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 if (Subtarget->is64Bit()) {
332 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
333 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
334 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
335 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
337 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
338 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
339 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
340 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
341 if (Subtarget->is64Bit()) {
342 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
343 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
344 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
347 if (Subtarget->hasSSE1())
348 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
350 if (!Subtarget->hasSSE2())
351 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
353 // Expand certain atomics
354 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
355 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
356 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
357 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
359 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
360 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
361 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
362 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
364 if (!Subtarget->is64Bit()) {
365 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
366 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
367 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
368 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
369 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
370 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
371 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
374 // Use the default ISD::DBG_STOPPOINT.
375 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
376 // FIXME - use subtarget debug flags
377 if (!Subtarget->isTargetDarwin() &&
378 !Subtarget->isTargetELF() &&
379 !Subtarget->isTargetCygMing()) {
380 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
381 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
384 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
385 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
386 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
387 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
388 if (Subtarget->is64Bit()) {
389 setExceptionPointerRegister(X86::RAX);
390 setExceptionSelectorRegister(X86::RDX);
392 setExceptionPointerRegister(X86::EAX);
393 setExceptionSelectorRegister(X86::EDX);
395 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
396 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
398 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
400 setOperationAction(ISD::TRAP, MVT::Other, Legal);
402 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
403 setOperationAction(ISD::VASTART , MVT::Other, Custom);
404 setOperationAction(ISD::VAEND , MVT::Other, Expand);
405 if (Subtarget->is64Bit()) {
406 setOperationAction(ISD::VAARG , MVT::Other, Custom);
407 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
409 setOperationAction(ISD::VAARG , MVT::Other, Expand);
410 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
413 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
414 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
415 if (Subtarget->is64Bit())
416 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
417 if (Subtarget->isTargetCygMing())
418 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
420 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
422 if (!UseSoftFloat && X86ScalarSSEf64) {
423 // f32 and f64 use SSE.
424 // Set up the FP register classes.
425 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
426 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
428 // Use ANDPD to simulate FABS.
429 setOperationAction(ISD::FABS , MVT::f64, Custom);
430 setOperationAction(ISD::FABS , MVT::f32, Custom);
432 // Use XORP to simulate FNEG.
433 setOperationAction(ISD::FNEG , MVT::f64, Custom);
434 setOperationAction(ISD::FNEG , MVT::f32, Custom);
436 // Use ANDPD and ORPD to simulate FCOPYSIGN.
437 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
438 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
440 // We don't support sin/cos/fmod
441 setOperationAction(ISD::FSIN , MVT::f64, Expand);
442 setOperationAction(ISD::FCOS , MVT::f64, Expand);
443 setOperationAction(ISD::FSIN , MVT::f32, Expand);
444 setOperationAction(ISD::FCOS , MVT::f32, Expand);
446 // Expand FP immediates into loads from the stack, except for the special
448 addLegalFPImmediate(APFloat(+0.0)); // xorpd
449 addLegalFPImmediate(APFloat(+0.0f)); // xorps
450 } else if (!UseSoftFloat && X86ScalarSSEf32) {
451 // Use SSE for f32, x87 for f64.
452 // Set up the FP register classes.
453 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
454 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
456 // Use ANDPS to simulate FABS.
457 setOperationAction(ISD::FABS , MVT::f32, Custom);
459 // Use XORP to simulate FNEG.
460 setOperationAction(ISD::FNEG , MVT::f32, Custom);
462 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
464 // Use ANDPS and ORPS to simulate FCOPYSIGN.
465 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
466 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
468 // We don't support sin/cos/fmod
469 setOperationAction(ISD::FSIN , MVT::f32, Expand);
470 setOperationAction(ISD::FCOS , MVT::f32, Expand);
472 // Special cases we handle for FP constants.
473 addLegalFPImmediate(APFloat(+0.0f)); // xorps
474 addLegalFPImmediate(APFloat(+0.0)); // FLD0
475 addLegalFPImmediate(APFloat(+1.0)); // FLD1
476 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
477 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
480 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
481 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
483 } else if (!UseSoftFloat) {
484 // f32 and f64 in x87.
485 // Set up the FP register classes.
486 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
487 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
489 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
490 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
491 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
492 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
495 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
496 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
498 addLegalFPImmediate(APFloat(+0.0)); // FLD0
499 addLegalFPImmediate(APFloat(+1.0)); // FLD1
500 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
501 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
502 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
503 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
504 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
505 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
508 // Long double always uses X87.
510 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
511 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
512 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
515 APFloat TmpFlt(+0.0);
516 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
518 addLegalFPImmediate(TmpFlt); // FLD0
520 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
521 APFloat TmpFlt2(+1.0);
522 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
524 addLegalFPImmediate(TmpFlt2); // FLD1
525 TmpFlt2.changeSign();
526 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
530 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
531 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
535 // Always use a library call for pow.
536 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
537 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
538 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
540 setOperationAction(ISD::FLOG, MVT::f80, Expand);
541 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
542 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
543 setOperationAction(ISD::FEXP, MVT::f80, Expand);
544 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
546 // First set operation action for all vector types to either promote
547 // (for widening) or expand (for scalarization). Then we will selectively
548 // turn on ones that can be effectively codegen'd.
549 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
550 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
551 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
566 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
567 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
575 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
576 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
577 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
578 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
579 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
580 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
581 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
582 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
583 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
584 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
585 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
586 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
587 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
588 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
589 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
590 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
591 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
592 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
593 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
594 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
595 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
596 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
597 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
598 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
601 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
602 // with -msoft-float, disable use of MMX as well.
603 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
604 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
605 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
606 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
607 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
608 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
610 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
611 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
612 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
613 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
615 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
616 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
617 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
618 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
620 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
621 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
623 setOperationAction(ISD::AND, MVT::v8i8, Promote);
624 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
625 setOperationAction(ISD::AND, MVT::v4i16, Promote);
626 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
627 setOperationAction(ISD::AND, MVT::v2i32, Promote);
628 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
629 setOperationAction(ISD::AND, MVT::v1i64, Legal);
631 setOperationAction(ISD::OR, MVT::v8i8, Promote);
632 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
633 setOperationAction(ISD::OR, MVT::v4i16, Promote);
634 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
635 setOperationAction(ISD::OR, MVT::v2i32, Promote);
636 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
637 setOperationAction(ISD::OR, MVT::v1i64, Legal);
639 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
640 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
641 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
642 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
643 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
644 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
645 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
647 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
648 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
649 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
650 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
651 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
652 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
653 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
654 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
655 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
657 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
658 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
659 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
660 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
661 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
663 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
664 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
665 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
666 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
668 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
669 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
670 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
671 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
673 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
675 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
676 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
677 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
678 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
679 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
680 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
681 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
682 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
683 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
686 if (!UseSoftFloat && Subtarget->hasSSE1()) {
687 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
689 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
690 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
691 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
692 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
693 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
694 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
695 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
696 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
697 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
698 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
699 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
700 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
703 if (!UseSoftFloat && Subtarget->hasSSE2()) {
704 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
706 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
707 // registers cannot be used even for integer operations.
708 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
709 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
710 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
711 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
713 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
714 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
715 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
716 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
717 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
718 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
719 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
720 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
721 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
722 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
723 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
724 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
725 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
726 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
727 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
728 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
730 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
731 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
732 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
733 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
735 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
736 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
737 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
738 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
739 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
741 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
742 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
743 EVT VT = (MVT::SimpleValueType)i;
744 // Do not attempt to custom lower non-power-of-2 vectors
745 if (!isPowerOf2_32(VT.getVectorNumElements()))
747 // Do not attempt to custom lower non-128-bit vectors
748 if (!VT.is128BitVector())
750 setOperationAction(ISD::BUILD_VECTOR,
751 VT.getSimpleVT().SimpleTy, Custom);
752 setOperationAction(ISD::VECTOR_SHUFFLE,
753 VT.getSimpleVT().SimpleTy, Custom);
754 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
755 VT.getSimpleVT().SimpleTy, Custom);
758 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
759 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
760 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
761 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
762 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
763 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
765 if (Subtarget->is64Bit()) {
766 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
767 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
770 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
771 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
772 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
775 // Do not attempt to promote non-128-bit vectors
776 if (!VT.is128BitVector()) {
779 setOperationAction(ISD::AND, SVT, Promote);
780 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
781 setOperationAction(ISD::OR, SVT, Promote);
782 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
783 setOperationAction(ISD::XOR, SVT, Promote);
784 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
785 setOperationAction(ISD::LOAD, SVT, Promote);
786 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
787 setOperationAction(ISD::SELECT, SVT, Promote);
788 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
791 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
793 // Custom lower v2i64 and v2f64 selects.
794 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
795 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
796 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
797 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
799 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
800 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
801 if (!DisableMMX && Subtarget->hasMMX()) {
802 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
803 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
807 if (Subtarget->hasSSE41()) {
808 // FIXME: Do we need to handle scalar-to-vector here?
809 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
811 // i8 and i16 vectors are custom , because the source register and source
812 // source memory operand types are not the same width. f32 vectors are
813 // custom since the immediate controlling the insert encodes additional
815 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
816 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
817 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
818 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
820 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
821 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
822 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
823 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
825 if (Subtarget->is64Bit()) {
826 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
827 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
831 if (Subtarget->hasSSE42()) {
832 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
835 if (!UseSoftFloat && Subtarget->hasAVX()) {
836 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
837 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
838 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
839 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
841 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
842 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
843 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
844 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
845 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
846 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
847 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
848 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
849 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
850 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
851 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
852 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
853 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
854 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
855 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
857 // Operations to consider commented out -v16i16 v32i8
858 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
859 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
860 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
861 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
862 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
863 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
864 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
865 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
866 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
867 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
868 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
869 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
870 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
871 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
873 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
874 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
875 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
876 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
878 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
879 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
880 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
881 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
882 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
884 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
885 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
886 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
887 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
888 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
889 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
892 // Not sure we want to do this since there are no 256-bit integer
895 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
896 // This includes 256-bit vectors
897 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
898 EVT VT = (MVT::SimpleValueType)i;
900 // Do not attempt to custom lower non-power-of-2 vectors
901 if (!isPowerOf2_32(VT.getVectorNumElements()))
904 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
905 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
906 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
909 if (Subtarget->is64Bit()) {
910 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
911 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
916 // Not sure we want to do this since there are no 256-bit integer
919 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
920 // Including 256-bit vectors
921 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
922 EVT VT = (MVT::SimpleValueType)i;
924 if (!VT.is256BitVector()) {
927 setOperationAction(ISD::AND, VT, Promote);
928 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
929 setOperationAction(ISD::OR, VT, Promote);
930 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
931 setOperationAction(ISD::XOR, VT, Promote);
932 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
933 setOperationAction(ISD::LOAD, VT, Promote);
934 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
935 setOperationAction(ISD::SELECT, VT, Promote);
936 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
939 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
943 // We want to custom lower some of our intrinsics.
944 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
946 // Add/Sub/Mul with overflow operations are custom lowered.
947 setOperationAction(ISD::SADDO, MVT::i32, Custom);
948 setOperationAction(ISD::SADDO, MVT::i64, Custom);
949 setOperationAction(ISD::UADDO, MVT::i32, Custom);
950 setOperationAction(ISD::UADDO, MVT::i64, Custom);
951 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
952 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
953 setOperationAction(ISD::USUBO, MVT::i32, Custom);
954 setOperationAction(ISD::USUBO, MVT::i64, Custom);
955 setOperationAction(ISD::SMULO, MVT::i32, Custom);
956 setOperationAction(ISD::SMULO, MVT::i64, Custom);
958 if (!Subtarget->is64Bit()) {
959 // These libcalls are not available in 32-bit.
960 setLibcallName(RTLIB::SHL_I128, 0);
961 setLibcallName(RTLIB::SRL_I128, 0);
962 setLibcallName(RTLIB::SRA_I128, 0);
965 // We have target-specific dag combine patterns for the following nodes:
966 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
967 setTargetDAGCombine(ISD::BUILD_VECTOR);
968 setTargetDAGCombine(ISD::SELECT);
969 setTargetDAGCombine(ISD::SHL);
970 setTargetDAGCombine(ISD::SRA);
971 setTargetDAGCombine(ISD::SRL);
972 setTargetDAGCombine(ISD::STORE);
973 setTargetDAGCombine(ISD::MEMBARRIER);
974 if (Subtarget->is64Bit())
975 setTargetDAGCombine(ISD::MUL);
977 computeRegisterProperties();
979 // FIXME: These should be based on subtarget info. Plus, the values should
980 // be smaller when we are in optimizing for size mode.
981 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
982 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
983 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
984 setPrefLoopAlignment(16);
985 benefitFromCodePlacementOpt = true;
989 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
994 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
995 /// the desired ByVal argument alignment.
996 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
999 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1000 if (VTy->getBitWidth() == 128)
1002 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1003 unsigned EltAlign = 0;
1004 getMaxByValAlign(ATy->getElementType(), EltAlign);
1005 if (EltAlign > MaxAlign)
1006 MaxAlign = EltAlign;
1007 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1008 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
1009 unsigned EltAlign = 0;
1010 getMaxByValAlign(STy->getElementType(i), EltAlign);
1011 if (EltAlign > MaxAlign)
1012 MaxAlign = EltAlign;
1020 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
1021 /// function arguments in the caller parameter area. For X86, aggregates
1022 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1023 /// are at 4-byte boundaries.
1024 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1025 if (Subtarget->is64Bit()) {
1026 // Max of 8 and alignment of type.
1027 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1034 if (Subtarget->hasSSE1())
1035 getMaxByValAlign(Ty, Align);
1039 /// getOptimalMemOpType - Returns the target specific optimal type for load
1040 /// and store operations as a result of memset, memcpy, and memmove
1041 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1044 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1045 bool isSrcConst, bool isSrcStr,
1046 SelectionDAG &DAG) const {
1047 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1048 // linux. This is because the stack realignment code can't handle certain
1049 // cases like PR2962. This should be removed when PR2962 is fixed.
1050 const Function *F = DAG.getMachineFunction().getFunction();
1051 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1052 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1053 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1055 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1058 if (Subtarget->is64Bit() && Size >= 8)
1063 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1065 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1066 SelectionDAG &DAG) const {
1067 if (usesGlobalOffsetTable())
1068 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1069 if (!Subtarget->is64Bit())
1070 // This doesn't have DebugLoc associated with it, but is not really the
1071 // same as a Register.
1072 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1077 /// getFunctionAlignment - Return the Log2 alignment of this function.
1078 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1079 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1082 //===----------------------------------------------------------------------===//
1083 // Return Value Calling Convention Implementation
1084 //===----------------------------------------------------------------------===//
1086 #include "X86GenCallingConv.inc"
1089 X86TargetLowering::LowerReturn(SDValue Chain,
1090 CallingConv::ID CallConv, bool isVarArg,
1091 const SmallVectorImpl<ISD::OutputArg> &Outs,
1092 DebugLoc dl, SelectionDAG &DAG) {
1094 SmallVector<CCValAssign, 16> RVLocs;
1095 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1096 RVLocs, *DAG.getContext());
1097 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1099 // If this is the first return lowered for this function, add the regs to the
1100 // liveout set for the function.
1101 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1102 for (unsigned i = 0; i != RVLocs.size(); ++i)
1103 if (RVLocs[i].isRegLoc())
1104 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1109 SmallVector<SDValue, 6> RetOps;
1110 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1111 // Operand #1 = Bytes To Pop
1112 RetOps.push_back(DAG.getTargetConstant(getBytesToPopOnReturn(), MVT::i16));
1114 // Copy the result values into the output registers.
1115 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1116 CCValAssign &VA = RVLocs[i];
1117 assert(VA.isRegLoc() && "Can only return in registers!");
1118 SDValue ValToCopy = Outs[i].Val;
1120 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1121 // the RET instruction and handled by the FP Stackifier.
1122 if (VA.getLocReg() == X86::ST0 ||
1123 VA.getLocReg() == X86::ST1) {
1124 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1125 // change the value to the FP stack register class.
1126 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1127 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1128 RetOps.push_back(ValToCopy);
1129 // Don't emit a copytoreg.
1133 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1134 // which is returned in RAX / RDX.
1135 if (Subtarget->is64Bit()) {
1136 EVT ValVT = ValToCopy.getValueType();
1137 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1138 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1139 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1140 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1144 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1145 Flag = Chain.getValue(1);
1148 // The x86-64 ABI for returning structs by value requires that we copy
1149 // the sret argument into %rax for the return. We saved the argument into
1150 // a virtual register in the entry block, so now we copy the value out
1152 if (Subtarget->is64Bit() &&
1153 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1154 MachineFunction &MF = DAG.getMachineFunction();
1155 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1156 unsigned Reg = FuncInfo->getSRetReturnReg();
1158 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1159 FuncInfo->setSRetReturnReg(Reg);
1161 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1163 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1164 Flag = Chain.getValue(1);
1166 // RAX now acts like a return value.
1167 MF.getRegInfo().addLiveOut(X86::RAX);
1170 RetOps[0] = Chain; // Update chain.
1172 // Add the flag if we have it.
1174 RetOps.push_back(Flag);
1176 return DAG.getNode(X86ISD::RET_FLAG, dl,
1177 MVT::Other, &RetOps[0], RetOps.size());
1180 /// LowerCallResult - Lower the result values of a call into the
1181 /// appropriate copies out of appropriate physical registers.
1184 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1185 CallingConv::ID CallConv, bool isVarArg,
1186 const SmallVectorImpl<ISD::InputArg> &Ins,
1187 DebugLoc dl, SelectionDAG &DAG,
1188 SmallVectorImpl<SDValue> &InVals) {
1190 // Assign locations to each value returned by this call.
1191 SmallVector<CCValAssign, 16> RVLocs;
1192 bool Is64Bit = Subtarget->is64Bit();
1193 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1194 RVLocs, *DAG.getContext());
1195 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1197 // Copy all of the result registers out of their specified physreg.
1198 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1199 CCValAssign &VA = RVLocs[i];
1200 EVT CopyVT = VA.getValVT();
1202 // If this is x86-64, and we disabled SSE, we can't return FP values
1203 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1204 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1205 llvm_report_error("SSE register return with SSE disabled");
1208 // If this is a call to a function that returns an fp value on the floating
1209 // point stack, but where we prefer to use the value in xmm registers, copy
1210 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1211 if ((VA.getLocReg() == X86::ST0 ||
1212 VA.getLocReg() == X86::ST1) &&
1213 isScalarFPTypeInSSEReg(VA.getValVT())) {
1218 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1219 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1220 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1221 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1222 MVT::v2i64, InFlag).getValue(1);
1223 Val = Chain.getValue(0);
1224 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1225 Val, DAG.getConstant(0, MVT::i64));
1227 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1228 MVT::i64, InFlag).getValue(1);
1229 Val = Chain.getValue(0);
1231 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1233 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1234 CopyVT, InFlag).getValue(1);
1235 Val = Chain.getValue(0);
1237 InFlag = Chain.getValue(2);
1239 if (CopyVT != VA.getValVT()) {
1240 // Round the F80 the right size, which also moves to the appropriate xmm
1242 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1243 // This truncation won't change the value.
1244 DAG.getIntPtrConstant(1));
1247 InVals.push_back(Val);
1254 //===----------------------------------------------------------------------===//
1255 // C & StdCall & Fast Calling Convention implementation
1256 //===----------------------------------------------------------------------===//
1257 // StdCall calling convention seems to be standard for many Windows' API
1258 // routines and around. It differs from C calling convention just a little:
1259 // callee should clean up the stack, not caller. Symbols should be also
1260 // decorated in some fancy way :) It doesn't support any vector arguments.
1261 // For info on fast calling convention see Fast Calling Convention (tail call)
1262 // implementation LowerX86_32FastCCCallTo.
1264 /// CallIsStructReturn - Determines whether a call uses struct return
1266 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1270 return Outs[0].Flags.isSRet();
1273 /// ArgsAreStructReturn - Determines whether a function uses struct
1274 /// return semantics.
1276 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1280 return Ins[0].Flags.isSRet();
1283 /// IsCalleePop - Determines whether the callee is required to pop its
1284 /// own arguments. Callee pop is necessary to support tail calls.
1285 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1289 switch (CallingConv) {
1292 case CallingConv::X86_StdCall:
1293 return !Subtarget->is64Bit();
1294 case CallingConv::X86_FastCall:
1295 return !Subtarget->is64Bit();
1296 case CallingConv::Fast:
1297 return PerformTailCallOpt;
1301 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1302 /// given CallingConvention value.
1303 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1304 if (Subtarget->is64Bit()) {
1305 if (Subtarget->isTargetWin64())
1306 return CC_X86_Win64_C;
1311 if (CC == CallingConv::X86_FastCall)
1312 return CC_X86_32_FastCall;
1313 else if (CC == CallingConv::Fast)
1314 return CC_X86_32_FastCC;
1319 /// NameDecorationForCallConv - Selects the appropriate decoration to
1320 /// apply to a MachineFunction containing a given calling convention.
1322 X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
1323 if (CallConv == CallingConv::X86_FastCall)
1325 else if (CallConv == CallingConv::X86_StdCall)
1331 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1332 /// by "Src" to address "Dst" with size and alignment information specified by
1333 /// the specific parameter attribute. The copy will be passed as a byval
1334 /// function parameter.
1336 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1337 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1339 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1340 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1341 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1345 X86TargetLowering::LowerMemArgument(SDValue Chain,
1346 CallingConv::ID CallConv,
1347 const SmallVectorImpl<ISD::InputArg> &Ins,
1348 DebugLoc dl, SelectionDAG &DAG,
1349 const CCValAssign &VA,
1350 MachineFrameInfo *MFI,
1353 // Create the nodes corresponding to a load from this parameter slot.
1354 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1355 bool AlwaysUseMutable = (CallConv==CallingConv::Fast) && PerformTailCallOpt;
1356 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1359 // If value is passed by pointer we have address passed instead of the value
1361 if (VA.getLocInfo() == CCValAssign::Indirect)
1362 ValVT = VA.getLocVT();
1364 ValVT = VA.getValVT();
1366 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1367 // changed with more analysis.
1368 // In case of tail call optimization mark all arguments mutable. Since they
1369 // could be overwritten by lowering of arguments in case of a tail call.
1370 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1371 VA.getLocMemOffset(), isImmutable);
1372 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1373 if (Flags.isByVal())
1375 return DAG.getLoad(ValVT, dl, Chain, FIN,
1376 PseudoSourceValue::getFixedStack(FI), 0);
1380 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1381 CallingConv::ID CallConv,
1383 const SmallVectorImpl<ISD::InputArg> &Ins,
1386 SmallVectorImpl<SDValue> &InVals) {
1388 MachineFunction &MF = DAG.getMachineFunction();
1389 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1391 const Function* Fn = MF.getFunction();
1392 if (Fn->hasExternalLinkage() &&
1393 Subtarget->isTargetCygMing() &&
1394 Fn->getName() == "main")
1395 FuncInfo->setForceFramePointer(true);
1397 // Decorate the function name.
1398 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1400 MachineFrameInfo *MFI = MF.getFrameInfo();
1401 bool Is64Bit = Subtarget->is64Bit();
1402 bool IsWin64 = Subtarget->isTargetWin64();
1404 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1405 "Var args not supported with calling convention fastcc");
1407 // Assign locations to all of the incoming arguments.
1408 SmallVector<CCValAssign, 16> ArgLocs;
1409 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1410 ArgLocs, *DAG.getContext());
1411 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1413 unsigned LastVal = ~0U;
1415 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1416 CCValAssign &VA = ArgLocs[i];
1417 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1419 assert(VA.getValNo() != LastVal &&
1420 "Don't support value assigned to multiple locs yet");
1421 LastVal = VA.getValNo();
1423 if (VA.isRegLoc()) {
1424 EVT RegVT = VA.getLocVT();
1425 TargetRegisterClass *RC = NULL;
1426 if (RegVT == MVT::i32)
1427 RC = X86::GR32RegisterClass;
1428 else if (Is64Bit && RegVT == MVT::i64)
1429 RC = X86::GR64RegisterClass;
1430 else if (RegVT == MVT::f32)
1431 RC = X86::FR32RegisterClass;
1432 else if (RegVT == MVT::f64)
1433 RC = X86::FR64RegisterClass;
1434 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1435 RC = X86::VR128RegisterClass;
1436 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1437 RC = X86::VR64RegisterClass;
1439 llvm_unreachable("Unknown argument type!");
1441 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1442 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1444 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1445 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1447 if (VA.getLocInfo() == CCValAssign::SExt)
1448 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1449 DAG.getValueType(VA.getValVT()));
1450 else if (VA.getLocInfo() == CCValAssign::ZExt)
1451 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1452 DAG.getValueType(VA.getValVT()));
1453 else if (VA.getLocInfo() == CCValAssign::BCvt)
1454 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1456 if (VA.isExtInLoc()) {
1457 // Handle MMX values passed in XMM regs.
1458 if (RegVT.isVector()) {
1459 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1460 ArgValue, DAG.getConstant(0, MVT::i64));
1461 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1463 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1466 assert(VA.isMemLoc());
1467 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1470 // If value is passed via pointer - do a load.
1471 if (VA.getLocInfo() == CCValAssign::Indirect)
1472 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1474 InVals.push_back(ArgValue);
1477 // The x86-64 ABI for returning structs by value requires that we copy
1478 // the sret argument into %rax for the return. Save the argument into
1479 // a virtual register so that we can access it from the return points.
1480 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1481 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1482 unsigned Reg = FuncInfo->getSRetReturnReg();
1484 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1485 FuncInfo->setSRetReturnReg(Reg);
1487 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1488 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1491 unsigned StackSize = CCInfo.getNextStackOffset();
1492 // align stack specially for tail calls
1493 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1494 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1496 // If the function takes variable number of arguments, make a frame index for
1497 // the start of the first vararg value... for expansion of llvm.va_start.
1499 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1500 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1503 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1505 // FIXME: We should really autogenerate these arrays
1506 static const unsigned GPR64ArgRegsWin64[] = {
1507 X86::RCX, X86::RDX, X86::R8, X86::R9
1509 static const unsigned XMMArgRegsWin64[] = {
1510 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1512 static const unsigned GPR64ArgRegs64Bit[] = {
1513 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1515 static const unsigned XMMArgRegs64Bit[] = {
1516 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1517 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1519 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1522 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1523 GPR64ArgRegs = GPR64ArgRegsWin64;
1524 XMMArgRegs = XMMArgRegsWin64;
1526 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1527 GPR64ArgRegs = GPR64ArgRegs64Bit;
1528 XMMArgRegs = XMMArgRegs64Bit;
1530 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1532 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1535 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1536 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1537 "SSE register cannot be used when SSE is disabled!");
1538 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1539 "SSE register cannot be used when SSE is disabled!");
1540 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1541 // Kernel mode asks for SSE to be disabled, so don't push them
1543 TotalNumXMMRegs = 0;
1545 // For X86-64, if there are vararg parameters that are passed via
1546 // registers, then we must store them to their spots on the stack so they
1547 // may be loaded by deferencing the result of va_next.
1548 VarArgsGPOffset = NumIntRegs * 8;
1549 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1550 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1551 TotalNumXMMRegs * 16, 16);
1553 // Store the integer parameter registers.
1554 SmallVector<SDValue, 8> MemOps;
1555 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1556 unsigned Offset = VarArgsGPOffset;
1557 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1558 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1559 DAG.getIntPtrConstant(Offset));
1560 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1561 X86::GR64RegisterClass);
1562 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1564 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1565 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1567 MemOps.push_back(Store);
1571 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1572 // Now store the XMM (fp + vector) parameter registers.
1573 SmallVector<SDValue, 11> SaveXMMOps;
1574 SaveXMMOps.push_back(Chain);
1576 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1577 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1578 SaveXMMOps.push_back(ALVal);
1580 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1581 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1583 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1584 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1585 X86::VR128RegisterClass);
1586 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1587 SaveXMMOps.push_back(Val);
1589 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1591 &SaveXMMOps[0], SaveXMMOps.size()));
1594 if (!MemOps.empty())
1595 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1596 &MemOps[0], MemOps.size());
1600 // Some CCs need callee pop.
1601 if (IsCalleePop(isVarArg, CallConv)) {
1602 BytesToPopOnReturn = StackSize; // Callee pops everything.
1603 BytesCallerReserves = 0;
1605 BytesToPopOnReturn = 0; // Callee pops nothing.
1606 // If this is an sret function, the return should pop the hidden pointer.
