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 "llvm/CallingConv.h"
20 #include "llvm/Constants.h"
21 #include "llvm/DerivedTypes.h"
22 #include "llvm/GlobalAlias.h"
23 #include "llvm/GlobalVariable.h"
24 #include "llvm/Function.h"
25 #include "llvm/Instructions.h"
26 #include "llvm/Intrinsics.h"
27 #include "llvm/LLVMContext.h"
28 #include "llvm/ADT/BitVector.h"
29 #include "llvm/ADT/VectorExtras.h"
30 #include "llvm/CodeGen/MachineFrameInfo.h"
31 #include "llvm/CodeGen/MachineFunction.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/PseudoSourceValue.h"
36 #include "llvm/Support/MathExtras.h"
37 #include "llvm/Support/Debug.h"
38 #include "llvm/Support/ErrorHandling.h"
39 #include "llvm/Target/TargetLoweringObjectFile.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 // Forward declarations.
51 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
54 static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
55 switch (TM.getSubtarget<X86Subtarget>().TargetType) {
56 default: llvm_unreachable("unknown subtarget type");
57 case X86Subtarget::isDarwin:
58 return new TargetLoweringObjectFileMachO();
59 case X86Subtarget::isELF:
60 return new TargetLoweringObjectFileELF();
61 case X86Subtarget::isMingw:
62 case X86Subtarget::isCygwin:
63 case X86Subtarget::isWindows:
64 return new TargetLoweringObjectFileCOFF();
69 X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
70 : TargetLowering(TM, createTLOF(TM)) {
71 Subtarget = &TM.getSubtarget<X86Subtarget>();
72 X86ScalarSSEf64 = Subtarget->hasSSE2();
73 X86ScalarSSEf32 = Subtarget->hasSSE1();
74 X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
76 RegInfo = TM.getRegisterInfo();
79 // Set up the TargetLowering object.
81 // X86 is weird, it always uses i8 for shift amounts and setcc results.
82 setShiftAmountType(MVT::i8);
83 setBooleanContents(ZeroOrOneBooleanContent);
84 setSchedulingPreference(SchedulingForRegPressure);
85 setStackPointerRegisterToSaveRestore(X86StackPtr);
87 if (Subtarget->isTargetDarwin()) {
88 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
89 setUseUnderscoreSetJmp(false);
90 setUseUnderscoreLongJmp(false);
91 } else if (Subtarget->isTargetMingw()) {
92 // MS runtime is weird: it exports _setjmp, but longjmp!
93 setUseUnderscoreSetJmp(true);
94 setUseUnderscoreLongJmp(false);
96 setUseUnderscoreSetJmp(true);
97 setUseUnderscoreLongJmp(true);
100 // Set up the register classes.
101 addRegisterClass(MVT::i8, X86::GR8RegisterClass);
102 addRegisterClass(MVT::i16, X86::GR16RegisterClass);
103 addRegisterClass(MVT::i32, X86::GR32RegisterClass);
104 if (Subtarget->is64Bit())
105 addRegisterClass(MVT::i64, X86::GR64RegisterClass);
107 setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
109 // We don't accept any truncstore of integer registers.
110 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
111 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
112 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
113 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
114 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
115 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
117 // SETOEQ and SETUNE require checking two conditions.
118 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
119 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
120 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
121 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
122 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
123 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
125 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
127 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
128 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
129 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
131 if (Subtarget->is64Bit()) {
132 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
133 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Expand);
134 } else if (!UseSoftFloat) {
135 if (X86ScalarSSEf64) {
136 // We have an impenetrably clever algorithm for ui64->double only.
137 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
139 // We have an algorithm for SSE2, and we turn this into a 64-bit
140 // FILD for other targets.
141 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
144 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
146 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
147 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
150 // SSE has no i16 to fp conversion, only i32
151 if (X86ScalarSSEf32) {
152 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
153 // f32 and f64 cases are Legal, f80 case is not
154 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
156 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
157 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
160 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
161 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
164 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
165 // are Legal, f80 is custom lowered.
166 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
167 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
169 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
171 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
172 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
174 if (X86ScalarSSEf32) {
175 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
176 // f32 and f64 cases are Legal, f80 case is not
177 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
179 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
180 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
183 // Handle FP_TO_UINT by promoting the destination to a larger signed
185 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
186 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
187 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
189 if (Subtarget->is64Bit()) {
190 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
191 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
192 } else if (!UseSoftFloat) {
193 if (X86ScalarSSEf32 && !Subtarget->hasSSE3())
194 // Expand FP_TO_UINT into a select.
195 // FIXME: We would like to use a Custom expander here eventually to do
196 // the optimal thing for SSE vs. the default expansion in the legalizer.
197 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
199 // With SSE3 we can use fisttpll to convert to a signed i64; without
200 // SSE, we're stuck with a fistpll.
201 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
204 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
205 if (!X86ScalarSSEf64) {
206 setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
207 setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
210 // Scalar integer divide and remainder are lowered to use operations that
211 // produce two results, to match the available instructions. This exposes
212 // the two-result form to trivial CSE, which is able to combine x/y and x%y
213 // into a single instruction.
215 // Scalar integer multiply-high is also lowered to use two-result
216 // operations, to match the available instructions. However, plain multiply
217 // (low) operations are left as Legal, as there are single-result
218 // instructions for this in x86. Using the two-result multiply instructions
219 // when both high and low results are needed must be arranged by dagcombine.
220 setOperationAction(ISD::MULHS , MVT::i8 , Expand);
221 setOperationAction(ISD::MULHU , MVT::i8 , Expand);
222 setOperationAction(ISD::SDIV , MVT::i8 , Expand);
223 setOperationAction(ISD::UDIV , MVT::i8 , Expand);
224 setOperationAction(ISD::SREM , MVT::i8 , Expand);
225 setOperationAction(ISD::UREM , MVT::i8 , Expand);
226 setOperationAction(ISD::MULHS , MVT::i16 , Expand);
227 setOperationAction(ISD::MULHU , MVT::i16 , Expand);
228 setOperationAction(ISD::SDIV , MVT::i16 , Expand);
229 setOperationAction(ISD::UDIV , MVT::i16 , Expand);
230 setOperationAction(ISD::SREM , MVT::i16 , Expand);
231 setOperationAction(ISD::UREM , MVT::i16 , Expand);
232 setOperationAction(ISD::MULHS , MVT::i32 , Expand);
233 setOperationAction(ISD::MULHU , MVT::i32 , Expand);
234 setOperationAction(ISD::SDIV , MVT::i32 , Expand);
235 setOperationAction(ISD::UDIV , MVT::i32 , Expand);
236 setOperationAction(ISD::SREM , MVT::i32 , Expand);
237 setOperationAction(ISD::UREM , MVT::i32 , Expand);
238 setOperationAction(ISD::MULHS , MVT::i64 , Expand);
239 setOperationAction(ISD::MULHU , MVT::i64 , Expand);
240 setOperationAction(ISD::SDIV , MVT::i64 , Expand);
241 setOperationAction(ISD::UDIV , MVT::i64 , Expand);
242 setOperationAction(ISD::SREM , MVT::i64 , Expand);
243 setOperationAction(ISD::UREM , MVT::i64 , Expand);
245 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
246 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
247 setOperationAction(ISD::BR_CC , MVT::Other, Expand);
248 setOperationAction(ISD::SELECT_CC , MVT::Other, Expand);
249 if (Subtarget->is64Bit())
250 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
251 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
252 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
253 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
254 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
255 setOperationAction(ISD::FREM , MVT::f32 , Expand);
256 setOperationAction(ISD::FREM , MVT::f64 , Expand);
257 setOperationAction(ISD::FREM , MVT::f80 , Expand);
258 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
260 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
261 setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
262 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
263 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
264 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
265 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
266 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
267 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
268 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
269 if (Subtarget->is64Bit()) {
270 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
271 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
272 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
275 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
276 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
278 // These should be promoted to a larger select which is supported.
279 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
280 // X86 wants to expand cmov itself.
