->addSuccessor(nextMBB);
7728 newMBB->addSuccessor(newMBB);
7730 DebugLoc dl = mInstr->getDebugLoc();
7731 // Insert instructions into newMBB based on incoming instruction
7732 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7733 "unexpected number of operands");
7734 MachineOperand& destOper = mInstr->getOperand(0);
7735 MachineOperand* argOpers[2 + X86AddrNumOperands];
7736 int numArgs = mInstr->getNumOperands() - 1;
7737 for (int i=0; i < numArgs; ++i)
7738 argOpers[i] = &mInstr->getOperand(i+1);
7740 // x86 address has 4 operands: base, index, scale, and displacement
7741 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7742 int valArgIndx = lastAddrIndx + 1;
7744 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7745 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7746 for (int i=0; i <= lastAddrIndx; ++i)
7747 (*MIB).addOperand(*argOpers[i]);
7749 // We only support register and immediate values
7750 assert((argOpers[valArgIndx]->isReg() ||
7751 argOpers[valArgIndx]->isImm()) &&
7754 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7755 if (argOpers[valArgIndx]->isReg())
7756 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7758 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7759 (*MIB).addOperand(*argOpers[valArgIndx]);
7761 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7764 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7769 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7770 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7774 // Cmp and exchange if none has modified the memory location
7775 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7776 for (int i=0; i <= lastAddrIndx; ++i)
7777 (*MIB).addOperand(*argOpers[i]);
7779 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7780 (*MIB).setMemRefs(mInstr->memoperands_begin(),
7781 mInstr->memoperands_end());
7783 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7784 MIB.addReg(X86::EAX);
7787 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7789 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7793 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
7794 // all of this code can be replaced with that in the .td file.
7796 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
7797 unsigned numArgs, bool memArg) const {
7799 MachineFunction *F = BB->getParent();
7800 DebugLoc dl = MI->getDebugLoc();
7801 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7805 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
7807 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
7809 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
7811 for (unsigned i = 0; i < numArgs; ++i) {
7812 MachineOperand &Op = MI->getOperand(i+1);
7814 if (!(Op.isReg() && Op.isImplicit()))
7818 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
7821 F->DeleteMachineInstr(MI);
7827 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
7829 MachineBasicBlock *MBB) const {
7830 // Emit code to save XMM registers to the stack. The ABI says that the
7831 // number of registers to save is given in %al, so it's theoretically
7832 // possible to do an indirect jump trick to avoid saving all of them,
7833 // however this code takes a simpler approach and just executes all
7834 // of the stores if %al is non-zero. It's less code, and it's probably
7835 // easier on the hardware branch predictor, and stores aren't all that
7836 // expensive anyway.
7838 // Create the new basic blocks. One block contains all the XMM stores,
7839 // and one block is the final destination regardless of whether any
7840 // stores were performed.
7841 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7842 MachineFunction *F = MBB->getParent();
7843 MachineFunction::iterator MBBIter = MBB;
7845 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
7846 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
7847 F->insert(MBBIter, XMMSaveMBB);
7848 F->insert(MBBIter, EndMBB);
7851 // Move any original successors of MBB to the end block.
7852 EndMBB->transferSuccessors(MBB);
7853 // The original block will now fall through to the XMM save block.
7854 MBB->addSuccessor(XMMSaveMBB);
7855 // The XMMSaveMBB will fall through to the end block.
7856 XMMSaveMBB->addSuccessor(EndMBB);
7858 // Now add the instructions.
7859 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7860 DebugLoc DL = MI->getDebugLoc();
7862 unsigned CountReg = MI->getOperand(0).getReg();
7863 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
7864 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
7866 if (!Subtarget->isTargetWin64()) {
7867 // If %al is 0, branch around the XMM save block.
7868 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
7869 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
7870 MBB->addSuccessor(EndMBB);
7873 // In the XMM save block, save all the XMM argument registers.
7874 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
7875 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
7876 MachineMemOperand *MMO =
7877 F->getMachineMemOperand(
7878 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
7879 MachineMemOperand::MOStore, Offset,
7880 /*Size=*/16, /*Align=*/16);
7881 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
7882 .addFrameIndex(RegSaveFrameIndex)
7883 .addImm(/*Scale=*/1)
7884 .addReg(/*IndexReg=*/0)
7885 .addImm(/*Disp=*/Offset)
7886 .addReg(/*Segment=*/0)
7887 .addReg(MI->getOperand(i).getReg())
7888 .addMemOperand(MMO);
7891 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7897 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
7898 MachineBasicBlock *BB,
7899 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
7900 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7901 DebugLoc DL = MI->getDebugLoc();
7903 // To "insert" a SELECT_CC instruction, we actually have to insert the
7904 // diamond control-flow pattern. The incoming instruction knows the
7905 // destination vreg to set, the condition code register to branch on, the
7906 // true/false values to select between, and a branch opcode to use.
7907 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7908 MachineFunction::iterator It = BB;
7914 // cmpTY ccX, r1, r2
7916 // fallthrough --> copy0MBB
7917 MachineBasicBlock *thisMBB = BB;
7918 MachineFunction *F = BB->getParent();
7919 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7920 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7922 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7923 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
7924 F->insert(It, copy0MBB);
7925 F->insert(It, sinkMBB);
7926 // Update machine-CFG edges by first adding all successors of the current
7927 // block to the new block which will contain the Phi node for the select.
7928 // Also inform sdisel of the edge changes.
7929 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
7930 E = BB->succ_end(); I != E; ++I) {
7931 EM->insert(std::make_pair(*I, sinkMBB));
7932 sinkMBB->addSuccessor(*I);
7934 // Next, remove all successors of the current block, and add the true
7935 // and fallthrough blocks as its successors.
7936 while (!BB->succ_empty())
7937 BB->removeSuccessor(BB->succ_begin());
7938 // Add the true and fallthrough blocks as its successors.
7939 BB->addSuccessor(copy0MBB);
7940 BB->addSuccessor(sinkMBB);
7943 // %FalseValue = ...
7944 // # fallthrough to sinkMBB
7947 // Update machine-CFG edges
7948 BB->addSuccessor(sinkMBB);
7951 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7954 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
7955 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7956 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7958 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7964 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7965 MachineBasicBlock *BB,
7966 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
7967 switch (MI->getOpcode()) {
7968 default: assert(false && "Unexpected instr type to insert");
7970 case X86::CMOV_V1I64:
7971 case X86::CMOV_FR32:
7972 case X86::CMOV_FR64:
7973 case X86::CMOV_V4F32:
7974 case X86::CMOV_V2F64:
7975 case X86::CMOV_V2I64:
7976 return EmitLoweredSelect(MI, BB, EM);
7978 case X86::FP32_TO_INT16_IN_MEM:
7979 case X86::FP32_TO_INT32_IN_MEM:
7980 case X86::FP32_TO_INT64_IN_MEM:
7981 case X86::FP64_TO_INT16_IN_MEM:
7982 case X86::FP64_TO_INT32_IN_MEM:
7983 case X86::FP64_TO_INT64_IN_MEM:
7984 case X86::FP80_TO_INT16_IN_MEM:
7985 case X86::FP80_TO_INT32_IN_MEM:
7986 case X86::FP80_TO_INT64_IN_MEM: {
7987 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7988 DebugLoc DL = MI->getDebugLoc();
7990 // Change the floating point control register to use "round towards zero"
7991 // mode when truncating to an integer value.