1607 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1608 BytesToPopOnReturn = 4;
1609 BytesCallerReserves = StackSize;
1613 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1614 if (CallConv == CallingConv::X86_FastCall)
1615 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1618 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1624 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1625 SDValue StackPtr, SDValue Arg,
1626 DebugLoc dl, SelectionDAG &DAG,
1627 const CCValAssign &VA,
1628 ISD::ArgFlagsTy Flags) {
1629 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1630 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1631 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1632 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1633 if (Flags.isByVal()) {
1634 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1636 return DAG.getStore(Chain, dl, Arg, PtrOff,
1637 PseudoSourceValue::getStack(), LocMemOffset);
1640 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1641 /// optimization is performed and it is required.
1643 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1644 SDValue &OutRetAddr,
1650 if (!IsTailCall || FPDiff==0) return Chain;
1652 // Adjust the Return address stack slot.
1653 EVT VT = getPointerTy();
1654 OutRetAddr = getReturnAddressFrameIndex(DAG);
1656 // Load the "old" Return address.
1657 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1658 return SDValue(OutRetAddr.getNode(), 1);
1661 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1662 /// optimization is performed and it is required (FPDiff!=0).
1664 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1665 SDValue Chain, SDValue RetAddrFrIdx,
1666 bool Is64Bit, int FPDiff, DebugLoc dl) {
1667 // Store the return address to the appropriate stack slot.
1668 if (!FPDiff) return Chain;
1669 // Calculate the new stack slot for the return address.
1670 int SlotSize = Is64Bit ? 8 : 4;
1671 int NewReturnAddrFI =
1672 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1673 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1674 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1675 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1676 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1681 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1682 CallingConv::ID CallConv, bool isVarArg,
1684 const SmallVectorImpl<ISD::OutputArg> &Outs,
1685 const SmallVectorImpl<ISD::InputArg> &Ins,
1686 DebugLoc dl, SelectionDAG &DAG,
1687 SmallVectorImpl<SDValue> &InVals) {
1689 MachineFunction &MF = DAG.getMachineFunction();
1690 bool Is64Bit = Subtarget->is64Bit();
1691 bool IsStructRet = CallIsStructReturn(Outs);
1693 assert((!isTailCall ||
1694 (CallConv == CallingConv::Fast && PerformTailCallOpt)) &&
1695 "IsEligibleForTailCallOptimization missed a case!");
1696 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1697 "Var args not supported with calling convention fastcc");
1699 // Analyze operands of the call, assigning locations to each operand.
1700 SmallVector<CCValAssign, 16> ArgLocs;
1701 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1702 ArgLocs, *DAG.getContext());
1703 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1705 // Get a count of how many bytes are to be pushed on the stack.
1706 unsigned NumBytes = CCInfo.getNextStackOffset();
1707 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1708 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1712 // Lower arguments at fp - stackoffset + fpdiff.
1713 unsigned NumBytesCallerPushed =
1714 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1715 FPDiff = NumBytesCallerPushed - NumBytes;
1717 // Set the delta of movement of the returnaddr stackslot.
1718 // But only set if delta is greater than previous delta.
1719 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1720 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1723 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1725 SDValue RetAddrFrIdx;
1726 // Load return adress for tail calls.
1727 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
1730 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1731 SmallVector<SDValue, 8> MemOpChains;
1734 // Walk the register/memloc assignments, inserting copies/loads. In the case
1735 // of tail call optimization arguments are handle later.
1736 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1737 CCValAssign &VA = ArgLocs[i];
1738 EVT RegVT = VA.getLocVT();
1739 SDValue Arg = Outs[i].Val;
1740 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1741 bool isByVal = Flags.isByVal();
1743 // Promote the value if needed.
1744 switch (VA.getLocInfo()) {
1745 default: llvm_unreachable("Unknown loc info!");
1746 case CCValAssign::Full: break;
1747 case CCValAssign::SExt:
1748 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1750 case CCValAssign::ZExt:
1751 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1753 case CCValAssign::AExt:
1754 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1755 // Special case: passing MMX values in XMM registers.
1756 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1757 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1758 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1760 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1762 case CCValAssign::BCvt:
1763 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1765 case CCValAssign::Indirect: {
1766 // Store the argument.
1767 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1768 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1769 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1770 PseudoSourceValue::getFixedStack(FI), 0);
1776 if (VA.isRegLoc()) {
1777 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1779 if (!isTailCall || (isTailCall && isByVal)) {
1780 assert(VA.isMemLoc());
1781 if (StackPtr.getNode() == 0)
1782 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1784 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1785 dl, DAG, VA, Flags));
1790 if (!MemOpChains.empty())
1791 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1792 &MemOpChains[0], MemOpChains.size());
1794 // Build a sequence of copy-to-reg nodes chained together with token chain
1795 // and flag operands which copy the outgoing args into registers.
1797 // Tail call byval lowering might overwrite argument registers so in case of
1798 // tail call optimization the copies to registers are lowered later.
1800 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1801 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1802 RegsToPass[i].second, InFlag);
1803 InFlag = Chain.getValue(1);
1807 if (Subtarget->isPICStyleGOT()) {
1808 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1811 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1812 DAG.getNode(X86ISD::GlobalBaseReg,
1813 DebugLoc::getUnknownLoc(),
1816 InFlag = Chain.getValue(1);
1818 // If we are tail calling and generating PIC/GOT style code load the
1819 // address of the callee into ECX. The value in ecx is used as target of
1820 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1821 // for tail calls on PIC/GOT architectures. Normally we would just put the
1822 // address of GOT into ebx and then call target@PLT. But for tail calls
1823 // ebx would be restored (since ebx is callee saved) before jumping to the
1826 // Note: The actual moving to ECX is done further down.
1827 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1828 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1829 !G->getGlobal()->hasProtectedVisibility())
1830 Callee = LowerGlobalAddress(Callee, DAG);
1831 else if (isa<ExternalSymbolSDNode>(Callee))
1832 Callee = LowerExternalSymbol(Callee, DAG);
1836 if (Is64Bit && isVarArg) {
1837 // From AMD64 ABI document:
1838 // For calls that may call functions that use varargs or stdargs
1839 // (prototype-less calls or calls to functions containing ellipsis (...) in
1840 // the declaration) %al is used as hidden argument to specify the number
1841 // of SSE registers used. The contents of %al do not need to match exactly
1842 // the number of registers, but must be an ubound on the number of SSE
1843 // registers used and is in the range 0 - 8 inclusive.
1845 // FIXME: Verify this on Win64
1846 // Count the number of XMM registers allocated.
1847 static const unsigned XMMArgRegs[] = {
1848 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1849 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1851 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1852 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1853 && "SSE registers cannot be used when SSE is disabled");
1855 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1856 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1857 InFlag = Chain.getValue(1);
1861 // For tail calls lower the arguments to the 'real' stack slot.
1863 // Force all the incoming stack arguments to be loaded from the stack
1864 // before any new outgoing arguments are stored to the stack, because the
1865 // outgoing stack slots may alias the incoming argument stack slots, and
1866 // the alias isn't otherwise explicit. This is slightly more conservative
1867 // than necessary, because it means that each store effectively depends
1868 // on every argument instead of just those arguments it would clobber.
1869 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1871 SmallVector<SDValue, 8> MemOpChains2;
1874 // Do not flag preceeding copytoreg stuff together with the following stuff.
1876 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1877 CCValAssign &VA = ArgLocs[i];
1878 if (!VA.isRegLoc()) {
1879 assert(VA.isMemLoc());
1880 SDValue Arg = Outs[i].Val;
1881 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1882 // Create frame index.
1883 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1884 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1885 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1886 FIN = DAG.getFrameIndex(FI, getPointerTy());
1888 if (Flags.isByVal()) {
1889 // Copy relative to framepointer.
1890 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1891 if (StackPtr.getNode() == 0)
1892 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1894 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1896 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1900 // Store relative to framepointer.
1901 MemOpChains2.push_back(
1902 DAG.getStore(ArgChain, dl, Arg, FIN,
1903 PseudoSourceValue::getFixedStack(FI), 0));
1908 if (!MemOpChains2.empty())
1909 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1910 &MemOpChains2[0], MemOpChains2.size());
1912 // Copy arguments to their registers.
1913 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1914 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1915 RegsToPass[i].second, InFlag);
1916 InFlag = Chain.getValue(1);
1920 // Store the return address to the appropriate stack slot.
1921 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1925 // If the callee is a GlobalAddress node (quite common, every direct call is)
1926 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1927 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1928 // We should use extra load for direct calls to dllimported functions in
1930 GlobalValue *GV = G->getGlobal();
1931 if (!GV->hasDLLImportLinkage()) {
1932 unsigned char OpFlags = 0;
1934 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1935 // external symbols most go through the PLT in PIC mode. If the symbol
1936 // has hidden or protected visibility, or if it is static or local, then
1937 // we don't need to use the PLT - we can directly call it.
1938 if (Subtarget->isTargetELF() &&
1939 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1940 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1941 OpFlags = X86II::MO_PLT;
1942 } else if (Subtarget->isPICStyleStubAny() &&
1943 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1944 Subtarget->getDarwinVers() < 9) {
1945 // PC-relative references to external symbols should go through $stub,
1946 // unless we're building with the leopard linker or later, which
1947 // automatically synthesizes these stubs.
1948 OpFlags = X86II::MO_DARWIN_STUB;
1951 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
1952 G->getOffset(), OpFlags);
1954 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1955 unsigned char OpFlags = 0;
1957 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
1958 // symbols should go through the PLT.
1959 if (Subtarget->isTargetELF() &&
1960 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1961 OpFlags = X86II::MO_PLT;
1962 } else if (Subtarget->isPICStyleStubAny() &&
1963 Subtarget->getDarwinVers() < 9) {
1964 // PC-relative references to external symbols should go through $stub,
1965 // unless we're building with the leopard linker or later, which
1966 // automatically synthesizes these stubs.
1967 OpFlags = X86II::MO_DARWIN_STUB;
1970 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
1972 } else if (isTailCall) {
1973 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
1975 Chain = DAG.getCopyToReg(Chain, dl,
1976 DAG.getRegister(Opc, getPointerTy()),
1978 Callee = DAG.getRegister(Opc, getPointerTy());
1979 // Add register as live out.
1980 MF.getRegInfo().addLiveOut(Opc);
1983 // Returns a chain & a flag for retval copy to use.
1984 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1985 SmallVector<SDValue, 8> Ops;
1988 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1989 DAG.getIntPtrConstant(0, true), InFlag);
1990 InFlag = Chain.getValue(1);
1993 Ops.push_back(Chain);
1994 Ops.push_back(Callee);
1997 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1999 // Add argument registers to the end of the list so that they are known live
2001 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
2002 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
2003 RegsToPass[i].second.getValueType()));
2005 // Add an implicit use GOT pointer in EBX.
2006 if (!isTailCall && Subtarget->isPICStyleGOT())
2007 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
2009 // Add an implicit use of AL for x86 vararg functions.
2010 if (Is64Bit && isVarArg)
2011 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
2013 if (InFlag.getNode())
2014 Ops.push_back(InFlag);
2017 // If this is the first return lowered for this function, add the regs
2018 // to the liveout set for the function.
2019 if (MF.getRegInfo().liveout_empty()) {
2020 SmallVector<CCValAssign, 16> RVLocs;
2021 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
2023 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2024 for (unsigned i = 0; i != RVLocs.size(); ++i)
2025 if (RVLocs[i].isRegLoc())
2026 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2029 assert(((Callee.getOpcode() == ISD::Register &&
2030 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2031 cast<RegisterSDNode>(Callee)->getReg() == X86::R9)) ||
2032 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2033 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2034 "Expecting an global address, external symbol, or register");
2036 return DAG.getNode(X86ISD::TC_RETURN, dl,
2037 NodeTys, &Ops[0], Ops.size());
2040 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2041 InFlag = Chain.getValue(1);
2043 // Create the CALLSEQ_END node.
2044 unsigned NumBytesForCalleeToPush;
2045 if (IsCalleePop(isVarArg, CallConv))
2046 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2047 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2048 // If this is is a call to a struct-return function, the callee
2049 // pops the hidden struct pointer, so we have to push it back.
2050 // This is common for Darwin/X86, Linux & Mingw32 targets.
2051 NumBytesForCalleeToPush = 4;
2053 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2055 // Returns a flag for retval copy to use.
2056 Chain = DAG.getCALLSEQ_END(Chain,
2057 DAG.getIntPtrConstant(NumBytes, true),
2058 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2061 InFlag = Chain.getValue(1);
2063 // Handle result values, copying them out of physregs into vregs that we
2065 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2066 Ins, dl, DAG, InVals);
2070 //===----------------------------------------------------------------------===//
2071 // Fast Calling Convention (tail call) implementation
2072 //===----------------------------------------------------------------------===//
2074 // Like std call, callee cleans arguments, convention except that ECX is
2075 // reserved for storing the tail called function address. Only 2 registers are
2076 // free for argument passing (inreg). Tail call optimization is performed
2078 // * tailcallopt is enabled
2079 // * caller/callee are fastcc
2080 // On X86_64 architecture with GOT-style position independent code only local
2081 // (within module) calls are supported at the moment.
2082 // To keep the stack aligned according to platform abi the function
2083 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2084 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2085 // If a tail called function callee has more arguments than the caller the
2086 // caller needs to make sure that there is room to move the RETADDR to. This is
2087 // achieved by reserving an area the size of the argument delta right after the
2088 // original REtADDR, but before the saved framepointer or the spilled registers
2089 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2101 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2102 /// for a 16 byte align requirement.
2103 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2104 SelectionDAG& DAG) {
2105 MachineFunction &MF = DAG.getMachineFunction();
2106 const TargetMachine &TM = MF.getTarget();
2107 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2108 unsigned StackAlignment = TFI.getStackAlignment();
2109 uint64_t AlignMask = StackAlignment - 1;
2110 int64_t Offset = StackSize;
2111 uint64_t SlotSize = TD->getPointerSize();
2112 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2113 // Number smaller than 12 so just add the difference.
2114 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2116 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2117 Offset = ((~AlignMask) & Offset) + StackAlignment +
2118 (StackAlignment-SlotSize);
2123 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2124 /// for tail call optimization. Targets which want to do tail call
2125 /// optimization should implement this function.
2127 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2128 CallingConv::ID CalleeCC,
2130 const SmallVectorImpl<ISD::InputArg> &Ins,
2131 SelectionDAG& DAG) const {
2132 MachineFunction &MF = DAG.getMachineFunction();
2133 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2134 return CalleeCC == CallingConv::Fast && CallerCC == CalleeCC;
2138 X86TargetLowering::createFastISel(MachineFunction &mf,
2139 MachineModuleInfo *mmo,
2141 DenseMap<const Value *, unsigned> &vm,
2142 DenseMap<const BasicBlock *,
2143 MachineBasicBlock *> &bm,
2144 DenseMap<const AllocaInst *, int> &am
2146 , SmallSet<Instruction*, 8> &cil
2149 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2157 //===----------------------------------------------------------------------===//
2158 // Other Lowering Hooks
2159 //===----------------------------------------------------------------------===//
2162 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2163 MachineFunction &MF = DAG.getMachineFunction();
2164 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2165 int ReturnAddrIndex = FuncInfo->getRAIndex();
2167 if (ReturnAddrIndex == 0) {
2168 // Set up a frame object for the return address.
2169 uint64_t SlotSize = TD->getPointerSize();
2170 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2171 FuncInfo->setRAIndex(ReturnAddrIndex);
2174 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2178 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2179 bool hasSymbolicDisplacement) {
2180 // Offset should fit into 32 bit immediate field.
2181 if (!isInt32(Offset))
2184 // If we don't have a symbolic displacement - we don't have any extra
2186 if (!hasSymbolicDisplacement)
2189 // FIXME: Some tweaks might be needed for medium code model.
2190 if (M != CodeModel::Small && M != CodeModel::Kernel)
2193 // For small code model we assume that latest object is 16MB before end of 31
2194 // bits boundary. We may also accept pretty large negative constants knowing
2195 // that all objects are in the positive half of address space.
2196 if (M == CodeModel::Small && Offset < 16*1024*1024)
2199 // For kernel code model we know that all object resist in the negative half
2200 // of 32bits address space. We may not accept negative offsets, since they may
2201 // be just off and we may accept pretty large positive ones.
2202 if (M == CodeModel::Kernel && Offset > 0)
2208 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2209 /// specific condition code, returning the condition code and the LHS/RHS of the
2210 /// comparison to make.
2211 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2212 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2214 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2215 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2216 // X > -1 -> X == 0, jump !sign.
2217 RHS = DAG.getConstant(0, RHS.getValueType());
2218 return X86::COND_NS;
2219 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2220 // X < 0 -> X == 0, jump on sign.
2222 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2224 RHS = DAG.getConstant(0, RHS.getValueType());
2225 return X86::COND_LE;
2229 switch (SetCCOpcode) {
2230 default: llvm_unreachable("Invalid integer condition!");
2231 case ISD::SETEQ: return X86::COND_E;
2232 case ISD::SETGT: return X86::COND_G;
2233 case ISD::SETGE: return X86::COND_GE;
2234 case ISD::SETLT: return X86::COND_L;
2235 case ISD::SETLE: return X86::COND_LE;
2236 case ISD::SETNE: return X86::COND_NE;
2237 case ISD::SETULT: return X86::COND_B;
2238 case ISD::SETUGT: return X86::COND_A;
2239 case ISD::SETULE: return X86::COND_BE;
2240 case ISD::SETUGE: return X86::COND_AE;
2244 // First determine if it is required or is profitable to flip the operands.
2246 // If LHS is a foldable load, but RHS is not, flip the condition.
2247 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2248 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2249 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2250 std::swap(LHS, RHS);
2253 switch (SetCCOpcode) {
2259 std::swap(LHS, RHS);
2263 // On a floating point condition, the flags are set as follows:
2265 // 0 | 0 | 0 | X > Y
2266 // 0 | 0 | 1 | X < Y
2267 // 1 | 0 | 0 | X == Y
2268 // 1 | 1 | 1 | unordered
2269 switch (SetCCOpcode) {
2270 default: llvm_unreachable("Condcode should be pre-legalized away");
2272 case ISD::SETEQ: return X86::COND_E;
2273 case ISD::SETOLT: // flipped
2275 case ISD::SETGT: return X86::COND_A;
2276 case ISD::SETOLE: // flipped
2278 case ISD::SETGE: return X86::COND_AE;
2279 case ISD::SETUGT: // flipped
2281 case ISD::SETLT: return X86::COND_B;
2282 case ISD::SETUGE: // flipped
2284 case ISD::SETLE: return X86::COND_BE;
2286 case ISD::SETNE: return X86::COND_NE;
2287 case ISD::SETUO: return X86::COND_P;
2288 case ISD::SETO: return X86::COND_NP;
2292 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2293 /// code. Current x86 isa includes the following FP cmov instructions:
2294 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2295 static bool hasFPCMov(unsigned X86CC) {
2311 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2312 /// the specified range (L, H].
2313 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2314 return (Val < 0) || (Val >= Low && Val < Hi);
2317 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2318 /// specified value.
2319 static bool isUndefOrEqual(int Val, int CmpVal) {
2320 if (Val < 0 || Val == CmpVal)
2325 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2326 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2327 /// the second operand.
2328 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2329 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2330 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2331 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2332 return (Mask[0] < 2 && Mask[1] < 2);
2336 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2337 SmallVector<int, 8> M;
2339 return ::isPSHUFDMask(M, N->getValueType(0));
2342 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2343 /// is suitable for input to PSHUFHW.
2344 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2345 if (VT != MVT::v8i16)
2348 // Lower quadword copied in order or undef.
2349 for (int i = 0; i != 4; ++i)
2350 if (Mask[i] >= 0 && Mask[i] != i)
2353 // Upper quadword shuffled.
2354 for (int i = 4; i != 8; ++i)
2355 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2361 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2362 SmallVector<int, 8> M;
2364 return ::isPSHUFHWMask(M, N->getValueType(0));
2367 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2368 /// is suitable for input to PSHUFLW.
2369 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2370 if (VT != MVT::v8i16)
2373 // Upper quadword copied in order.
2374 for (int i = 4; i != 8; ++i)
2375 if (Mask[i] >= 0 && Mask[i] != i)
2378 // Lower quadword shuffled.
2379 for (int i = 0; i != 4; ++i)
2386 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2387 SmallVector<int, 8> M;
2389 return ::isPSHUFLWMask(M, N->getValueType(0));
2392 /// isPALIGNRMask - Return true if the node specifies a shuffle of elements that
2393 /// is suitable for input to PALIGNR.
2394 static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
2396 int i, e = VT.getVectorNumElements();
2398 // Do not handle v2i64 / v2f64 shuffles with palignr.
2399 if (e < 4 || !hasSSSE3)
2402 for (i = 0; i != e; ++i)
2406 // All undef, not a palignr.
2410 // Determine if it's ok to perform a palignr with only the LHS, since we
2411 // don't have access to the actual shuffle elements to see if RHS is undef.
2412 bool Unary = Mask[i] < (int)e;
2413 bool NeedsUnary = false;
2415 int s = Mask[i] - i;
2417 // Check the rest of the elements to see if they are consecutive.
2418 for (++i; i != e; ++i) {
2423 Unary = Unary && (m < (int)e);
2424 NeedsUnary = NeedsUnary || (m < s);
2426 if (NeedsUnary && !Unary)
2428 if (Unary && m != ((s+i) & (e-1)))
2430 if (!Unary && m != (s+i))
2436 bool X86::isPALIGNRMask(ShuffleVectorSDNode *N) {
2437 SmallVector<int, 8> M;
2439 return ::isPALIGNRMask(M, N->getValueType(0), true);
2442 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2443 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2444 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2445 int NumElems = VT.getVectorNumElements();
2446 if (NumElems != 2 && NumElems != 4)
2449 int Half = NumElems / 2;
2450 for (int i = 0; i < Half; ++i)
2451 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2453 for (int i = Half; i < NumElems; ++i)
2454 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2460 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2461 SmallVector<int, 8> M;
2463 return ::isSHUFPMask(M, N->getValueType(0));
2466 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2467 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2468 /// half elements to come from vector 1 (which would equal the dest.) and
2469 /// the upper half to come from vector 2.
2470 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2471 int NumElems = VT.getVectorNumElements();
2473 if (NumElems != 2 && NumElems != 4)
2476 int Half = NumElems / 2;
2477 for (int i = 0; i < Half; ++i)
2478 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2480 for (int i = Half; i < NumElems; ++i)
2481 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2486 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2487 SmallVector<int, 8> M;
2489 return isCommutedSHUFPMask(M, N->getValueType(0));
2492 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2493 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2494 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2495 if (N->getValueType(0).getVectorNumElements() != 4)
2498 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2499 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2500 isUndefOrEqual(N->getMaskElt(1), 7) &&
2501 isUndefOrEqual(N->getMaskElt(2), 2) &&
2502 isUndefOrEqual(N->getMaskElt(3), 3);
2505 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2506 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2507 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2508 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2510 if (NumElems != 2 && NumElems != 4)
2513 for (unsigned i = 0; i < NumElems/2; ++i)
2514 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2517 for (unsigned i = NumElems/2; i < NumElems; ++i)
2518 if (!isUndefOrEqual(N->getMaskElt(i), i))
2524 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2525 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2527 bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
2528 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2530 if (NumElems != 2 && NumElems != 4)
2533 for (unsigned i = 0; i < NumElems/2; ++i)
2534 if (!isUndefOrEqual(N->getMaskElt(i), i))
2537 for (unsigned i = 0; i < NumElems/2; ++i)
2538 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2544 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2545 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2547 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2548 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2553 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2554 isUndefOrEqual(N->getMaskElt(1), 3) &&
2555 isUndefOrEqual(N->getMaskElt(2), 2) &&
2556 isUndefOrEqual(N->getMaskElt(3), 3);
2559 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2560 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2561 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2562 bool V2IsSplat = false) {
2563 int NumElts = VT.getVectorNumElements();
2564 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2567 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2569 int BitI1 = Mask[i+1];
2570 if (!isUndefOrEqual(BitI, j))
2573 if (!isUndefOrEqual(BitI1, NumElts))
2576 if (!isUndefOrEqual(BitI1, j + NumElts))
2583 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2584 SmallVector<int, 8> M;
2586 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2589 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2590 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2591 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2592 bool V2IsSplat = false) {
2593 int NumElts = VT.getVectorNumElements();
2594 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2597 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2599 int BitI1 = Mask[i+1];
2600 if (!isUndefOrEqual(BitI, j + NumElts/2))
2603 if (isUndefOrEqual(BitI1, NumElts))
2606 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2613 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2614 SmallVector<int, 8> M;
2616 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2619 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2620 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2622 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2623 int NumElems = VT.getVectorNumElements();
2624 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2627 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2629 int BitI1 = Mask[i+1];
2630 if (!isUndefOrEqual(BitI, j))
2632 if (!isUndefOrEqual(BitI1, j))
2638 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2639 SmallVector<int, 8> M;
2641 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2644 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2645 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2647 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2648 int NumElems = VT.getVectorNumElements();
2649 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2652 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2654 int BitI1 = Mask[i+1];
2655 if (!isUndefOrEqual(BitI, j))
2657 if (!isUndefOrEqual(BitI1, j))
2663 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2664 SmallVector<int, 8> M;
2666 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2669 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2670 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2671 /// MOVSD, and MOVD, i.e. setting the lowest element.
2672 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2673 if (VT.getVectorElementType().getSizeInBits() < 32)
2676 int NumElts = VT.getVectorNumElements();
2678 if (!isUndefOrEqual(Mask[0], NumElts))
2681 for (int i = 1; i < NumElts; ++i)
2682 if (!isUndefOrEqual(Mask[i], i))
2688 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2689 SmallVector<int, 8> M;
2691 return ::isMOVLMask(M, N->getValueType(0));
2694 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2695 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2696 /// element of vector 2 and the other elements to come from vector 1 in order.
2697 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2698 bool V2IsSplat = false, bool V2IsUndef = false) {
2699 int NumOps = VT.getVectorNumElements();
2700 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2703 if (!isUndefOrEqual(Mask[0], 0))
2706 for (int i = 1; i < NumOps; ++i)
2707 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2708 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2709 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2715 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2716 bool V2IsUndef = false) {
2717 SmallVector<int, 8> M;
2719 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2722 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2723 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2724 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2725 if (N->getValueType(0).getVectorNumElements() != 4)
2728 // Expect 1, 1, 3, 3
2729 for (unsigned i = 0; i < 2; ++i) {
2730 int Elt = N->getMaskElt(i);
2731 if (Elt >= 0 && Elt != 1)
2736 for (unsigned i = 2; i < 4; ++i) {
2737 int Elt = N->getMaskElt(i);
2738 if (Elt >= 0 && Elt != 3)
2743 // Don't use movshdup if it can be done with a shufps.
2744 // FIXME: verify that matching u, u, 3, 3 is what we want.
2748 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2749 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2750 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2751 if (N->getValueType(0).getVectorNumElements() != 4)
2754 // Expect 0, 0, 2, 2
2755 for (unsigned i = 0; i < 2; ++i)
2756 if (N->getMaskElt(i) > 0)
2760 for (unsigned i = 2; i < 4; ++i) {
2761 int Elt = N->getMaskElt(i);
2762 if (Elt >= 0 && Elt != 2)
2767 // Don't use movsldup if it can be done with a shufps.
2771 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2772 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2773 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2774 int e = N->getValueType(0).getVectorNumElements() / 2;
2776 for (int i = 0; i < e; ++i)
2777 if (!isUndefOrEqual(N->getMaskElt(i), i))
2779 for (int i = 0; i < e; ++i)
2780 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2785 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2786 /// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
2787 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2788 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2789 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2791 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2793 for (int i = 0; i < NumOperands; ++i) {
2794 int Val = SVOp->getMaskElt(NumOperands-i-1);
2795 if (Val < 0) Val = 0;
2796 if (Val >= NumOperands) Val -= NumOperands;
2798 if (i != NumOperands - 1)
2804 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2805 /// the specified VECTOR_SHUFFLE mask with the PSHUFHW instruction.
2806 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2807 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2809 // 8 nodes, but we only care about the last 4.
2810 for (unsigned i = 7; i >= 4; --i) {
2811 int Val = SVOp->getMaskElt(i);
2820 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2821 /// the specified VECTOR_SHUFFLE mask with the PSHUFLW instruction.
2822 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2823 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2825 // 8 nodes, but we only care about the first 4.
2826 for (int i = 3; i >= 0; --i) {
2827 int Val = SVOp->getMaskElt(i);
2836 /// getShufflePALIGNRImmediate - Return the appropriate immediate to shuffle
2837 /// the specified VECTOR_SHUFFLE mask with the PALIGNR instruction.