281 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
282 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
283 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
284 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
285 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
286 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
287 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
288 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
289 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
290 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
291 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
292 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
293 if (Subtarget->is64Bit()) {
294 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
295 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
297 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
300 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
301 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
302 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
303 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
304 if (Subtarget->is64Bit())
305 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
306 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
307 if (Subtarget->is64Bit()) {
308 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
309 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
310 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
311 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
313 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
314 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
315 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
316 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
317 if (Subtarget->is64Bit()) {
318 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
319 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
320 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
323 if (Subtarget->hasSSE1())
324 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
326 if (!Subtarget->hasSSE2())
327 setOperationAction(ISD::MEMBARRIER , MVT::Other, Expand);
329 // Expand certain atomics
330 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
331 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
332 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
333 setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
335 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
336 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
337 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
338 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
340 if (!Subtarget->is64Bit()) {
341 setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
342 setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
343 setOperationAction(ISD::ATOMIC_LOAD_AND, MVT::i64, Custom);
344 setOperationAction(ISD::ATOMIC_LOAD_OR, MVT::i64, Custom);
345 setOperationAction(ISD::ATOMIC_LOAD_XOR, MVT::i64, Custom);
346 setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i64, Custom);
347 setOperationAction(ISD::ATOMIC_SWAP, MVT::i64, Custom);
350 // Use the default ISD::DBG_STOPPOINT.
351 setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
352 // FIXME - use subtarget debug flags
353 if (!Subtarget->isTargetDarwin() &&
354 !Subtarget->isTargetELF() &&
355 !Subtarget->isTargetCygMing()) {
356 setOperationAction(ISD::DBG_LABEL, MVT::Other, Expand);
357 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
360 setOperationAction(ISD::EXCEPTIONADDR, MVT::i64, Expand);
361 setOperationAction(ISD::EHSELECTION, MVT::i64, Expand);
362 setOperationAction(ISD::EXCEPTIONADDR, MVT::i32, Expand);
363 setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
364 if (Subtarget->is64Bit()) {
365 setExceptionPointerRegister(X86::RAX);
366 setExceptionSelectorRegister(X86::RDX);
368 setExceptionPointerRegister(X86::EAX);
369 setExceptionSelectorRegister(X86::EDX);
371 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
372 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
374 setOperationAction(ISD::TRAMPOLINE, MVT::Other, Custom);
376 setOperationAction(ISD::TRAP, MVT::Other, Legal);
378 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
379 setOperationAction(ISD::VASTART , MVT::Other, Custom);
380 setOperationAction(ISD::VAEND , MVT::Other, Expand);
381 if (Subtarget->is64Bit()) {
382 setOperationAction(ISD::VAARG , MVT::Other, Custom);
383 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
385 setOperationAction(ISD::VAARG , MVT::Other, Expand);
386 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
389 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
390 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
391 if (Subtarget->is64Bit())
392 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
393 if (Subtarget->isTargetCygMing())
394 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
396 setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
398 if (!UseSoftFloat && X86ScalarSSEf64) {
399 // f32 and f64 use SSE.
400 // Set up the FP register classes.
401 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
402 addRegisterClass(MVT::f64, X86::FR64RegisterClass);
404 // Use ANDPD to simulate FABS.
405 setOperationAction(ISD::FABS , MVT::f64, Custom);
406 setOperationAction(ISD::FABS , MVT::f32, Custom);
408 // Use XORP to simulate FNEG.
409 setOperationAction(ISD::FNEG , MVT::f64, Custom);
410 setOperationAction(ISD::FNEG , MVT::f32, Custom);
412 // Use ANDPD and ORPD to simulate FCOPYSIGN.
413 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
414 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
416 // We don't support sin/cos/fmod
417 setOperationAction(ISD::FSIN , MVT::f64, Expand);
418 setOperationAction(ISD::FCOS , MVT::f64, Expand);
419 setOperationAction(ISD::FSIN , MVT::f32, Expand);
420 setOperationAction(ISD::FCOS , MVT::f32, Expand);
422 // Expand FP immediates into loads from the stack, except for the special
424 addLegalFPImmediate(APFloat(+0.0)); // xorpd
425 addLegalFPImmediate(APFloat(+0.0f)); // xorps
426 } else if (!UseSoftFloat && X86ScalarSSEf32) {
427 // Use SSE for f32, x87 for f64.
428 // Set up the FP register classes.
429 addRegisterClass(MVT::f32, X86::FR32RegisterClass);
430 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
432 // Use ANDPS to simulate FABS.
433 setOperationAction(ISD::FABS , MVT::f32, Custom);
435 // Use XORP to simulate FNEG.
436 setOperationAction(ISD::FNEG , MVT::f32, Custom);
438 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
440 // Use ANDPS and ORPS to simulate FCOPYSIGN.
441 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
442 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
444 // We don't support sin/cos/fmod
445 setOperationAction(ISD::FSIN , MVT::f32, Expand);
446 setOperationAction(ISD::FCOS , MVT::f32, Expand);
448 // Special cases we handle for FP constants.
449 addLegalFPImmediate(APFloat(+0.0f)); // xorps
450 addLegalFPImmediate(APFloat(+0.0)); // FLD0
451 addLegalFPImmediate(APFloat(+1.0)); // FLD1
452 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
453 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
456 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
457 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
459 } else if (!UseSoftFloat) {
460 // f32 and f64 in x87.
461 // Set up the FP register classes.
462 addRegisterClass(MVT::f64, X86::RFP64RegisterClass);
463 addRegisterClass(MVT::f32, X86::RFP32RegisterClass);
465 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
466 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
467 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
468 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
471 setOperationAction(ISD::FSIN , MVT::f64 , Expand);
472 setOperationAction(ISD::FCOS , MVT::f64 , Expand);
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
478 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
479 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
480 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
481 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
484 // Long double always uses X87.
486 addRegisterClass(MVT::f80, X86::RFP80RegisterClass);
487 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
488 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
491 APFloat TmpFlt(+0.0);
492 TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
494 addLegalFPImmediate(TmpFlt); // FLD0
496 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
497 APFloat TmpFlt2(+1.0);
498 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
500 addLegalFPImmediate(TmpFlt2); // FLD1
501 TmpFlt2.changeSign();
502 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
506 setOperationAction(ISD::FSIN , MVT::f80 , Expand);
507 setOperationAction(ISD::FCOS , MVT::f80 , Expand);
511 // Always use a library call for pow.
512 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
513 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
514 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
516 setOperationAction(ISD::FLOG, MVT::f80, Expand);
517 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
518 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
519 setOperationAction(ISD::FEXP, MVT::f80, Expand);
520 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
522 // First set operation action for all vector types to either promote
523 // (for widening) or expand (for scalarization). Then we will selectively
524 // turn on ones that can be effectively codegen'd.
525 for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
526 VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
527 setOperationAction(ISD::ADD , (MVT::SimpleValueType)VT, Expand);
528 setOperationAction(ISD::SUB , (MVT::SimpleValueType)VT, Expand);
529 setOperationAction(ISD::FADD, (MVT::SimpleValueType)VT, Expand);
530 setOperationAction(ISD::FNEG, (MVT::SimpleValueType)VT, Expand);
531 setOperationAction(ISD::FSUB, (MVT::SimpleValueType)VT, Expand);
532 setOperationAction(ISD::MUL , (MVT::SimpleValueType)VT, Expand);
533 setOperationAction(ISD::FMUL, (MVT::SimpleValueType)VT, Expand);
534 setOperationAction(ISD::SDIV, (MVT::SimpleValueType)VT, Expand);
535 setOperationAction(ISD::UDIV, (MVT::SimpleValueType)VT, Expand);
536 setOperationAction(ISD::FDIV, (MVT::SimpleValueType)VT, Expand);
537 setOperationAction(ISD::SREM, (MVT::SimpleValueType)VT, Expand);
538 setOperationAction(ISD::UREM, (MVT::SimpleValueType)VT, Expand);
539 setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
540 setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
541 setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
542 setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
543 setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
544 setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
545 setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
546 setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
547 setOperationAction(ISD::FREM, (MVT::SimpleValueType)VT, Expand);
548 setOperationAction(ISD::FPOWI, (MVT::SimpleValueType)VT, Expand);
549 setOperationAction(ISD::FSQRT, (MVT::SimpleValueType)VT, Expand);
550 setOperationAction(ISD::FCOPYSIGN, (MVT::SimpleValueType)VT, Expand);
551 setOperationAction(ISD::SMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
552 setOperationAction(ISD::UMUL_LOHI, (MVT::SimpleValueType)VT, Expand);
553 setOperationAction(ISD::SDIVREM, (MVT::SimpleValueType)VT, Expand);
554 setOperationAction(ISD::UDIVREM, (MVT::SimpleValueType)VT, Expand);
555 setOperationAction(ISD::FPOW, (MVT::SimpleValueType)VT, Expand);
556 setOperationAction(ISD::CTPOP, (MVT::SimpleValueType)VT, Expand);
557 setOperationAction(ISD::CTTZ, (MVT::SimpleValueType)VT, Expand);
558 setOperationAction(ISD::CTLZ, (MVT::SimpleValueType)VT, Expand);
559 setOperationAction(ISD::SHL, (MVT::SimpleValueType)VT, Expand);
560 setOperationAction(ISD::SRA, (MVT::SimpleValueType)VT, Expand);
561 setOperationAction(ISD::SRL, (MVT::SimpleValueType)VT, Expand);
562 setOperationAction(ISD::ROTL, (MVT::SimpleValueType)VT, Expand);
563 setOperationAction(ISD::ROTR, (MVT::SimpleValueType)VT, Expand);
564 setOperationAction(ISD::BSWAP, (MVT::SimpleValueType)VT, Expand);
565 setOperationAction(ISD::VSETCC, (MVT::SimpleValueType)VT, Expand);
566 setOperationAction(ISD::FLOG, (MVT::SimpleValueType)VT, Expand);
567 setOperationAction(ISD::FLOG2, (MVT::SimpleValueType)VT, Expand);
568 setOperationAction(ISD::FLOG10, (MVT::SimpleValueType)VT, Expand);
569 setOperationAction(ISD::FEXP, (MVT::SimpleValueType)VT, Expand);
570 setOperationAction(ISD::FEXP2, (MVT::SimpleValueType)VT, Expand);
571 setOperationAction(ISD::FP_TO_UINT, (MVT::SimpleValueType)VT, Expand);
572 setOperationAction(ISD::FP_TO_SINT, (MVT::SimpleValueType)VT, Expand);
573 setOperationAction(ISD::UINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
574 setOperationAction(ISD::SINT_TO_FP, (MVT::SimpleValueType)VT, Expand);
577 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
578 // with -msoft-float, disable use of MMX as well.