7992 MachineFunction *F = BB->getParent();
7993 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
7994 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7996 // Load the old value of the high byte of the control word...
7998 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7999 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
8002 // Set the high part to be round to zero...
8003 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
8006 // Reload the modified control word now...
8007 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8009 // Restore the memory image of control word to original value
8010 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
8013 // Get the X86 opcode to use.
8015 switch (MI->getOpcode()) {
8016 default: llvm_unreachable("illegal opcode!");
8017 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
8018 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
8019 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
8020 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
8021 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
8022 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
8023 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
8024 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
8025 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
8029 MachineOperand &Op = MI->getOperand(0);
8031 AM.BaseType = X86AddressMode::RegBase;
8032 AM.Base.Reg = Op.getReg();
8034 AM.BaseType = X86AddressMode::FrameIndexBase;
8035 AM.Base.FrameIndex = Op.getIndex();
8037 Op = MI->getOperand(1);
8039 AM.Scale = Op.getImm();
8040 Op = MI->getOperand(2);
8042 AM.IndexReg = Op.getImm();
8043 Op = MI->getOperand(3);
8044 if (Op.isGlobal()) {
8045 AM.GV = Op.getGlobal();
8047 AM.Disp = Op.getImm();
8049 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
8050 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
8052 // Reload the original control word now.
8053 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
8055 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8058 // String/text processing lowering.
8059 case X86::PCMPISTRM128REG:
8060 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8061 case X86::PCMPISTRM128MEM:
8062 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8063 case X86::PCMPESTRM128REG:
8064 return EmitPCMP(MI, BB, 5, false /* in mem */);
8065 case X86::PCMPESTRM128MEM:
8066 return EmitPCMP(MI, BB, 5, true /* in mem */);
8069 case X86::ATOMAND32:
8070 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8071 X86::AND32ri, X86::MOV32rm,
8072 X86::LCMPXCHG32, X86::MOV32rr,
8073 X86::NOT32r, X86::EAX,
8074 X86::GR32RegisterClass);
8076 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8077 X86::OR32ri, X86::MOV32rm,
8078 X86::LCMPXCHG32, X86::MOV32rr,
8079 X86::NOT32r, X86::EAX,
8080 X86::GR32RegisterClass);
8081 case X86::ATOMXOR32:
8082 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8083 X86::XOR32ri, X86::MOV32rm,
8084 X86::LCMPXCHG32, X86::MOV32rr,
8085 X86::NOT32r, X86::EAX,
8086 X86::GR32RegisterClass);
8087 case X86::ATOMNAND32:
8088 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8089 X86::AND32ri, X86::MOV32rm,
8090 X86::LCMPXCHG32, X86::MOV32rr,
8091 X86::NOT32r, X86::EAX,
8092 X86::GR32RegisterClass, true);
8093 case X86::ATOMMIN32:
8094 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8095 case X86::ATOMMAX32:
8096 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8097 case X86::ATOMUMIN32:
8098 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8099 case X86::ATOMUMAX32:
8100 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8102 case X86::ATOMAND16:
8103 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8104 X86::AND16ri, X86::MOV16rm,
8105 X86::LCMPXCHG16, X86::MOV16rr,
8106 X86::NOT16r, X86::AX,
8107 X86::GR16RegisterClass);
8109 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8110 X86::OR16ri, X86::MOV16rm,
8111 X86::LCMPXCHG16, X86::MOV16rr,
8112 X86::NOT16r, X86::AX,
8113 X86::GR16RegisterClass);
8114 case X86::ATOMXOR16:
8115 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8116 X86::XOR16ri, X86::MOV16rm,
8117 X86::LCMPXCHG16, X86::MOV16rr,
8118 X86::NOT16r, X86::AX,
8119 X86::GR16RegisterClass);
8120 case X86::ATOMNAND16:
8121 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8122 X86::AND16ri, X86::MOV16rm,
8123 X86::LCMPXCHG16, X86::MOV16rr,
8124 X86::NOT16r, X86::AX,
8125 X86::GR16RegisterClass, true);
8126 case X86::ATOMMIN16:
8127 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8128 case X86::ATOMMAX16:
8129 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8130 case X86::ATOMUMIN16:
8131 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8132 case X86::ATOMUMAX16:
8133 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8136 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8137 X86::AND8ri, X86::MOV8rm,
8138 X86::LCMPXCHG8, X86::MOV8rr,
8139 X86::NOT8r, X86::AL,
8140 X86::GR8RegisterClass);
8142 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8143 X86::OR8ri, X86::MOV8rm,
8144 X86::LCMPXCHG8, X86::MOV8rr,
8145 X86::NOT8r, X86::AL,
8146 X86::GR8RegisterClass);
8148 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8149 X86::XOR8ri, X86::MOV8rm,
8150 X86::LCMPXCHG8, X86::MOV8rr,
8151 X86::NOT8r, X86::AL,
8152 X86::GR8RegisterClass);
8153 case X86::ATOMNAND8:
8154 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8155 X86::AND8ri, X86::MOV8rm,
8156 X86::LCMPXCHG8, X86::MOV8rr,
8157 X86::NOT8r, X86::AL,
8158 X86::GR8RegisterClass, true);
8159 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8160 // This group is for 64-bit host.
8161 case X86::ATOMAND64:
8162 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8163 X86::AND64ri32, X86::MOV64rm,
8164 X86::LCMPXCHG64, X86::MOV64rr,
8165 X86::NOT64r, X86::RAX,
8166 X86::GR64RegisterClass);
8168 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8169 X86::OR64ri32, X86::MOV64rm,
8170 X86::LCMPXCHG64, X86::MOV64rr,
8171 X86::NOT64r, X86::RAX,
8172 X86::GR64RegisterClass);
8173 case X86::ATOMXOR64:
8174 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8175 X86::XOR64ri32, X86::MOV64rm,
8176 X86::LCMPXCHG64, X86::MOV64rr,
8177 X86::NOT64r, X86::RAX,
8178 X86::GR64RegisterClass);
8179 case X86::ATOMNAND64:
8180 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8181 X86::AND64ri32, X86::MOV64rm,
8182 X86::LCMPXCHG64, X86::MOV64rr,
8183 X86::NOT64r, X86::RAX,
8184 X86::GR64RegisterClass, true);
8185 case X86::ATOMMIN64:
8186 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8187 case X86::ATOMMAX64:
8188 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8189 case X86::ATOMUMIN64:
8190 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8191 case X86::ATOMUMAX64:
8192 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8194 // This group does 64-bit operations on a 32-bit host.