2838 unsigned X86::getShufflePALIGNRImmediate(SDNode *N) {
2839 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2840 EVT VVT = N->getValueType(0);
2841 unsigned EltSize = VVT.getVectorElementType().getSizeInBits() >> 3;
2845 for (i = 0, e = VVT.getVectorNumElements(); i != e; ++i) {
2846 Val = SVOp->getMaskElt(i);
2850 return (Val - i) * EltSize;
2853 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2855 bool X86::isZeroNode(SDValue Elt) {
2856 return ((isa<ConstantSDNode>(Elt) &&
2857 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2858 (isa<ConstantFPSDNode>(Elt) &&
2859 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2862 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2863 /// their permute mask.
2864 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2865 SelectionDAG &DAG) {
2866 EVT VT = SVOp->getValueType(0);
2867 unsigned NumElems = VT.getVectorNumElements();
2868 SmallVector<int, 8> MaskVec;
2870 for (unsigned i = 0; i != NumElems; ++i) {
2871 int idx = SVOp->getMaskElt(i);
2873 MaskVec.push_back(idx);
2874 else if (idx < (int)NumElems)
2875 MaskVec.push_back(idx + NumElems);
2877 MaskVec.push_back(idx - NumElems);
2879 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2880 SVOp->getOperand(0), &MaskVec[0]);
2883 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2884 /// the two vector operands have swapped position.
2885 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
2886 unsigned NumElems = VT.getVectorNumElements();
2887 for (unsigned i = 0; i != NumElems; ++i) {
2891 else if (idx < (int)NumElems)
2892 Mask[i] = idx + NumElems;
2894 Mask[i] = idx - NumElems;
2898 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2899 /// match movhlps. The lower half elements should come from upper half of
2900 /// V1 (and in order), and the upper half elements should come from the upper
2901 /// half of V2 (and in order).
2902 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2903 if (Op->getValueType(0).getVectorNumElements() != 4)
2905 for (unsigned i = 0, e = 2; i != e; ++i)
2906 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2908 for (unsigned i = 2; i != 4; ++i)
2909 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2914 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2915 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2917 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2918 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2920 N = N->getOperand(0).getNode();
2921 if (!ISD::isNON_EXTLoad(N))
2924 *LD = cast<LoadSDNode>(N);
2928 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2929 /// match movlp{s|d}. The lower half elements should come from lower half of
2930 /// V1 (and in order), and the upper half elements should come from the upper
2931 /// half of V2 (and in order). And since V1 will become the source of the
2932 /// MOVLP, it must be either a vector load or a scalar load to vector.
2933 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2934 ShuffleVectorSDNode *Op) {
2935 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2937 // Is V2 is a vector load, don't do this transformation. We will try to use
2938 // load folding shufps op.
2939 if (ISD::isNON_EXTLoad(V2))
2942 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2944 if (NumElems != 2 && NumElems != 4)
2946 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2947 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2949 for (unsigned i = NumElems/2; i != NumElems; ++i)
2950 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2955 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2957 static bool isSplatVector(SDNode *N) {
2958 if (N->getOpcode() != ISD::BUILD_VECTOR)
2961 SDValue SplatValue = N->getOperand(0);
2962 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2963 if (N->getOperand(i) != SplatValue)
2968 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2969 /// to an zero vector.
2970 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
2971 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
2972 SDValue V1 = N->getOperand(0);
2973 SDValue V2 = N->getOperand(1);
2974 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2975 for (unsigned i = 0; i != NumElems; ++i) {
2976 int Idx = N->getMaskElt(i);
2977 if (Idx >= (int)NumElems) {
2978 unsigned Opc = V2.getOpcode();
2979 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2981 if (Opc != ISD::BUILD_VECTOR ||
2982 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
2984 } else if (Idx >= 0) {
2985 unsigned Opc = V1.getOpcode();
2986 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2988 if (Opc != ISD::BUILD_VECTOR ||
2989 !X86::isZeroNode(V1.getOperand(Idx)))
2996 /// getZeroVector - Returns a vector of specified type with all zero elements.
2998 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
3000 assert(VT.isVector() && "Expected a vector type");
3002 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3003 // type. This ensures they get CSE'd.
3005 if (VT.getSizeInBits() == 64) { // MMX
3006 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3007 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3008 } else if (HasSSE2) { // SSE2
3009 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
3010 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3012 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
3013 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
3015 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3018 /// getOnesVector - Returns a vector of specified type with all bits set.
3020 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
3021 assert(VT.isVector() && "Expected a vector type");
3023 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
3024 // type. This ensures they get CSE'd.
3025 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
3027 if (VT.getSizeInBits() == 64) // MMX
3028 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
3030 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
3031 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
3035 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
3036 /// that point to V2 points to its first element.
3037 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3038 EVT VT = SVOp->getValueType(0);
3039 unsigned NumElems = VT.getVectorNumElements();
3041 bool Changed = false;
3042 SmallVector<int, 8> MaskVec;
3043 SVOp->getMask(MaskVec);
3045 for (unsigned i = 0; i != NumElems; ++i) {
3046 if (MaskVec[i] > (int)NumElems) {
3047 MaskVec[i] = NumElems;
3052 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
3053 SVOp->getOperand(1), &MaskVec[0]);
3054 return SDValue(SVOp, 0);
3057 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
3058 /// operation of specified width.
3059 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3061 unsigned NumElems = VT.getVectorNumElements();
3062 SmallVector<int, 8> Mask;
3063 Mask.push_back(NumElems);
3064 for (unsigned i = 1; i != NumElems; ++i)
3066 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3069 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
3070 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3072 unsigned NumElems = VT.getVectorNumElements();
3073 SmallVector<int, 8> Mask;
3074 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
3076 Mask.push_back(i + NumElems);
3078 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3081 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
3082 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
3084 unsigned NumElems = VT.getVectorNumElements();
3085 unsigned Half = NumElems/2;
3086 SmallVector<int, 8> Mask;
3087 for (unsigned i = 0; i != Half; ++i) {
3088 Mask.push_back(i + Half);
3089 Mask.push_back(i + NumElems + Half);
3091 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3094 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3095 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3097 if (SV->getValueType(0).getVectorNumElements() <= 4)
3098 return SDValue(SV, 0);
3100 EVT PVT = MVT::v4f32;
3101 EVT VT = SV->getValueType(0);
3102 DebugLoc dl = SV->getDebugLoc();
3103 SDValue V1 = SV->getOperand(0);
3104 int NumElems = VT.getVectorNumElements();
3105 int EltNo = SV->getSplatIndex();
3107 // unpack elements to the correct location
3108 while (NumElems > 4) {
3109 if (EltNo < NumElems/2) {
3110 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3112 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3113 EltNo -= NumElems/2;
3118 // Perform the splat.
3119 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3120 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3121 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3122 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3125 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3126 /// vector of zero or undef vector. This produces a shuffle where the low
3127 /// element of V2 is swizzled into the zero/undef vector, landing at element
3128 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3129 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3130 bool isZero, bool HasSSE2,
3131 SelectionDAG &DAG) {
3132 EVT VT = V2.getValueType();
3134 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3135 unsigned NumElems = VT.getVectorNumElements();
3136 SmallVector<int, 16> MaskVec;
3137 for (unsigned i = 0; i != NumElems; ++i)
3138 // If this is the insertion idx, put the low elt of V2 here.
3139 MaskVec.push_back(i == Idx ? NumElems : i);
3140 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3143 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3144 /// a shuffle that is zero.
3146 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3147 bool Low, SelectionDAG &DAG) {
3148 unsigned NumZeros = 0;
3149 for (int i = 0; i < NumElems; ++i) {
3150 unsigned Index = Low ? i : NumElems-i-1;
3151 int Idx = SVOp->getMaskElt(Index);
3156 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3157 if (Elt.getNode() && X86::isZeroNode(Elt))
3165 /// isVectorShift - Returns true if the shuffle can be implemented as a
3166 /// logical left or right shift of a vector.
3167 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3168 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3169 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3170 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3173 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3176 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3180 bool SeenV1 = false;
3181 bool SeenV2 = false;
3182 for (int i = NumZeros; i < NumElems; ++i) {
3183 int Val = isLeft ? (i - NumZeros) : i;
3184 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3196 if (SeenV1 && SeenV2)
3199 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3205 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3207 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3208 unsigned NumNonZero, unsigned NumZero,
3209 SelectionDAG &DAG, TargetLowering &TLI) {
3213 DebugLoc dl = Op.getDebugLoc();
3216 for (unsigned i = 0; i < 16; ++i) {
3217 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3218 if (ThisIsNonZero && First) {
3220 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3222 V = DAG.getUNDEF(MVT::v8i16);
3227 SDValue ThisElt(0, 0), LastElt(0, 0);
3228 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3229 if (LastIsNonZero) {
3230 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3231 MVT::i16, Op.getOperand(i-1));
3233 if (ThisIsNonZero) {
3234 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3235 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3236 ThisElt, DAG.getConstant(8, MVT::i8));
3238 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3242 if (ThisElt.getNode())
3243 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3244 DAG.getIntPtrConstant(i/2));
3248 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3251 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3253 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3254 unsigned NumNonZero, unsigned NumZero,
3255 SelectionDAG &DAG, TargetLowering &TLI) {
3259 DebugLoc dl = Op.getDebugLoc();
3262 for (unsigned i = 0; i < 8; ++i) {
3263 bool isNonZero = (NonZeros & (1 << i)) != 0;
3267 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3269 V = DAG.getUNDEF(MVT::v8i16);
3272 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3273 MVT::v8i16, V, Op.getOperand(i),
3274 DAG.getIntPtrConstant(i));
3281 /// getVShift - Return a vector logical shift node.
3283 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3284 unsigned NumBits, SelectionDAG &DAG,
3285 const TargetLowering &TLI, DebugLoc dl) {
3286 bool isMMX = VT.getSizeInBits() == 64;
3287 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3288 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3289 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3290 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3291 DAG.getNode(Opc, dl, ShVT, SrcOp,
3292 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3296 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3297 DebugLoc dl = Op.getDebugLoc();
3298 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3299 if (ISD::isBuildVectorAllZeros(Op.getNode())
3300 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3301 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3302 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3303 // eliminated on x86-32 hosts.
3304 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3307 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3308 return getOnesVector(Op.getValueType(), DAG, dl);
3309 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3312 EVT VT = Op.getValueType();
3313 EVT ExtVT = VT.getVectorElementType();
3314 unsigned EVTBits = ExtVT.getSizeInBits();
3316 unsigned NumElems = Op.getNumOperands();
3317 unsigned NumZero = 0;
3318 unsigned NumNonZero = 0;
3319 unsigned NonZeros = 0;
3320 bool IsAllConstants = true;
3321 SmallSet<SDValue, 8> Values;
3322 for (unsigned i = 0; i < NumElems; ++i) {
3323 SDValue Elt = Op.getOperand(i);
3324 if (Elt.getOpcode() == ISD::UNDEF)
3327 if (Elt.getOpcode() != ISD::Constant &&
3328 Elt.getOpcode() != ISD::ConstantFP)
3329 IsAllConstants = false;
3330 if (X86::isZeroNode(Elt))
3333 NonZeros |= (1 << i);
3338 if (NumNonZero == 0) {
3339 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3340 return DAG.getUNDEF(VT);
3343 // Special case for single non-zero, non-undef, element.
3344 if (NumNonZero == 1) {
3345 unsigned Idx = CountTrailingZeros_32(NonZeros);
3346 SDValue Item = Op.getOperand(Idx);
3348 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3349 // the value are obviously zero, truncate the value to i32 and do the
3350 // insertion that way. Only do this if the value is non-constant or if the
3351 // value is a constant being inserted into element 0. It is cheaper to do
3352 // a constant pool load than it is to do a movd + shuffle.
3353 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3354 (!IsAllConstants || Idx == 0)) {
3355 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3356 // Handle MMX and SSE both.
3357 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3358 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3360 // Truncate the value (which may itself be a constant) to i32, and
3361 // convert it to a vector with movd (S2V+shuffle to zero extend).
3362 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3363 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3364 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3365 Subtarget->hasSSE2(), DAG);
3367 // Now we have our 32-bit value zero extended in the low element of
3368 // a vector. If Idx != 0, swizzle it into place.
3370 SmallVector<int, 4> Mask;
3371 Mask.push_back(Idx);
3372 for (unsigned i = 1; i != VecElts; ++i)
3374 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3375 DAG.getUNDEF(Item.getValueType()),
3378 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3382 // If we have a constant or non-constant insertion into the low element of
3383 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3384 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3385 // depending on what the source datatype is.
3388 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3389 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3390 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3391 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3392 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3393 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3395 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3396 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3397 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3398 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3399 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3400 Subtarget->hasSSE2(), DAG);
3401 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3405 // Is it a vector logical left shift?
3406 if (NumElems == 2 && Idx == 1 &&
3407 X86::isZeroNode(Op.getOperand(0)) &&
3408 !X86::isZeroNode(Op.getOperand(1))) {
3409 unsigned NumBits = VT.getSizeInBits();
3410 return getVShift(true, VT,
3411 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3412 VT, Op.getOperand(1)),
3413 NumBits/2, DAG, *this, dl);
3416 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3419 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3420 // is a non-constant being inserted into an element other than the low one,
3421 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3422 // movd/movss) to move this into the low element, then shuffle it into
3424 if (EVTBits == 32) {
3425 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3427 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3428 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3429 Subtarget->hasSSE2(), DAG);
3430 SmallVector<int, 8> MaskVec;
3431 for (unsigned i = 0; i < NumElems; i++)
3432 MaskVec.push_back(i == Idx ? 0 : 1);
3433 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3437 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3438 if (Values.size() == 1)
3441 // A vector full of immediates; various special cases are already
3442 // handled, so this is best done with a single constant-pool load.
3446 // Let legalizer expand 2-wide build_vectors.
3447 if (EVTBits == 64) {
3448 if (NumNonZero == 1) {
3449 // One half is zero or undef.
3450 unsigned Idx = CountTrailingZeros_32(NonZeros);
3451 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3452 Op.getOperand(Idx));
3453 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3454 Subtarget->hasSSE2(), DAG);
3459 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3460 if (EVTBits == 8 && NumElems == 16) {
3461 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3463 if (V.getNode()) return V;
3466 if (EVTBits == 16 && NumElems == 8) {
3467 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3469 if (V.getNode()) return V;
3472 // If element VT is == 32 bits, turn it into a number of shuffles.
3473 SmallVector<SDValue, 8> V;
3475 if (NumElems == 4 && NumZero > 0) {
3476 for (unsigned i = 0; i < 4; ++i) {
3477 bool isZero = !(NonZeros & (1 << i));
3479 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3481 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3484 for (unsigned i = 0; i < 2; ++i) {
3485 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3488 V[i] = V[i*2]; // Must be a zero vector.
3491 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3494 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3497 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3502 SmallVector<int, 8> MaskVec;
3503 bool Reverse = (NonZeros & 0x3) == 2;
3504 for (unsigned i = 0; i < 2; ++i)
3505 MaskVec.push_back(Reverse ? 1-i : i);
3506 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3507 for (unsigned i = 0; i < 2; ++i)
3508 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3509 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3512 if (Values.size() > 2) {
3513 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3514 // values to be inserted is equal to the number of elements, in which case
3515 // use the unpack code below in the hopes of matching the consecutive elts
3516 // load merge pattern for shuffles.
3517 // FIXME: We could probably just check that here directly.
3518 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3519 getSubtarget()->hasSSE41()) {
3520 V[0] = DAG.getUNDEF(VT);
3521 for (unsigned i = 0; i < NumElems; ++i)
3522 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3523 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3524 Op.getOperand(i), DAG.getIntPtrConstant(i));
3527 // Expand into a number of unpckl*.
3529 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3530 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3531 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3532 for (unsigned i = 0; i < NumElems; ++i)
3533 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3535 while (NumElems != 0) {
3536 for (unsigned i = 0; i < NumElems; ++i)
3537 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3546 // v8i16 shuffles - Prefer shuffles in the following order:
3547 // 1. [all] pshuflw, pshufhw, optional move
3548 // 2. [ssse3] 1 x pshufb
3549 // 3. [ssse3] 2 x pshufb + 1 x por
3550 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3552 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3553 SelectionDAG &DAG, X86TargetLowering &TLI) {
3554 SDValue V1 = SVOp->getOperand(0);
3555 SDValue V2 = SVOp->getOperand(1);
3556 DebugLoc dl = SVOp->getDebugLoc();
3557 SmallVector<int, 8> MaskVals;
3559 // Determine if more than 1 of the words in each of the low and high quadwords
3560 // of the result come from the same quadword of one of the two inputs. Undef
3561 // mask values count as coming from any quadword, for better codegen.
3562 SmallVector<unsigned, 4> LoQuad(4);
3563 SmallVector<unsigned, 4> HiQuad(4);
3564 BitVector InputQuads(4);
3565 for (unsigned i = 0; i < 8; ++i) {
3566 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3567 int EltIdx = SVOp->getMaskElt(i);
3568 MaskVals.push_back(EltIdx);
3577 InputQuads.set(EltIdx / 4);
3580 int BestLoQuad = -1;
3581 unsigned MaxQuad = 1;
3582 for (unsigned i = 0; i < 4; ++i) {
3583 if (LoQuad[i] > MaxQuad) {
3585 MaxQuad = LoQuad[i];
3589 int BestHiQuad = -1;
3591 for (unsigned i = 0; i < 4; ++i) {
3592 if (HiQuad[i] > MaxQuad) {
3594 MaxQuad = HiQuad[i];
3598 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3599 // of the two input vectors, shuffle them into one input vector so only a
3600 // single pshufb instruction is necessary. If There are more than 2 input
3601 // quads, disable the next transformation since it does not help SSSE3.
3602 bool V1Used = InputQuads[0] || InputQuads[1];
3603 bool V2Used = InputQuads[2] || InputQuads[3];
3604 if (TLI.getSubtarget()->hasSSSE3()) {
3605 if (InputQuads.count() == 2 && V1Used && V2Used) {
3606 BestLoQuad = InputQuads.find_first();
3607 BestHiQuad = InputQuads.find_next(BestLoQuad);
3609 if (InputQuads.count() > 2) {
3615 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3616 // the shuffle mask. If a quad is scored as -1, that means that it contains
3617 // words from all 4 input quadwords.
3619 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3620 SmallVector<int, 8> MaskV;
3621 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3622 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3623 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3624 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3625 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3626 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3628 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3629 // source words for the shuffle, to aid later transformations.
3630 bool AllWordsInNewV = true;
3631 bool InOrder[2] = { true, true };
3632 for (unsigned i = 0; i != 8; ++i) {
3633 int idx = MaskVals[i];
3635 InOrder[i/4] = false;
3636 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3638 AllWordsInNewV = false;
3642 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3643 if (AllWordsInNewV) {
3644 for (int i = 0; i != 8; ++i) {
3645 int idx = MaskVals[i];
3648 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3649 if ((idx != i) && idx < 4)
3651 if ((idx != i) && idx > 3)
3660 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3661 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3662 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3663 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3664 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3668 // If we have SSSE3, and all words of the result are from 1 input vector,
3669 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3670 // is present, fall back to case 4.
3671 if (TLI.getSubtarget()->hasSSSE3()) {
3672 SmallVector<SDValue,16> pshufbMask;
3674 // If we have elements from both input vectors, set the high bit of the
3675 // shuffle mask element to zero out elements that come from V2 in the V1
3676 // mask, and elements that come from V1 in the V2 mask, so that the two
3677 // results can be OR'd together.
3678 bool TwoInputs = V1Used && V2Used;
3679 for (unsigned i = 0; i != 8; ++i) {
3680 int EltIdx = MaskVals[i] * 2;
3681 if (TwoInputs && (EltIdx >= 16)) {
3682 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3683 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3686 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3687 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3689 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3690 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3691 DAG.getNode(ISD::BUILD_VECTOR, dl,
3692 MVT::v16i8, &pshufbMask[0], 16));
3694 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3696 // Calculate the shuffle mask for the second input, shuffle it, and
3697 // OR it with the first shuffled input.
3699 for (unsigned i = 0; i != 8; ++i) {
3700 int EltIdx = MaskVals[i] * 2;
3702 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3703 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3706 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3707 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3709 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3710 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3711 DAG.getNode(ISD::BUILD_VECTOR, dl,
3712 MVT::v16i8, &pshufbMask[0], 16));
3713 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3714 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3717 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3718 // and update MaskVals with new element order.
3719 BitVector InOrder(8);
3720 if (BestLoQuad >= 0) {
3721 SmallVector<int, 8> MaskV;
3722 for (int i = 0; i != 4; ++i) {
3723 int idx = MaskVals[i];
3725 MaskV.push_back(-1);
3727 } else if ((idx / 4) == BestLoQuad) {
3728 MaskV.push_back(idx & 3);
3731 MaskV.push_back(-1);
3734 for (unsigned i = 4; i != 8; ++i)
3736 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3740 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3741 // and update MaskVals with the new element order.
3742 if (BestHiQuad >= 0) {
3743 SmallVector<int, 8> MaskV;
3744 for (unsigned i = 0; i != 4; ++i)
3746 for (unsigned i = 4; i != 8; ++i) {
3747 int idx = MaskVals[i];
3749 MaskV.push_back(-1);
3751 } else if ((idx / 4) == BestHiQuad) {
3752 MaskV.push_back((idx & 3) + 4);
3755 MaskV.push_back(-1);
3758 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3762 // In case BestHi & BestLo were both -1, which means each quadword has a word
3763 // from each of the four input quadwords, calculate the InOrder bitvector now
3764 // before falling through to the insert/extract cleanup.
3765 if (BestLoQuad == -1 && BestHiQuad == -1) {
3767 for (int i = 0; i != 8; ++i)
3768 if (MaskVals[i] < 0 || MaskVals[i] == i)
3772 // The other elements are put in the right place using pextrw and pinsrw.
3773 for (unsigned i = 0; i != 8; ++i) {
3776 int EltIdx = MaskVals[i];
3779 SDValue ExtOp = (EltIdx < 8)
3780 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3781 DAG.getIntPtrConstant(EltIdx))
3782 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3783 DAG.getIntPtrConstant(EltIdx - 8));
3784 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3785 DAG.getIntPtrConstant(i));
3790 // v16i8 shuffles - Prefer shuffles in the following order:
3791 // 1. [ssse3] 1 x pshufb
3792 // 2. [ssse3] 2 x pshufb + 1 x por
3793 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3795 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3796 SelectionDAG &DAG, X86TargetLowering &TLI) {
3797 SDValue V1 = SVOp->getOperand(0);
3798 SDValue V2 = SVOp->getOperand(1);
3799 DebugLoc dl = SVOp->getDebugLoc();
3800 SmallVector<int, 16> MaskVals;
3801 SVOp->getMask(MaskVals);
3803 // If we have SSSE3, case 1 is generated when all result bytes come from
3804 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3805 // present, fall back to case 3.
3806 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3809 for (unsigned i = 0; i < 16; ++i) {
3810 int EltIdx = MaskVals[i];
3819 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3820 if (TLI.getSubtarget()->hasSSSE3()) {
3821 SmallVector<SDValue,16> pshufbMask;
3823 // If all result elements are from one input vector, then only translate
3824 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3826 // Otherwise, we have elements from both input vectors, and must zero out
3827 // elements that come from V2 in the first mask, and V1 in the second mask
3828 // so that we can OR them together.
3829 bool TwoInputs = !(V1Only || V2Only);
3830 for (unsigned i = 0; i != 16; ++i) {
3831 int EltIdx = MaskVals[i];
3832 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3833 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3836 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3838 // If all the elements are from V2, assign it to V1 and return after
3839 // building the first pshufb.
3842 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3843 DAG.getNode(ISD::BUILD_VECTOR, dl,
3844 MVT::v16i8, &pshufbMask[0], 16));
3848 // Calculate the shuffle mask for the second input, shuffle it, and
3849 // OR it with the first shuffled input.
3851 for (unsigned i = 0; i != 16; ++i) {
3852 int EltIdx = MaskVals[i];
3854 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3857 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3859 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3860 DAG.getNode(ISD::BUILD_VECTOR, dl,
3861 MVT::v16i8, &pshufbMask[0], 16));
3862 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3865 // No SSSE3 - Calculate in place words and then fix all out of place words
3866 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3867 // the 16 different words that comprise the two doublequadword input vectors.
3868 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3869 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3870 SDValue NewV = V2Only ? V2 : V1;
3871 for (int i = 0; i != 8; ++i) {
3872 int Elt0 = MaskVals[i*2];
3873 int Elt1 = MaskVals[i*2+1];
3875 // This word of the result is all undef, skip it.
3876 if (Elt0 < 0 && Elt1 < 0)
3879 // This word of the result is already in the correct place, skip it.
3880 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3882 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3885 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3886 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3889 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3890 // using a single extract together, load it and store it.
3891 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3892 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3893 DAG.getIntPtrConstant(Elt1 / 2));
3894 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3895 DAG.getIntPtrConstant(i));
3899 // If Elt1 is defined, extract it from the appropriate source. If the
3900 // source byte is not also odd, shift the extracted word left 8 bits
3901 // otherwise clear the bottom 8 bits if we need to do an or.
3903 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3904 DAG.getIntPtrConstant(Elt1 / 2));
3905 if ((Elt1 & 1) == 0)
3906 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3907 DAG.getConstant(8, TLI.getShiftAmountTy()));
3909 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3910 DAG.getConstant(0xFF00, MVT::i16));
3912 // If Elt0 is defined, extract it from the appropriate source. If the
3913 // source byte is not also even, shift the extracted word right 8 bits. If
3914 // Elt1 was also defined, OR the extracted values together before
3915 // inserting them in the result.
3917 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3918 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3919 if ((Elt0 & 1) != 0)
3920 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3921 DAG.getConstant(8, TLI.getShiftAmountTy()));
3923 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3924 DAG.getConstant(0x00FF, MVT::i16));
3925 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3928 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3929 DAG.getIntPtrConstant(i));
3931 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3934 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3935 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3936 /// done when every pair / quad of shuffle mask elements point to elements in
3937 /// the right sequence. e.g.
3938 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3940 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3942 TargetLowering &TLI, DebugLoc dl) {
3943 EVT VT = SVOp->getValueType(0);
3944 SDValue V1 = SVOp->getOperand(0);
3945 SDValue V2 = SVOp->getOperand(1);
3946 unsigned NumElems = VT.getVectorNumElements();
3947 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3948 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3949 EVT MaskEltVT = MaskVT.getVectorElementType();
3951 switch (VT.getSimpleVT().SimpleTy) {
3952 default: assert(false && "Unexpected!");
3953 case MVT::v4f32: NewVT = MVT::v2f64; break;
3954 case MVT::v4i32: NewVT = MVT::v2i64; break;
3955 case MVT::v8i16: NewVT = MVT::v4i32; break;
3956 case MVT::v16i8: NewVT = MVT::v4i32; break;
3959 if (NewWidth == 2) {
3965 int Scale = NumElems / NewWidth;
3966 SmallVector<int, 8> MaskVec;
3967 for (unsigned i = 0; i < NumElems; i += Scale) {
3969 for (int j = 0; j < Scale; ++j) {
3970 int EltIdx = SVOp->getMaskElt(i+j);
3974 StartIdx = EltIdx - (EltIdx % Scale);
3975 if (EltIdx != StartIdx + j)
3979 MaskVec.push_back(-1);
3981 MaskVec.push_back(StartIdx / Scale);
3984 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3985 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3986 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
3989 /// getVZextMovL - Return a zero-extending vector move low node.
3991 static SDValue getVZextMovL(EVT VT, EVT OpVT,
3992 SDValue SrcOp, SelectionDAG &DAG,
3993 const X86Subtarget *Subtarget, DebugLoc dl) {
3994 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3995 LoadSDNode *LD = NULL;
3996 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3997 LD = dyn_cast<LoadSDNode>(SrcOp);
3999 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
4001 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
4002 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
4003 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
4004 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
4005 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
4007 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
4008 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4009 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4010 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4018 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4019 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
4020 DAG.getNode(ISD::BIT_CONVERT, dl,
4024 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
4027 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
4028 SDValue V1 = SVOp->getOperand(0);
4029 SDValue V2 = SVOp->getOperand(1);
4030 DebugLoc dl = SVOp->getDebugLoc();
4031 EVT VT = SVOp->getValueType(0);
4033 SmallVector<std::pair<int, int>, 8> Locs;
4035 SmallVector<int, 8> Mask1(4U, -1);
4036 SmallVector<int, 8> PermMask;
4037 SVOp->getMask(PermMask);
4041 for (unsigned i = 0; i != 4; ++i) {
4042 int Idx = PermMask[i];
4044 Locs[i] = std::make_pair(-1, -1);
4046 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
4048 Locs[i] = std::make_pair(0, NumLo);
4052 Locs[i] = std::make_pair(1, NumHi);
4054 Mask1[2+NumHi] = Idx;
4060 if (NumLo <= 2 && NumHi <= 2) {
4061 // If no more than two elements come from either vector. This can be
4062 // implemented with two shuffles. First shuffle gather the elements.