579 if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
580 addRegisterClass(MVT::v8i8, X86::VR64RegisterClass);
581 addRegisterClass(MVT::v4i16, X86::VR64RegisterClass);
582 addRegisterClass(MVT::v2i32, X86::VR64RegisterClass);
583 addRegisterClass(MVT::v2f32, X86::VR64RegisterClass);
584 addRegisterClass(MVT::v1i64, X86::VR64RegisterClass);
586 setOperationAction(ISD::ADD, MVT::v8i8, Legal);
587 setOperationAction(ISD::ADD, MVT::v4i16, Legal);
588 setOperationAction(ISD::ADD, MVT::v2i32, Legal);
589 setOperationAction(ISD::ADD, MVT::v1i64, Legal);
591 setOperationAction(ISD::SUB, MVT::v8i8, Legal);
592 setOperationAction(ISD::SUB, MVT::v4i16, Legal);
593 setOperationAction(ISD::SUB, MVT::v2i32, Legal);
594 setOperationAction(ISD::SUB, MVT::v1i64, Legal);
596 setOperationAction(ISD::MULHS, MVT::v4i16, Legal);
597 setOperationAction(ISD::MUL, MVT::v4i16, Legal);
599 setOperationAction(ISD::AND, MVT::v8i8, Promote);
600 AddPromotedToType (ISD::AND, MVT::v8i8, MVT::v1i64);
601 setOperationAction(ISD::AND, MVT::v4i16, Promote);
602 AddPromotedToType (ISD::AND, MVT::v4i16, MVT::v1i64);
603 setOperationAction(ISD::AND, MVT::v2i32, Promote);
604 AddPromotedToType (ISD::AND, MVT::v2i32, MVT::v1i64);
605 setOperationAction(ISD::AND, MVT::v1i64, Legal);
607 setOperationAction(ISD::OR, MVT::v8i8, Promote);
608 AddPromotedToType (ISD::OR, MVT::v8i8, MVT::v1i64);
609 setOperationAction(ISD::OR, MVT::v4i16, Promote);
610 AddPromotedToType (ISD::OR, MVT::v4i16, MVT::v1i64);
611 setOperationAction(ISD::OR, MVT::v2i32, Promote);
612 AddPromotedToType (ISD::OR, MVT::v2i32, MVT::v1i64);
613 setOperationAction(ISD::OR, MVT::v1i64, Legal);
615 setOperationAction(ISD::XOR, MVT::v8i8, Promote);
616 AddPromotedToType (ISD::XOR, MVT::v8i8, MVT::v1i64);
617 setOperationAction(ISD::XOR, MVT::v4i16, Promote);
618 AddPromotedToType (ISD::XOR, MVT::v4i16, MVT::v1i64);
619 setOperationAction(ISD::XOR, MVT::v2i32, Promote);
620 AddPromotedToType (ISD::XOR, MVT::v2i32, MVT::v1i64);
621 setOperationAction(ISD::XOR, MVT::v1i64, Legal);
623 setOperationAction(ISD::LOAD, MVT::v8i8, Promote);
624 AddPromotedToType (ISD::LOAD, MVT::v8i8, MVT::v1i64);
625 setOperationAction(ISD::LOAD, MVT::v4i16, Promote);
626 AddPromotedToType (ISD::LOAD, MVT::v4i16, MVT::v1i64);
627 setOperationAction(ISD::LOAD, MVT::v2i32, Promote);
628 AddPromotedToType (ISD::LOAD, MVT::v2i32, MVT::v1i64);
629 setOperationAction(ISD::LOAD, MVT::v2f32, Promote);
630 AddPromotedToType (ISD::LOAD, MVT::v2f32, MVT::v1i64);
631 setOperationAction(ISD::LOAD, MVT::v1i64, Legal);
633 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i8, Custom);
634 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i16, Custom);
635 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i32, Custom);
636 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f32, Custom);
637 setOperationAction(ISD::BUILD_VECTOR, MVT::v1i64, Custom);
639 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i8, Custom);
640 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i16, Custom);
641 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i32, Custom);
642 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v1i64, Custom);
644 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v2f32, Custom);
645 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i8, Custom);
646 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v4i16, Custom);
647 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v1i64, Custom);
649 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i16, Custom);
651 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Expand);
652 setOperationAction(ISD::TRUNCATE, MVT::v8i8, Expand);
653 setOperationAction(ISD::SELECT, MVT::v8i8, Promote);
654 setOperationAction(ISD::SELECT, MVT::v4i16, Promote);
655 setOperationAction(ISD::SELECT, MVT::v2i32, Promote);
656 setOperationAction(ISD::SELECT, MVT::v1i64, Custom);
657 setOperationAction(ISD::VSETCC, MVT::v8i8, Custom);
658 setOperationAction(ISD::VSETCC, MVT::v4i16, Custom);
659 setOperationAction(ISD::VSETCC, MVT::v2i32, Custom);
662 if (!UseSoftFloat && Subtarget->hasSSE1()) {
663 addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
665 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
666 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
667 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
668 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
669 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
670 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
671 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
672 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
673 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
674 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
675 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
676 setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
679 if (!UseSoftFloat && Subtarget->hasSSE2()) {
680 addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
682 // FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
683 // registers cannot be used even for integer operations.
684 addRegisterClass(MVT::v16i8, X86::VR128RegisterClass);
685 addRegisterClass(MVT::v8i16, X86::VR128RegisterClass);
686 addRegisterClass(MVT::v4i32, X86::VR128RegisterClass);
687 addRegisterClass(MVT::v2i64, X86::VR128RegisterClass);
689 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
690 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
691 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
692 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
693 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
694 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
695 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
696 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
697 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
698 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
699 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
700 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
701 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
702 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
703 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
704 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
706 setOperationAction(ISD::VSETCC, MVT::v2f64, Custom);
707 setOperationAction(ISD::VSETCC, MVT::v16i8, Custom);
708 setOperationAction(ISD::VSETCC, MVT::v8i16, Custom);
709 setOperationAction(ISD::VSETCC, MVT::v4i32, Custom);
711 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
712 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
713 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
714 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
715 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
717 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
718 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; ++i) {
719 EVT VT = (MVT::SimpleValueType)i;
720 // Do not attempt to custom lower non-power-of-2 vectors
721 if (!isPowerOf2_32(VT.getVectorNumElements()))
723 // Do not attempt to custom lower non-128-bit vectors
724 if (!VT.is128BitVector())
726 setOperationAction(ISD::BUILD_VECTOR,
727 VT.getSimpleVT().SimpleTy, Custom);
728 setOperationAction(ISD::VECTOR_SHUFFLE,
729 VT.getSimpleVT().SimpleTy, Custom);
730 setOperationAction(ISD::EXTRACT_VECTOR_ELT,
731 VT.getSimpleVT().SimpleTy, Custom);
734 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
735 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
736 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
737 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
738 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
739 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
741 if (Subtarget->is64Bit()) {
742 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
743 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
746 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
747 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v2i64; i++) {
748 MVT::SimpleValueType SVT = (MVT::SimpleValueType)i;
751 // Do not attempt to promote non-128-bit vectors
752 if (!VT.is128BitVector()) {
755 setOperationAction(ISD::AND, SVT, Promote);
756 AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
757 setOperationAction(ISD::OR, SVT, Promote);
758 AddPromotedToType (ISD::OR, SVT, MVT::v2i64);
759 setOperationAction(ISD::XOR, SVT, Promote);
760 AddPromotedToType (ISD::XOR, SVT, MVT::v2i64);
761 setOperationAction(ISD::LOAD, SVT, Promote);
762 AddPromotedToType (ISD::LOAD, SVT, MVT::v2i64);
763 setOperationAction(ISD::SELECT, SVT, Promote);
764 AddPromotedToType (ISD::SELECT, SVT, MVT::v2i64);
767 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
769 // Custom lower v2i64 and v2f64 selects.