8195 case X86::ATOMAND6432:
8196 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8197 X86::AND32rr, X86::AND32rr,
8198 X86::AND32ri, X86::AND32ri,
8200 case X86::ATOMOR6432:
8201 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8202 X86::OR32rr, X86::OR32rr,
8203 X86::OR32ri, X86::OR32ri,
8205 case X86::ATOMXOR6432:
8206 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8207 X86::XOR32rr, X86::XOR32rr,
8208 X86::XOR32ri, X86::XOR32ri,
8210 case X86::ATOMNAND6432:
8211 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8212 X86::AND32rr, X86::AND32rr,
8213 X86::AND32ri, X86::AND32ri,
8215 case X86::ATOMADD6432:
8216 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8217 X86::ADD32rr, X86::ADC32rr,
8218 X86::ADD32ri, X86::ADC32ri,
8220 case X86::ATOMSUB6432:
8221 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8222 X86::SUB32rr, X86::SBB32rr,
8223 X86::SUB32ri, X86::SBB32ri,
8225 case X86::ATOMSWAP6432:
8226 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8227 X86::MOV32rr, X86::MOV32rr,
8228 X86::MOV32ri, X86::MOV32ri,
8230 case X86::VASTART_SAVE_XMM_REGS:
8231 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8235 //===----------------------------------------------------------------------===//
8236 // X86 Optimization Hooks
8237 //===----------------------------------------------------------------------===//
8239 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8243 const SelectionDAG &DAG,
8244 unsigned Depth) const {
8245 unsigned Opc = Op.getOpcode();
8246 assert((Opc >= ISD::BUILTIN_OP_END ||
8247 Opc == ISD::INTRINSIC_WO_CHAIN ||
8248 Opc == ISD::INTRINSIC_W_CHAIN ||
8249 Opc == ISD::INTRINSIC_VOID) &&
8250 "Should use MaskedValueIsZero if you don't know whether Op"
8251 " is a target node!");
8253 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8265 // These nodes' second result is a boolean.
8266 if (Op.getResNo() == 0)
8270 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8271 Mask.getBitWidth() - 1);
8276 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8277 /// node is a GlobalAddress + offset.
8278 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8279 GlobalValue* &GA, int64_t &Offset) const{
8280 if (N->getOpcode() == X86ISD::Wrapper) {
8281 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8282 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8283 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8287 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8290 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
8291 const TargetLowering &TLI) {
8294 if (TLI.isGAPlusOffset(Base, GV, Offset))
8295 return (GV->getAlignment() >= N && (Offset % N) == 0);
8296 // DAG combine handles the stack object case.
8300 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8301 EVT EltVT, LoadSDNode *&LDBase,
8302 unsigned &LastLoadedElt,
8303 SelectionDAG &DAG, MachineFrameInfo *MFI,
8304 const TargetLowering &TLI) {
8306 LastLoadedElt = -1U;
8307 for (unsigned i = 0; i < NumElems; ++i) {
8308 if (N->getMaskElt(i) < 0) {
8314 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8315 if (!Elt.getNode() ||
8316 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8319 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8321 LDBase = cast<LoadSDNode>(Elt.getNode());
8325 if (Elt.getOpcode() == ISD::UNDEF)
8328 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8329 if (!TLI.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i, MFI))
8336 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8337 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8338 /// if the load addresses are consecutive, non-overlapping, and in the right
8339 /// order. In the case of v2i64, it will see if it can rewrite the
8340 /// shuffle to be an appropriate build vector so it can take advantage of
8341 // performBuildVectorCombine.
8342 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8343 const TargetLowering &TLI) {
8344 DebugLoc dl = N->getDebugLoc();
8345 EVT VT = N->getValueType(0);
8346 EVT EltVT = VT.getVectorElementType();
8347 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8348 unsigned NumElems = VT.getVectorNumElements();
8350 if (VT.getSizeInBits() != 128)
8353 // Try to combine a vector_shuffle into a 128-bit load.
8354 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8355 LoadSDNode *LD = NULL;
8356 unsigned LastLoadedElt;
8357 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8361 if (LastLoadedElt == NumElems - 1) {
8362 if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
8363 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8364 LD->getSrcValue(), LD->getSrcValueOffset(),
8366 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8367 LD->getSrcValue(), LD->getSrcValueOffset(),
8368 LD->isVolatile(), LD->getAlignment());
8369 } else if (NumElems == 4 && LastLoadedElt == 1) {
8370 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8371 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8372 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8373 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8378 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8379 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8380 const X86Subtarget *Subtarget) {
8381 DebugLoc DL = N->getDebugLoc();
8382 SDValue Cond = N->getOperand(0);
8383 // Get the LHS/RHS of the select.
8384 SDValue LHS = N->getOperand(1);
8385 SDValue RHS = N->getOperand(2);
8387 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8388 // instructions have the peculiarity that if either operand is a NaN,
8389 // they chose what we call the RHS operand (and as such are not symmetric).
8390 // It happens that this matches the semantics of the common C idiom
8391 // x<y?x:y and related forms, so we can recognize these cases.
8392 if (Subtarget->hasSSE2() &&
8393 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8394 Cond.getOpcode() == ISD::SETCC) {
8395 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8397 unsigned Opcode = 0;
8398 // Check for x CC y ? x : y.
8399 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8403 // This can be a min if we can prove that at least one of the operands
8405 if (!FiniteOnlyFPMath()) {
8406 if (DAG.isKnownNeverNaN(RHS)) {
8407 // Put the potential NaN in the RHS so that SSE will preserve it.
8408 std::swap(LHS, RHS);
8409 } else if (!DAG.isKnownNeverNaN(LHS))
8412 Opcode = X86ISD::FMIN;
8415 // This can be a min if we can prove that at least one of the operands
8417 if (!FiniteOnlyFPMath()) {
8418 if (DAG.isKnownNeverNaN(LHS)) {
8419 // Put the potential NaN in the RHS so that SSE will preserve it.
8420 std::swap(LHS, RHS);
8421 } else if (!DAG.isKnownNeverNaN(RHS))
8424 Opcode = X86ISD::FMIN;
8427 // This can be a min, but if either operand is a NaN we need it to
8428 // preserve the original LHS.
8429 std::swap(LHS, RHS);
8433 Opcode = X86ISD::FMIN;
8437 // This can be a max if we can prove that at least one of the operands
8439 if (!FiniteOnlyFPMath()) {
8440 if (DAG.isKnownNeverNaN(LHS)) {
8441 // Put the potential NaN in the RHS so that SSE will preserve it.