4063 // The second shuffle, which takes the first shuffle as both of its
4064 // vector operands, put the elements into the right order.
4065 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4067 SmallVector<int, 8> Mask2(4U, -1);
4069 for (unsigned i = 0; i != 4; ++i) {
4070 if (Locs[i].first == -1)
4073 unsigned Idx = (i < 2) ? 0 : 4;
4074 Idx += Locs[i].first * 2 + Locs[i].second;
4079 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
4080 } else if (NumLo == 3 || NumHi == 3) {
4081 // Otherwise, we must have three elements from one vector, call it X, and
4082 // one element from the other, call it Y. First, use a shufps to build an
4083 // intermediate vector with the one element from Y and the element from X
4084 // that will be in the same half in the final destination (the indexes don't
4085 // matter). Then, use a shufps to build the final vector, taking the half
4086 // containing the element from Y from the intermediate, and the other half
4089 // Normalize it so the 3 elements come from V1.
4090 CommuteVectorShuffleMask(PermMask, VT);
4094 // Find the element from V2.
4096 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4097 int Val = PermMask[HiIndex];
4104 Mask1[0] = PermMask[HiIndex];
4106 Mask1[2] = PermMask[HiIndex^1];
4108 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4111 Mask1[0] = PermMask[0];
4112 Mask1[1] = PermMask[1];
4113 Mask1[2] = HiIndex & 1 ? 6 : 4;
4114 Mask1[3] = HiIndex & 1 ? 4 : 6;
4115 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4117 Mask1[0] = HiIndex & 1 ? 2 : 0;
4118 Mask1[1] = HiIndex & 1 ? 0 : 2;
4119 Mask1[2] = PermMask[2];
4120 Mask1[3] = PermMask[3];
4125 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4129 // Break it into (shuffle shuffle_hi, shuffle_lo).
4131 SmallVector<int,8> LoMask(4U, -1);
4132 SmallVector<int,8> HiMask(4U, -1);
4134 SmallVector<int,8> *MaskPtr = &LoMask;
4135 unsigned MaskIdx = 0;
4138 for (unsigned i = 0; i != 4; ++i) {
4145 int Idx = PermMask[i];
4147 Locs[i] = std::make_pair(-1, -1);
4148 } else if (Idx < 4) {
4149 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4150 (*MaskPtr)[LoIdx] = Idx;
4153 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4154 (*MaskPtr)[HiIdx] = Idx;
4159 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4160 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4161 SmallVector<int, 8> MaskOps;
4162 for (unsigned i = 0; i != 4; ++i) {
4163 if (Locs[i].first == -1) {
4164 MaskOps.push_back(-1);
4166 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4167 MaskOps.push_back(Idx);
4170 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4174 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4175 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4176 SDValue V1 = Op.getOperand(0);
4177 SDValue V2 = Op.getOperand(1);
4178 EVT VT = Op.getValueType();
4179 DebugLoc dl = Op.getDebugLoc();
4180 unsigned NumElems = VT.getVectorNumElements();
4181 bool isMMX = VT.getSizeInBits() == 64;
4182 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4183 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4184 bool V1IsSplat = false;
4185 bool V2IsSplat = false;
4187 if (isZeroShuffle(SVOp))
4188 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4190 // Promote splats to v4f32.
4191 if (SVOp->isSplat()) {
4192 if (isMMX || NumElems < 4)
4194 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4197 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4199 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4200 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4201 if (NewOp.getNode())
4202 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4203 LowerVECTOR_SHUFFLE(NewOp, DAG));
4204 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4205 // FIXME: Figure out a cleaner way to do this.
4206 // Try to make use of movq to zero out the top part.
4207 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4208 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4209 if (NewOp.getNode()) {
4210 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4211 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4212 DAG, Subtarget, dl);
4214 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4215 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4216 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4217 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4218 DAG, Subtarget, dl);
4222 if (X86::isPSHUFDMask(SVOp))
4225 // Check if this can be converted into a logical shift.
4226 bool isLeft = false;
4229 bool isShift = getSubtarget()->hasSSE2() &&
4230 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4231 if (isShift && ShVal.hasOneUse()) {
4232 // If the shifted value has multiple uses, it may be cheaper to use
4233 // v_set0 + movlhps or movhlps, etc.
4234 EVT EltVT = VT.getVectorElementType();
4235 ShAmt *= EltVT.getSizeInBits();
4236 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4239 if (X86::isMOVLMask(SVOp)) {
4242 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4243 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4248 // FIXME: fold these into legal mask.
4249 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4250 X86::isMOVSLDUPMask(SVOp) ||
4251 X86::isMOVHLPSMask(SVOp) ||
4252 X86::isMOVHPMask(SVOp) ||
4253 X86::isMOVLPMask(SVOp)))
4256 if (ShouldXformToMOVHLPS(SVOp) ||
4257 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4258 return CommuteVectorShuffle(SVOp, DAG);
4261 // No better options. Use a vshl / vsrl.
4262 EVT EltVT = VT.getVectorElementType();
4263 ShAmt *= EltVT.getSizeInBits();
4264 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4267 bool Commuted = false;
4268 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4269 // 1,1,1,1 -> v8i16 though.
4270 V1IsSplat = isSplatVector(V1.getNode());
4271 V2IsSplat = isSplatVector(V2.getNode());
4273 // Canonicalize the splat or undef, if present, to be on the RHS.
4274 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4275 Op = CommuteVectorShuffle(SVOp, DAG);
4276 SVOp = cast<ShuffleVectorSDNode>(Op);
4277 V1 = SVOp->getOperand(0);
4278 V2 = SVOp->getOperand(1);
4279 std::swap(V1IsSplat, V2IsSplat);
4280 std::swap(V1IsUndef, V2IsUndef);
4284 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4285 // Shuffling low element of v1 into undef, just return v1.
4288 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4289 // the instruction selector will not match, so get a canonical MOVL with
4290 // swapped operands to undo the commute.
4291 return getMOVL(DAG, dl, VT, V2, V1);
4294 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4295 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4296 X86::isUNPCKLMask(SVOp) ||
4297 X86::isUNPCKHMask(SVOp))
4301 // Normalize mask so all entries that point to V2 points to its first
4302 // element then try to match unpck{h|l} again. If match, return a
4303 // new vector_shuffle with the corrected mask.
4304 SDValue NewMask = NormalizeMask(SVOp, DAG);
4305 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4306 if (NSVOp != SVOp) {
4307 if (X86::isUNPCKLMask(NSVOp, true)) {
4309 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4316 // Commute is back and try unpck* again.
4317 // FIXME: this seems wrong.
4318 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4319 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4320 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4321 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4322 X86::isUNPCKLMask(NewSVOp) ||
4323 X86::isUNPCKHMask(NewSVOp))
4327 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4329 // Normalize the node to match x86 shuffle ops if needed
4330 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4331 return CommuteVectorShuffle(SVOp, DAG);
4333 // Check for legal shuffle and return?
4334 SmallVector<int, 16> PermMask;
4335 SVOp->getMask(PermMask);
4336 if (isShuffleMaskLegal(PermMask, VT))
4339 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4340 if (VT == MVT::v8i16) {
4341 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4342 if (NewOp.getNode())
4346 if (VT == MVT::v16i8) {
4347 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4348 if (NewOp.getNode())
4352 // Handle all 4 wide cases with a number of shuffles except for MMX.
4353 if (NumElems == 4 && !isMMX)
4354 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4360 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4361 SelectionDAG &DAG) {
4362 EVT VT = Op.getValueType();
4363 DebugLoc dl = Op.getDebugLoc();
4364 if (VT.getSizeInBits() == 8) {
4365 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4366 Op.getOperand(0), Op.getOperand(1));
4367 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4368 DAG.getValueType(VT));
4369 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4370 } else if (VT.getSizeInBits() == 16) {
4371 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4372 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4374 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4375 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4376 DAG.getNode(ISD::BIT_CONVERT, dl,
4380 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4381 Op.getOperand(0), Op.getOperand(1));
4382 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4383 DAG.getValueType(VT));
4384 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4385 } else if (VT == MVT::f32) {
4386 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4387 // the result back to FR32 register. It's only worth matching if the
4388 // result has a single use which is a store or a bitcast to i32. And in
4389 // the case of a store, it's not worth it if the index is a constant 0,
4390 // because a MOVSSmr can be used instead, which is smaller and faster.
4391 if (!Op.hasOneUse())
4393 SDNode *User = *Op.getNode()->use_begin();
4394 if ((User->getOpcode() != ISD::STORE ||
4395 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4396 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4397 (User->getOpcode() != ISD::BIT_CONVERT ||
4398 User->getValueType(0) != MVT::i32))
4400 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4401 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4404 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4405 } else if (VT == MVT::i32) {
4406 // ExtractPS works with constant index.
4407 if (isa<ConstantSDNode>(Op.getOperand(1)))
4415 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4416 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4419 if (Subtarget->hasSSE41()) {
4420 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4425 EVT VT = Op.getValueType();
4426 DebugLoc dl = Op.getDebugLoc();
4427 // TODO: handle v16i8.
4428 if (VT.getSizeInBits() == 16) {
4429 SDValue Vec = Op.getOperand(0);
4430 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4432 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4433 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4434 DAG.getNode(ISD::BIT_CONVERT, dl,
4437 // Transform it so it match pextrw which produces a 32-bit result.
4438 EVT EltVT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy+1);
4439 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
4440 Op.getOperand(0), Op.getOperand(1));
4441 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
4442 DAG.getValueType(VT));
4443 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4444 } else if (VT.getSizeInBits() == 32) {
4445 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4449 // SHUFPS the element to the lowest double word, then movss.
4450 int Mask[4] = { Idx, -1, -1, -1 };
4451 EVT VVT = Op.getOperand(0).getValueType();
4452 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4453 DAG.getUNDEF(VVT), Mask);
4454 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4455 DAG.getIntPtrConstant(0));
4456 } else if (VT.getSizeInBits() == 64) {
4457 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4458 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4459 // to match extract_elt for f64.
4460 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4464 // UNPCKHPD the element to the lowest double word, then movsd.
4465 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4466 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4467 int Mask[2] = { 1, -1 };
4468 EVT VVT = Op.getOperand(0).getValueType();
4469 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4470 DAG.getUNDEF(VVT), Mask);
4471 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4472 DAG.getIntPtrConstant(0));
4479 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4480 EVT VT = Op.getValueType();
4481 EVT EltVT = VT.getVectorElementType();
4482 DebugLoc dl = Op.getDebugLoc();
4484 SDValue N0 = Op.getOperand(0);
4485 SDValue N1 = Op.getOperand(1);
4486 SDValue N2 = Op.getOperand(2);
4488 if ((EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) &&
4489 isa<ConstantSDNode>(N2)) {
4490 unsigned Opc = (EltVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4492 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4494 if (N1.getValueType() != MVT::i32)
4495 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4496 if (N2.getValueType() != MVT::i32)
4497 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4498 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4499 } else if (EltVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4500 // Bits [7:6] of the constant are the source select. This will always be
4501 // zero here. The DAG Combiner may combine an extract_elt index into these
4502 // bits. For example (insert (extract, 3), 2) could be matched by putting
4503 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4504 // Bits [5:4] of the constant are the destination select. This is the
4505 // value of the incoming immediate.
4506 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4507 // combine either bitwise AND or insert of float 0.0 to set these bits.
4508 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4509 // Create this as a scalar to vector..
4510 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4511 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4512 } else if (EltVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4513 // PINSR* works with constant index.
4520 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4521 EVT VT = Op.getValueType();
4522 EVT EltVT = VT.getVectorElementType();
4524 if (Subtarget->hasSSE41())
4525 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4527 if (EltVT == MVT::i8)
4530 DebugLoc dl = Op.getDebugLoc();
4531 SDValue N0 = Op.getOperand(0);
4532 SDValue N1 = Op.getOperand(1);
4533 SDValue N2 = Op.getOperand(2);
4535 if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4536 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4537 // as its second argument.
4538 if (N1.getValueType() != MVT::i32)
4539 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4540 if (N2.getValueType() != MVT::i32)
4541 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4542 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4548 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4549 DebugLoc dl = Op.getDebugLoc();
4550 if (Op.getValueType() == MVT::v2f32)
4551 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4552 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4553 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4554 Op.getOperand(0))));
4556 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4557 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4559 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4560 EVT VT = MVT::v2i32;
4561 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4568 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4569 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4572 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4573 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4574 // one of the above mentioned nodes. It has to be wrapped because otherwise
4575 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4576 // be used to form addressing mode. These wrapped nodes will be selected
4579 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4580 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4582 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4584 unsigned char OpFlag = 0;
4585 unsigned WrapperKind = X86ISD::Wrapper;
4586 CodeModel::Model M = getTargetMachine().getCodeModel();
4588 if (Subtarget->isPICStyleRIPRel() &&
4589 (M == CodeModel::Small || M == CodeModel::Kernel))
4590 WrapperKind = X86ISD::WrapperRIP;
4591 else if (Subtarget->isPICStyleGOT())
4592 OpFlag = X86II::MO_GOTOFF;
4593 else if (Subtarget->isPICStyleStubPIC())
4594 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4596 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4598 CP->getOffset(), OpFlag);
4599 DebugLoc DL = CP->getDebugLoc();
4600 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4601 // With PIC, the address is actually $g + Offset.
4603 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4604 DAG.getNode(X86ISD::GlobalBaseReg,
4605 DebugLoc::getUnknownLoc(), getPointerTy()),
4612 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4613 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4615 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4617 unsigned char OpFlag = 0;
4618 unsigned WrapperKind = X86ISD::Wrapper;
4619 CodeModel::Model M = getTargetMachine().getCodeModel();
4621 if (Subtarget->isPICStyleRIPRel() &&
4622 (M == CodeModel::Small || M == CodeModel::Kernel))
4623 WrapperKind = X86ISD::WrapperRIP;
4624 else if (Subtarget->isPICStyleGOT())
4625 OpFlag = X86II::MO_GOTOFF;
4626 else if (Subtarget->isPICStyleStubPIC())
4627 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4629 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4631 DebugLoc DL = JT->getDebugLoc();
4632 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4634 // With PIC, the address is actually $g + Offset.
4636 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4637 DAG.getNode(X86ISD::GlobalBaseReg,
4638 DebugLoc::getUnknownLoc(), getPointerTy()),
4646 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4647 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4649 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4651 unsigned char OpFlag = 0;
4652 unsigned WrapperKind = X86ISD::Wrapper;
4653 CodeModel::Model M = getTargetMachine().getCodeModel();
4655 if (Subtarget->isPICStyleRIPRel() &&
4656 (M == CodeModel::Small || M == CodeModel::Kernel))
4657 WrapperKind = X86ISD::WrapperRIP;
4658 else if (Subtarget->isPICStyleGOT())
4659 OpFlag = X86II::MO_GOTOFF;
4660 else if (Subtarget->isPICStyleStubPIC())
4661 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4663 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4665 DebugLoc DL = Op.getDebugLoc();
4666 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4669 // With PIC, the address is actually $g + Offset.
4670 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4671 !Subtarget->is64Bit()) {
4672 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4673 DAG.getNode(X86ISD::GlobalBaseReg,
4674 DebugLoc::getUnknownLoc(),
4683 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4685 SelectionDAG &DAG) const {
4686 // Create the TargetGlobalAddress node, folding in the constant
4687 // offset if it is legal.
4688 unsigned char OpFlags =
4689 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4690 CodeModel::Model M = getTargetMachine().getCodeModel();
4692 if (OpFlags == X86II::MO_NO_FLAG &&
4693 X86::isOffsetSuitableForCodeModel(Offset, M)) {
4694 // A direct static reference to a global.
4695 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4698 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4701 if (Subtarget->isPICStyleRIPRel() &&
4702 (M == CodeModel::Small || M == CodeModel::Kernel))
4703 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4705 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4707 // With PIC, the address is actually $g + Offset.
4708 if (isGlobalRelativeToPICBase(OpFlags)) {
4709 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4710 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4714 // For globals that require a load from a stub to get the address, emit the
4716 if (isGlobalStubReference(OpFlags))
4717 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4718 PseudoSourceValue::getGOT(), 0);
4720 // If there was a non-zero offset that we didn't fold, create an explicit
4723 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4724 DAG.getConstant(Offset, getPointerTy()));
4730 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4731 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4732 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4733 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4737 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4738 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
4739 unsigned char OperandFlags) {
4740 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4741 DebugLoc dl = GA->getDebugLoc();
4742 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4743 GA->getValueType(0),
4747 SDValue Ops[] = { Chain, TGA, *InFlag };
4748 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4750 SDValue Ops[] = { Chain, TGA };
4751 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4753 SDValue Flag = Chain.getValue(1);
4754 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4757 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4759 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4762 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4763 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4764 DAG.getNode(X86ISD::GlobalBaseReg,
4765 DebugLoc::getUnknownLoc(),
4767 InFlag = Chain.getValue(1);
4769 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4772 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4774 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4776 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4777 X86::RAX, X86II::MO_TLSGD);
4780 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4781 // "local exec" model.
4782 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4783 const EVT PtrVT, TLSModel::Model model,
4785 DebugLoc dl = GA->getDebugLoc();
4786 // Get the Thread Pointer
4787 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4788 DebugLoc::getUnknownLoc(), PtrVT,
4789 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4792 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4795 unsigned char OperandFlags = 0;
4796 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
4798 unsigned WrapperKind = X86ISD::Wrapper;
4799 if (model == TLSModel::LocalExec) {
4800 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
4801 } else if (is64Bit) {
4802 assert(model == TLSModel::InitialExec);
4803 OperandFlags = X86II::MO_GOTTPOFF;
4804 WrapperKind = X86ISD::WrapperRIP;
4806 assert(model == TLSModel::InitialExec);
4807 OperandFlags = X86II::MO_INDNTPOFF;
4810 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4812 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4813 GA->getOffset(), OperandFlags);
4814 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
4816 if (model == TLSModel::InitialExec)
4817 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4818 PseudoSourceValue::getGOT(), 0);
4820 // The address of the thread local variable is the add of the thread
4821 // pointer with the offset of the variable.
4822 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4826 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4827 // TODO: implement the "local dynamic" model
4828 // TODO: implement the "initial exec"model for pic executables
4829 assert(Subtarget->isTargetELF() &&
4830 "TLS not implemented for non-ELF targets");
4831 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4832 const GlobalValue *GV = GA->getGlobal();
4834 // If GV is an alias then use the aliasee for determining
4835 // thread-localness.
4836 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
4837 GV = GA->resolveAliasedGlobal(false);
4839 TLSModel::Model model = getTLSModel(GV,
4840 getTargetMachine().getRelocationModel());
4843 case TLSModel::GeneralDynamic:
4844 case TLSModel::LocalDynamic: // not implemented
4845 if (Subtarget->is64Bit())
4846 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4847 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4849 case TLSModel::InitialExec:
4850 case TLSModel::LocalExec:
4851 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
4852 Subtarget->is64Bit());
4855 llvm_unreachable("Unreachable");
4860 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4861 /// take a 2 x i32 value to shift plus a shift amount.
4862 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4863 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4864 EVT VT = Op.getValueType();
4865 unsigned VTBits = VT.getSizeInBits();
4866 DebugLoc dl = Op.getDebugLoc();
4867 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4868 SDValue ShOpLo = Op.getOperand(0);
4869 SDValue ShOpHi = Op.getOperand(1);
4870 SDValue ShAmt = Op.getOperand(2);
4871 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4872 DAG.getConstant(VTBits - 1, MVT::i8))
4873 : DAG.getConstant(0, VT);
4876 if (Op.getOpcode() == ISD::SHL_PARTS) {
4877 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4878 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4880 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4881 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4884 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4885 DAG.getConstant(VTBits, MVT::i8));
4886 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4887 AndNode, DAG.getConstant(0, MVT::i8));
4890 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4891 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4892 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4894 if (Op.getOpcode() == ISD::SHL_PARTS) {
4895 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4896 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4898 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4899 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4902 SDValue Ops[2] = { Lo, Hi };
4903 return DAG.getMergeValues(Ops, 2, dl);
4906 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4907 EVT SrcVT = Op.getOperand(0).getValueType();
4909 if (SrcVT.isVector()) {
4910 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
4916 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4917 "Unknown SINT_TO_FP to lower!");
4919 // These are really Legal; return the operand so the caller accepts it as
4921 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4923 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
4924 Subtarget->is64Bit()) {
4928 DebugLoc dl = Op.getDebugLoc();
4929 unsigned Size = SrcVT.getSizeInBits()/8;
4930 MachineFunction &MF = DAG.getMachineFunction();
4931 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4932 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4933 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4935 PseudoSourceValue::getFixedStack(SSFI), 0);
4936 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
4939 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
4941 SelectionDAG &DAG) {
4943 DebugLoc dl = Op.getDebugLoc();
4945 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4947 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4949 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4950 SmallVector<SDValue, 8> Ops;
4951 Ops.push_back(Chain);
4952 Ops.push_back(StackSlot);
4953 Ops.push_back(DAG.getValueType(SrcVT));
4954 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4955 Tys, &Ops[0], Ops.size());
4958 Chain = Result.getValue(1);
4959 SDValue InFlag = Result.getValue(2);
4961 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4962 // shouldn't be necessary except that RFP cannot be live across
4963 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4964 MachineFunction &MF = DAG.getMachineFunction();
4965 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4966 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4967 Tys = DAG.getVTList(MVT::Other);
4968 SmallVector<SDValue, 8> Ops;
4969 Ops.push_back(Chain);
4970 Ops.push_back(Result);
4971 Ops.push_back(StackSlot);
4972 Ops.push_back(DAG.getValueType(Op.getValueType()));
4973 Ops.push_back(InFlag);
4974 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4975 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4976 PseudoSourceValue::getFixedStack(SSFI), 0);
4982 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4983 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4984 // This algorithm is not obvious. Here it is in C code, more or less:
4986 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4987 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4988 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4990 // Copy ints to xmm registers.
4991 __m128i xh = _mm_cvtsi32_si128( hi );
4992 __m128i xl = _mm_cvtsi32_si128( lo );
4994 // Combine into low half of a single xmm register.
4995 __m128i x = _mm_unpacklo_epi32( xh, xl );
4999 // Merge in appropriate exponents to give the integer bits the right
5001 x = _mm_unpacklo_epi32( x, exp );
5003 // Subtract away the biases to deal with the IEEE-754 double precision
5005 d = _mm_sub_pd( (__m128d) x, bias );
5007 // All conversions up to here are exact. The correctly rounded result is
5008 // calculated using the current rounding mode using the following
5010 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
5011 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
5012 // store doesn't really need to be here (except
5013 // maybe to zero the other double)
5018 DebugLoc dl = Op.getDebugLoc();
5019 LLVMContext *Context = DAG.getContext();
5021 // Build some magic constants.
5022 std::vector<Constant*> CV0;
5023 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
5024 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
5025 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5026 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
5027 Constant *C0 = ConstantVector::get(CV0);
5028 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
5030 std::vector<Constant*> CV1;
5032 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
5034 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
5035 Constant *C1 = ConstantVector::get(CV1);
5036 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
5038 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5039 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5041 DAG.getIntPtrConstant(1)));
5042 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5043 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5045 DAG.getIntPtrConstant(0)));
5046 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
5047 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
5048 PseudoSourceValue::getConstantPool(), 0,
5050 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
5051 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
5052 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
5053 PseudoSourceValue::getConstantPool(), 0,
5055 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
5057 // Add the halves; easiest way is to swap them into another reg first.
5058 int ShufMask[2] = { 1, -1 };
5059 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
5060 DAG.getUNDEF(MVT::v2f64), ShufMask);
5061 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
5062 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
5063 DAG.getIntPtrConstant(0));
5066 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
5067 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
5068 DebugLoc dl = Op.getDebugLoc();
5069 // FP constant to bias correct the final result.
5070 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
5073 // Load the 32-bit value into an XMM register.
5074 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
5075 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
5077 DAG.getIntPtrConstant(0)));
5079 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5080 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
5081 DAG.getIntPtrConstant(0));
5083 // Or the load with the bias.
5084 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
5085 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5086 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5088 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5089 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
5090 MVT::v2f64, Bias)));
5091 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5092 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5093 DAG.getIntPtrConstant(0));
5095 // Subtract the bias.
5096 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5098 // Handle final rounding.
5099 EVT DestVT = Op.getValueType();
5101 if (DestVT.bitsLT(MVT::f64)) {
5102 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5103 DAG.getIntPtrConstant(0));
5104 } else if (DestVT.bitsGT(MVT::f64)) {
5105 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5108 // Handle final rounding.
5112 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5113 SDValue N0 = Op.getOperand(0);
5114 DebugLoc dl = Op.getDebugLoc();
5116 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5117 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5118 // the optimization here.
5119 if (DAG.SignBitIsZero(N0))
5120 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5122 EVT SrcVT = N0.getValueType();
5123 if (SrcVT == MVT::i64) {
5124 // We only handle SSE2 f64 target here; caller can expand the rest.
5125 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5128 return LowerUINT_TO_FP_i64(Op, DAG);
5129 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5130 return LowerUINT_TO_FP_i32(Op, DAG);
5133 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5135 // Make a 64-bit buffer, and use it to build an FILD.
5136 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5137 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5138 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5139 getPointerTy(), StackSlot, WordOff);
5140 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5141 StackSlot, NULL, 0);
5142 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5143 OffsetSlot, NULL, 0);
5144 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5147 std::pair<SDValue,SDValue> X86TargetLowering::
5148 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5149 DebugLoc dl = Op.getDebugLoc();
5151 EVT DstTy = Op.getValueType();
5154 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5158 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5159 DstTy.getSimpleVT() >= MVT::i16 &&
5160 "Unknown FP_TO_SINT to lower!");
5162 // These are really Legal.
5163 if (DstTy == MVT::i32 &&
5164 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5165 return std::make_pair(SDValue(), SDValue());
5166 if (Subtarget->is64Bit() &&
5167 DstTy == MVT::i64 &&
5168 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5169 return std::make_pair(SDValue(), SDValue());
5171 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5173 MachineFunction &MF = DAG.getMachineFunction();
5174 unsigned MemSize = DstTy.getSizeInBits()/8;
5175 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5176 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5179 switch (DstTy.getSimpleVT().SimpleTy) {
5180 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5181 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5182 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5183 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5186 SDValue Chain = DAG.getEntryNode();
5187 SDValue Value = Op.getOperand(0);
5188 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5189 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5190 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5191 PseudoSourceValue::getFixedStack(SSFI), 0);
5192 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5194 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5196 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5197 Chain = Value.getValue(1);
5198 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5199 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5202 // Build the FP_TO_INT*_IN_MEM
5203 SDValue Ops[] = { Chain, Value, StackSlot };
5204 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5206 return std::make_pair(FIST, StackSlot);
5209 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5210 if (Op.getValueType().isVector()) {
5211 if (Op.getValueType() == MVT::v2i32 &&
5212 Op.getOperand(0).getValueType() == MVT::v2f64) {
5218 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5219 SDValue FIST = Vals.first, StackSlot = Vals.second;
5220 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5221 if (FIST.getNode() == 0) return Op;
5224 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5225 FIST, StackSlot, NULL, 0);
5228 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5229 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5230 SDValue FIST = Vals.first, StackSlot = Vals.second;
5231 assert(FIST.getNode() && "Unexpected failure");
5234 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5235 FIST, StackSlot, NULL, 0);
5238 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5239 LLVMContext *Context = DAG.getContext();
5240 DebugLoc dl = Op.getDebugLoc();
5241 EVT VT = Op.getValueType();
5244 EltVT = VT.getVectorElementType();
5245 std::vector<Constant*> CV;
5246 if (EltVT == MVT::f64) {
5247 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5251 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5257 Constant *C = ConstantVector::get(CV);
5258 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5259 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5260 PseudoSourceValue::getConstantPool(), 0,
5262 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5265 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5266 LLVMContext *Context = DAG.getContext();
5267 DebugLoc dl = Op.getDebugLoc();
5268 EVT VT = Op.getValueType();
5271 EltVT = VT.getVectorElementType();
5272 std::vector<Constant*> CV;
5273 if (EltVT == MVT::f64) {
5274 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5278 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5284 Constant *C = ConstantVector::get(CV);
5285 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5286 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5287 PseudoSourceValue::getConstantPool(), 0,
5289 if (VT.isVector()) {
5290 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5291 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5292 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5294 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5296 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5300 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5301 LLVMContext *Context = DAG.getContext();
5302 SDValue Op0 = Op.getOperand(0);
5303 SDValue Op1 = Op.getOperand(1);
5304 DebugLoc dl = Op.getDebugLoc();
5305 EVT VT = Op.getValueType();
5306 EVT SrcVT = Op1.getValueType();
5308 // If second operand is smaller, extend it first.