770 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
771 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
772 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
773 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
775 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
776 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
777 if (!DisableMMX && Subtarget->hasMMX()) {
778 setOperationAction(ISD::FP_TO_SINT, MVT::v2i32, Custom);
779 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
783 if (Subtarget->hasSSE41()) {
784 // FIXME: Do we need to handle scalar-to-vector here?
785 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
787 // i8 and i16 vectors are custom , because the source register and source
788 // source memory operand types are not the same width. f32 vectors are
789 // custom since the immediate controlling the insert encodes additional
791 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
792 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
793 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
794 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
796 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
797 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
798 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
799 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
801 if (Subtarget->is64Bit()) {
802 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Legal);
803 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Legal);
807 if (Subtarget->hasSSE42()) {
808 setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
811 if (!UseSoftFloat && Subtarget->hasAVX()) {
812 addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
813 addRegisterClass(MVT::v4f64, X86::VR256RegisterClass);
814 addRegisterClass(MVT::v8i32, X86::VR256RegisterClass);
815 addRegisterClass(MVT::v4i64, X86::VR256RegisterClass);
817 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
818 setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
819 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
820 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
821 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
822 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
823 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
824 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
825 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
826 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
827 //setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
828 //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
829 //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
830 //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
831 //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
833 // Operations to consider commented out -v16i16 v32i8
834 //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
835 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
836 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
837 //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
838 //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
839 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
840 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
841 //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
842 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
843 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
844 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
845 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
846 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
847 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
849 setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
850 // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
851 // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
852 setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
854 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
855 // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
856 // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
857 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
858 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
860 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
861 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
862 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
863 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
864 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
865 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
868 // Not sure we want to do this since there are no 256-bit integer
871 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
872 // This includes 256-bit vectors
873 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
874 EVT VT = (MVT::SimpleValueType)i;
876 // Do not attempt to custom lower non-power-of-2 vectors
877 if (!isPowerOf2_32(VT.getVectorNumElements()))
880 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
881 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
882 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
885 if (Subtarget->is64Bit()) {
886 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
887 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
892 // Not sure we want to do this since there are no 256-bit integer
895 // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
896 // Including 256-bit vectors
897 for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
898 EVT VT = (MVT::SimpleValueType)i;
900 if (!VT.is256BitVector()) {
903 setOperationAction(ISD::AND, VT, Promote);
904 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
905 setOperationAction(ISD::OR, VT, Promote);
906 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
907 setOperationAction(ISD::XOR, VT, Promote);
908 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
909 setOperationAction(ISD::LOAD, VT, Promote);
910 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
911 setOperationAction(ISD::SELECT, VT, Promote);
912 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
915 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
919 // We want to custom lower some of our intrinsics.
920 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
922 // Add/Sub/Mul with overflow operations are custom lowered.
923 setOperationAction(ISD::SADDO, MVT::i32, Custom);
924 setOperationAction(ISD::SADDO, MVT::i64, Custom);
925 setOperationAction(ISD::UADDO, MVT::i32, Custom);
926 setOperationAction(ISD::UADDO, MVT::i64, Custom);
927 setOperationAction(ISD::SSUBO, MVT::i32, Custom);
928 setOperationAction(ISD::SSUBO, MVT::i64, Custom);
929 setOperationAction(ISD::USUBO, MVT::i32, Custom);
930 setOperationAction(ISD::USUBO, MVT::i64, Custom);
931 setOperationAction(ISD::SMULO, MVT::i32, Custom);
932 setOperationAction(ISD::SMULO, MVT::i64, Custom);
934 if (!Subtarget->is64Bit()) {
935 // These libcalls are not available in 32-bit.
936 setLibcallName(RTLIB::SHL_I128, 0);
937 setLibcallName(RTLIB::SRL_I128, 0);
938 setLibcallName(RTLIB::SRA_I128, 0);
941 // We have target-specific dag combine patterns for the following nodes:
942 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
943 setTargetDAGCombine(ISD::BUILD_VECTOR);
944 setTargetDAGCombine(ISD::SELECT);
945 setTargetDAGCombine(ISD::SHL);
946 setTargetDAGCombine(ISD::SRA);
947 setTargetDAGCombine(ISD::SRL);
948 setTargetDAGCombine(ISD::STORE);
949 setTargetDAGCombine(ISD::MEMBARRIER);
950 if (Subtarget->is64Bit())
951 setTargetDAGCombine(ISD::MUL);
953 computeRegisterProperties();
955 // FIXME: These should be based on subtarget info. Plus, the values should
956 // be smaller when we are in optimizing for size mode.
957 maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
958 maxStoresPerMemcpy = 16; // For @llvm.memcpy -> sequence of stores
959 maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
960 setPrefLoopAlignment(16);
961 benefitFromCodePlacementOpt = true;
965 MVT::SimpleValueType X86TargetLowering::getSetCCResultType(EVT VT) const {
970 /// getMaxByValAlign - Helper for getByValTypeAlignment to determine
971 /// the desired ByVal argument alignment.
972 static void getMaxByValAlign(const Type *Ty, unsigned &MaxAlign) {
975 if (const VectorType *VTy = dyn_cast<VectorType>(Ty)) {
976 if (VTy->getBitWidth() == 128)
978 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
979 unsigned EltAlign = 0;
980 getMaxByValAlign(ATy->getElementType(), EltAlign);
981 if (EltAlign > MaxAlign)
983 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
984 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
985 unsigned EltAlign = 0;
986 getMaxByValAlign(STy->getElementType(i), EltAlign);
987 if (EltAlign > MaxAlign)
996 /// getByValTypeAlignment - Return the desired alignment for ByVal aggregate
997 /// function arguments in the caller parameter area. For X86, aggregates
998 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
999 /// are at 4-byte boundaries.
1000 unsigned X86TargetLowering::getByValTypeAlignment(const Type *Ty) const {
1001 if (Subtarget->is64Bit()) {
1002 // Max of 8 and alignment of type.
1003 unsigned TyAlign = TD->getABITypeAlignment(Ty);
1010 if (Subtarget->hasSSE1())
1011 getMaxByValAlign(Ty, Align);
1015 /// getOptimalMemOpType - Returns the target specific optimal type for load
1016 /// and store operations as a result of memset, memcpy, and memmove
1017 /// lowering. It returns MVT::iAny if SelectionDAG should be responsible for
1020 X86TargetLowering::getOptimalMemOpType(uint64_t Size, unsigned Align,
1021 bool isSrcConst, bool isSrcStr,
1022 SelectionDAG &DAG) const {
1023 // FIXME: This turns off use of xmm stores for memset/memcpy on targets like
1024 // linux. This is because the stack realignment code can't handle certain
1025 // cases like PR2962. This should be removed when PR2962 is fixed.
1026 const Function *F = DAG.getMachineFunction().getFunction();
1027 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
1028 if (!NoImplicitFloatOps && Subtarget->getStackAlignment() >= 16) {
1029 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE2() && Size >= 16)
1031 if ((isSrcConst || isSrcStr) && Subtarget->hasSSE1() && Size >= 16)
1034 if (Subtarget->is64Bit() && Size >= 8)
1039 /// getPICJumpTableRelocaBase - Returns relocation base for the given PIC
1041 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1042 SelectionDAG &DAG) const {
1043 if (usesGlobalOffsetTable())
1044 return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy());
1045 if (!Subtarget->is64Bit())
1046 // This doesn't have DebugLoc associated with it, but is not really the
1047 // same as a Register.
1048 return DAG.getNode(X86ISD::GlobalBaseReg, DebugLoc::getUnknownLoc(),
1053 /// getFunctionAlignment - Return the Log2 alignment of this function.