8442 std::swap(LHS, RHS);
8443 } else if (!DAG.isKnownNeverNaN(RHS))
8446 Opcode = X86ISD::FMAX;
8449 // This can be a max if we can prove that at least one of the operands
8451 if (!FiniteOnlyFPMath()) {
8452 if (DAG.isKnownNeverNaN(RHS)) {
8453 // Put the potential NaN in the RHS so that SSE will preserve it.
8454 std::swap(LHS, RHS);
8455 } else if (!DAG.isKnownNeverNaN(LHS))
8458 Opcode = X86ISD::FMAX;
8461 // This can be a max, but if either operand is a NaN we need it to
8462 // preserve the original LHS.
8463 std::swap(LHS, RHS);
8467 Opcode = X86ISD::FMAX;
8470 // Check for x CC y ? y : x -- a min/max with reversed arms.
8471 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8475 // This can be a min if we can prove that at least one of the operands
8477 if (!FiniteOnlyFPMath()) {
8478 if (DAG.isKnownNeverNaN(RHS)) {
8479 // Put the potential NaN in the RHS so that SSE will preserve it.
8480 std::swap(LHS, RHS);
8481 } else if (!DAG.isKnownNeverNaN(LHS))
8484 Opcode = X86ISD::FMIN;
8487 // This can be a min if we can prove that at least one of the operands
8489 if (!FiniteOnlyFPMath()) {
8490 if (DAG.isKnownNeverNaN(LHS)) {
8491 // Put the potential NaN in the RHS so that SSE will preserve it.
8492 std::swap(LHS, RHS);
8493 } else if (!DAG.isKnownNeverNaN(RHS))
8496 Opcode = X86ISD::FMIN;
8499 // This can be a min, but if either operand is a NaN we need it to
8500 // preserve the original LHS.
8501 std::swap(LHS, RHS);
8505 Opcode = X86ISD::FMIN;
8509 // This can be a max if we can prove that at least one of the operands
8511 if (!FiniteOnlyFPMath()) {
8512 if (DAG.isKnownNeverNaN(LHS)) {
8513 // Put the potential NaN in the RHS so that SSE will preserve it.
8514 std::swap(LHS, RHS);
8515 } else if (!DAG.isKnownNeverNaN(RHS))
8518 Opcode = X86ISD::FMAX;
8521 // This can be a max if we can prove that at least one of the operands
8523 if (!FiniteOnlyFPMath()) {
8524 if (DAG.isKnownNeverNaN(RHS)) {
8525 // Put the potential NaN in the RHS so that SSE will preserve it.
8526 std::swap(LHS, RHS);
8527 } else if (!DAG.isKnownNeverNaN(LHS))
8530 Opcode = X86ISD::FMAX;
8533 // This can be a max, but if either operand is a NaN we need it to
8534 // preserve the original LHS.
8535 std::swap(LHS, RHS);
8539 Opcode = X86ISD::FMAX;
8545 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8548 // If this is a select between two integer constants, try to do some
8550 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8551 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8552 // Don't do this for crazy integer types.
8553 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8554 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8555 // so that TrueC (the true value) is larger than FalseC.
8556 bool NeedsCondInvert = false;
8558 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8559 // Efficiently invertible.
8560 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8561 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8562 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8563 NeedsCondInvert = true;
8564 std::swap(TrueC, FalseC);
8567 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8568 if (FalseC->getAPIntValue() == 0 &&
8569 TrueC->getAPIntValue().isPowerOf2()) {
8570 if (NeedsCondInvert) // Invert the condition if needed.
8571 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8572 DAG.getConstant(1, Cond.getValueType()));
8574 // Zero extend the condition if needed.
8575 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8577 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8578 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8579 DAG.getConstant(ShAmt, MVT::i8));
8582 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8583 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8584 if (NeedsCondInvert) // Invert the condition if needed.
8585 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8586 DAG.getConstant(1, Cond.getValueType()));
8588 // Zero extend the condition if needed.
8589 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8590 FalseC->getValueType(0), Cond);
8591 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8592 SDValue(FalseC, 0));
8595 // Optimize cases that will turn into an LEA instruction. This requires
8596 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8597 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8598 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8599 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8601 bool isFastMultiplier = false;
8603 switch ((unsigned char)Diff) {
8605 case 1: // result = add base, cond
8606 case 2: // result = lea base( , cond*2)
8607 case 3: // result = lea base(cond, cond*2)
8608 case 4: // result = lea base( , cond*4)
8609 case 5: // result = lea base(cond, cond*4)
8610 case 8: // result = lea base( , cond*8)
8611 case 9: // result = lea base(cond, cond*8)
8612 isFastMultiplier = true;
8617 if (isFastMultiplier) {
8618 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8619 if (NeedsCondInvert) // Invert the condition if needed.
8620 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8621 DAG.getConstant(1, Cond.getValueType()));
8623 // Zero extend the condition if needed.
8624 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8626 // Scale the condition by the difference.
8628 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8629 DAG.getConstant(Diff, Cond.getValueType()));
8631 // Add the base if non-zero.
8632 if (FalseC->getAPIntValue() != 0)
8633 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8634 SDValue(FalseC, 0));
8644 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8645 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8646 TargetLowering::DAGCombinerInfo &DCI) {
8647 DebugLoc DL = N->getDebugLoc();
8649 // If the flag operand isn't dead, don't touch this CMOV.
8650 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8653 // If this is a select between two integer constants, try to do some
8654 // optimizations. Note that the operands are ordered the opposite of SELECT
8656 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8657 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8658 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8659 // larger than FalseC (the false value).
8660 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8662 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8663 CC = X86::GetOppositeBranchCondition(CC);
8664 std::swap(TrueC, FalseC);
8667 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8668 // This is efficient for any integer data type (including i8/i16) and
8670 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8671 SDValue Cond = N->getOperand(3);
8672 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8673 DAG.getConstant(CC, MVT::i8), Cond);
8675 // Zero extend the condition if needed.
8676 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8678 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8679 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8680 DAG.getConstant(ShAmt, MVT::i8));
8681 if (N->getNumValues() == 2) // Dead flag value?
8682 return DCI.CombineTo(N, Cond, SDValue());
8686 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8687 // for any integer data type, including i8/i16.
8688 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8689 SDValue Cond = N->getOperand(3);
8690 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8691 DAG.getConstant(CC, MVT::i8), Cond);
8693 // Zero extend the condition if needed.
8694 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8695 FalseC->getValueType(0), Cond);
8696 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8697 SDValue(FalseC, 0));
8699 if (N->getNumValues() == 2) // Dead flag value?