5309 if (SrcVT.bitsLT(VT)) {
5310 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5313 // And if it is bigger, shrink it first.
5314 if (SrcVT.bitsGT(VT)) {
5315 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5319 // At this point the operands and the result should have the same
5320 // type, and that won't be f80 since that is not custom lowered.
5322 // First get the sign bit of second operand.
5323 std::vector<Constant*> CV;
5324 if (SrcVT == MVT::f64) {
5325 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5326 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5328 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5329 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5330 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5331 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5333 Constant *C = ConstantVector::get(CV);
5334 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5335 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5336 PseudoSourceValue::getConstantPool(), 0,
5338 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5340 // Shift sign bit right or left if the two operands have different types.
5341 if (SrcVT.bitsGT(VT)) {
5342 // Op0 is MVT::f32, Op1 is MVT::f64.
5343 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5344 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5345 DAG.getConstant(32, MVT::i32));
5346 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5347 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5348 DAG.getIntPtrConstant(0));
5351 // Clear first operand sign bit.
5353 if (VT == MVT::f64) {
5354 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5355 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5357 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5358 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5359 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5360 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5362 C = ConstantVector::get(CV);
5363 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5364 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5365 PseudoSourceValue::getConstantPool(), 0,
5367 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5369 // Or the value with the sign bit.
5370 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5373 /// Emit nodes that will be selected as "test Op0,Op0", or something
5375 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5376 SelectionDAG &DAG) {
5377 DebugLoc dl = Op.getDebugLoc();
5379 // CF and OF aren't always set the way we want. Determine which
5380 // of these we need.
5381 bool NeedCF = false;
5382 bool NeedOF = false;
5384 case X86::COND_A: case X86::COND_AE:
5385 case X86::COND_B: case X86::COND_BE:
5388 case X86::COND_G: case X86::COND_GE:
5389 case X86::COND_L: case X86::COND_LE:
5390 case X86::COND_O: case X86::COND_NO:
5396 // See if we can use the EFLAGS value from the operand instead of
5397 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5398 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5399 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5400 unsigned Opcode = 0;
5401 unsigned NumOperands = 0;
5402 switch (Op.getNode()->getOpcode()) {
5404 // Due to an isel shortcoming, be conservative if this add is likely to
5405 // be selected as part of a load-modify-store instruction. When the root
5406 // node in a match is a store, isel doesn't know how to remap non-chain
5407 // non-flag uses of other nodes in the match, such as the ADD in this
5408 // case. This leads to the ADD being left around and reselected, with
5409 // the result being two adds in the output.
5410 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5411 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5412 if (UI->getOpcode() == ISD::STORE)
5414 if (ConstantSDNode *C =
5415 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5416 // An add of one will be selected as an INC.
5417 if (C->getAPIntValue() == 1) {
5418 Opcode = X86ISD::INC;
5422 // An add of negative one (subtract of one) will be selected as a DEC.
5423 if (C->getAPIntValue().isAllOnesValue()) {
5424 Opcode = X86ISD::DEC;
5429 // Otherwise use a regular EFLAGS-setting add.
5430 Opcode = X86ISD::ADD;
5434 // If the primary and result isn't used, don't bother using X86ISD::AND,
5435 // because a TEST instruction will be better.
5436 bool NonFlagUse = false;
5437 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5438 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5439 if (UI->getOpcode() != ISD::BRCOND &&
5440 UI->getOpcode() != ISD::SELECT &&
5441 UI->getOpcode() != ISD::SETCC) {
5452 // Due to the ISEL shortcoming noted above, be conservative if this op is
5453 // likely to be selected as part of a load-modify-store instruction.
5454 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5455 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5456 if (UI->getOpcode() == ISD::STORE)
5458 // Otherwise use a regular EFLAGS-setting instruction.
5459 switch (Op.getNode()->getOpcode()) {
5460 case ISD::SUB: Opcode = X86ISD::SUB; break;
5461 case ISD::OR: Opcode = X86ISD::OR; break;
5462 case ISD::XOR: Opcode = X86ISD::XOR; break;
5463 case ISD::AND: Opcode = X86ISD::AND; break;
5464 default: llvm_unreachable("unexpected operator!");
5475 return SDValue(Op.getNode(), 1);
5481 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5482 SmallVector<SDValue, 4> Ops;
5483 for (unsigned i = 0; i != NumOperands; ++i)
5484 Ops.push_back(Op.getOperand(i));
5485 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5486 DAG.ReplaceAllUsesWith(Op, New);
5487 return SDValue(New.getNode(), 1);
5491 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5492 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5493 DAG.getConstant(0, Op.getValueType()));
5496 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5498 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5499 SelectionDAG &DAG) {
5500 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5501 if (C->getAPIntValue() == 0)
5502 return EmitTest(Op0, X86CC, DAG);
5504 DebugLoc dl = Op0.getDebugLoc();
5505 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5508 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5509 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5510 SDValue Op0 = Op.getOperand(0);
5511 SDValue Op1 = Op.getOperand(1);
5512 DebugLoc dl = Op.getDebugLoc();
5513 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5515 // Lower (X & (1 << N)) == 0 to BT(X, N).
5516 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5517 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5518 if (Op0.getOpcode() == ISD::AND &&
5520 Op1.getOpcode() == ISD::Constant &&
5521 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5522 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5524 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5525 if (ConstantSDNode *Op010C =
5526 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5527 if (Op010C->getZExtValue() == 1) {
5528 LHS = Op0.getOperand(0);
5529 RHS = Op0.getOperand(1).getOperand(1);
5531 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5532 if (ConstantSDNode *Op000C =
5533 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5534 if (Op000C->getZExtValue() == 1) {
5535 LHS = Op0.getOperand(1);
5536 RHS = Op0.getOperand(0).getOperand(1);
5538 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5539 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5540 SDValue AndLHS = Op0.getOperand(0);
5541 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5542 LHS = AndLHS.getOperand(0);
5543 RHS = AndLHS.getOperand(1);
5547 if (LHS.getNode()) {
5548 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5549 // instruction. Since the shift amount is in-range-or-undefined, we know
5550 // that doing a bittest on the i16 value is ok. We extend to i32 because
5551 // the encoding for the i16 version is larger than the i32 version.
5552 if (LHS.getValueType() == MVT::i8)
5553 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5555 // If the operand types disagree, extend the shift amount to match. Since
5556 // BT ignores high bits (like shifts) we can use anyextend.
5557 if (LHS.getValueType() != RHS.getValueType())
5558 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5560 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5561 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5562 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5563 DAG.getConstant(Cond, MVT::i8), BT);
5567 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5568 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5570 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5571 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5572 DAG.getConstant(X86CC, MVT::i8), Cond);
5575 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5577 SDValue Op0 = Op.getOperand(0);
5578 SDValue Op1 = Op.getOperand(1);
5579 SDValue CC = Op.getOperand(2);
5580 EVT VT = Op.getValueType();
5581 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5582 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5583 DebugLoc dl = Op.getDebugLoc();
5587 EVT VT0 = Op0.getValueType();
5588 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5589 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5592 switch (SetCCOpcode) {
5595 case ISD::SETEQ: SSECC = 0; break;
5597 case ISD::SETGT: Swap = true; // Fallthrough
5599 case ISD::SETOLT: SSECC = 1; break;
5601 case ISD::SETGE: Swap = true; // Fallthrough
5603 case ISD::SETOLE: SSECC = 2; break;
5604 case ISD::SETUO: SSECC = 3; break;
5606 case ISD::SETNE: SSECC = 4; break;
5607 case ISD::SETULE: Swap = true;
5608 case ISD::SETUGE: SSECC = 5; break;
5609 case ISD::SETULT: Swap = true;
5610 case ISD::SETUGT: SSECC = 6; break;
5611 case ISD::SETO: SSECC = 7; break;
5614 std::swap(Op0, Op1);
5616 // In the two special cases we can't handle, emit two comparisons.
5618 if (SetCCOpcode == ISD::SETUEQ) {
5620 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5621 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5622 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5624 else if (SetCCOpcode == ISD::SETONE) {
5626 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5627 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5628 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5630 llvm_unreachable("Illegal FP comparison");
5632 // Handle all other FP comparisons here.
5633 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5636 // We are handling one of the integer comparisons here. Since SSE only has
5637 // GT and EQ comparisons for integer, swapping operands and multiple
5638 // operations may be required for some comparisons.
5639 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5640 bool Swap = false, Invert = false, FlipSigns = false;
5642 switch (VT.getSimpleVT().SimpleTy) {
5645 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5647 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5649 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5650 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5653 switch (SetCCOpcode) {
5655 case ISD::SETNE: Invert = true;
5656 case ISD::SETEQ: Opc = EQOpc; break;
5657 case ISD::SETLT: Swap = true;
5658 case ISD::SETGT: Opc = GTOpc; break;
5659 case ISD::SETGE: Swap = true;
5660 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5661 case ISD::SETULT: Swap = true;
5662 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5663 case ISD::SETUGE: Swap = true;
5664 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5667 std::swap(Op0, Op1);
5669 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5670 // bits of the inputs before performing those operations.
5672 EVT EltVT = VT.getVectorElementType();
5673 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5675 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5676 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5678 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5679 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5682 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5684 // If the logical-not of the result is required, perform that now.
5686 Result = DAG.getNOT(dl, Result, VT);
5691 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5692 static bool isX86LogicalCmp(SDValue Op) {
5693 unsigned Opc = Op.getNode()->getOpcode();
5694 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5696 if (Op.getResNo() == 1 &&
5697 (Opc == X86ISD::ADD ||
5698 Opc == X86ISD::SUB ||
5699 Opc == X86ISD::SMUL ||
5700 Opc == X86ISD::UMUL ||
5701 Opc == X86ISD::INC ||
5702 Opc == X86ISD::DEC ||
5703 Opc == X86ISD::OR ||
5704 Opc == X86ISD::XOR ||
5705 Opc == X86ISD::AND))
5711 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5712 bool addTest = true;
5713 SDValue Cond = Op.getOperand(0);
5714 DebugLoc dl = Op.getDebugLoc();
5717 if (Cond.getOpcode() == ISD::SETCC)
5718 Cond = LowerSETCC(Cond, DAG);
5720 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5721 // setting operand in place of the X86ISD::SETCC.
5722 if (Cond.getOpcode() == X86ISD::SETCC) {
5723 CC = Cond.getOperand(0);
5725 SDValue Cmp = Cond.getOperand(1);
5726 unsigned Opc = Cmp.getOpcode();
5727 EVT VT = Op.getValueType();
5729 bool IllegalFPCMov = false;
5730 if (VT.isFloatingPoint() && !VT.isVector() &&
5731 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5732 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5734 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5735 Opc == X86ISD::BT) { // FIXME
5742 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5743 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5746 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5747 SmallVector<SDValue, 4> Ops;
5748 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5749 // condition is true.
5750 Ops.push_back(Op.getOperand(2));
5751 Ops.push_back(Op.getOperand(1));
5753 Ops.push_back(Cond);
5754 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5757 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5758 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5759 // from the AND / OR.
5760 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5761 Opc = Op.getOpcode();
5762 if (Opc != ISD::OR && Opc != ISD::AND)
5764 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5765 Op.getOperand(0).hasOneUse() &&
5766 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5767 Op.getOperand(1).hasOneUse());
5770 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5771 // 1 and that the SETCC node has a single use.
5772 static bool isXor1OfSetCC(SDValue Op) {
5773 if (Op.getOpcode() != ISD::XOR)
5775 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5776 if (N1C && N1C->getAPIntValue() == 1) {
5777 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5778 Op.getOperand(0).hasOneUse();
5783 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5784 bool addTest = true;
5785 SDValue Chain = Op.getOperand(0);
5786 SDValue Cond = Op.getOperand(1);
5787 SDValue Dest = Op.getOperand(2);
5788 DebugLoc dl = Op.getDebugLoc();
5791 if (Cond.getOpcode() == ISD::SETCC)
5792 Cond = LowerSETCC(Cond, DAG);
5794 // FIXME: LowerXALUO doesn't handle these!!
5795 else if (Cond.getOpcode() == X86ISD::ADD ||
5796 Cond.getOpcode() == X86ISD::SUB ||
5797 Cond.getOpcode() == X86ISD::SMUL ||
5798 Cond.getOpcode() == X86ISD::UMUL)
5799 Cond = LowerXALUO(Cond, DAG);
5802 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5803 // setting operand in place of the X86ISD::SETCC.
5804 if (Cond.getOpcode() == X86ISD::SETCC) {
5805 CC = Cond.getOperand(0);
5807 SDValue Cmp = Cond.getOperand(1);
5808 unsigned Opc = Cmp.getOpcode();
5809 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5810 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5814 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5818 // These can only come from an arithmetic instruction with overflow,
5819 // e.g. SADDO, UADDO.
5820 Cond = Cond.getNode()->getOperand(1);
5827 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5828 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5829 if (CondOpc == ISD::OR) {
5830 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5831 // two branches instead of an explicit OR instruction with a
5833 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5834 isX86LogicalCmp(Cmp)) {
5835 CC = Cond.getOperand(0).getOperand(0);
5836 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5837 Chain, Dest, CC, Cmp);
5838 CC = Cond.getOperand(1).getOperand(0);
5842 } else { // ISD::AND
5843 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5844 // two branches instead of an explicit AND instruction with a
5845 // separate test. However, we only do this if this block doesn't
5846 // have a fall-through edge, because this requires an explicit
5847 // jmp when the condition is false.
5848 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5849 isX86LogicalCmp(Cmp) &&
5850 Op.getNode()->hasOneUse()) {
5851 X86::CondCode CCode =
5852 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5853 CCode = X86::GetOppositeBranchCondition(CCode);
5854 CC = DAG.getConstant(CCode, MVT::i8);
5855 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5856 // Look for an unconditional branch following this conditional branch.
5857 // We need this because we need to reverse the successors in order
5858 // to implement FCMP_OEQ.
5859 if (User.getOpcode() == ISD::BR) {
5860 SDValue FalseBB = User.getOperand(1);
5862 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5863 assert(NewBR == User);
5866 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5867 Chain, Dest, CC, Cmp);
5868 X86::CondCode CCode =
5869 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5870 CCode = X86::GetOppositeBranchCondition(CCode);
5871 CC = DAG.getConstant(CCode, MVT::i8);
5877 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5878 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5879 // It should be transformed during dag combiner except when the condition
5880 // is set by a arithmetics with overflow node.
5881 X86::CondCode CCode =
5882 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5883 CCode = X86::GetOppositeBranchCondition(CCode);
5884 CC = DAG.getConstant(CCode, MVT::i8);
5885 Cond = Cond.getOperand(0).getOperand(1);
5891 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5892 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5894 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5895 Chain, Dest, CC, Cond);
5899 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5900 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5901 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5902 // that the guard pages used by the OS virtual memory manager are allocated in
5903 // correct sequence.
5905 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5906 SelectionDAG &DAG) {
5907 assert(Subtarget->isTargetCygMing() &&
5908 "This should be used only on Cygwin/Mingw targets");
5909 DebugLoc dl = Op.getDebugLoc();
5912 SDValue Chain = Op.getOperand(0);
5913 SDValue Size = Op.getOperand(1);
5914 // FIXME: Ensure alignment here
5918 EVT IntPtr = getPointerTy();
5919 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5921 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5923 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5924 Flag = Chain.getValue(1);
5926 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5927 SDValue Ops[] = { Chain,
5928 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5929 DAG.getRegister(X86::EAX, IntPtr),
5930 DAG.getRegister(X86StackPtr, SPTy),
5932 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5933 Flag = Chain.getValue(1);
5935 Chain = DAG.getCALLSEQ_END(Chain,
5936 DAG.getIntPtrConstant(0, true),
5937 DAG.getIntPtrConstant(0, true),
5940 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5942 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5943 return DAG.getMergeValues(Ops1, 2, dl);
5947 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5949 SDValue Dst, SDValue Src,
5950 SDValue Size, unsigned Align,
5952 uint64_t DstSVOff) {
5953 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5955 // If not DWORD aligned or size is more than the threshold, call the library.
5956 // The libc version is likely to be faster for these cases. It can use the
5957 // address value and run time information about the CPU.
5958 if ((Align & 3) != 0 ||
5960 ConstantSize->getZExtValue() >
5961 getSubtarget()->getMaxInlineSizeThreshold()) {
5962 SDValue InFlag(0, 0);
5964 // Check to see if there is a specialized entry-point for memory zeroing.
5965 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5967 if (const char *bzeroEntry = V &&
5968 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5969 EVT IntPtr = getPointerTy();
5970 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
5971 TargetLowering::ArgListTy Args;
5972 TargetLowering::ArgListEntry Entry;
5974 Entry.Ty = IntPtrTy;
5975 Args.push_back(Entry);
5977 Args.push_back(Entry);
5978 std::pair<SDValue,SDValue> CallResult =
5979 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
5980 false, false, false, false,
5981 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
5982 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5983 return CallResult.second;
5986 // Otherwise have the target-independent code call memset.
5990 uint64_t SizeVal = ConstantSize->getZExtValue();
5991 SDValue InFlag(0, 0);
5994 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5995 unsigned BytesLeft = 0;
5996 bool TwoRepStos = false;
5999 uint64_t Val = ValC->getZExtValue() & 255;
6001 // If the value is a constant, then we can potentially use larger sets.
6002 switch (Align & 3) {
6003 case 2: // WORD aligned
6006 Val = (Val << 8) | Val;
6008 case 0: // DWORD aligned
6011 Val = (Val << 8) | Val;
6012 Val = (Val << 16) | Val;
6013 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
6016 Val = (Val << 32) | Val;
6019 default: // Byte aligned
6022 Count = DAG.getIntPtrConstant(SizeVal);
6026 if (AVT.bitsGT(MVT::i8)) {
6027 unsigned UBytes = AVT.getSizeInBits() / 8;
6028 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
6029 BytesLeft = SizeVal % UBytes;
6032 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
6034 InFlag = Chain.getValue(1);
6037 Count = DAG.getIntPtrConstant(SizeVal);
6038 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
6039 InFlag = Chain.getValue(1);
6042 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6045 InFlag = Chain.getValue(1);
6046 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6049 InFlag = Chain.getValue(1);
6051 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6052 SmallVector<SDValue, 8> Ops;
6053 Ops.push_back(Chain);
6054 Ops.push_back(DAG.getValueType(AVT));
6055 Ops.push_back(InFlag);
6056 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6059 InFlag = Chain.getValue(1);
6061 EVT CVT = Count.getValueType();
6062 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
6063 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
6064 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
6067 InFlag = Chain.getValue(1);
6068 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6070 Ops.push_back(Chain);
6071 Ops.push_back(DAG.getValueType(MVT::i8));
6072 Ops.push_back(InFlag);
6073 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
6074 } else if (BytesLeft) {
6075 // Handle the last 1 - 7 bytes.
6076 unsigned Offset = SizeVal - BytesLeft;
6077 EVT AddrVT = Dst.getValueType();
6078 EVT SizeVT = Size.getValueType();
6080 Chain = DAG.getMemset(Chain, dl,
6081 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
6082 DAG.getConstant(Offset, AddrVT)),
6084 DAG.getConstant(BytesLeft, SizeVT),
6085 Align, DstSV, DstSVOff + Offset);
6088 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
6093 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
6094 SDValue Chain, SDValue Dst, SDValue Src,
6095 SDValue Size, unsigned Align,
6097 const Value *DstSV, uint64_t DstSVOff,
6098 const Value *SrcSV, uint64_t SrcSVOff) {
6099 // This requires the copy size to be a constant, preferrably
6100 // within a subtarget-specific limit.
6101 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
6104 uint64_t SizeVal = ConstantSize->getZExtValue();
6105 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
6108 /// If not DWORD aligned, call the library.
6109 if ((Align & 3) != 0)
6114 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
6117 unsigned UBytes = AVT.getSizeInBits() / 8;
6118 unsigned CountVal = SizeVal / UBytes;
6119 SDValue Count = DAG.getIntPtrConstant(CountVal);
6120 unsigned BytesLeft = SizeVal % UBytes;
6122 SDValue InFlag(0, 0);
6123 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6126 InFlag = Chain.getValue(1);
6127 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6130 InFlag = Chain.getValue(1);
6131 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6134 InFlag = Chain.getValue(1);
6136 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6137 SmallVector<SDValue, 8> Ops;
6138 Ops.push_back(Chain);
6139 Ops.push_back(DAG.getValueType(AVT));
6140 Ops.push_back(InFlag);
6141 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
6143 SmallVector<SDValue, 4> Results;
6144 Results.push_back(RepMovs);
6146 // Handle the last 1 - 7 bytes.
6147 unsigned Offset = SizeVal - BytesLeft;
6148 EVT DstVT = Dst.getValueType();
6149 EVT SrcVT = Src.getValueType();
6150 EVT SizeVT = Size.getValueType();
6151 Results.push_back(DAG.getMemcpy(Chain, dl,
6152 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6153 DAG.getConstant(Offset, DstVT)),
6154 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6155 DAG.getConstant(Offset, SrcVT)),
6156 DAG.getConstant(BytesLeft, SizeVT),
6157 Align, AlwaysInline,
6158 DstSV, DstSVOff + Offset,
6159 SrcSV, SrcSVOff + Offset));
6162 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6163 &Results[0], Results.size());
6166 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6167 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6168 DebugLoc dl = Op.getDebugLoc();
6170 if (!Subtarget->is64Bit()) {
6171 // vastart just stores the address of the VarArgsFrameIndex slot into the
6172 // memory location argument.
6173 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6174 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6178 // gp_offset (0 - 6 * 8)
6179 // fp_offset (48 - 48 + 8 * 16)
6180 // overflow_arg_area (point to parameters coming in memory).
6182 SmallVector<SDValue, 8> MemOps;
6183 SDValue FIN = Op.getOperand(1);
6185 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6186 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6188 MemOps.push_back(Store);
6191 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6192 FIN, DAG.getIntPtrConstant(4));
6193 Store = DAG.getStore(Op.getOperand(0), dl,
6194 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6196 MemOps.push_back(Store);
6198 // Store ptr to overflow_arg_area
6199 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6200 FIN, DAG.getIntPtrConstant(4));
6201 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6202 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6203 MemOps.push_back(Store);
6205 // Store ptr to reg_save_area.
6206 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6207 FIN, DAG.getIntPtrConstant(8));
6208 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6209 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6210 MemOps.push_back(Store);
6211 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6212 &MemOps[0], MemOps.size());
6215 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6216 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6217 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6218 SDValue Chain = Op.getOperand(0);
6219 SDValue SrcPtr = Op.getOperand(1);
6220 SDValue SrcSV = Op.getOperand(2);
6222 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6226 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6227 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6228 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6229 SDValue Chain = Op.getOperand(0);
6230 SDValue DstPtr = Op.getOperand(1);
6231 SDValue SrcPtr = Op.getOperand(2);
6232 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6233 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6234 DebugLoc dl = Op.getDebugLoc();
6236 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6237 DAG.getIntPtrConstant(24), 8, false,
6238 DstSV, 0, SrcSV, 0);
6242 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6243 DebugLoc dl = Op.getDebugLoc();
6244 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6246 default: return SDValue(); // Don't custom lower most intrinsics.
6247 // Comparison intrinsics.
6248 case Intrinsic::x86_sse_comieq_ss:
6249 case Intrinsic::x86_sse_comilt_ss:
6250 case Intrinsic::x86_sse_comile_ss:
6251 case Intrinsic::x86_sse_comigt_ss:
6252 case Intrinsic::x86_sse_comige_ss:
6253 case Intrinsic::x86_sse_comineq_ss:
6254 case Intrinsic::x86_sse_ucomieq_ss:
6255 case Intrinsic::x86_sse_ucomilt_ss:
6256 case Intrinsic::x86_sse_ucomile_ss:
6257 case Intrinsic::x86_sse_ucomigt_ss:
6258 case Intrinsic::x86_sse_ucomige_ss:
6259 case Intrinsic::x86_sse_ucomineq_ss:
6260 case Intrinsic::x86_sse2_comieq_sd:
6261 case Intrinsic::x86_sse2_comilt_sd:
6262 case Intrinsic::x86_sse2_comile_sd:
6263 case Intrinsic::x86_sse2_comigt_sd:
6264 case Intrinsic::x86_sse2_comige_sd:
6265 case Intrinsic::x86_sse2_comineq_sd:
6266 case Intrinsic::x86_sse2_ucomieq_sd:
6267 case Intrinsic::x86_sse2_ucomilt_sd:
6268 case Intrinsic::x86_sse2_ucomile_sd:
6269 case Intrinsic::x86_sse2_ucomigt_sd:
6270 case Intrinsic::x86_sse2_ucomige_sd:
6271 case Intrinsic::x86_sse2_ucomineq_sd: {
6273 ISD::CondCode CC = ISD::SETCC_INVALID;
6276 case Intrinsic::x86_sse_comieq_ss:
6277 case Intrinsic::x86_sse2_comieq_sd:
6281 case Intrinsic::x86_sse_comilt_ss:
6282 case Intrinsic::x86_sse2_comilt_sd:
6286 case Intrinsic::x86_sse_comile_ss:
6287 case Intrinsic::x86_sse2_comile_sd:
6291 case Intrinsic::x86_sse_comigt_ss:
6292 case Intrinsic::x86_sse2_comigt_sd:
6296 case Intrinsic::x86_sse_comige_ss:
6297 case Intrinsic::x86_sse2_comige_sd:
6301 case Intrinsic::x86_sse_comineq_ss:
6302 case Intrinsic::x86_sse2_comineq_sd:
6306 case Intrinsic::x86_sse_ucomieq_ss:
6307 case Intrinsic::x86_sse2_ucomieq_sd:
6308 Opc = X86ISD::UCOMI;
6311 case Intrinsic::x86_sse_ucomilt_ss:
6312 case Intrinsic::x86_sse2_ucomilt_sd:
6313 Opc = X86ISD::UCOMI;
6316 case Intrinsic::x86_sse_ucomile_ss:
6317 case Intrinsic::x86_sse2_ucomile_sd:
6318 Opc = X86ISD::UCOMI;
6321 case Intrinsic::x86_sse_ucomigt_ss:
6322 case Intrinsic::x86_sse2_ucomigt_sd:
6323 Opc = X86ISD::UCOMI;
6326 case Intrinsic::x86_sse_ucomige_ss:
6327 case Intrinsic::x86_sse2_ucomige_sd:
6328 Opc = X86ISD::UCOMI;
6331 case Intrinsic::x86_sse_ucomineq_ss:
6332 case Intrinsic::x86_sse2_ucomineq_sd:
6333 Opc = X86ISD::UCOMI;
6338 SDValue LHS = Op.getOperand(1);
6339 SDValue RHS = Op.getOperand(2);
6340 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6341 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6342 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6343 DAG.getConstant(X86CC, MVT::i8), Cond);
6344 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6346 // ptest intrinsics. The intrinsic these come from are designed to return
6347 // an integer value, not just an instruction so lower it to the ptest
6348 // pattern and a setcc for the result.