1054 unsigned X86TargetLowering::getFunctionAlignment(const Function *F) const {
1055 return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
1058 //===----------------------------------------------------------------------===//
1059 // Return Value Calling Convention Implementation
1060 //===----------------------------------------------------------------------===//
1062 #include "X86GenCallingConv.inc"
1065 X86TargetLowering::LowerReturn(SDValue Chain,
1066 CallingConv::ID CallConv, bool isVarArg,
1067 const SmallVectorImpl<ISD::OutputArg> &Outs,
1068 DebugLoc dl, SelectionDAG &DAG) {
1070 SmallVector<CCValAssign, 16> RVLocs;
1071 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1072 RVLocs, *DAG.getContext());
1073 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
1075 // If this is the first return lowered for this function, add the regs to the
1076 // liveout set for the function.
1077 if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
1078 for (unsigned i = 0; i != RVLocs.size(); ++i)
1079 if (RVLocs[i].isRegLoc())
1080 DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
1085 SmallVector<SDValue, 6> RetOps;
1086 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
1087 // Operand #1 = Bytes To Pop
1088 RetOps.push_back(DAG.getConstant(getBytesToPopOnReturn(), MVT::i16));
1090 // Copy the result values into the output registers.
1091 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1092 CCValAssign &VA = RVLocs[i];
1093 assert(VA.isRegLoc() && "Can only return in registers!");
1094 SDValue ValToCopy = Outs[i].Val;
1096 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
1097 // the RET instruction and handled by the FP Stackifier.
1098 if (VA.getLocReg() == X86::ST0 ||
1099 VA.getLocReg() == X86::ST1) {
1100 // If this is a copy from an xmm register to ST(0), use an FPExtend to
1101 // change the value to the FP stack register class.
1102 if (isScalarFPTypeInSSEReg(VA.getValVT()))
1103 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
1104 RetOps.push_back(ValToCopy);
1105 // Don't emit a copytoreg.
1109 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
1110 // which is returned in RAX / RDX.
1111 if (Subtarget->is64Bit()) {
1112 EVT ValVT = ValToCopy.getValueType();
1113 if (ValVT.isVector() && ValVT.getSizeInBits() == 64) {
1114 ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
1115 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1)
1116 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, ValToCopy);
1120 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
1121 Flag = Chain.getValue(1);
1124 // The x86-64 ABI for returning structs by value requires that we copy
1125 // the sret argument into %rax for the return. We saved the argument into
1126 // a virtual register in the entry block, so now we copy the value out
1128 if (Subtarget->is64Bit() &&
1129 DAG.getMachineFunction().getFunction()->hasStructRetAttr()) {
1130 MachineFunction &MF = DAG.getMachineFunction();
1131 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1132 unsigned Reg = FuncInfo->getSRetReturnReg();
1134 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1135 FuncInfo->setSRetReturnReg(Reg);
1137 SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
1139 Chain = DAG.getCopyToReg(Chain, dl, X86::RAX, Val, Flag);
1140 Flag = Chain.getValue(1);
1143 RetOps[0] = Chain; // Update chain.
1145 // Add the flag if we have it.
1147 RetOps.push_back(Flag);
1149 return DAG.getNode(X86ISD::RET_FLAG, dl,
1150 MVT::Other, &RetOps[0], RetOps.size());
1153 /// LowerCallResult - Lower the result values of a call into the
1154 /// appropriate copies out of appropriate physical registers.
1157 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
1158 CallingConv::ID CallConv, bool isVarArg,
1159 const SmallVectorImpl<ISD::InputArg> &Ins,
1160 DebugLoc dl, SelectionDAG &DAG,
1161 SmallVectorImpl<SDValue> &InVals) {
1163 // Assign locations to each value returned by this call.
1164 SmallVector<CCValAssign, 16> RVLocs;
1165 bool Is64Bit = Subtarget->is64Bit();
1166 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1167 RVLocs, *DAG.getContext());
1168 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1170 // Copy all of the result registers out of their specified physreg.
1171 for (unsigned i = 0; i != RVLocs.size(); ++i) {
1172 CCValAssign &VA = RVLocs[i];
1173 EVT CopyVT = VA.getValVT();
1175 // If this is x86-64, and we disabled SSE, we can't return FP values
1176 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
1177 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
1178 llvm_report_error("SSE register return with SSE disabled");
1181 // If this is a call to a function that returns an fp value on the floating
1182 // point stack, but where we prefer to use the value in xmm registers, copy
1183 // it out as F80 and use a truncate to move it from fp stack reg to xmm reg.
1184 if ((VA.getLocReg() == X86::ST0 ||
1185 VA.getLocReg() == X86::ST1) &&
1186 isScalarFPTypeInSSEReg(VA.getValVT())) {
1191 if (Is64Bit && CopyVT.isVector() && CopyVT.getSizeInBits() == 64) {
1192 // For x86-64, MMX values are returned in XMM0 / XMM1 except for v1i64.
1193 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
1194 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1195 MVT::v2i64, InFlag).getValue(1);
1196 Val = Chain.getValue(0);
1197 Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1198 Val, DAG.getConstant(0, MVT::i64));
1200 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1201 MVT::i64, InFlag).getValue(1);
1202 Val = Chain.getValue(0);
1204 Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
1206 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
1207 CopyVT, InFlag).getValue(1);
1208 Val = Chain.getValue(0);
1210 InFlag = Chain.getValue(2);
1212 if (CopyVT != VA.getValVT()) {
1213 // Round the F80 the right size, which also moves to the appropriate xmm
1215 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
1216 // This truncation won't change the value.
1217 DAG.getIntPtrConstant(1));
1220 InVals.push_back(Val);
1227 //===----------------------------------------------------------------------===//
1228 // C & StdCall & Fast Calling Convention implementation
1229 //===----------------------------------------------------------------------===//
1230 // StdCall calling convention seems to be standard for many Windows' API
1231 // routines and around. It differs from C calling convention just a little:
1232 // callee should clean up the stack, not caller. Symbols should be also
1233 // decorated in some fancy way :) It doesn't support any vector arguments.
1234 // For info on fast calling convention see Fast Calling Convention (tail call)
1235 // implementation LowerX86_32FastCCCallTo.
1237 /// CallIsStructReturn - Determines whether a call uses struct return
1239 static bool CallIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
1243 return Outs[0].Flags.isSRet();
1246 /// ArgsAreStructReturn - Determines whether a function uses struct
1247 /// return semantics.
1249 ArgsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
1253 return Ins[0].Flags.isSRet();
1256 /// IsCalleePop - Determines whether the callee is required to pop its
1257 /// own arguments. Callee pop is necessary to support tail calls.
1258 bool X86TargetLowering::IsCalleePop(bool IsVarArg, CallingConv::ID CallingConv){
1262 switch (CallingConv) {
1265 case CallingConv::X86_StdCall:
1266 return !Subtarget->is64Bit();
1267 case CallingConv::X86_FastCall:
1268 return !Subtarget->is64Bit();
1269 case CallingConv::Fast:
1270 return PerformTailCallOpt;
1274 /// CCAssignFnForNode - Selects the correct CCAssignFn for a the
1275 /// given CallingConvention value.
1276 CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
1277 if (Subtarget->is64Bit()) {
1278 if (Subtarget->isTargetWin64())
1279 return CC_X86_Win64_C;
1284 if (CC == CallingConv::X86_FastCall)
1285 return CC_X86_32_FastCall;
1286 else if (CC == CallingConv::Fast)
1287 return CC_X86_32_FastCC;
1292 /// NameDecorationForCallConv - Selects the appropriate decoration to
1293 /// apply to a MachineFunction containing a given calling convention.
1295 X86TargetLowering::NameDecorationForCallConv(CallingConv::ID CallConv) {
1296 if (CallConv == CallingConv::X86_FastCall)
1298 else if (CallConv == CallingConv::X86_StdCall)
1304 /// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
1305 /// by "Src" to address "Dst" with size and alignment information specified by
1306 /// the specific parameter attribute. The copy will be passed as a byval
1307 /// function parameter.
1309 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
1310 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
1312 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
1313 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
1314 /*AlwaysInline=*/true, NULL, 0, NULL, 0);
1318 X86TargetLowering::LowerMemArgument(SDValue Chain,
1319 CallingConv::ID CallConv,
1320 const SmallVectorImpl<ISD::InputArg> &Ins,
1321 DebugLoc dl, SelectionDAG &DAG,
1322 const CCValAssign &VA,
1323 MachineFrameInfo *MFI,
1326 // Create the nodes corresponding to a load from this parameter slot.
1327 ISD::ArgFlagsTy Flags = Ins[i].Flags;
1328 bool AlwaysUseMutable = (CallConv==CallingConv::Fast) && PerformTailCallOpt;
1329 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
1332 // If value is passed by pointer we have address passed instead of the value
1334 if (VA.getLocInfo() == CCValAssign::Indirect)
1335 ValVT = VA.getLocVT();
1337 ValVT = VA.getValVT();
1339 // FIXME: For now, all byval parameter objects are marked mutable. This can be
1340 // changed with more analysis.