8700 return DCI.CombineTo(N, Cond, SDValue());
8704 // Optimize cases that will turn into an LEA instruction. This requires
8705 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8706 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8707 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8708 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8710 bool isFastMultiplier = false;
8712 switch ((unsigned char)Diff) {
8714 case 1: // result = add base, cond
8715 case 2: // result = lea base( , cond*2)
8716 case 3: // result = lea base(cond, cond*2)
8717 case 4: // result = lea base( , cond*4)
8718 case 5: // result = lea base(cond, cond*4)
8719 case 8: // result = lea base( , cond*8)
8720 case 9: // result = lea base(cond, cond*8)
8721 isFastMultiplier = true;
8726 if (isFastMultiplier) {
8727 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8728 SDValue Cond = N->getOperand(3);
8729 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8730 DAG.getConstant(CC, MVT::i8), Cond);
8731 // Zero extend the condition if needed.
8732 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8734 // Scale the condition by the difference.
8736 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8737 DAG.getConstant(Diff, Cond.getValueType()));
8739 // Add the base if non-zero.
8740 if (FalseC->getAPIntValue() != 0)
8741 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8742 SDValue(FalseC, 0));
8743 if (N->getNumValues() == 2) // Dead flag value?
8744 return DCI.CombineTo(N, Cond, SDValue());
8754 /// PerformMulCombine - Optimize a single multiply with constant into two
8755 /// in order to implement it with two cheaper instructions, e.g.
8756 /// LEA + SHL, LEA + LEA.
8757 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8758 TargetLowering::DAGCombinerInfo &DCI) {
8759 if (DAG.getMachineFunction().
8760 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8763 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8766 EVT VT = N->getValueType(0);
8770 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8773 uint64_t MulAmt = C->getZExtValue();
8774 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8777 uint64_t MulAmt1 = 0;
8778 uint64_t MulAmt2 = 0;
8779 if ((MulAmt % 9) == 0) {
8781 MulAmt2 = MulAmt / 9;
8782 } else if ((MulAmt % 5) == 0) {
8784 MulAmt2 = MulAmt / 5;
8785 } else if ((MulAmt % 3) == 0) {
8787 MulAmt2 = MulAmt / 3;
8790 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8791 DebugLoc DL = N->getDebugLoc();
8793 if (isPowerOf2_64(MulAmt2) &&
8794 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8795 // If second multiplifer is pow2, issue it first. We want the multiply by
8796 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8798 std::swap(MulAmt1, MulAmt2);
8801 if (isPowerOf2_64(MulAmt1))
8802 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8803 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8805 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8806 DAG.getConstant(MulAmt1, VT));
8808 if (isPowerOf2_64(MulAmt2))
8809 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8810 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8812 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8813 DAG.getConstant(MulAmt2, VT));
8815 // Do not add new nodes to DAG combiner worklist.
8816 DCI.CombineTo(N, NewMul, false);
8822 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8824 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8825 const X86Subtarget *Subtarget) {
8826 // On X86 with SSE2 support, we can transform this to a vector shift if
8827 // all elements are shifted by the same amount. We can't do this in legalize
8828 // because the a constant vector is typically transformed to a constant pool
8829 // so we have no knowledge of the shift amount.
8830 if (!Subtarget->hasSSE2())
8833 EVT VT = N->getValueType(0);
8834 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8837 SDValue ShAmtOp = N->getOperand(1);
8838 EVT EltVT = VT.getVectorElementType();
8839 DebugLoc DL = N->getDebugLoc();
8840 SDValue BaseShAmt = SDValue();
8841 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8842 unsigned NumElts = VT.getVectorNumElements();
8844 for (; i != NumElts; ++i) {
8845 SDValue Arg = ShAmtOp.getOperand(i);
8846 if (Arg.getOpcode() == ISD::UNDEF) continue;
8850 for (; i != NumElts; ++i) {
8851 SDValue Arg = ShAmtOp.getOperand(i);
8852 if (Arg.getOpcode() == ISD::UNDEF) continue;
8853 if (Arg != BaseShAmt) {
8857 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8858 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8859 SDValue InVec = ShAmtOp.getOperand(0);
8860 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
8861 unsigned NumElts = InVec.getValueType().getVectorNumElements();
8863 for (; i != NumElts; ++i) {
8864 SDValue Arg = InVec.getOperand(i);
8865 if (Arg.getOpcode() == ISD::UNDEF) continue;
8869 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
8870 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
8871 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
8872 if (C->getZExtValue() == SplatIdx)
8873 BaseShAmt = InVec.getOperand(1);
8876 if (BaseShAmt.getNode() == 0)
8877 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8878 DAG.getIntPtrConstant(0));
8882 // The shift amount is an i32.
8883 if (EltVT.bitsGT(MVT::i32))
8884 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8885 else if (EltVT.bitsLT(MVT::i32))
8886 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
8888 // The shift amount is identical so we can do a vector shift.
8889 SDValue ValOp = N->getOperand(0);
8890 switch (N->getOpcode()) {
8892 llvm_unreachable("Unknown shift opcode!");
8895 if (VT == MVT::v2i64)
8896 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8897 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8899 if (VT == MVT::v4i32)
8900 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8901 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8903 if (VT == MVT::v8i16)
8904 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8905 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8909 if (VT == MVT::v4i32)
8910 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8911 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8913 if (VT == MVT::v8i16)
8914 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8915 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8919 if (VT == MVT::v2i64)
8920 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8921 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8923 if (VT == MVT::v4i32)
8924 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8925 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8927 if (VT == MVT::v8i16)
8928 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8929 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8936 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8937 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8938 const X86Subtarget *Subtarget) {
8939 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8940 // the FP state in cases where an emms may be missing.
8941 // A preferable solution to the general problem is to figure out the right
8942 // places to insert EMMS. This qualifies as a quick hack.
8944 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8945 StoreSDNode *St = cast<StoreSDNode>(N);
8946 EVT VT = St->getValue().getValueType();
8947 if (VT.getSizeInBits() != 64)
8950 const Function *F = DAG.getMachineFunction().getFunction();
8951 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
8952 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
8953 && Subtarget->hasSSE2();
8954 if ((VT.isVector() ||
8955 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8956 isa<LoadSDNode>(St->getValue()) &&
8957 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8958 St->getChain().hasOneUse() && !St->isVolatile()) {
8959 SDNode* LdVal = St->getValue().getNode();
8961 int TokenFactorIndex = -1;
8962 SmallVector<SDValue, 8> Ops;
8963 SDNode* ChainVal = St->getChain().getNode();
8964 // Must be a store of a load. We currently handle two cases: the load
8965 // is a direct child, and it's under an intervening TokenFactor. It is
8966 // possible to dig deeper under nested TokenFactors.
8967 if (ChainVal == LdVal)
8968 Ld = cast<LoadSDNode>(St->getChain());
8969 else if (St->getValue().hasOneUse() &&
8970 ChainVal->getOpcode() == ISD::TokenFactor) {
8971 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8972 if (ChainVal->getOperand(i).getNode() == LdVal) {
8973 TokenFactorIndex = i;
8974 Ld = cast<LoadSDNode>(St->getValue());
8976 Ops.push_back(ChainVal->getOperand(i));
8980 if (!Ld || !ISD::isNormalLoad(Ld))
8983 // If this is not the MMX case, i.e. we are just turning i64 load/store
8984 // into f64 load/store, avoid the transformation if there are multiple
8985 // uses of the loaded value.