6349 case Intrinsic::x86_sse41_ptestz:
6350 case Intrinsic::x86_sse41_ptestc:
6351 case Intrinsic::x86_sse41_ptestnzc:{
6354 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6355 case Intrinsic::x86_sse41_ptestz:
6357 X86CC = X86::COND_E;
6359 case Intrinsic::x86_sse41_ptestc:
6361 X86CC = X86::COND_B;
6363 case Intrinsic::x86_sse41_ptestnzc:
6365 X86CC = X86::COND_A;
6369 SDValue LHS = Op.getOperand(1);
6370 SDValue RHS = Op.getOperand(2);
6371 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6372 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6373 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6374 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6377 // Fix vector shift instructions where the last operand is a non-immediate
6379 case Intrinsic::x86_sse2_pslli_w:
6380 case Intrinsic::x86_sse2_pslli_d:
6381 case Intrinsic::x86_sse2_pslli_q:
6382 case Intrinsic::x86_sse2_psrli_w:
6383 case Intrinsic::x86_sse2_psrli_d:
6384 case Intrinsic::x86_sse2_psrli_q:
6385 case Intrinsic::x86_sse2_psrai_w:
6386 case Intrinsic::x86_sse2_psrai_d:
6387 case Intrinsic::x86_mmx_pslli_w:
6388 case Intrinsic::x86_mmx_pslli_d:
6389 case Intrinsic::x86_mmx_pslli_q:
6390 case Intrinsic::x86_mmx_psrli_w:
6391 case Intrinsic::x86_mmx_psrli_d:
6392 case Intrinsic::x86_mmx_psrli_q:
6393 case Intrinsic::x86_mmx_psrai_w:
6394 case Intrinsic::x86_mmx_psrai_d: {
6395 SDValue ShAmt = Op.getOperand(2);
6396 if (isa<ConstantSDNode>(ShAmt))
6399 unsigned NewIntNo = 0;
6400 EVT ShAmtVT = MVT::v4i32;
6402 case Intrinsic::x86_sse2_pslli_w:
6403 NewIntNo = Intrinsic::x86_sse2_psll_w;
6405 case Intrinsic::x86_sse2_pslli_d:
6406 NewIntNo = Intrinsic::x86_sse2_psll_d;
6408 case Intrinsic::x86_sse2_pslli_q:
6409 NewIntNo = Intrinsic::x86_sse2_psll_q;
6411 case Intrinsic::x86_sse2_psrli_w:
6412 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6414 case Intrinsic::x86_sse2_psrli_d:
6415 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6417 case Intrinsic::x86_sse2_psrli_q:
6418 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6420 case Intrinsic::x86_sse2_psrai_w:
6421 NewIntNo = Intrinsic::x86_sse2_psra_w;
6423 case Intrinsic::x86_sse2_psrai_d:
6424 NewIntNo = Intrinsic::x86_sse2_psra_d;
6427 ShAmtVT = MVT::v2i32;
6429 case Intrinsic::x86_mmx_pslli_w:
6430 NewIntNo = Intrinsic::x86_mmx_psll_w;
6432 case Intrinsic::x86_mmx_pslli_d:
6433 NewIntNo = Intrinsic::x86_mmx_psll_d;
6435 case Intrinsic::x86_mmx_pslli_q:
6436 NewIntNo = Intrinsic::x86_mmx_psll_q;
6438 case Intrinsic::x86_mmx_psrli_w:
6439 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6441 case Intrinsic::x86_mmx_psrli_d:
6442 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6444 case Intrinsic::x86_mmx_psrli_q:
6445 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6447 case Intrinsic::x86_mmx_psrai_w:
6448 NewIntNo = Intrinsic::x86_mmx_psra_w;
6450 case Intrinsic::x86_mmx_psrai_d:
6451 NewIntNo = Intrinsic::x86_mmx_psra_d;
6453 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6459 // The vector shift intrinsics with scalars uses 32b shift amounts but
6460 // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
6464 ShOps[1] = DAG.getConstant(0, MVT::i32);
6465 if (ShAmtVT == MVT::v4i32) {
6466 ShOps[2] = DAG.getUNDEF(MVT::i32);
6467 ShOps[3] = DAG.getUNDEF(MVT::i32);
6468 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 4);
6470 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
6473 EVT VT = Op.getValueType();
6474 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
6475 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6476 DAG.getConstant(NewIntNo, MVT::i32),
6477 Op.getOperand(1), ShAmt);
6482 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6483 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6484 DebugLoc dl = Op.getDebugLoc();
6487 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6489 DAG.getConstant(TD->getPointerSize(),
6490 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6491 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6492 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6497 // Just load the return address.
6498 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6499 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6500 RetAddrFI, NULL, 0);
6503 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6504 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6505 MFI->setFrameAddressIsTaken(true);
6506 EVT VT = Op.getValueType();
6507 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6508 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6509 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6510 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6512 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6516 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6517 SelectionDAG &DAG) {
6518 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6521 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6523 MachineFunction &MF = DAG.getMachineFunction();
6524 SDValue Chain = Op.getOperand(0);
6525 SDValue Offset = Op.getOperand(1);
6526 SDValue Handler = Op.getOperand(2);
6527 DebugLoc dl = Op.getDebugLoc();
6529 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6531 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6533 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6534 DAG.getIntPtrConstant(-TD->getPointerSize()));
6535 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6536 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6537 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6538 MF.getRegInfo().addLiveOut(StoreAddrReg);
6540 return DAG.getNode(X86ISD::EH_RETURN, dl,
6542 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6545 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6546 SelectionDAG &DAG) {
6547 SDValue Root = Op.getOperand(0);
6548 SDValue Trmp = Op.getOperand(1); // trampoline
6549 SDValue FPtr = Op.getOperand(2); // nested function
6550 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6551 DebugLoc dl = Op.getDebugLoc();
6553 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6555 const X86InstrInfo *TII =
6556 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6558 if (Subtarget->is64Bit()) {
6559 SDValue OutChains[6];
6561 // Large code-model.
6563 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6564 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6566 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6567 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6569 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6571 // Load the pointer to the nested function into R11.
6572 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6573 SDValue Addr = Trmp;
6574 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6577 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6578 DAG.getConstant(2, MVT::i64));
6579 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6581 // Load the 'nest' parameter value into R10.
6582 // R10 is specified in X86CallingConv.td
6583 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6584 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6585 DAG.getConstant(10, MVT::i64));
6586 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6587 Addr, TrmpAddr, 10);
6589 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6590 DAG.getConstant(12, MVT::i64));
6591 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6593 // Jump to the nested function.
6594 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6595 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6596 DAG.getConstant(20, MVT::i64));
6597 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6598 Addr, TrmpAddr, 20);
6600 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6601 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6602 DAG.getConstant(22, MVT::i64));
6603 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6607 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6608 return DAG.getMergeValues(Ops, 2, dl);
6610 const Function *Func =
6611 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6612 CallingConv::ID CC = Func->getCallingConv();
6617 llvm_unreachable("Unsupported calling convention");
6618 case CallingConv::C:
6619 case CallingConv::X86_StdCall: {
6620 // Pass 'nest' parameter in ECX.
6621 // Must be kept in sync with X86CallingConv.td
6624 // Check that ECX wasn't needed by an 'inreg' parameter.
6625 const FunctionType *FTy = Func->getFunctionType();
6626 const AttrListPtr &Attrs = Func->getAttributes();
6628 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6629 unsigned InRegCount = 0;
6632 for (FunctionType::param_iterator I = FTy->param_begin(),
6633 E = FTy->param_end(); I != E; ++I, ++Idx)
6634 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6635 // FIXME: should only count parameters that are lowered to integers.
6636 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6638 if (InRegCount > 2) {
6639 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6644 case CallingConv::X86_FastCall:
6645 case CallingConv::Fast:
6646 // Pass 'nest' parameter in EAX.
6647 // Must be kept in sync with X86CallingConv.td
6652 SDValue OutChains[4];
6655 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6656 DAG.getConstant(10, MVT::i32));
6657 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6659 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6660 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6661 OutChains[0] = DAG.getStore(Root, dl,
6662 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6665 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6666 DAG.getConstant(1, MVT::i32));
6667 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6669 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6670 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6671 DAG.getConstant(5, MVT::i32));
6672 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6673 TrmpAddr, 5, false, 1);
6675 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6676 DAG.getConstant(6, MVT::i32));
6677 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6680 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6681 return DAG.getMergeValues(Ops, 2, dl);
6685 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6687 The rounding mode is in bits 11:10 of FPSR, and has the following
6694 FLT_ROUNDS, on the other hand, expects the following:
6701 To perform the conversion, we do:
6702 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6705 MachineFunction &MF = DAG.getMachineFunction();
6706 const TargetMachine &TM = MF.getTarget();
6707 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6708 unsigned StackAlignment = TFI.getStackAlignment();
6709 EVT VT = Op.getValueType();
6710 DebugLoc dl = Op.getDebugLoc();
6712 // Save FP Control Word to stack slot
6713 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6714 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6716 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6717 DAG.getEntryNode(), StackSlot);
6719 // Load FP Control Word from stack slot
6720 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6722 // Transform as necessary
6724 DAG.getNode(ISD::SRL, dl, MVT::i16,
6725 DAG.getNode(ISD::AND, dl, MVT::i16,
6726 CWD, DAG.getConstant(0x800, MVT::i16)),
6727 DAG.getConstant(11, MVT::i8));
6729 DAG.getNode(ISD::SRL, dl, MVT::i16,
6730 DAG.getNode(ISD::AND, dl, MVT::i16,
6731 CWD, DAG.getConstant(0x400, MVT::i16)),
6732 DAG.getConstant(9, MVT::i8));
6735 DAG.getNode(ISD::AND, dl, MVT::i16,
6736 DAG.getNode(ISD::ADD, dl, MVT::i16,
6737 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6738 DAG.getConstant(1, MVT::i16)),
6739 DAG.getConstant(3, MVT::i16));
6742 return DAG.getNode((VT.getSizeInBits() < 16 ?
6743 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6746 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6747 EVT VT = Op.getValueType();
6749 unsigned NumBits = VT.getSizeInBits();
6750 DebugLoc dl = Op.getDebugLoc();
6752 Op = Op.getOperand(0);
6753 if (VT == MVT::i8) {
6754 // Zero extend to i32 since there is not an i8 bsr.
6756 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6759 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6760 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6761 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6763 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6764 SmallVector<SDValue, 4> Ops;
6766 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6767 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6768 Ops.push_back(Op.getValue(1));
6769 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6771 // Finally xor with NumBits-1.
6772 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6775 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6779 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6780 EVT VT = Op.getValueType();
6782 unsigned NumBits = VT.getSizeInBits();
6783 DebugLoc dl = Op.getDebugLoc();
6785 Op = Op.getOperand(0);
6786 if (VT == MVT::i8) {
6788 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6791 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6792 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6793 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6795 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6796 SmallVector<SDValue, 4> Ops;
6798 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6799 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6800 Ops.push_back(Op.getValue(1));
6801 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6804 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6808 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6809 EVT VT = Op.getValueType();
6810 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6811 DebugLoc dl = Op.getDebugLoc();
6813 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6814 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6815 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6816 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6817 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6819 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6820 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6821 // return AloBlo + AloBhi + AhiBlo;
6823 SDValue A = Op.getOperand(0);
6824 SDValue B = Op.getOperand(1);
6826 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6827 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6828 A, DAG.getConstant(32, MVT::i32));
6829 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6830 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6831 B, DAG.getConstant(32, MVT::i32));
6832 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6833 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6835 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6836 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6838 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6839 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6841 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6842 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6843 AloBhi, DAG.getConstant(32, MVT::i32));
6844 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6845 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6846 AhiBlo, DAG.getConstant(32, MVT::i32));
6847 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6848 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6853 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6854 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6855 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6856 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6857 // has only one use.
6858 SDNode *N = Op.getNode();
6859 SDValue LHS = N->getOperand(0);
6860 SDValue RHS = N->getOperand(1);
6861 unsigned BaseOp = 0;
6863 DebugLoc dl = Op.getDebugLoc();
6865 switch (Op.getOpcode()) {
6866 default: llvm_unreachable("Unknown ovf instruction!");
6868 // A subtract of one will be selected as a INC. Note that INC doesn't
6869 // set CF, so we can't do this for UADDO.
6870 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6871 if (C->getAPIntValue() == 1) {
6872 BaseOp = X86ISD::INC;
6876 BaseOp = X86ISD::ADD;
6880 BaseOp = X86ISD::ADD;
6884 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6885 // set CF, so we can't do this for USUBO.
6886 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6887 if (C->getAPIntValue() == 1) {
6888 BaseOp = X86ISD::DEC;
6892 BaseOp = X86ISD::SUB;
6896 BaseOp = X86ISD::SUB;
6900 BaseOp = X86ISD::SMUL;
6904 BaseOp = X86ISD::UMUL;
6909 // Also sets EFLAGS.
6910 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6911 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6914 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6915 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6917 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6921 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6922 EVT T = Op.getValueType();
6923 DebugLoc dl = Op.getDebugLoc();
6926 switch(T.getSimpleVT().SimpleTy) {
6928 assert(false && "Invalid value type!");
6929 case MVT::i8: Reg = X86::AL; size = 1; break;
6930 case MVT::i16: Reg = X86::AX; size = 2; break;
6931 case MVT::i32: Reg = X86::EAX; size = 4; break;
6933 assert(Subtarget->is64Bit() && "Node not type legal!");
6934 Reg = X86::RAX; size = 8;
6937 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6938 Op.getOperand(2), SDValue());
6939 SDValue Ops[] = { cpIn.getValue(0),
6942 DAG.getTargetConstant(size, MVT::i8),
6944 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6945 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6947 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6951 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6952 SelectionDAG &DAG) {
6953 assert(Subtarget->is64Bit() && "Result not type legalized?");
6954 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6955 SDValue TheChain = Op.getOperand(0);
6956 DebugLoc dl = Op.getDebugLoc();
6957 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6958 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6959 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6961 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6962 DAG.getConstant(32, MVT::i8));
6964 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6967 return DAG.getMergeValues(Ops, 2, dl);
6970 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6971 SDNode *Node = Op.getNode();
6972 DebugLoc dl = Node->getDebugLoc();
6973 EVT T = Node->getValueType(0);
6974 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6975 DAG.getConstant(0, T), Node->getOperand(2));
6976 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6977 cast<AtomicSDNode>(Node)->getMemoryVT(),
6978 Node->getOperand(0),
6979 Node->getOperand(1), negOp,
6980 cast<AtomicSDNode>(Node)->getSrcValue(),
6981 cast<AtomicSDNode>(Node)->getAlignment());
6984 /// LowerOperation - Provide custom lowering hooks for some operations.
6986 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6987 switch (Op.getOpcode()) {
6988 default: llvm_unreachable("Should not custom lower this!");
6989 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6990 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6991 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6992 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6993 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6994 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6995 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6996 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6997 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6998 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6999 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
7000 case ISD::SHL_PARTS:
7001 case ISD::SRA_PARTS:
7002 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
7003 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
7004 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
7005 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
7006 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
7007 case ISD::FABS: return LowerFABS(Op, DAG);
7008 case ISD::FNEG: return LowerFNEG(Op, DAG);
7009 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
7010 case ISD::SETCC: return LowerSETCC(Op, DAG);
7011 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
7012 case ISD::SELECT: return LowerSELECT(Op, DAG);
7013 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
7014 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
7015 case ISD::VASTART: return LowerVASTART(Op, DAG);
7016 case ISD::VAARG: return LowerVAARG(Op, DAG);
7017 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
7018 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
7019 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
7020 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
7021 case ISD::FRAME_TO_ARGS_OFFSET:
7022 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
7023 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
7024 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
7025 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
7026 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
7027 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
7028 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
7029 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
7035 case ISD::UMULO: return LowerXALUO(Op, DAG);
7036 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
7040 void X86TargetLowering::
7041 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
7042 SelectionDAG &DAG, unsigned NewOp) {
7043 EVT T = Node->getValueType(0);
7044 DebugLoc dl = Node->getDebugLoc();
7045 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
7047 SDValue Chain = Node->getOperand(0);
7048 SDValue In1 = Node->getOperand(1);
7049 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7050 Node->getOperand(2), DAG.getIntPtrConstant(0));
7051 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
7052 Node->getOperand(2), DAG.getIntPtrConstant(1));
7053 SDValue Ops[] = { Chain, In1, In2L, In2H };
7054 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
7056 DAG.getMemIntrinsicNode(NewOp, dl, Tys, Ops, 4, MVT::i64,
7057 cast<MemSDNode>(Node)->getMemOperand());
7058 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
7059 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7060 Results.push_back(Result.getValue(2));
7063 /// ReplaceNodeResults - Replace a node with an illegal result type
7064 /// with a new node built out of custom code.
7065 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
7066 SmallVectorImpl<SDValue>&Results,
7067 SelectionDAG &DAG) {
7068 DebugLoc dl = N->getDebugLoc();
7069 switch (N->getOpcode()) {
7071 assert(false && "Do not know how to custom type legalize this operation!");
7073 case ISD::FP_TO_SINT: {
7074 std::pair<SDValue,SDValue> Vals =
7075 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
7076 SDValue FIST = Vals.first, StackSlot = Vals.second;
7077 if (FIST.getNode() != 0) {
7078 EVT VT = N->getValueType(0);
7079 // Return a load from the stack slot.
7080 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
7084 case ISD::READCYCLECOUNTER: {
7085 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7086 SDValue TheChain = N->getOperand(0);
7087 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
7088 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
7090 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
7092 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
7093 SDValue Ops[] = { eax, edx };
7094 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
7095 Results.push_back(edx.getValue(1));
7098 case ISD::ATOMIC_CMP_SWAP: {
7099 EVT T = N->getValueType(0);
7100 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
7101 SDValue cpInL, cpInH;
7102 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7103 DAG.getConstant(0, MVT::i32));
7104 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
7105 DAG.getConstant(1, MVT::i32));
7106 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
7107 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
7109 SDValue swapInL, swapInH;
7110 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7111 DAG.getConstant(0, MVT::i32));
7112 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
7113 DAG.getConstant(1, MVT::i32));
7114 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
7116 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
7117 swapInL.getValue(1));
7118 SDValue Ops[] = { swapInH.getValue(0),
7120 swapInH.getValue(1) };
7121 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
7122 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
7123 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
7124 MVT::i32, Result.getValue(1));
7125 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
7126 MVT::i32, cpOutL.getValue(2));
7127 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
7128 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
7129 Results.push_back(cpOutH.getValue(1));
7132 case ISD::ATOMIC_LOAD_ADD:
7133 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7135 case ISD::ATOMIC_LOAD_AND:
7136 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7138 case ISD::ATOMIC_LOAD_NAND:
7139 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7141 case ISD::ATOMIC_LOAD_OR:
7142 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7144 case ISD::ATOMIC_LOAD_SUB:
7145 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7147 case ISD::ATOMIC_LOAD_XOR:
7148 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7150 case ISD::ATOMIC_SWAP:
7151 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7156 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7158 default: return NULL;
7159 case X86ISD::BSF: return "X86ISD::BSF";
7160 case X86ISD::BSR: return "X86ISD::BSR";
7161 case X86ISD::SHLD: return "X86ISD::SHLD";
7162 case X86ISD::SHRD: return "X86ISD::SHRD";
7163 case X86ISD::FAND: return "X86ISD::FAND";
7164 case X86ISD::FOR: return "X86ISD::FOR";
7165 case X86ISD::FXOR: return "X86ISD::FXOR";
7166 case X86ISD::FSRL: return "X86ISD::FSRL";
7167 case X86ISD::FILD: return "X86ISD::FILD";
7168 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7169 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7170 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7171 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7172 case X86ISD::FLD: return "X86ISD::FLD";
7173 case X86ISD::FST: return "X86ISD::FST";
7174 case X86ISD::CALL: return "X86ISD::CALL";
7175 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7176 case X86ISD::BT: return "X86ISD::BT";
7177 case X86ISD::CMP: return "X86ISD::CMP";
7178 case X86ISD::COMI: return "X86ISD::COMI";
7179 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7180 case X86ISD::SETCC: return "X86ISD::SETCC";
7181 case X86ISD::CMOV: return "X86ISD::CMOV";
7182 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7183 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7184 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7185 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7186 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7187 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7188 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7189 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7190 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7191 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7192 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7193 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7194 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7195 case X86ISD::FMAX: return "X86ISD::FMAX";
7196 case X86ISD::FMIN: return "X86ISD::FMIN";
7197 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7198 case X86ISD::FRCP: return "X86ISD::FRCP";
7199 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7200 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7201 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7202 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7203 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7204 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7205 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7206 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7207 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7208 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7209 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7210 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7211 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7212 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7213 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7214 case X86ISD::VSHL: return "X86ISD::VSHL";
7215 case X86ISD::VSRL: return "X86ISD::VSRL";
7216 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7217 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7218 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7219 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7220 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7221 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7222 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7223 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7224 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7225 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7226 case X86ISD::ADD: return "X86ISD::ADD";
7227 case X86ISD::SUB: return "X86ISD::SUB";
7228 case X86ISD::SMUL: return "X86ISD::SMUL";
7229 case X86ISD::UMUL: return "X86ISD::UMUL";
7230 case X86ISD::INC: return "X86ISD::INC";
7231 case X86ISD::DEC: return "X86ISD::DEC";
7232 case X86ISD::OR: return "X86ISD::OR";
7233 case X86ISD::XOR: return "X86ISD::XOR";
7234 case X86ISD::AND: return "X86ISD::AND";
7235 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7236 case X86ISD::PTEST: return "X86ISD::PTEST";
7237 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7241 // isLegalAddressingMode - Return true if the addressing mode represented
7242 // by AM is legal for this target, for a load/store of the specified type.
7243 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7244 const Type *Ty) const {
7245 // X86 supports extremely general addressing modes.
7246 CodeModel::Model M = getTargetMachine().getCodeModel();
7248 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7249 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7254 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7256 // If a reference to this global requires an extra load, we can't fold it.
7257 if (isGlobalStubReference(GVFlags))
7260 // If BaseGV requires a register for the PIC base, we cannot also have a
7261 // BaseReg specified.
7262 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7265 // If lower 4G is not available, then we must use rip-relative addressing.
7266 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7276 // These scales always work.
7281 // These scales are formed with basereg+scalereg. Only accept if there is
7286 default: // Other stuff never works.
7294 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7295 if (!Ty1->isInteger() || !Ty2->isInteger())
7297 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7298 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7299 if (NumBits1 <= NumBits2)
7301 return Subtarget->is64Bit() || NumBits1 < 64;
7304 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7305 if (!VT1.isInteger() || !VT2.isInteger())
7307 unsigned NumBits1 = VT1.getSizeInBits();
7308 unsigned NumBits2 = VT2.getSizeInBits();
7309 if (NumBits1 <= NumBits2)
7311 return Subtarget->is64Bit() || NumBits1 < 64;
7314 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7315 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7316 return Ty1 == Type::getInt32Ty(Ty1->getContext()) &&
7317 Ty2 == Type::getInt64Ty(Ty1->getContext()) && Subtarget->is64Bit();
7320 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7321 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7322 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7325 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7326 // i16 instructions are longer (0x66 prefix) and potentially slower.
7327 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7330 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7331 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7332 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7333 /// are assumed to be legal.
7335 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7337 // Only do shuffles on 128-bit vector types for now.
7338 if (VT.getSizeInBits() == 64)
7341 // FIXME: pshufb, blends, shifts.
7342 return (VT.getVectorNumElements() == 2 ||
7343 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7344 isMOVLMask(M, VT) ||
7345 isSHUFPMask(M, VT) ||
7346 isPSHUFDMask(M, VT) ||
7347 isPSHUFHWMask(M, VT) ||
7348 isPSHUFLWMask(M, VT) ||
7349 isPALIGNRMask(M, VT, Subtarget->hasSSSE3()) ||
7350 isUNPCKLMask(M, VT) ||
7351 isUNPCKHMask(M, VT) ||
7352 isUNPCKL_v_undef_Mask(M, VT) ||
7353 isUNPCKH_v_undef_Mask(M, VT));
7357 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7359 unsigned NumElts = VT.getVectorNumElements();
7360 // FIXME: This collection of masks seems suspect.
7363 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7364 return (isMOVLMask(Mask, VT) ||
7365 isCommutedMOVLMask(Mask, VT, true) ||
7366 isSHUFPMask(Mask, VT) ||
7367 isCommutedSHUFPMask(Mask, VT));
7372 //===----------------------------------------------------------------------===//
7373 // X86 Scheduler Hooks
7374 //===----------------------------------------------------------------------===//
7376 // private utility function
7378 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7379 MachineBasicBlock *MBB,
7387 TargetRegisterClass *RC,
7388 bool invSrc) const {
7389 // For the atomic bitwise operator, we generate
7392 // ld t1 = [bitinstr.addr]
7393 // op t2 = t1, [bitinstr.val]
7395 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7397 // fallthrough -->nextMBB
7398 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7399 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7400 MachineFunction::iterator MBBIter = MBB;
7403 /// First build the CFG
7404 MachineFunction *F = MBB->getParent();
7405 MachineBasicBlock *thisMBB = MBB;
7406 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7407 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7408 F->insert(MBBIter, newMBB);
7409 F->insert(MBBIter, nextMBB);
7411 // Move all successors to thisMBB to nextMBB
7412 nextMBB->transferSuccessors(thisMBB);
7414 // Update thisMBB to fall through to newMBB
7415 thisMBB->addSuccessor(newMBB);
7417 // newMBB jumps to itself and fall through to nextMBB
7418 newMBB->addSuccessor(nextMBB);
7419 newMBB->addSuccessor(newMBB);
7421 // Insert instructions into newMBB based on incoming instruction
7422 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7423 "unexpected number of operands");
7424 DebugLoc dl = bInstr->getDebugLoc();
7425 MachineOperand& destOper = bInstr->getOperand(0);
7426 MachineOperand* argOpers[2 + X86AddrNumOperands];
7427 int numArgs = bInstr->getNumOperands() - 1;
7428 for (int i=0; i < numArgs; ++i)
7429 argOpers[i] = &bInstr->getOperand(i+1);
7431 // x86 address has 4 operands: base, index, scale, and displacement
7432 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7433 int valArgIndx = lastAddrIndx + 1;
7435 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7436 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7437 for (int i=0; i <= lastAddrIndx; ++i)
7438 (*MIB).addOperand(*argOpers[i]);
7440 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7442 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7447 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7448 assert((argOpers[valArgIndx]->isReg() ||
7449 argOpers[valArgIndx]->isImm()) &&
7451 if (argOpers[valArgIndx]->isReg())
7452 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7454 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7456 (*MIB).addOperand(*argOpers[valArgIndx]);
7458 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7461 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7462 for (int i=0; i <= lastAddrIndx; ++i)
7463 (*MIB).addOperand(*argOpers[i]);
7465 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7466 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7467 bInstr->memoperands_end());
7469 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7473 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7475 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7479 // private utility function: 64 bit atomics on 32 bit host.
7481 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7482 MachineBasicBlock *MBB,
7487 bool invSrc) const {
7488 // For the atomic bitwise operator, we generate
7489 // thisMBB (instructions are in pairs, except cmpxchg8b)
7490 // ld t1,t2 = [bitinstr.addr]
7492 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7493 // op t5, t6 <- out1, out2, [bitinstr.val]
7494 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7495 // mov ECX, EBX <- t5, t6
7496 // mov EAX, EDX <- t1, t2
7497 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7498 // mov t3, t4 <- EAX, EDX
7500 // result in out1, out2
7501 // fallthrough -->nextMBB
7503 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7504 const unsigned LoadOpc = X86::MOV32rm;
7505 const unsigned copyOpc = X86::MOV32rr;
7506 const unsigned NotOpc = X86::NOT32r;
7507 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7508 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7509 MachineFunction::iterator MBBIter = MBB;
7512 /// First build the CFG
7513 MachineFunction *F = MBB->getParent();
7514 MachineBasicBlock *thisMBB = MBB;
7515 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7516 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7517 F->insert(MBBIter, newMBB);
7518 F->insert(MBBIter, nextMBB);
7520 // Move all successors to thisMBB to nextMBB
7521 nextMBB->transferSuccessors(thisMBB);
7523 // Update thisMBB to fall through to newMBB
7524 thisMBB->addSuccessor(newMBB);
7526 // newMBB jumps to itself and fall through to nextMBB
7527 newMBB->addSuccessor(nextMBB);
7528 newMBB->addSuccessor(newMBB);
7530 DebugLoc dl = bInstr->getDebugLoc();
7531 // Insert instructions into newMBB based on incoming instruction
7532 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7533 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7534 "unexpected number of operands");
7535 MachineOperand& dest1Oper = bInstr->getOperand(0);
7536 MachineOperand& dest2Oper = bInstr->getOperand(1);
7537 MachineOperand* argOpers[2 + X86AddrNumOperands];
7538 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7539 argOpers[i] = &bInstr->getOperand(i+2);
7541 // x86 address has 4 operands: base, index, scale, and displacement
7542 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7544 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7545 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7546 for (int i=0; i <= lastAddrIndx; ++i)
7547 (*MIB).addOperand(*argOpers[i]);
7548 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7549 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7550 // add 4 to displacement.