1341 // In case of tail call optimization mark all arguments mutable. Since they
1342 // could be overwritten by lowering of arguments in case of a tail call.
1343 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
1344 VA.getLocMemOffset(), isImmutable);
1345 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
1346 if (Flags.isByVal())
1348 return DAG.getLoad(ValVT, dl, Chain, FIN,
1349 PseudoSourceValue::getFixedStack(FI), 0);
1353 X86TargetLowering::LowerFormalArguments(SDValue Chain,
1354 CallingConv::ID CallConv,
1356 const SmallVectorImpl<ISD::InputArg> &Ins,
1359 SmallVectorImpl<SDValue> &InVals) {
1361 MachineFunction &MF = DAG.getMachineFunction();
1362 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1364 const Function* Fn = MF.getFunction();
1365 if (Fn->hasExternalLinkage() &&
1366 Subtarget->isTargetCygMing() &&
1367 Fn->getName() == "main")
1368 FuncInfo->setForceFramePointer(true);
1370 // Decorate the function name.
1371 FuncInfo->setDecorationStyle(NameDecorationForCallConv(CallConv));
1373 MachineFrameInfo *MFI = MF.getFrameInfo();
1374 bool Is64Bit = Subtarget->is64Bit();
1375 bool IsWin64 = Subtarget->isTargetWin64();
1377 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1378 "Var args not supported with calling convention fastcc");
1380 // Assign locations to all of the incoming arguments.
1381 SmallVector<CCValAssign, 16> ArgLocs;
1382 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1383 ArgLocs, *DAG.getContext());
1384 CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
1386 unsigned LastVal = ~0U;
1388 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1389 CCValAssign &VA = ArgLocs[i];
1390 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
1392 assert(VA.getValNo() != LastVal &&
1393 "Don't support value assigned to multiple locs yet");
1394 LastVal = VA.getValNo();
1396 if (VA.isRegLoc()) {
1397 EVT RegVT = VA.getLocVT();
1398 TargetRegisterClass *RC = NULL;
1399 if (RegVT == MVT::i32)
1400 RC = X86::GR32RegisterClass;
1401 else if (Is64Bit && RegVT == MVT::i64)
1402 RC = X86::GR64RegisterClass;
1403 else if (RegVT == MVT::f32)
1404 RC = X86::FR32RegisterClass;
1405 else if (RegVT == MVT::f64)
1406 RC = X86::FR64RegisterClass;
1407 else if (RegVT.isVector() && RegVT.getSizeInBits() == 128)
1408 RC = X86::VR128RegisterClass;
1409 else if (RegVT.isVector() && RegVT.getSizeInBits() == 64)
1410 RC = X86::VR64RegisterClass;
1412 llvm_unreachable("Unknown argument type!");
1414 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
1415 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
1417 // If this is an 8 or 16-bit value, it is really passed promoted to 32
1418 // bits. Insert an assert[sz]ext to capture this, then truncate to the
1420 if (VA.getLocInfo() == CCValAssign::SExt)
1421 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
1422 DAG.getValueType(VA.getValVT()));
1423 else if (VA.getLocInfo() == CCValAssign::ZExt)
1424 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
1425 DAG.getValueType(VA.getValVT()));
1426 else if (VA.getLocInfo() == CCValAssign::BCvt)
1427 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1429 if (VA.isExtInLoc()) {
1430 // Handle MMX values passed in XMM regs.
1431 if (RegVT.isVector()) {
1432 ArgValue = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64,
1433 ArgValue, DAG.getConstant(0, MVT::i64));
1434 ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
1436 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
1439 assert(VA.isMemLoc());
1440 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
1443 // If value is passed via pointer - do a load.
1444 if (VA.getLocInfo() == CCValAssign::Indirect)
1445 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue, NULL, 0);
1447 InVals.push_back(ArgValue);
1450 // The x86-64 ABI for returning structs by value requires that we copy
1451 // the sret argument into %rax for the return. Save the argument into
1452 // a virtual register so that we can access it from the return points.
1453 if (Is64Bit && MF.getFunction()->hasStructRetAttr()) {
1454 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
1455 unsigned Reg = FuncInfo->getSRetReturnReg();
1457 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(MVT::i64));
1458 FuncInfo->setSRetReturnReg(Reg);
1460 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
1461 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
1464 unsigned StackSize = CCInfo.getNextStackOffset();
1465 // align stack specially for tail calls
1466 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1467 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
1469 // If the function takes variable number of arguments, make a frame index for
1470 // the start of the first vararg value... for expansion of llvm.va_start.
1472 if (Is64Bit || CallConv != CallingConv::X86_FastCall) {
1473 VarArgsFrameIndex = MFI->CreateFixedObject(1, StackSize);
1476 unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
1478 // FIXME: We should really autogenerate these arrays
1479 static const unsigned GPR64ArgRegsWin64[] = {
1480 X86::RCX, X86::RDX, X86::R8, X86::R9
1482 static const unsigned XMMArgRegsWin64[] = {
1483 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3
1485 static const unsigned GPR64ArgRegs64Bit[] = {
1486 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
1488 static const unsigned XMMArgRegs64Bit[] = {
1489 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1490 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1492 const unsigned *GPR64ArgRegs, *XMMArgRegs;
1495 TotalNumIntRegs = 4; TotalNumXMMRegs = 4;
1496 GPR64ArgRegs = GPR64ArgRegsWin64;
1497 XMMArgRegs = XMMArgRegsWin64;
1499 TotalNumIntRegs = 6; TotalNumXMMRegs = 8;
1500 GPR64ArgRegs = GPR64ArgRegs64Bit;
1501 XMMArgRegs = XMMArgRegs64Bit;
1503 unsigned NumIntRegs = CCInfo.getFirstUnallocated(GPR64ArgRegs,
1505 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs,
1508 bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
1509 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
1510 "SSE register cannot be used when SSE is disabled!");
1511 assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
1512 "SSE register cannot be used when SSE is disabled!");
1513 if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
1514 // Kernel mode asks for SSE to be disabled, so don't push them
1516 TotalNumXMMRegs = 0;
1518 // For X86-64, if there are vararg parameters that are passed via
1519 // registers, then we must store them to their spots on the stack so they
1520 // may be loaded by deferencing the result of va_next.
1521 VarArgsGPOffset = NumIntRegs * 8;
1522 VarArgsFPOffset = TotalNumIntRegs * 8 + NumXMMRegs * 16;
1523 RegSaveFrameIndex = MFI->CreateStackObject(TotalNumIntRegs * 8 +
1524 TotalNumXMMRegs * 16, 16);
1526 // Store the integer parameter registers.
1527 SmallVector<SDValue, 8> MemOps;
1528 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
1529 unsigned Offset = VarArgsGPOffset;
1530 for (; NumIntRegs != TotalNumIntRegs; ++NumIntRegs) {
1531 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), RSFIN,
1532 DAG.getIntPtrConstant(Offset));
1533 unsigned VReg = MF.addLiveIn(GPR64ArgRegs[NumIntRegs],
1534 X86::GR64RegisterClass);
1535 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i64);
1537 DAG.getStore(Val.getValue(1), dl, Val, FIN,
1538 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
1540 MemOps.push_back(Store);
1544 if (TotalNumXMMRegs != 0 && NumXMMRegs != TotalNumXMMRegs) {
1545 // Now store the XMM (fp + vector) parameter registers.
1546 SmallVector<SDValue, 11> SaveXMMOps;
1547 SaveXMMOps.push_back(Chain);
1549 unsigned AL = MF.addLiveIn(X86::AL, X86::GR8RegisterClass);
1550 SDValue ALVal = DAG.getCopyFromReg(DAG.getEntryNode(), dl, AL, MVT::i8);
1551 SaveXMMOps.push_back(ALVal);
1553 SaveXMMOps.push_back(DAG.getIntPtrConstant(RegSaveFrameIndex));
1554 SaveXMMOps.push_back(DAG.getIntPtrConstant(VarArgsFPOffset));
1556 for (; NumXMMRegs != TotalNumXMMRegs; ++NumXMMRegs) {
1557 unsigned VReg = MF.addLiveIn(XMMArgRegs[NumXMMRegs],
1558 X86::VR128RegisterClass);
1559 SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::v4f32);
1560 SaveXMMOps.push_back(Val);
1562 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
1564 &SaveXMMOps[0], SaveXMMOps.size()));
1567 if (!MemOps.empty())
1568 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1569 &MemOps[0], MemOps.size());
1573 // Some CCs need callee pop.