8986 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8989 DebugLoc LdDL = Ld->getDebugLoc();
8990 DebugLoc StDL = N->getDebugLoc();
8991 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8992 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8994 if (Subtarget->is64Bit() || F64IsLegal) {
8995 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8996 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8997 Ld->getBasePtr(), Ld->getSrcValue(),
8998 Ld->getSrcValueOffset(), Ld->isVolatile(),
8999 Ld->getAlignment());
9000 SDValue NewChain = NewLd.getValue(1);
9001 if (TokenFactorIndex != -1) {
9002 Ops.push_back(NewChain);
9003 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9006 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
9007 St->getSrcValue(), St->getSrcValueOffset(),
9008 St->isVolatile(), St->getAlignment());
9011 // Otherwise, lower to two pairs of 32-bit loads / stores.
9012 SDValue LoAddr = Ld->getBasePtr();
9013 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
9014 DAG.getConstant(4, MVT::i32));
9016 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
9017 Ld->getSrcValue(), Ld->getSrcValueOffset(),
9018 Ld->isVolatile(), Ld->getAlignment());
9019 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
9020 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
9022 MinAlign(Ld->getAlignment(), 4));
9024 SDValue NewChain = LoLd.getValue(1);
9025 if (TokenFactorIndex != -1) {
9026 Ops.push_back(LoLd);
9027 Ops.push_back(HiLd);
9028 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
9032 LoAddr = St->getBasePtr();
9033 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
9034 DAG.getConstant(4, MVT::i32));
9036 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
9037 St->getSrcValue(), St->getSrcValueOffset(),
9038 St->isVolatile(), St->getAlignment());
9039 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
9041 St->getSrcValueOffset() + 4,
9043 MinAlign(St->getAlignment(), 4));
9044 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
9049 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
9050 /// X86ISD::FXOR nodes.
9051 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
9052 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
9053 // F[X]OR(0.0, x) -> x
9054 // F[X]OR(x, 0.0) -> x
9055 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9056 if (C->getValueAPF().isPosZero())
9057 return N->getOperand(1);
9058 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9059 if (C->getValueAPF().isPosZero())
9060 return N->getOperand(0);
9064 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9065 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9066 // FAND(0.0, x) -> 0.0
9067 // FAND(x, 0.0) -> 0.0
9068 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9069 if (C->getValueAPF().isPosZero())
9070 return N->getOperand(0);
9071 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9072 if (C->getValueAPF().isPosZero())
9073 return N->getOperand(1);
9077 static SDValue PerformBTCombine(SDNode *N,
9079 TargetLowering::DAGCombinerInfo &DCI) {
9080 // BT ignores high bits in the bit index operand.
9081 SDValue Op1 = N->getOperand(1);
9082 if (Op1.hasOneUse()) {
9083 unsigned BitWidth = Op1.getValueSizeInBits();
9084 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9085 APInt KnownZero, KnownOne;
9086 TargetLowering::TargetLoweringOpt TLO(DAG);
9087 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9088 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9089 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9090 DCI.CommitTargetLoweringOpt(TLO);
9095 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9096 SDValue Op = N->getOperand(0);
9097 if (Op.getOpcode() == ISD::BIT_CONVERT)
9098 Op = Op.getOperand(0);
9099 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9100 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9101 VT.getVectorElementType().getSizeInBits() ==
9102 OpVT.getVectorElementType().getSizeInBits()) {
9103 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9108 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9109 // Locked instructions, in turn, have implicit fence semantics (all memory
9110 // operations are flushed before issuing the locked instruction, and the
9111 // are not buffered), so we can fold away the common pattern of
9112 // fence-atomic-fence.
9113 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9114 SDValue atomic = N->getOperand(0);
9115 switch (atomic.getOpcode()) {
9116 case ISD::ATOMIC_CMP_SWAP:
9117 case ISD::ATOMIC_SWAP:
9118 case ISD::ATOMIC_LOAD_ADD:
9119 case ISD::ATOMIC_LOAD_SUB:
9120 case ISD::ATOMIC_LOAD_AND:
9121 case ISD::ATOMIC_LOAD_OR:
9122 case ISD::ATOMIC_LOAD_XOR:
9123 case ISD::ATOMIC_LOAD_NAND:
9124 case ISD::ATOMIC_LOAD_MIN:
9125 case ISD::ATOMIC_LOAD_MAX:
9126 case ISD::ATOMIC_LOAD_UMIN:
9127 case ISD::ATOMIC_LOAD_UMAX:
9133 SDValue fence = atomic.getOperand(0);
9134 if (fence.getOpcode() != ISD::MEMBARRIER)
9137 switch (atomic.getOpcode()) {
9138 case ISD::ATOMIC_CMP_SWAP:
9139 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9140 atomic.getOperand(1), atomic.getOperand(2),
9141 atomic.getOperand(3));
9142 case ISD::ATOMIC_SWAP:
9143 case ISD::ATOMIC_LOAD_ADD:
9144 case ISD::ATOMIC_LOAD_SUB:
9145 case ISD::ATOMIC_LOAD_AND:
9146 case ISD::ATOMIC_LOAD_OR:
9147 case ISD::ATOMIC_LOAD_XOR:
9148 case ISD::ATOMIC_LOAD_NAND:
9149 case ISD::ATOMIC_LOAD_MIN:
9150 case ISD::ATOMIC_LOAD_MAX:
9151 case ISD::ATOMIC_LOAD_UMIN:
9152 case ISD::ATOMIC_LOAD_UMAX:
9153 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9154 atomic.getOperand(1), atomic.getOperand(2));
9160 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9161 DAGCombinerInfo &DCI) const {
9162 SelectionDAG &DAG = DCI.DAG;
9163 switch (N->getOpcode()) {
9165 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9166 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9167 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9168 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9171 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9172 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9174 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9175 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9176 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9177 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9178 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9184 //===----------------------------------------------------------------------===//
9185 // X86 Inline Assembly Support
9186 //===----------------------------------------------------------------------===//
9188 static bool LowerToBSwap(CallInst *CI) {
9189 // FIXME: this should verify that we are targetting a 486 or better. If not,
9190 // we will turn this bswap into something that will be lowered to logical ops
9191 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9192 // so don't worry about this.
9194 // Verify this is a simple bswap.
9195 if (CI->getNumOperands() != 2 ||
9196 CI->getType() != CI->getOperand(1)->getType() ||
9197 !CI->getType()->isInteger())
9200 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9201 if (!Ty || Ty->getBitWidth() % 16 != 0)
9204 // Okay, we can do this xform, do so now.