7551 for (int i=0; i <= lastAddrIndx-2; ++i)
7552 (*MIB).addOperand(*argOpers[i]);
7553 MachineOperand newOp3 = *(argOpers[3]);
7555 newOp3.setImm(newOp3.getImm()+4);
7557 newOp3.setOffset(newOp3.getOffset()+4);
7558 (*MIB).addOperand(newOp3);
7559 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7561 // t3/4 are defined later, at the bottom of the loop
7562 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7563 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7564 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7565 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7566 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7567 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7569 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7570 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7572 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7573 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7579 int valArgIndx = lastAddrIndx + 1;
7580 assert((argOpers[valArgIndx]->isReg() ||
7581 argOpers[valArgIndx]->isImm()) &&
7583 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7584 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7585 if (argOpers[valArgIndx]->isReg())
7586 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7588 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7589 if (regOpcL != X86::MOV32rr)
7591 (*MIB).addOperand(*argOpers[valArgIndx]);
7592 assert(argOpers[valArgIndx + 1]->isReg() ==
7593 argOpers[valArgIndx]->isReg());
7594 assert(argOpers[valArgIndx + 1]->isImm() ==
7595 argOpers[valArgIndx]->isImm());
7596 if (argOpers[valArgIndx + 1]->isReg())
7597 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7599 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7600 if (regOpcH != X86::MOV32rr)
7602 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7604 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7606 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7609 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7611 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7614 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7615 for (int i=0; i <= lastAddrIndx; ++i)
7616 (*MIB).addOperand(*argOpers[i]);
7618 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7619 (*MIB).setMemRefs(bInstr->memoperands_begin(),
7620 bInstr->memoperands_end());
7622 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7623 MIB.addReg(X86::EAX);
7624 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7625 MIB.addReg(X86::EDX);
7628 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7630 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7634 // private utility function
7636 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7637 MachineBasicBlock *MBB,
7638 unsigned cmovOpc) const {
7639 // For the atomic min/max operator, we generate
7642 // ld t1 = [min/max.addr]
7643 // mov t2 = [min/max.val]
7645 // cmov[cond] t2 = t1
7647 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7649 // fallthrough -->nextMBB
7651 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7652 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7653 MachineFunction::iterator MBBIter = MBB;
7656 /// First build the CFG
7657 MachineFunction *F = MBB->getParent();
7658 MachineBasicBlock *thisMBB = MBB;
7659 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7660 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7661 F->insert(MBBIter, newMBB);
7662 F->insert(MBBIter, nextMBB);
7664 // Move all successors of thisMBB to nextMBB
7665 nextMBB->transferSuccessors(thisMBB);
7667 // Update thisMBB to fall through to newMBB
7668 thisMBB->addSuccessor(newMBB);
7670 // newMBB jumps to newMBB and fall through to nextMBB
7671 newMBB->addSuccessor(nextMBB);
7672 newMBB->addSuccessor(newMBB);
7674 DebugLoc dl = mInstr->getDebugLoc();
7675 // Insert instructions into newMBB based on incoming instruction
7676 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7677 "unexpected number of operands");
7678 MachineOperand& destOper = mInstr->getOperand(0);
7679 MachineOperand* argOpers[2 + X86AddrNumOperands];
7680 int numArgs = mInstr->getNumOperands() - 1;
7681 for (int i=0; i < numArgs; ++i)
7682 argOpers[i] = &mInstr->getOperand(i+1);
7684 // x86 address has 4 operands: base, index, scale, and displacement
7685 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7686 int valArgIndx = lastAddrIndx + 1;
7688 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7689 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7690 for (int i=0; i <= lastAddrIndx; ++i)
7691 (*MIB).addOperand(*argOpers[i]);
7693 // We only support register and immediate values
7694 assert((argOpers[valArgIndx]->isReg() ||
7695 argOpers[valArgIndx]->isImm()) &&
7698 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7699 if (argOpers[valArgIndx]->isReg())
7700 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7702 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7703 (*MIB).addOperand(*argOpers[valArgIndx]);
7705 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7708 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7713 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7714 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7718 // Cmp and exchange if none has modified the memory location
7719 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7720 for (int i=0; i <= lastAddrIndx; ++i)
7721 (*MIB).addOperand(*argOpers[i]);
7723 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7724 (*MIB).setMemRefs(mInstr->memoperands_begin(),
7725 mInstr->memoperands_end());
7727 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7728 MIB.addReg(X86::EAX);
7731 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7733 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7737 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
7738 // all of this code can be replaced with that in the .td file.
7740 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
7741 unsigned numArgs, bool memArg) const {
7743 MachineFunction *F = BB->getParent();
7744 DebugLoc dl = MI->getDebugLoc();
7745 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7749 Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
7751 Opc = numArgs == 3 ? X86::PCMPISTRM128rr : X86::PCMPESTRM128rr;
7753 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
7755 for (unsigned i = 0; i < numArgs; ++i) {
7756 MachineOperand &Op = MI->getOperand(i+1);
7758 if (!(Op.isReg() && Op.isImplicit()))
7762 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
7765 F->DeleteMachineInstr(MI);
7771 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
7773 MachineBasicBlock *MBB) const {
7774 // Emit code to save XMM registers to the stack. The ABI says that the
7775 // number of registers to save is given in %al, so it's theoretically
7776 // possible to do an indirect jump trick to avoid saving all of them,
7777 // however this code takes a simpler approach and just executes all
7778 // of the stores if %al is non-zero. It's less code, and it's probably
7779 // easier on the hardware branch predictor, and stores aren't all that
7780 // expensive anyway.
7782 // Create the new basic blocks. One block contains all the XMM stores,
7783 // and one block is the final destination regardless of whether any
7784 // stores were performed.
7785 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7786 MachineFunction *F = MBB->getParent();
7787 MachineFunction::iterator MBBIter = MBB;
7789 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
7790 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
7791 F->insert(MBBIter, XMMSaveMBB);
7792 F->insert(MBBIter, EndMBB);
7795 // Move any original successors of MBB to the end block.
7796 EndMBB->transferSuccessors(MBB);
7797 // The original block will now fall through to the XMM save block.
7798 MBB->addSuccessor(XMMSaveMBB);
7799 // The XMMSaveMBB will fall through to the end block.
7800 XMMSaveMBB->addSuccessor(EndMBB);
7802 // Now add the instructions.
7803 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7804 DebugLoc DL = MI->getDebugLoc();
7806 unsigned CountReg = MI->getOperand(0).getReg();
7807 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
7808 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
7810 if (!Subtarget->isTargetWin64()) {
7811 // If %al is 0, branch around the XMM save block.
7812 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
7813 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
7814 MBB->addSuccessor(EndMBB);
7817 // In the XMM save block, save all the XMM argument registers.
7818 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
7819 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
7820 MachineMemOperand *MMO =
7821 F->getMachineMemOperand(
7822 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
7823 MachineMemOperand::MOStore, Offset,
7824 /*Size=*/16, /*Align=*/16);
7825 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
7826 .addFrameIndex(RegSaveFrameIndex)
7827 .addImm(/*Scale=*/1)
7828 .addReg(/*IndexReg=*/0)
7829 .addImm(/*Disp=*/Offset)
7830 .addReg(/*Segment=*/0)
7831 .addReg(MI->getOperand(i).getReg())
7832 .addMemOperand(MMO);
7835 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7841 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
7842 MachineBasicBlock *BB,
7843 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
7844 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7845 DebugLoc DL = MI->getDebugLoc();
7847 // To "insert" a SELECT_CC instruction, we actually have to insert the
7848 // diamond control-flow pattern. The incoming instruction knows the
7849 // destination vreg to set, the condition code register to branch on, the
7850 // true/false values to select between, and a branch opcode to use.
7851 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7852 MachineFunction::iterator It = BB;
7858 // cmpTY ccX, r1, r2
7860 // fallthrough --> copy0MBB
7861 MachineBasicBlock *thisMBB = BB;
7862 MachineFunction *F = BB->getParent();
7863 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7864 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7866 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7867 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
7868 F->insert(It, copy0MBB);
7869 F->insert(It, sinkMBB);
7870 // Update machine-CFG edges by first adding all successors of the current
7871 // block to the new block which will contain the Phi node for the select.
7872 // Also inform sdisel of the edge changes.
7873 for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
7874 E = BB->succ_end(); I != E; ++I) {
7875 EM->insert(std::make_pair(*I, sinkMBB));
7876 sinkMBB->addSuccessor(*I);
7878 // Next, remove all successors of the current block, and add the true
7879 // and fallthrough blocks as its successors.
7880 while (!BB->succ_empty())
7881 BB->removeSuccessor(BB->succ_begin());
7882 // Add the true and fallthrough blocks as its successors.
7883 BB->addSuccessor(copy0MBB);
7884 BB->addSuccessor(sinkMBB);
7887 // %FalseValue = ...
7888 // # fallthrough to sinkMBB
7891 // Update machine-CFG edges
7892 BB->addSuccessor(sinkMBB);
7895 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7898 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
7899 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7900 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7902 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7908 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7909 MachineBasicBlock *BB,
7910 DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
7911 switch (MI->getOpcode()) {
7912 default: assert(false && "Unexpected instr type to insert");
7914 case X86::CMOV_V1I64:
7915 case X86::CMOV_FR32:
7916 case X86::CMOV_FR64:
7917 case X86::CMOV_V4F32:
7918 case X86::CMOV_V2F64:
7919 case X86::CMOV_V2I64:
7920 return EmitLoweredSelect(MI, BB, EM);
7922 case X86::FP32_TO_INT16_IN_MEM:
7923 case X86::FP32_TO_INT32_IN_MEM:
7924 case X86::FP32_TO_INT64_IN_MEM:
7925 case X86::FP64_TO_INT16_IN_MEM:
7926 case X86::FP64_TO_INT32_IN_MEM:
7927 case X86::FP64_TO_INT64_IN_MEM:
7928 case X86::FP80_TO_INT16_IN_MEM:
7929 case X86::FP80_TO_INT32_IN_MEM:
7930 case X86::FP80_TO_INT64_IN_MEM: {
7931 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7932 DebugLoc DL = MI->getDebugLoc();
7934 // Change the floating point control register to use "round towards zero"
7935 // mode when truncating to an integer value.
7936 MachineFunction *F = BB->getParent();
7937 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7938 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7940 // Load the old value of the high byte of the control word...
7942 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7943 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
7946 // Set the high part to be round to zero...
7947 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
7950 // Reload the modified control word now...
7951 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
7953 // Restore the memory image of control word to original value
7954 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
7957 // Get the X86 opcode to use.
7959 switch (MI->getOpcode()) {
7960 default: llvm_unreachable("illegal opcode!");
7961 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7962 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7963 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7964 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7965 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7966 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7967 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7968 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7969 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7973 MachineOperand &Op = MI->getOperand(0);
7975 AM.BaseType = X86AddressMode::RegBase;
7976 AM.Base.Reg = Op.getReg();
7978 AM.BaseType = X86AddressMode::FrameIndexBase;
7979 AM.Base.FrameIndex = Op.getIndex();
7981 Op = MI->getOperand(1);
7983 AM.Scale = Op.getImm();
7984 Op = MI->getOperand(2);
7986 AM.IndexReg = Op.getImm();
7987 Op = MI->getOperand(3);
7988 if (Op.isGlobal()) {
7989 AM.GV = Op.getGlobal();
7991 AM.Disp = Op.getImm();
7993 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
7994 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7996 // Reload the original control word now.
7997 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
7999 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
8002 // String/text processing lowering.
8003 case X86::PCMPISTRM128REG:
8004 return EmitPCMP(MI, BB, 3, false /* in-mem */);
8005 case X86::PCMPISTRM128MEM:
8006 return EmitPCMP(MI, BB, 3, true /* in-mem */);
8007 case X86::PCMPESTRM128REG:
8008 return EmitPCMP(MI, BB, 5, false /* in mem */);
8009 case X86::PCMPESTRM128MEM:
8010 return EmitPCMP(MI, BB, 5, true /* in mem */);
8013 case X86::ATOMAND32:
8014 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8015 X86::AND32ri, X86::MOV32rm,
8016 X86::LCMPXCHG32, X86::MOV32rr,
8017 X86::NOT32r, X86::EAX,
8018 X86::GR32RegisterClass);
8020 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
8021 X86::OR32ri, X86::MOV32rm,
8022 X86::LCMPXCHG32, X86::MOV32rr,
8023 X86::NOT32r, X86::EAX,
8024 X86::GR32RegisterClass);
8025 case X86::ATOMXOR32:
8026 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
8027 X86::XOR32ri, X86::MOV32rm,
8028 X86::LCMPXCHG32, X86::MOV32rr,
8029 X86::NOT32r, X86::EAX,
8030 X86::GR32RegisterClass);
8031 case X86::ATOMNAND32:
8032 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
8033 X86::AND32ri, X86::MOV32rm,
8034 X86::LCMPXCHG32, X86::MOV32rr,
8035 X86::NOT32r, X86::EAX,
8036 X86::GR32RegisterClass, true);
8037 case X86::ATOMMIN32:
8038 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
8039 case X86::ATOMMAX32:
8040 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
8041 case X86::ATOMUMIN32:
8042 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
8043 case X86::ATOMUMAX32:
8044 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
8046 case X86::ATOMAND16:
8047 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8048 X86::AND16ri, X86::MOV16rm,
8049 X86::LCMPXCHG16, X86::MOV16rr,
8050 X86::NOT16r, X86::AX,
8051 X86::GR16RegisterClass);
8053 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
8054 X86::OR16ri, X86::MOV16rm,
8055 X86::LCMPXCHG16, X86::MOV16rr,
8056 X86::NOT16r, X86::AX,
8057 X86::GR16RegisterClass);
8058 case X86::ATOMXOR16:
8059 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
8060 X86::XOR16ri, X86::MOV16rm,
8061 X86::LCMPXCHG16, X86::MOV16rr,
8062 X86::NOT16r, X86::AX,
8063 X86::GR16RegisterClass);
8064 case X86::ATOMNAND16:
8065 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
8066 X86::AND16ri, X86::MOV16rm,
8067 X86::LCMPXCHG16, X86::MOV16rr,
8068 X86::NOT16r, X86::AX,
8069 X86::GR16RegisterClass, true);
8070 case X86::ATOMMIN16:
8071 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
8072 case X86::ATOMMAX16:
8073 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
8074 case X86::ATOMUMIN16:
8075 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
8076 case X86::ATOMUMAX16:
8077 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
8080 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8081 X86::AND8ri, X86::MOV8rm,
8082 X86::LCMPXCHG8, X86::MOV8rr,
8083 X86::NOT8r, X86::AL,
8084 X86::GR8RegisterClass);
8086 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
8087 X86::OR8ri, X86::MOV8rm,
8088 X86::LCMPXCHG8, X86::MOV8rr,
8089 X86::NOT8r, X86::AL,
8090 X86::GR8RegisterClass);
8092 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
8093 X86::XOR8ri, X86::MOV8rm,
8094 X86::LCMPXCHG8, X86::MOV8rr,
8095 X86::NOT8r, X86::AL,
8096 X86::GR8RegisterClass);
8097 case X86::ATOMNAND8:
8098 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
8099 X86::AND8ri, X86::MOV8rm,
8100 X86::LCMPXCHG8, X86::MOV8rr,
8101 X86::NOT8r, X86::AL,
8102 X86::GR8RegisterClass, true);
8103 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
8104 // This group is for 64-bit host.
8105 case X86::ATOMAND64:
8106 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8107 X86::AND64ri32, X86::MOV64rm,
8108 X86::LCMPXCHG64, X86::MOV64rr,
8109 X86::NOT64r, X86::RAX,
8110 X86::GR64RegisterClass);
8112 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
8113 X86::OR64ri32, X86::MOV64rm,
8114 X86::LCMPXCHG64, X86::MOV64rr,
8115 X86::NOT64r, X86::RAX,
8116 X86::GR64RegisterClass);
8117 case X86::ATOMXOR64:
8118 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
8119 X86::XOR64ri32, X86::MOV64rm,
8120 X86::LCMPXCHG64, X86::MOV64rr,
8121 X86::NOT64r, X86::RAX,
8122 X86::GR64RegisterClass);
8123 case X86::ATOMNAND64:
8124 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
8125 X86::AND64ri32, X86::MOV64rm,
8126 X86::LCMPXCHG64, X86::MOV64rr,
8127 X86::NOT64r, X86::RAX,
8128 X86::GR64RegisterClass, true);
8129 case X86::ATOMMIN64:
8130 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
8131 case X86::ATOMMAX64:
8132 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
8133 case X86::ATOMUMIN64:
8134 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
8135 case X86::ATOMUMAX64:
8136 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
8138 // This group does 64-bit operations on a 32-bit host.
8139 case X86::ATOMAND6432:
8140 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8141 X86::AND32rr, X86::AND32rr,
8142 X86::AND32ri, X86::AND32ri,
8144 case X86::ATOMOR6432:
8145 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8146 X86::OR32rr, X86::OR32rr,
8147 X86::OR32ri, X86::OR32ri,
8149 case X86::ATOMXOR6432:
8150 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8151 X86::XOR32rr, X86::XOR32rr,
8152 X86::XOR32ri, X86::XOR32ri,
8154 case X86::ATOMNAND6432:
8155 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8156 X86::AND32rr, X86::AND32rr,
8157 X86::AND32ri, X86::AND32ri,
8159 case X86::ATOMADD6432:
8160 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8161 X86::ADD32rr, X86::ADC32rr,
8162 X86::ADD32ri, X86::ADC32ri,
8164 case X86::ATOMSUB6432:
8165 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8166 X86::SUB32rr, X86::SBB32rr,
8167 X86::SUB32ri, X86::SBB32ri,
8169 case X86::ATOMSWAP6432:
8170 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8171 X86::MOV32rr, X86::MOV32rr,
8172 X86::MOV32ri, X86::MOV32ri,
8174 case X86::VASTART_SAVE_XMM_REGS:
8175 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8179 //===----------------------------------------------------------------------===//
8180 // X86 Optimization Hooks
8181 //===----------------------------------------------------------------------===//
8183 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8187 const SelectionDAG &DAG,
8188 unsigned Depth) const {
8189 unsigned Opc = Op.getOpcode();
8190 assert((Opc >= ISD::BUILTIN_OP_END ||
8191 Opc == ISD::INTRINSIC_WO_CHAIN ||
8192 Opc == ISD::INTRINSIC_W_CHAIN ||
8193 Opc == ISD::INTRINSIC_VOID) &&
8194 "Should use MaskedValueIsZero if you don't know whether Op"
8195 " is a target node!");
8197 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8209 // These nodes' second result is a boolean.
8210 if (Op.getResNo() == 0)
8214 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8215 Mask.getBitWidth() - 1);
8220 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8221 /// node is a GlobalAddress + offset.
8222 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8223 GlobalValue* &GA, int64_t &Offset) const{
8224 if (N->getOpcode() == X86ISD::Wrapper) {
8225 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8226 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8227 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8231 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8234 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
8235 const TargetLowering &TLI) {
8238 if (TLI.isGAPlusOffset(Base, GV, Offset))
8239 return (GV->getAlignment() >= N && (Offset % N) == 0);
8240 // DAG combine handles the stack object case.
8244 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8245 EVT EltVT, LoadSDNode *&LDBase,
8246 unsigned &LastLoadedElt,
8247 SelectionDAG &DAG, MachineFrameInfo *MFI,
8248 const TargetLowering &TLI) {
8250 LastLoadedElt = -1U;
8251 for (unsigned i = 0; i < NumElems; ++i) {
8252 if (N->getMaskElt(i) < 0) {
8258 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8259 if (!Elt.getNode() ||
8260 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8263 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8265 LDBase = cast<LoadSDNode>(Elt.getNode());
8269 if (Elt.getOpcode() == ISD::UNDEF)
8272 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8273 if (!TLI.isConsecutiveLoad(LD, LDBase, EltVT.getSizeInBits()/8, i, MFI))
8280 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8281 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8282 /// if the load addresses are consecutive, non-overlapping, and in the right
8283 /// order. In the case of v2i64, it will see if it can rewrite the
8284 /// shuffle to be an appropriate build vector so it can take advantage of
8285 // performBuildVectorCombine.
8286 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8287 const TargetLowering &TLI) {
8288 DebugLoc dl = N->getDebugLoc();
8289 EVT VT = N->getValueType(0);
8290 EVT EltVT = VT.getVectorElementType();
8291 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8292 unsigned NumElems = VT.getVectorNumElements();
8294 if (VT.getSizeInBits() != 128)
8297 // Try to combine a vector_shuffle into a 128-bit load.
8298 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8299 LoadSDNode *LD = NULL;
8300 unsigned LastLoadedElt;
8301 if (!EltsFromConsecutiveLoads(SVN, NumElems, EltVT, LD, LastLoadedElt, DAG,
8305 if (LastLoadedElt == NumElems - 1) {
8306 if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
8307 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8308 LD->getSrcValue(), LD->getSrcValueOffset(),
8310 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8311 LD->getSrcValue(), LD->getSrcValueOffset(),
8312 LD->isVolatile(), LD->getAlignment());
8313 } else if (NumElems == 4 && LastLoadedElt == 1) {
8314 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8315 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8316 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8317 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8322 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8323 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8324 const X86Subtarget *Subtarget) {
8325 DebugLoc DL = N->getDebugLoc();
8326 SDValue Cond = N->getOperand(0);
8327 // Get the LHS/RHS of the select.
8328 SDValue LHS = N->getOperand(1);
8329 SDValue RHS = N->getOperand(2);
8331 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
8332 // instructions have the peculiarity that if either operand is a NaN,
8333 // they chose what we call the RHS operand (and as such are not symmetric).
8334 // It happens that this matches the semantics of the common C idiom
8335 // x<y?x:y and related forms, so we can recognize these cases.
8336 if (Subtarget->hasSSE2() &&
8337 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8338 Cond.getOpcode() == ISD::SETCC) {
8339 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8341 unsigned Opcode = 0;
8342 // Check for x CC y ? x : y.
8343 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8347 // This can be a min if we can prove that at least one of the operands
8349 if (!FiniteOnlyFPMath()) {
8350 if (DAG.isKnownNeverNaN(RHS)) {
8351 // Put the potential NaN in the RHS so that SSE will preserve it.
8352 std::swap(LHS, RHS);
8353 } else if (!DAG.isKnownNeverNaN(LHS))
8356 Opcode = X86ISD::FMIN;
8359 // This can be a min if we can prove that at least one of the operands
8361 if (!FiniteOnlyFPMath()) {
8362 if (DAG.isKnownNeverNaN(LHS)) {
8363 // Put the potential NaN in the RHS so that SSE will preserve it.
8364 std::swap(LHS, RHS);
8365 } else if (!DAG.isKnownNeverNaN(RHS))
8368 Opcode = X86ISD::FMIN;
8371 // This can be a min, but if either operand is a NaN we need it to
8372 // preserve the original LHS.
8373 std::swap(LHS, RHS);
8377 Opcode = X86ISD::FMIN;
8381 // This can be a max if we can prove that at least one of the operands
8383 if (!FiniteOnlyFPMath()) {
8384 if (DAG.isKnownNeverNaN(LHS)) {
8385 // Put the potential NaN in the RHS so that SSE will preserve it.
8386 std::swap(LHS, RHS);
8387 } else if (!DAG.isKnownNeverNaN(RHS))
8390 Opcode = X86ISD::FMAX;
8393 // This can be a max if we can prove that at least one of the operands
8395 if (!FiniteOnlyFPMath()) {
8396 if (DAG.isKnownNeverNaN(RHS)) {
8397 // Put the potential NaN in the RHS so that SSE will preserve it.
8398 std::swap(LHS, RHS);
8399 } else if (!DAG.isKnownNeverNaN(LHS))
8402 Opcode = X86ISD::FMAX;
8405 // This can be a max, but if either operand is a NaN we need it to
8406 // preserve the original LHS.
8407 std::swap(LHS, RHS);
8411 Opcode = X86ISD::FMAX;
8414 // Check for x CC y ? y : x -- a min/max with reversed arms.
8415 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8419 // This can be a min if we can prove that at least one of the operands
8421 if (!FiniteOnlyFPMath()) {
8422 if (DAG.isKnownNeverNaN(RHS)) {
8423 // Put the potential NaN in the RHS so that SSE will preserve it.
8424 std::swap(LHS, RHS);
8425 } else if (!DAG.isKnownNeverNaN(LHS))
8428 Opcode = X86ISD::FMIN;
8431 // This can be a min if we can prove that at least one of the operands
8433 if (!FiniteOnlyFPMath()) {
8434 if (DAG.isKnownNeverNaN(LHS)) {
8435 // Put the potential NaN in the RHS so that SSE will preserve it.
8436 std::swap(LHS, RHS);
8437 } else if (!DAG.isKnownNeverNaN(RHS))
8440 Opcode = X86ISD::FMIN;
8443 // This can be a min, but if either operand is a NaN we need it to
8444 // preserve the original LHS.
8445 std::swap(LHS, RHS);
8449 Opcode = X86ISD::FMIN;
8453 // This can be a max if we can prove that at least one of the operands
8455 if (!FiniteOnlyFPMath()) {
8456 if (DAG.isKnownNeverNaN(LHS)) {
8457 // Put the potential NaN in the RHS so that SSE will preserve it.
8458 std::swap(LHS, RHS);
8459 } else if (!DAG.isKnownNeverNaN(RHS))
8462 Opcode = X86ISD::FMAX;
8465 // This can be a max if we can prove that at least one of the operands
8467 if (!FiniteOnlyFPMath()) {
8468 if (DAG.isKnownNeverNaN(RHS)) {
8469 // Put the potential NaN in the RHS so that SSE will preserve it.
8470 std::swap(LHS, RHS);
8471 } else if (!DAG.isKnownNeverNaN(LHS))
8474 Opcode = X86ISD::FMAX;
8477 // This can be a max, but if either operand is a NaN we need it to
8478 // preserve the original LHS.
8479 std::swap(LHS, RHS);
8483 Opcode = X86ISD::FMAX;
8489 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8492 // If this is a select between two integer constants, try to do some
8494 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8495 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8496 // Don't do this for crazy integer types.
8497 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8498 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8499 // so that TrueC (the true value) is larger than FalseC.
8500 bool NeedsCondInvert = false;
8502 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8503 // Efficiently invertible.
8504 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8505 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8506 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8507 NeedsCondInvert = true;
8508 std::swap(TrueC, FalseC);
8511 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8512 if (FalseC->getAPIntValue() == 0 &&
8513 TrueC->getAPIntValue().isPowerOf2()) {
8514 if (NeedsCondInvert) // Invert the condition if needed.
8515 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8516 DAG.getConstant(1, Cond.getValueType()));
8518 // Zero extend the condition if needed.
8519 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8521 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8522 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8523 DAG.getConstant(ShAmt, MVT::i8));
8526 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8527 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8528 if (NeedsCondInvert) // Invert the condition if needed.
8529 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8530 DAG.getConstant(1, Cond.getValueType()));
8532 // Zero extend the condition if needed.
8533 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8534 FalseC->getValueType(0), Cond);
8535 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8536 SDValue(FalseC, 0));
8539 // Optimize cases that will turn into an LEA instruction. This requires
8540 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8541 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8542 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8543 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8545 bool isFastMultiplier = false;
8547 switch ((unsigned char)Diff) {
8549 case 1: // result = add base, cond
8550 case 2: // result = lea base( , cond*2)
8551 case 3: // result = lea base(cond, cond*2)
8552 case 4: // result = lea base( , cond*4)
8553 case 5: // result = lea base(cond, cond*4)
8554 case 8: // result = lea base( , cond*8)
8555 case 9: // result = lea base(cond, cond*8)
8556 isFastMultiplier = true;
8561 if (isFastMultiplier) {
8562 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8563 if (NeedsCondInvert) // Invert the condition if needed.
8564 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8565 DAG.getConstant(1, Cond.getValueType()));
8567 // Zero extend the condition if needed.
8568 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8570 // Scale the condition by the difference.
8572 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8573 DAG.getConstant(Diff, Cond.getValueType()));
8575 // Add the base if non-zero.
8576 if (FalseC->getAPIntValue() != 0)
8577 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8578 SDValue(FalseC, 0));
8588 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8589 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8590 TargetLowering::DAGCombinerInfo &DCI) {
8591 DebugLoc DL = N->getDebugLoc();
8593 // If the flag operand isn't dead, don't touch this CMOV.
8594 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8597 // If this is a select between two integer constants, try to do some
8598 // optimizations. Note that the operands are ordered the opposite of SELECT
8600 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8601 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8602 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8603 // larger than FalseC (the false value).
8604 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8606 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8607 CC = X86::GetOppositeBranchCondition(CC);
8608 std::swap(TrueC, FalseC);
8611 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8612 // This is efficient for any integer data type (including i8/i16) and
8614 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8615 SDValue Cond = N->getOperand(3);
8616 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8617 DAG.getConstant(CC, MVT::i8), Cond);
8619 // Zero extend the condition if needed.