1574 if (IsCalleePop(isVarArg, CallConv)) {
1575 BytesToPopOnReturn = StackSize; // Callee pops everything.
1576 BytesCallerReserves = 0;
1578 BytesToPopOnReturn = 0; // Callee pops nothing.
1579 // If this is an sret function, the return should pop the hidden pointer.
1580 if (!Is64Bit && CallConv != CallingConv::Fast && ArgsAreStructReturn(Ins))
1581 BytesToPopOnReturn = 4;
1582 BytesCallerReserves = StackSize;
1586 RegSaveFrameIndex = 0xAAAAAAA; // RegSaveFrameIndex is X86-64 only.
1587 if (CallConv == CallingConv::X86_FastCall)
1588 VarArgsFrameIndex = 0xAAAAAAA; // fastcc functions can't have varargs.
1591 FuncInfo->setBytesToPopOnReturn(BytesToPopOnReturn);
1597 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
1598 SDValue StackPtr, SDValue Arg,
1599 DebugLoc dl, SelectionDAG &DAG,
1600 const CCValAssign &VA,
1601 ISD::ArgFlagsTy Flags) {
1602 const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
1603 unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
1604 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
1605 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
1606 if (Flags.isByVal()) {
1607 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
1609 return DAG.getStore(Chain, dl, Arg, PtrOff,
1610 PseudoSourceValue::getStack(), LocMemOffset);
1613 /// EmitTailCallLoadRetAddr - Emit a load of return address if tail call
1614 /// optimization is performed and it is required.
1616 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
1617 SDValue &OutRetAddr,
1623 if (!IsTailCall || FPDiff==0) return Chain;
1625 // Adjust the Return address stack slot.
1626 EVT VT = getPointerTy();
1627 OutRetAddr = getReturnAddressFrameIndex(DAG);
1629 // Load the "old" Return address.
1630 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, NULL, 0);
1631 return SDValue(OutRetAddr.getNode(), 1);
1634 /// EmitTailCallStoreRetAddr - Emit a store of the return adress if tail call
1635 /// optimization is performed and it is required (FPDiff!=0).
1637 EmitTailCallStoreRetAddr(SelectionDAG & DAG, MachineFunction &MF,
1638 SDValue Chain, SDValue RetAddrFrIdx,
1639 bool Is64Bit, int FPDiff, DebugLoc dl) {
1640 // Store the return address to the appropriate stack slot.
1641 if (!FPDiff) return Chain;
1642 // Calculate the new stack slot for the return address.
1643 int SlotSize = Is64Bit ? 8 : 4;
1644 int NewReturnAddrFI =
1645 MF.getFrameInfo()->CreateFixedObject(SlotSize, FPDiff-SlotSize);
1646 EVT VT = Is64Bit ? MVT::i64 : MVT::i32;
1647 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, VT);
1648 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
1649 PseudoSourceValue::getFixedStack(NewReturnAddrFI), 0);
1654 X86TargetLowering::LowerCall(SDValue Chain, SDValue Callee,
1655 CallingConv::ID CallConv, bool isVarArg,
1657 const SmallVectorImpl<ISD::OutputArg> &Outs,
1658 const SmallVectorImpl<ISD::InputArg> &Ins,
1659 DebugLoc dl, SelectionDAG &DAG,
1660 SmallVectorImpl<SDValue> &InVals) {
1662 MachineFunction &MF = DAG.getMachineFunction();
1663 bool Is64Bit = Subtarget->is64Bit();
1664 bool IsStructRet = CallIsStructReturn(Outs);
1666 assert((!isTailCall ||
1667 (CallConv == CallingConv::Fast && PerformTailCallOpt)) &&
1668 "IsEligibleForTailCallOptimization missed a case!");
1669 assert(!(isVarArg && CallConv == CallingConv::Fast) &&
1670 "Var args not supported with calling convention fastcc");
1672 // Analyze operands of the call, assigning locations to each operand.
1673 SmallVector<CCValAssign, 16> ArgLocs;
1674 CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
1675 ArgLocs, *DAG.getContext());
1676 CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
1678 // Get a count of how many bytes are to be pushed on the stack.
1679 unsigned NumBytes = CCInfo.getNextStackOffset();
1680 if (PerformTailCallOpt && CallConv == CallingConv::Fast)
1681 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
1685 // Lower arguments at fp - stackoffset + fpdiff.
1686 unsigned NumBytesCallerPushed =
1687 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn();
1688 FPDiff = NumBytesCallerPushed - NumBytes;
1690 // Set the delta of movement of the returnaddr stackslot.
1691 // But only set if delta is greater than previous delta.
1692 if (FPDiff < (MF.getInfo<X86MachineFunctionInfo>()->getTCReturnAddrDelta()))
1693 MF.getInfo<X86MachineFunctionInfo>()->setTCReturnAddrDelta(FPDiff);
1696 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));
1698 SDValue RetAddrFrIdx;
1699 // Load return adress for tail calls.
1700 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall, Is64Bit,
1703 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
1704 SmallVector<SDValue, 8> MemOpChains;
1707 // Walk the register/memloc assignments, inserting copies/loads. In the case
1708 // of tail call optimization arguments are handle later.
1709 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1710 CCValAssign &VA = ArgLocs[i];
1711 EVT RegVT = VA.getLocVT();
1712 SDValue Arg = Outs[i].Val;
1713 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1714 bool isByVal = Flags.isByVal();
1716 // Promote the value if needed.
1717 switch (VA.getLocInfo()) {
1718 default: llvm_unreachable("Unknown loc info!");
1719 case CCValAssign::Full: break;
1720 case CCValAssign::SExt:
1721 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
1723 case CCValAssign::ZExt:
1724 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
1726 case CCValAssign::AExt:
1727 if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
1728 // Special case: passing MMX values in XMM registers.
1729 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
1730 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
1731 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
1733 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
1735 case CCValAssign::BCvt:
1736 Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
1738 case CCValAssign::Indirect: {
1739 // Store the argument.
1740 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
1741 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
1742 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
1743 PseudoSourceValue::getFixedStack(FI), 0);
1749 if (VA.isRegLoc()) {
1750 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
1752 if (!isTailCall || (isTailCall && isByVal)) {
1753 assert(VA.isMemLoc());
1754 if (StackPtr.getNode() == 0)
1755 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr, getPointerTy());
1757 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
1758 dl, DAG, VA, Flags));
1763 if (!MemOpChains.empty())
1764 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1765 &MemOpChains[0], MemOpChains.size());
1767 // Build a sequence of copy-to-reg nodes chained together with token chain
1768 // and flag operands which copy the outgoing args into registers.
1770 // Tail call byval lowering might overwrite argument registers so in case of
1771 // tail call optimization the copies to registers are lowered later.
1773 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1774 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1775 RegsToPass[i].second, InFlag);
1776 InFlag = Chain.getValue(1);
1780 if (Subtarget->isPICStyleGOT()) {
1781 // ELF / PIC requires GOT in the EBX register before function calls via PLT
1784 Chain = DAG.getCopyToReg(Chain, dl, X86::EBX,
1785 DAG.getNode(X86ISD::GlobalBaseReg,
1786 DebugLoc::getUnknownLoc(),
1789 InFlag = Chain.getValue(1);
1791 // If we are tail calling and generating PIC/GOT style code load the
1792 // address of the callee into ECX. The value in ecx is used as target of
1793 // the tail jump. This is done to circumvent the ebx/callee-saved problem
1794 // for tail calls on PIC/GOT architectures. Normally we would just put the
1795 // address of GOT into ebx and then call target@PLT. But for tail calls
1796 // ebx would be restored (since ebx is callee saved) before jumping to the
1799 // Note: The actual moving to ECX is done further down.
1800 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
1801 if (G && !G->getGlobal()->hasHiddenVisibility() &&
1802 !G->getGlobal()->hasProtectedVisibility())
1803 Callee = LowerGlobalAddress(Callee, DAG);
1804 else if (isa<ExternalSymbolSDNode>(Callee))
1805 Callee = LowerExternalSymbol(Callee, DAG);
1809 if (Is64Bit && isVarArg) {
1810 // From AMD64 ABI document:
1811 // For calls that may call functions that use varargs or stdargs
1812 // (prototype-less calls or calls to functions containing ellipsis (...) in
1813 // the declaration) %al is used as hidden argument to specify the number
1814 // of SSE registers used. The contents of %al do not need to match exactly
1815 // the number of registers, but must be an ubound on the number of SSE
1816 // registers used and is in the range 0 - 8 inclusive.
1818 // FIXME: Verify this on Win64
1819 // Count the number of XMM registers allocated.