9205 const Type *Tys[] = { Ty };
9206 Module *M = CI->getParent()->getParent()->getParent();
9207 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9209 Value *Op = CI->getOperand(1);
9210 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9212 CI->replaceAllUsesWith(Op);
9213 CI->eraseFromParent();
9217 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9218 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9219 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9221 std::string AsmStr = IA->getAsmString();
9223 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9224 std::vector<std::string> AsmPieces;
9225 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9227 switch (AsmPieces.size()) {
9228 default: return false;
9230 AsmStr = AsmPieces[0];
9232 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9235 if (AsmPieces.size() == 2 &&
9236 (AsmPieces[0] == "bswap" ||
9237 AsmPieces[0] == "bswapq" ||
9238 AsmPieces[0] == "bswapl") &&
9239 (AsmPieces[1] == "$0" ||
9240 AsmPieces[1] == "${0:q}")) {
9241 // No need to check constraints, nothing other than the equivalent of
9242 // "=r,0" would be valid here.
9243 return LowerToBSwap(CI);
9245 // rorw $$8, ${0:w} --> llvm.bswap.i16
9246 if (CI->getType() == Type::getInt16Ty(CI->getContext()) &&
9247 AsmPieces.size() == 3 &&
9248 AsmPieces[0] == "rorw" &&
9249 AsmPieces[1] == "$$8," &&
9250 AsmPieces[2] == "${0:w}" &&
9251 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9252 return LowerToBSwap(CI);
9256 if (CI->getType() == Type::getInt64Ty(CI->getContext()) &&
9257 Constraints.size() >= 2 &&
9258 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9259 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9260 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9261 std::vector<std::string> Words;
9262 SplitString(AsmPieces[0], Words, " \t");
9263 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9265 SplitString(AsmPieces[1], Words, " \t");
9266 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9268 SplitString(AsmPieces[2], Words, " \t,");
9269 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9270 Words[2] == "%edx") {
9271 return LowerToBSwap(CI);
9283 /// getConstraintType - Given a constraint letter, return the type of
9284 /// constraint it is for this target.
9285 X86TargetLowering::ConstraintType
9286 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9287 if (Constraint.size() == 1) {
9288 switch (Constraint[0]) {
9300 return C_RegisterClass;
9308 return TargetLowering::getConstraintType(Constraint);
9311 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9312 /// with another that has more specific requirements based on the type of the
9313 /// corresponding operand.
9314 const char *X86TargetLowering::
9315 LowerXConstraint(EVT ConstraintVT) const {
9316 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9317 // 'f' like normal targets.
9318 if (ConstraintVT.isFloatingPoint()) {
9319 if (Subtarget->hasSSE2())
9321 if (Subtarget->hasSSE1())
9325 return TargetLowering::LowerXConstraint(ConstraintVT);
9328 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9329 /// vector. If it is invalid, don't add anything to Ops.
9330 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9333 std::vector<SDValue>&Ops,
9334 SelectionDAG &DAG) const {
9335 SDValue Result(0, 0);
9337 switch (Constraint) {
9340 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9341 if (C->getZExtValue() <= 31) {
9342 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9348 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9349 if (C->getZExtValue() <= 63) {
9350 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9356 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9357 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9358 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9364 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9365 if (C->getZExtValue() <= 255) {
9366 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9372 // 32-bit signed value
9373 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9374 const ConstantInt *CI = C->getConstantIntValue();
9375 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9376 C->getSExtValue())) {
9377 // Widen to 64 bits here to get it sign extended.
9378 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9381 // FIXME gcc accepts some relocatable values here too, but only in certain
9382 // memory models; it's complicated.
9387 // 32-bit unsigned value
9388 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9389 const ConstantInt *CI = C->getConstantIntValue();
9390 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9391 C->getZExtValue())) {
9392 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9396 // FIXME gcc accepts some relocatable values here too, but only in certain
9397 // memory models; it's complicated.
9401 // Literal immediates are always ok.
9402 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9403 // Widen to 64 bits here to get it sign extended.
9404 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9408 // If we are in non-pic codegen mode, we allow the address of a global (with
9409 // an optional displacement) to be used with 'i'.
9410 GlobalAddressSDNode *GA = 0;
9413 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9415 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9416 Offset += GA->getOffset();
9418 } else if (Op.getOpcode() == ISD::ADD) {
9419 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9420 Offset += C->getZExtValue();
9421 Op = Op.getOperand(0);
9424 } else if (Op.getOpcode() == ISD::SUB) {
9425 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9426 Offset += -C->getZExtValue();
9427 Op = Op.getOperand(0);
9432 // Otherwise, this isn't something we can handle, reject it.
9436 GlobalValue *GV = GA->getGlobal();
9437 // If we require an extra load to get this address, as in PIC mode, we
9439 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9440 getTargetMachine())))
9444 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9446 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9452 if (Result.getNode()) {
9453 Ops.push_back(Result);
9456 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9460 std::vector<unsigned> X86TargetLowering::
9461 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9463 if (Constraint.size() == 1) {
9464 // FIXME: not handling fp-stack yet!
9465 switch (Constraint[0]) { // GCC X86 Constraint Letters
9466 default: break; // Unknown constraint letter
9467 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9468 if (Subtarget->is64Bit()) {
9470 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9471 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9472 X86::R10D,X86::R11D,X86::R12D,
9473 X86::R13D,X86::R14D,X86::R15D,
9474 X86::EBP, X86::ESP, 0);
9475 else if (VT == MVT::i16)
9476 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9477 X86::SI, X86::DI, X86::R8W,X86::R9W,
9478 X86::R10W,X86::R11W,X86::R12W,
9479 X86::R13W,X86::R14W,X86::R15W,
9480 X86::BP, X86::SP, 0);
9481 else if (VT == MVT::i8)
9482 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9483 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9484 X86::R10B,X86::R11B,X86::R12B,
9485 X86::R13B,X86::R14B,X86::R15B,
9486 X86::BPL, X86::SPL, 0);
9488 else if (VT == MVT::i64)
9489 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9490 X86::RSI, X86::RDI, X86::R8, X86::R9,
9491 X86::R10, X86::R11, X86::R12,
9492 X86::R13, X86::R14, X86::R15,
9493 X86::RBP, X86::RSP, 0);
9497 // 32-bit fallthrough
9500 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9501 else if (VT == MVT::i16)
9502 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9503 else if (VT == MVT::i8)
9504 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9505 else if (VT == MVT::i64)
9506 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9511 return std::vector<unsigned>();
9514 std::pair<unsigned, const TargetRegisterClass*>
9515 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9517 // First, see if this is a constraint that directly corresponds to an LLVM
9519 if (Constraint.size() == 1) {
9520 // GCC Constraint Letters
9521 switch (Constraint[0]) {
9523 case 'r': // GENERAL_REGS
9524 case 'l': // INDEX_REGS
9526 return std::make_pair(0U, X86::GR8RegisterClass);
9528 return std::make_pair(0U, X86::GR16RegisterClass);
9529 if (VT == MVT::i32 || !Subtarget->is64Bit())
9530 return std::make_pair(0U, X86::GR32RegisterClass);
9531 return std::make_pair(0U, X86::GR64RegisterClass);
9532 case 'R': // LEGACY_REGS
9534 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
9536 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
9537 if (VT == MVT::i32 || !Subtarget->is64Bit())
9538 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
9539 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
9540 case 'f': // FP Stack registers.