8620 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8622 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8623 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8624 DAG.getConstant(ShAmt, MVT::i8));
8625 if (N->getNumValues() == 2) // Dead flag value?
8626 return DCI.CombineTo(N, Cond, SDValue());
8630 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8631 // for any integer data type, including i8/i16.
8632 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8633 SDValue Cond = N->getOperand(3);
8634 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8635 DAG.getConstant(CC, MVT::i8), Cond);
8637 // Zero extend the condition if needed.
8638 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8639 FalseC->getValueType(0), Cond);
8640 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8641 SDValue(FalseC, 0));
8643 if (N->getNumValues() == 2) // Dead flag value?
8644 return DCI.CombineTo(N, Cond, SDValue());
8648 // Optimize cases that will turn into an LEA instruction. This requires
8649 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8650 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8651 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8652 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8654 bool isFastMultiplier = false;
8656 switch ((unsigned char)Diff) {
8658 case 1: // result = add base, cond
8659 case 2: // result = lea base( , cond*2)
8660 case 3: // result = lea base(cond, cond*2)
8661 case 4: // result = lea base( , cond*4)
8662 case 5: // result = lea base(cond, cond*4)
8663 case 8: // result = lea base( , cond*8)
8664 case 9: // result = lea base(cond, cond*8)
8665 isFastMultiplier = true;
8670 if (isFastMultiplier) {
8671 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8672 SDValue Cond = N->getOperand(3);
8673 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8674 DAG.getConstant(CC, MVT::i8), Cond);
8675 // Zero extend the condition if needed.
8676 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8678 // Scale the condition by the difference.
8680 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8681 DAG.getConstant(Diff, Cond.getValueType()));
8683 // Add the base if non-zero.
8684 if (FalseC->getAPIntValue() != 0)
8685 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8686 SDValue(FalseC, 0));
8687 if (N->getNumValues() == 2) // Dead flag value?
8688 return DCI.CombineTo(N, Cond, SDValue());
8698 /// PerformMulCombine - Optimize a single multiply with constant into two
8699 /// in order to implement it with two cheaper instructions, e.g.
8700 /// LEA + SHL, LEA + LEA.
8701 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8702 TargetLowering::DAGCombinerInfo &DCI) {
8703 if (DAG.getMachineFunction().
8704 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8707 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8710 EVT VT = N->getValueType(0);
8714 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8717 uint64_t MulAmt = C->getZExtValue();
8718 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8721 uint64_t MulAmt1 = 0;
8722 uint64_t MulAmt2 = 0;
8723 if ((MulAmt % 9) == 0) {
8725 MulAmt2 = MulAmt / 9;
8726 } else if ((MulAmt % 5) == 0) {
8728 MulAmt2 = MulAmt / 5;
8729 } else if ((MulAmt % 3) == 0) {
8731 MulAmt2 = MulAmt / 3;
8734 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8735 DebugLoc DL = N->getDebugLoc();
8737 if (isPowerOf2_64(MulAmt2) &&
8738 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8739 // If second multiplifer is pow2, issue it first. We want the multiply by
8740 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8742 std::swap(MulAmt1, MulAmt2);
8745 if (isPowerOf2_64(MulAmt1))
8746 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8747 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8749 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8750 DAG.getConstant(MulAmt1, VT));
8752 if (isPowerOf2_64(MulAmt2))
8753 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8754 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8756 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8757 DAG.getConstant(MulAmt2, VT));
8759 // Do not add new nodes to DAG combiner worklist.
8760 DCI.CombineTo(N, NewMul, false);
8766 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8768 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8769 const X86Subtarget *Subtarget) {
8770 // On X86 with SSE2 support, we can transform this to a vector shift if
8771 // all elements are shifted by the same amount. We can't do this in legalize
8772 // because the a constant vector is typically transformed to a constant pool
8773 // so we have no knowledge of the shift amount.
8774 if (!Subtarget->hasSSE2())
8777 EVT VT = N->getValueType(0);
8778 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8781 SDValue ShAmtOp = N->getOperand(1);
8782 EVT EltVT = VT.getVectorElementType();
8783 DebugLoc DL = N->getDebugLoc();
8784 SDValue BaseShAmt = SDValue();
8785 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8786 unsigned NumElts = VT.getVectorNumElements();
8788 for (; i != NumElts; ++i) {
8789 SDValue Arg = ShAmtOp.getOperand(i);
8790 if (Arg.getOpcode() == ISD::UNDEF) continue;
8794 for (; i != NumElts; ++i) {
8795 SDValue Arg = ShAmtOp.getOperand(i);
8796 if (Arg.getOpcode() == ISD::UNDEF) continue;
8797 if (Arg != BaseShAmt) {
8801 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8802 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8803 SDValue InVec = ShAmtOp.getOperand(0);
8804 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
8805 unsigned NumElts = InVec.getValueType().getVectorNumElements();
8807 for (; i != NumElts; ++i) {
8808 SDValue Arg = InVec.getOperand(i);
8809 if (Arg.getOpcode() == ISD::UNDEF) continue;
8813 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
8814 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
8815 unsigned SplatIdx = cast<ShuffleVectorSDNode>(ShAmtOp)->getSplatIndex();
8816 if (C->getZExtValue() == SplatIdx)
8817 BaseShAmt = InVec.getOperand(1);
8820 if (BaseShAmt.getNode() == 0)
8821 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8822 DAG.getIntPtrConstant(0));
8826 // The shift amount is an i32.
8827 if (EltVT.bitsGT(MVT::i32))
8828 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8829 else if (EltVT.bitsLT(MVT::i32))
8830 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, BaseShAmt);
8832 // The shift amount is identical so we can do a vector shift.
8833 SDValue ValOp = N->getOperand(0);
8834 switch (N->getOpcode()) {
8836 llvm_unreachable("Unknown shift opcode!");
8839 if (VT == MVT::v2i64)
8840 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8841 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8843 if (VT == MVT::v4i32)
8844 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8845 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8847 if (VT == MVT::v8i16)
8848 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8849 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8853 if (VT == MVT::v4i32)
8854 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8855 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8857 if (VT == MVT::v8i16)
8858 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8859 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8863 if (VT == MVT::v2i64)
8864 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8865 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8867 if (VT == MVT::v4i32)
8868 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8869 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8871 if (VT == MVT::v8i16)
8872 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8873 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8880 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8881 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8882 const X86Subtarget *Subtarget) {
8883 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8884 // the FP state in cases where an emms may be missing.
8885 // A preferable solution to the general problem is to figure out the right
8886 // places to insert EMMS. This qualifies as a quick hack.
8888 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8889 StoreSDNode *St = cast<StoreSDNode>(N);
8890 EVT VT = St->getValue().getValueType();
8891 if (VT.getSizeInBits() != 64)
8894 const Function *F = DAG.getMachineFunction().getFunction();
8895 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
8896 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
8897 && Subtarget->hasSSE2();
8898 if ((VT.isVector() ||
8899 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8900 isa<LoadSDNode>(St->getValue()) &&
8901 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8902 St->getChain().hasOneUse() && !St->isVolatile()) {
8903 SDNode* LdVal = St->getValue().getNode();
8905 int TokenFactorIndex = -1;
8906 SmallVector<SDValue, 8> Ops;
8907 SDNode* ChainVal = St->getChain().getNode();
8908 // Must be a store of a load. We currently handle two cases: the load
8909 // is a direct child, and it's under an intervening TokenFactor. It is
8910 // possible to dig deeper under nested TokenFactors.
8911 if (ChainVal == LdVal)
8912 Ld = cast<LoadSDNode>(St->getChain());
8913 else if (St->getValue().hasOneUse() &&
8914 ChainVal->getOpcode() == ISD::TokenFactor) {
8915 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8916 if (ChainVal->getOperand(i).getNode() == LdVal) {
8917 TokenFactorIndex = i;
8918 Ld = cast<LoadSDNode>(St->getValue());
8920 Ops.push_back(ChainVal->getOperand(i));
8924 if (!Ld || !ISD::isNormalLoad(Ld))
8927 // If this is not the MMX case, i.e. we are just turning i64 load/store
8928 // into f64 load/store, avoid the transformation if there are multiple
8929 // uses of the loaded value.
8930 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8933 DebugLoc LdDL = Ld->getDebugLoc();
8934 DebugLoc StDL = N->getDebugLoc();
8935 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8936 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8938 if (Subtarget->is64Bit() || F64IsLegal) {
8939 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8940 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8941 Ld->getBasePtr(), Ld->getSrcValue(),
8942 Ld->getSrcValueOffset(), Ld->isVolatile(),
8943 Ld->getAlignment());
8944 SDValue NewChain = NewLd.getValue(1);
8945 if (TokenFactorIndex != -1) {
8946 Ops.push_back(NewChain);
8947 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8950 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8951 St->getSrcValue(), St->getSrcValueOffset(),
8952 St->isVolatile(), St->getAlignment());
8955 // Otherwise, lower to two pairs of 32-bit loads / stores.
8956 SDValue LoAddr = Ld->getBasePtr();
8957 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8958 DAG.getConstant(4, MVT::i32));
8960 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8961 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8962 Ld->isVolatile(), Ld->getAlignment());
8963 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8964 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8966 MinAlign(Ld->getAlignment(), 4));
8968 SDValue NewChain = LoLd.getValue(1);
8969 if (TokenFactorIndex != -1) {
8970 Ops.push_back(LoLd);
8971 Ops.push_back(HiLd);
8972 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8976 LoAddr = St->getBasePtr();
8977 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8978 DAG.getConstant(4, MVT::i32));
8980 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8981 St->getSrcValue(), St->getSrcValueOffset(),
8982 St->isVolatile(), St->getAlignment());
8983 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8985 St->getSrcValueOffset() + 4,
8987 MinAlign(St->getAlignment(), 4));
8988 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8993 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8994 /// X86ISD::FXOR nodes.
8995 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8996 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8997 // F[X]OR(0.0, x) -> x
8998 // F[X]OR(x, 0.0) -> x
8999 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9000 if (C->getValueAPF().isPosZero())
9001 return N->getOperand(1);
9002 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9003 if (C->getValueAPF().isPosZero())
9004 return N->getOperand(0);
9008 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
9009 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
9010 // FAND(0.0, x) -> 0.0
9011 // FAND(x, 0.0) -> 0.0
9012 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
9013 if (C->getValueAPF().isPosZero())
9014 return N->getOperand(0);
9015 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
9016 if (C->getValueAPF().isPosZero())
9017 return N->getOperand(1);
9021 static SDValue PerformBTCombine(SDNode *N,
9023 TargetLowering::DAGCombinerInfo &DCI) {
9024 // BT ignores high bits in the bit index operand.
9025 SDValue Op1 = N->getOperand(1);
9026 if (Op1.hasOneUse()) {
9027 unsigned BitWidth = Op1.getValueSizeInBits();
9028 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
9029 APInt KnownZero, KnownOne;
9030 TargetLowering::TargetLoweringOpt TLO(DAG);
9031 TargetLowering &TLI = DAG.getTargetLoweringInfo();
9032 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
9033 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
9034 DCI.CommitTargetLoweringOpt(TLO);
9039 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
9040 SDValue Op = N->getOperand(0);
9041 if (Op.getOpcode() == ISD::BIT_CONVERT)
9042 Op = Op.getOperand(0);
9043 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
9044 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
9045 VT.getVectorElementType().getSizeInBits() ==
9046 OpVT.getVectorElementType().getSizeInBits()) {
9047 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
9052 // On X86 and X86-64, atomic operations are lowered to locked instructions.
9053 // Locked instructions, in turn, have implicit fence semantics (all memory
9054 // operations are flushed before issuing the locked instruction, and the
9055 // are not buffered), so we can fold away the common pattern of
9056 // fence-atomic-fence.
9057 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
9058 SDValue atomic = N->getOperand(0);
9059 switch (atomic.getOpcode()) {
9060 case ISD::ATOMIC_CMP_SWAP:
9061 case ISD::ATOMIC_SWAP:
9062 case ISD::ATOMIC_LOAD_ADD:
9063 case ISD::ATOMIC_LOAD_SUB:
9064 case ISD::ATOMIC_LOAD_AND:
9065 case ISD::ATOMIC_LOAD_OR:
9066 case ISD::ATOMIC_LOAD_XOR:
9067 case ISD::ATOMIC_LOAD_NAND:
9068 case ISD::ATOMIC_LOAD_MIN:
9069 case ISD::ATOMIC_LOAD_MAX:
9070 case ISD::ATOMIC_LOAD_UMIN:
9071 case ISD::ATOMIC_LOAD_UMAX:
9077 SDValue fence = atomic.getOperand(0);
9078 if (fence.getOpcode() != ISD::MEMBARRIER)
9081 switch (atomic.getOpcode()) {
9082 case ISD::ATOMIC_CMP_SWAP:
9083 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9084 atomic.getOperand(1), atomic.getOperand(2),
9085 atomic.getOperand(3));
9086 case ISD::ATOMIC_SWAP:
9087 case ISD::ATOMIC_LOAD_ADD:
9088 case ISD::ATOMIC_LOAD_SUB:
9089 case ISD::ATOMIC_LOAD_AND:
9090 case ISD::ATOMIC_LOAD_OR:
9091 case ISD::ATOMIC_LOAD_XOR:
9092 case ISD::ATOMIC_LOAD_NAND:
9093 case ISD::ATOMIC_LOAD_MIN:
9094 case ISD::ATOMIC_LOAD_MAX:
9095 case ISD::ATOMIC_LOAD_UMIN:
9096 case ISD::ATOMIC_LOAD_UMAX:
9097 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
9098 atomic.getOperand(1), atomic.getOperand(2));
9104 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
9105 DAGCombinerInfo &DCI) const {
9106 SelectionDAG &DAG = DCI.DAG;
9107 switch (N->getOpcode()) {
9109 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
9110 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
9111 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
9112 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
9115 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
9116 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
9118 case X86ISD::FOR: return PerformFORCombine(N, DAG);
9119 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
9120 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
9121 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
9122 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
9128 //===----------------------------------------------------------------------===//
9129 // X86 Inline Assembly Support
9130 //===----------------------------------------------------------------------===//
9132 static bool LowerToBSwap(CallInst *CI) {
9133 // FIXME: this should verify that we are targetting a 486 or better. If not,
9134 // we will turn this bswap into something that will be lowered to logical ops
9135 // instead of emitting the bswap asm. For now, we don't support 486 or lower
9136 // so don't worry about this.
9138 // Verify this is a simple bswap.
9139 if (CI->getNumOperands() != 2 ||
9140 CI->getType() != CI->getOperand(1)->getType() ||
9141 !CI->getType()->isInteger())
9144 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
9145 if (!Ty || Ty->getBitWidth() % 16 != 0)
9148 // Okay, we can do this xform, do so now.
9149 const Type *Tys[] = { Ty };
9150 Module *M = CI->getParent()->getParent()->getParent();
9151 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
9153 Value *Op = CI->getOperand(1);
9154 Op = CallInst::Create(Int, Op, CI->getName(), CI);
9156 CI->replaceAllUsesWith(Op);
9157 CI->eraseFromParent();
9161 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
9162 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
9163 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
9165 std::string AsmStr = IA->getAsmString();
9167 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
9168 std::vector<std::string> AsmPieces;
9169 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
9171 switch (AsmPieces.size()) {
9172 default: return false;
9174 AsmStr = AsmPieces[0];
9176 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
9179 if (AsmPieces.size() == 2 &&
9180 (AsmPieces[0] == "bswap" ||
9181 AsmPieces[0] == "bswapq" ||
9182 AsmPieces[0] == "bswapl") &&
9183 (AsmPieces[1] == "$0" ||
9184 AsmPieces[1] == "${0:q}")) {
9185 // No need to check constraints, nothing other than the equivalent of
9186 // "=r,0" would be valid here.
9187 return LowerToBSwap(CI);
9189 // rorw $$8, ${0:w} --> llvm.bswap.i16
9190 if (CI->getType() == Type::getInt16Ty(CI->getContext()) &&
9191 AsmPieces.size() == 3 &&
9192 AsmPieces[0] == "rorw" &&
9193 AsmPieces[1] == "$$8," &&
9194 AsmPieces[2] == "${0:w}" &&
9195 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
9196 return LowerToBSwap(CI);
9200 if (CI->getType() == Type::getInt64Ty(CI->getContext()) &&
9201 Constraints.size() >= 2 &&
9202 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
9203 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
9204 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
9205 std::vector<std::string> Words;
9206 SplitString(AsmPieces[0], Words, " \t");
9207 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
9209 SplitString(AsmPieces[1], Words, " \t");
9210 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
9212 SplitString(AsmPieces[2], Words, " \t,");
9213 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
9214 Words[2] == "%edx") {
9215 return LowerToBSwap(CI);
9227 /// getConstraintType - Given a constraint letter, return the type of
9228 /// constraint it is for this target.
9229 X86TargetLowering::ConstraintType
9230 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
9231 if (Constraint.size() == 1) {
9232 switch (Constraint[0]) {
9244 return C_RegisterClass;
9252 return TargetLowering::getConstraintType(Constraint);
9255 /// LowerXConstraint - try to replace an X constraint, which matches anything,
9256 /// with another that has more specific requirements based on the type of the
9257 /// corresponding operand.
9258 const char *X86TargetLowering::
9259 LowerXConstraint(EVT ConstraintVT) const {
9260 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
9261 // 'f' like normal targets.
9262 if (ConstraintVT.isFloatingPoint()) {
9263 if (Subtarget->hasSSE2())
9265 if (Subtarget->hasSSE1())
9269 return TargetLowering::LowerXConstraint(ConstraintVT);
9272 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9273 /// vector. If it is invalid, don't add anything to Ops.
9274 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9277 std::vector<SDValue>&Ops,
9278 SelectionDAG &DAG) const {
9279 SDValue Result(0, 0);
9281 switch (Constraint) {
9284 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9285 if (C->getZExtValue() <= 31) {
9286 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9292 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9293 if (C->getZExtValue() <= 63) {
9294 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9300 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9301 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9302 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9308 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9309 if (C->getZExtValue() <= 255) {
9310 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9316 // 32-bit signed value
9317 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9318 const ConstantInt *CI = C->getConstantIntValue();
9319 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9320 C->getSExtValue())) {
9321 // Widen to 64 bits here to get it sign extended.
9322 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9325 // FIXME gcc accepts some relocatable values here too, but only in certain
9326 // memory models; it's complicated.
9331 // 32-bit unsigned value
9332 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9333 const ConstantInt *CI = C->getConstantIntValue();
9334 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9335 C->getZExtValue())) {
9336 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9340 // FIXME gcc accepts some relocatable values here too, but only in certain
9341 // memory models; it's complicated.
9345 // Literal immediates are always ok.
9346 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9347 // Widen to 64 bits here to get it sign extended.
9348 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9352 // If we are in non-pic codegen mode, we allow the address of a global (with
9353 // an optional displacement) to be used with 'i'.
9354 GlobalAddressSDNode *GA = 0;
9357 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9359 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9360 Offset += GA->getOffset();
9362 } else if (Op.getOpcode() == ISD::ADD) {
9363 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9364 Offset += C->getZExtValue();
9365 Op = Op.getOperand(0);
9368 } else if (Op.getOpcode() == ISD::SUB) {
9369 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9370 Offset += -C->getZExtValue();
9371 Op = Op.getOperand(0);
9376 // Otherwise, this isn't something we can handle, reject it.
9380 GlobalValue *GV = GA->getGlobal();
9381 // If we require an extra load to get this address, as in PIC mode, we
9383 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9384 getTargetMachine())))
9388 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9390 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9396 if (Result.getNode()) {
9397 Ops.push_back(Result);
9400 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9404 std::vector<unsigned> X86TargetLowering::
9405 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9407 if (Constraint.size() == 1) {
9408 // FIXME: not handling fp-stack yet!
9409 switch (Constraint[0]) { // GCC X86 Constraint Letters
9410 default: break; // Unknown constraint letter
9411 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9412 if (Subtarget->is64Bit()) {
9414 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9415 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9416 X86::R10D,X86::R11D,X86::R12D,
9417 X86::R13D,X86::R14D,X86::R15D,
9418 X86::EBP, X86::ESP, 0);
9419 else if (VT == MVT::i16)
9420 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9421 X86::SI, X86::DI, X86::R8W,X86::R9W,
9422 X86::R10W,X86::R11W,X86::R12W,
9423 X86::R13W,X86::R14W,X86::R15W,
9424 X86::BP, X86::SP, 0);
9425 else if (VT == MVT::i8)
9426 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9427 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9428 X86::R10B,X86::R11B,X86::R12B,
9429 X86::R13B,X86::R14B,X86::R15B,
9430 X86::BPL, X86::SPL, 0);
9432 else if (VT == MVT::i64)
9433 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9434 X86::RSI, X86::RDI, X86::R8, X86::R9,
9435 X86::R10, X86::R11, X86::R12,
9436 X86::R13, X86::R14, X86::R15,
9437 X86::RBP, X86::RSP, 0);
9441 // 32-bit fallthrough
9444 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9445 else if (VT == MVT::i16)
9446 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9447 else if (VT == MVT::i8)
9448 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9449 else if (VT == MVT::i64)
9450 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9455 return std::vector<unsigned>();
9458 std::pair<unsigned, const TargetRegisterClass*>
9459 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9461 // First, see if this is a constraint that directly corresponds to an LLVM
9463 if (Constraint.size() == 1) {
9464 // GCC Constraint Letters
9465 switch (Constraint[0]) {
9467 case 'r': // GENERAL_REGS
9468 case 'l': // INDEX_REGS
9470 return std::make_pair(0U, X86::GR8RegisterClass);
9472 return std::make_pair(0U, X86::GR16RegisterClass);
9473 if (VT == MVT::i32 || !Subtarget->is64Bit())
9474 return std::make_pair(0U, X86::GR32RegisterClass);
9475 return std::make_pair(0U, X86::GR64RegisterClass);
9476 case 'R': // LEGACY_REGS
9478 return std::make_pair(0U, X86::GR8_NOREXRegisterClass);
9480 return std::make_pair(0U, X86::GR16_NOREXRegisterClass);
9481 if (VT == MVT::i32 || !Subtarget->is64Bit())
9482 return std::make_pair(0U, X86::GR32_NOREXRegisterClass);
9483 return std::make_pair(0U, X86::GR64_NOREXRegisterClass);
9484 case 'f': // FP Stack registers.
9485 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
9486 // value to the correct fpstack register class.
9487 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9488 return std::make_pair(0U, X86::RFP32RegisterClass);
9489 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9490 return std::make_pair(0U, X86::RFP64RegisterClass);
9491 return std::make_pair(0U, X86::RFP80RegisterClass);
9492 case 'y': // MMX_REGS if MMX allowed.
9493 if (!Subtarget->hasMMX()) break;
9494 return std::make_pair(0U, X86::VR64RegisterClass);
9495 case 'Y': // SSE_REGS if SSE2 allowed
9496 if (!Subtarget->hasSSE2()) break;
9498 case 'x': // SSE_REGS if SSE1 allowed
9499 if (!Subtarget->hasSSE1()) break;
9501 switch (VT.getSimpleVT().SimpleTy) {
9503 // Scalar SSE types.
9506 return std::make_pair(0U, X86::FR32RegisterClass);
9509 return std::make_pair(0U, X86::FR64RegisterClass);
9517 return std::make_pair(0U, X86::VR128RegisterClass);
9523 // Use the default implementation in TargetLowering to convert the register
9524 // constraint into a member of a register class.
9525 std::pair<unsigned, const TargetRegisterClass*> Res;
9526 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9528 // Not found as a standard register?
9529 if (Res.second == 0) {
9530 // Map st(0) -> st(7) -> ST0
9531 if (Constraint.size() == 7 && Constraint[0] == '{' &&
9532 tolower(Constraint[1]) == 's' &&
9533 tolower(Constraint[2]) == 't' &&
9534 Constraint[3] == '(' &&
9535 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
9536 Constraint[5] == ')' &&
9537 Constraint[6] == '}') {
9539 Res.first = X86::ST0+Constraint[4]-'0';
9540 Res.second = X86::RFP80RegisterClass;
9544 // GCC allows "st(0)" to be called just plain "st".
9545 if (StringsEqualNoCase("{st}", Constraint)) {
9546 Res.first = X86::ST0;
9547 Res.second = X86::RFP80RegisterClass;
9552 if (StringsEqualNoCase("{flags}", Constraint)) {
9553 Res.first = X86::EFLAGS;
9554 Res.second = X86::CCRRegisterClass;
9558 // 'A' means EAX + EDX.
9559 if (Constraint == "A") {
9560 Res.first = X86::EAX;
9561 Res.second = X86::GR32_ADRegisterClass;
9567 // Otherwise, check to see if this is a register class of the wrong value
9568 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
9569 // turn into {ax},{dx}.
9570 if (Res.second->hasType(VT))
9571 return Res; // Correct type already, nothing to do.
9573 // All of the single-register GCC register classes map their values onto
9574 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
9575 // really want an 8-bit or 32-bit register, map to the appropriate register
9576 // class and return the appropriate register.
9577 if (Res.second == X86::GR16RegisterClass) {
9578 if (VT == MVT::i8) {
9579 unsigned DestReg = 0;
9580 switch (Res.first) {
9582 case X86::AX: DestReg = X86::AL; break;
9583 case X86::DX: DestReg = X86::DL; break;
9584 case X86::CX: DestReg = X86::CL; break;
9585 case X86::BX: DestReg = X86::BL; break;
9588 Res.first = DestReg;
9589 Res.second = X86::GR8RegisterClass;
9591 } else if (VT == MVT::i32) {
9592 unsigned DestReg = 0;
9593 switch (Res.first) {
9595 case X86::AX: DestReg = X86::EAX; break;
9596 case X86::DX: DestReg = X86::EDX; break;
9597 case X86::CX: DestReg = X86::ECX; break;
9598 case X86::BX: DestReg = X86::EBX; break;
9599 case X86::SI: DestReg = X86::ESI; break;
9600 case X86::DI: DestReg = X86::EDI; break;
9601 case X86::BP: DestReg = X86::EBP; break;
9602 case X86::SP: DestReg = X86::ESP; break;
9605 Res.first = DestReg;
9606 Res.second = X86::GR32RegisterClass;
9608 } else if (VT == MVT::i64) {
9609 unsigned DestReg = 0;
9610 switch (Res.first) {
9612 case X86::AX: DestReg = X86::RAX; break;
9613 case X86::DX: DestReg = X86::RDX; break;
9614 case X86::CX: DestReg = X86::RCX; break;
9615 case X86::BX: DestReg = X86::RBX; break;
9616 case X86::SI: DestReg = X86::RSI; break;
9617 case X86::DI: DestReg = X86::RDI; break;
9618 case X86::BP: DestReg = X86::RBP; break;
9619 case X86::SP: DestReg = X86::RSP; break;
9622 Res.first = DestReg;
9623 Res.second = X86::GR64RegisterClass;
9626 } else if (Res.second == X86::FR32RegisterClass ||
9627 Res.second == X86::FR64RegisterClass ||
9628 Res.second == X86::VR128RegisterClass) {
9629 // Handle references to XMM physical registers that got mapped into the
9630 // wrong class. This can happen with constraints like {xmm0} where the
9631 // target independent register mapper will just pick the first match it can
9632 // find, ignoring the required type.
9634 Res.second = X86::FR32RegisterClass;
9635 else if (VT == MVT::f64)
9636 Res.second = X86::FR64RegisterClass;
9637 else if (X86::VR128RegisterClass->hasType(VT))
9638 Res.second = X86::VR128RegisterClass;
9644 //===----------------------------------------------------------------------===//
9645 // X86 Widen vector type
9646 //===----------------------------------------------------------------------===//
9648 /// getWidenVectorType: given a vector type, returns the type to widen
9649 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9650 /// If there is no vector type that we want to widen to, returns MVT::Other
9651 /// When and where to widen is target dependent based on the cost of
9652 /// scalarizing vs using the wider vector type.
9654 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
9655 assert(VT.isVector());
9656 if (isTypeLegal(VT))
9659 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9660 // type based on element type. This would speed up our search (though
9661 // it may not be worth it since the size of the list is relatively
9663 EVT EltVT = VT.getVectorElementType();
9664 unsigned NElts = VT.getVectorNumElements();
9666 // On X86, it make sense to widen any vector wider than 1
9670 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9671 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9672 EVT SVT = (MVT::SimpleValueType)nVT;
9674 if (isTypeLegal(SVT) &&
9675 SVT.getVectorElementType() == EltVT &&
9676 SVT.getVectorNumElements() > NElts)