9541 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
9542 // value to the correct fpstack register class.
9543 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9544 return std::make_pair(0U, X86::RFP32RegisterClass);
9545 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9546 return std::make_pair(0U, X86::RFP64RegisterClass);
9547 return std::make_pair(0U, X86::RFP80RegisterClass);
9548 case 'y': // MMX_REGS if MMX allowed.
9549 if (!Subtarget->hasMMX()) break;
9550 return std::make_pair(0U, X86::VR64RegisterClass);
9551 case 'Y': // SSE_REGS if SSE2 allowed
9552 if (!Subtarget->hasSSE2()) break;
9554 case 'x': // SSE_REGS if SSE1 allowed
9555 if (!Subtarget->hasSSE1()) break;
9557 switch (VT.getSimpleVT().SimpleTy) {
9559 // Scalar SSE types.
9562 return std::make_pair(0U, X86::FR32RegisterClass);
9565 return std::make_pair(0U, X86::FR64RegisterClass);
9573 return std::make_pair(0U, X86::VR128RegisterClass);
9579 // Use the default implementation in TargetLowering to convert the register
9580 // constraint into a member of a register class.
9581 std::pair<unsigned, const TargetRegisterClass*> Res;
9582 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9584 // Not found as a standard register?
9585 if (Res.second == 0) {
9586 // Map st(0) -> st(7) -> ST0
9587 if (Constraint.size() == 7 && Constraint[0] == '{' &&
9588 tolower(Constraint[1]) == 's' &&
9589 tolower(Constraint[2]) == 't' &&
9590 Constraint[3] == '(' &&
9591 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
9592 Constraint[5] == ')' &&
9593 Constraint[6] == '}') {
9595 Res.first = X86::ST0+Constraint[4]-'0';
9596 Res.second = X86::RFP80RegisterClass;
9600 // GCC allows "st(0)" to be called just plain "st".
9601 if (StringRef("{st}").equals_lower(Constraint)) {
9602 Res.first = X86::ST0;
9603 Res.second = X86::RFP80RegisterClass;
9608 if (StringRef("{flags}").equals_lower(Constraint)) {
9609 Res.first = X86::EFLAGS;
9610 Res.second = X86::CCRRegisterClass;
9614 // 'A' means EAX + EDX.
9615 if (Constraint == "A") {
9616 Res.first = X86::EAX;
9617 Res.second = X86::GR32_ADRegisterClass;
9623 // Otherwise, check to see if this is a register class of the wrong value
9624 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
9625 // turn into {ax},{dx}.
9626 if (Res.second->hasType(VT))
9627 return Res; // Correct type already, nothing to do.
9629 // All of the single-register GCC register classes map their values onto
9630 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
9631 // really want an 8-bit or 32-bit register, map to the appropriate register
9632 // class and return the appropriate register.
9633 if (Res.second == X86::GR16RegisterClass) {
9634 if (VT == MVT::i8) {
9635 unsigned DestReg = 0;
9636 switch (Res.first) {
9638 case X86::AX: DestReg = X86::AL; break;
9639 case X86::DX: DestReg = X86::DL; break;
9640 case X86::CX: DestReg = X86::CL; break;
9641 case X86::BX: DestReg = X86::BL; break;
9644 Res.first = DestReg;
9645 Res.second = X86::GR8RegisterClass;
9647 } else if (VT == MVT::i32) {
9648 unsigned DestReg = 0;
9649 switch (Res.first) {
9651 case X86::AX: DestReg = X86::EAX; break;
9652 case X86::DX: DestReg = X86::EDX; break;
9653 case X86::CX: DestReg = X86::ECX; break;
9654 case X86::BX: DestReg = X86::EBX; break;
9655 case X86::SI: DestReg = X86::ESI; break;
9656 case X86::DI: DestReg = X86::EDI; break;
9657 case X86::BP: DestReg = X86::EBP; break;
9658 case X86::SP: DestReg = X86::ESP; break;
9661 Res.first = DestReg;
9662 Res.second = X86::GR32RegisterClass;
9664 } else if (VT == MVT::i64) {
9665 unsigned DestReg = 0;
9666 switch (Res.first) {
9668 case X86::AX: DestReg = X86::RAX; break;
9669 case X86::DX: DestReg = X86::RDX; break;
9670 case X86::CX: DestReg = X86::RCX; break;
9671 case X86::BX: DestReg = X86::RBX; break;
9672 case X86::SI: DestReg = X86::RSI; break;
9673 case X86::DI: DestReg = X86::RDI; break;
9674 case X86::BP: DestReg = X86::RBP; break;
9675 case X86::SP: DestReg = X86::RSP; break;
9678 Res.first = DestReg;
9679 Res.second = X86::GR64RegisterClass;
9682 } else if (Res.second == X86::FR32RegisterClass ||
9683 Res.second == X86::FR64RegisterClass ||
9684 Res.second == X86::VR128RegisterClass) {
9685 // Handle references to XMM physical registers that got mapped into the
9686 // wrong class. This can happen with constraints like {xmm0} where the
9687 // target independent register mapper will just pick the first match it can
9688 // find, ignoring the required type.
9690 Res.second = X86::FR32RegisterClass;
9691 else if (VT == MVT::f64)
9692 Res.second = X86::FR64RegisterClass;
9693 else if (X86::VR128RegisterClass->hasType(VT))
9694 Res.second = X86::VR128RegisterClass;
9700 //===----------------------------------------------------------------------===//
9701 // X86 Widen vector type
9702 //===----------------------------------------------------------------------===//
9704 /// getWidenVectorType: given a vector type, returns the type to widen
9705 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9706 /// If there is no vector type that we want to widen to, returns MVT::Other
9707 /// When and where to widen is target dependent based on the cost of
9708 /// scalarizing vs using the wider vector type.
9710 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
9711 assert(VT.isVector());
9712 if (isTypeLegal(VT))
9715 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9716 // type based on element type. This would speed up our search (though
9717 // it may not be worth it since the size of the list is relatively
9719 EVT EltVT = VT.getVectorElementType();
9720 unsigned NElts = VT.getVectorNumElements();
9722 // On X86, it make sense to widen any vector wider than 1
9726 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9727 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9728 EVT SVT = (MVT::SimpleValueType)nVT;
9730 if (isTypeLegal(SVT) &&
9731 SVT.getVectorElementType() == EltVT &&
9732 SVT.getVectorNumElements() > NElts)