1820 static const unsigned XMMArgRegs[] = {
1821 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
1822 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
1824 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
1825 assert((Subtarget->hasSSE1() || !NumXMMRegs)
1826 && "SSE registers cannot be used when SSE is disabled");
1828 Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
1829 DAG.getConstant(NumXMMRegs, MVT::i8), InFlag);
1830 InFlag = Chain.getValue(1);
1834 // For tail calls lower the arguments to the 'real' stack slot.
1836 // Force all the incoming stack arguments to be loaded from the stack
1837 // before any new outgoing arguments are stored to the stack, because the
1838 // outgoing stack slots may alias the incoming argument stack slots, and
1839 // the alias isn't otherwise explicit. This is slightly more conservative
1840 // than necessary, because it means that each store effectively depends
1841 // on every argument instead of just those arguments it would clobber.
1842 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
1844 SmallVector<SDValue, 8> MemOpChains2;
1847 // Do not flag preceeding copytoreg stuff together with the following stuff.
1849 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
1850 CCValAssign &VA = ArgLocs[i];
1851 if (!VA.isRegLoc()) {
1852 assert(VA.isMemLoc());
1853 SDValue Arg = Outs[i].Val;
1854 ISD::ArgFlagsTy Flags = Outs[i].Flags;
1855 // Create frame index.
1856 int32_t Offset = VA.getLocMemOffset()+FPDiff;
1857 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
1858 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset);
1859 FIN = DAG.getFrameIndex(FI, getPointerTy());
1861 if (Flags.isByVal()) {
1862 // Copy relative to framepointer.
1863 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
1864 if (StackPtr.getNode() == 0)
1865 StackPtr = DAG.getCopyFromReg(Chain, dl, X86StackPtr,
1867 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, Source);
1869 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
1873 // Store relative to framepointer.
1874 MemOpChains2.push_back(
1875 DAG.getStore(ArgChain, dl, Arg, FIN,
1876 PseudoSourceValue::getFixedStack(FI), 0));
1881 if (!MemOpChains2.empty())
1882 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
1883 &MemOpChains2[0], MemOpChains2.size());
1885 // Copy arguments to their registers.
1886 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
1887 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
1888 RegsToPass[i].second, InFlag);
1889 InFlag = Chain.getValue(1);
1893 // Store the return address to the appropriate stack slot.
1894 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx, Is64Bit,
1898 // If the callee is a GlobalAddress node (quite common, every direct call is)
1899 // turn it into a TargetGlobalAddress node so that legalize doesn't hack it.
1900 if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
1901 // We should use extra load for direct calls to dllimported functions in
1903 GlobalValue *GV = G->getGlobal();
1904 if (!GV->hasDLLImportLinkage()) {
1905 unsigned char OpFlags = 0;
1907 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
1908 // external symbols most go through the PLT in PIC mode. If the symbol
1909 // has hidden or protected visibility, or if it is static or local, then
1910 // we don't need to use the PLT - we can directly call it.
1911 if (Subtarget->isTargetELF() &&
1912 getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1913 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
1914 OpFlags = X86II::MO_PLT;
1915 } else if (Subtarget->isPICStyleStubAny() &&
1916 (GV->isDeclaration() || GV->isWeakForLinker()) &&
1917 Subtarget->getDarwinVers() < 9) {
1918 // PC-relative references to external symbols should go through $stub,
1919 // unless we're building with the leopard linker or later, which
1920 // automatically synthesizes these stubs.
1921 OpFlags = X86II::MO_DARWIN_STUB;
1924 Callee = DAG.getTargetGlobalAddress(GV, getPointerTy(),
1925 G->getOffset(), OpFlags);
1927 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
1928 unsigned char OpFlags = 0;
1930 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
1931 // symbols should go through the PLT.
1932 if (Subtarget->isTargetELF() &&
1933 getTargetMachine().getRelocationModel() == Reloc::PIC_) {
1934 OpFlags = X86II::MO_PLT;
1935 } else if (Subtarget->isPICStyleStubAny() &&
1936 Subtarget->getDarwinVers() < 9) {
1937 // PC-relative references to external symbols should go through $stub,
1938 // unless we're building with the leopard linker or later, which
1939 // automatically synthesizes these stubs.
1940 OpFlags = X86II::MO_DARWIN_STUB;
1943 Callee = DAG.getTargetExternalSymbol(S->getSymbol(), getPointerTy(),
1945 } else if (isTailCall) {
1946 unsigned Opc = Is64Bit ? X86::R11 : X86::EAX;
1948 Chain = DAG.getCopyToReg(Chain, dl,
1949 DAG.getRegister(Opc, getPointerTy()),
1951 Callee = DAG.getRegister(Opc, getPointerTy());
1952 // Add register as live out.
1953 MF.getRegInfo().addLiveOut(Opc);
1956 // Returns a chain & a flag for retval copy to use.
1957 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
1958 SmallVector<SDValue, 8> Ops;
1961 Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
1962 DAG.getIntPtrConstant(0, true), InFlag);
1963 InFlag = Chain.getValue(1);
1966 Ops.push_back(Chain);
1967 Ops.push_back(Callee);
1970 Ops.push_back(DAG.getConstant(FPDiff, MVT::i32));
1972 // Add argument registers to the end of the list so that they are known live
1974 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
1975 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
1976 RegsToPass[i].second.getValueType()));
1978 // Add an implicit use GOT pointer in EBX.
1979 if (!isTailCall && Subtarget->isPICStyleGOT())
1980 Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
1982 // Add an implicit use of AL for x86 vararg functions.
1983 if (Is64Bit && isVarArg)
1984 Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
1986 if (InFlag.getNode())
1987 Ops.push_back(InFlag);
1990 // If this is the first return lowered for this function, add the regs
1991 // to the liveout set for the function.
1992 if (MF.getRegInfo().liveout_empty()) {
1993 SmallVector<CCValAssign, 16> RVLocs;
1994 CCState CCInfo(CallConv, isVarArg, getTargetMachine(), RVLocs,
1996 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
1997 for (unsigned i = 0; i != RVLocs.size(); ++i)
1998 if (RVLocs[i].isRegLoc())
1999 MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
2002 assert(((Callee.getOpcode() == ISD::Register &&
2003 (cast<RegisterSDNode>(Callee)->getReg() == X86::EAX ||
2004 cast<RegisterSDNode>(Callee)->getReg() == X86::R9)) ||
2005 Callee.getOpcode() == ISD::TargetExternalSymbol ||
2006 Callee.getOpcode() == ISD::TargetGlobalAddress) &&
2007 "Expecting an global address, external symbol, or register");
2009 return DAG.getNode(X86ISD::TC_RETURN, dl,
2010 NodeTys, &Ops[0], Ops.size());
2013 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, &Ops[0], Ops.size());
2014 InFlag = Chain.getValue(1);
2016 // Create the CALLSEQ_END node.
2017 unsigned NumBytesForCalleeToPush;
2018 if (IsCalleePop(isVarArg, CallConv))
2019 NumBytesForCalleeToPush = NumBytes; // Callee pops everything
2020 else if (!Is64Bit && CallConv != CallingConv::Fast && IsStructRet)
2021 // If this is is a call to a struct-return function, the callee
2022 // pops the hidden struct pointer, so we have to push it back.
2023 // This is common for Darwin/X86, Linux & Mingw32 targets.
2024 NumBytesForCalleeToPush = 4;
2026 NumBytesForCalleeToPush = 0; // Callee pops nothing.
2028 // Returns a flag for retval copy to use.
2029 Chain = DAG.getCALLSEQ_END(Chain,
2030 DAG.getIntPtrConstant(NumBytes, true),
2031 DAG.getIntPtrConstant(NumBytesForCalleeToPush,
2034 InFlag = Chain.getValue(1);
2036 // Handle result values, copying them out of physregs into vregs that we
2038 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
2039 Ins, dl, DAG, InVals);
2043 //===----------------------------------------------------------------------===//
2044 // Fast Calling Convention (tail call) implementation
2045 //===----------------------------------------------------------------------===//
2047 // Like std call, callee cleans arguments, convention except that ECX is
2048 // reserved for storing the tail called function address. Only 2 registers are
2049 // free for argument passing (inreg). Tail call optimization is performed
2051 // * tailcallopt is enabled
2052 // * caller/callee are fastcc
2053 // On X86_64 architecture with GOT-style position independent code only local
2054 // (within module) calls are supported at the moment.
2055 // To keep the stack aligned according to platform abi the function
2056 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
2057 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
2058 // If a tail called function callee has more arguments than the caller the
2059 // caller needs to make sure that there is room to move the RETADDR to. This is
2060 // achieved by reserving an area the size of the argument delta right after the
2061 // original REtADDR, but before the saved framepointer or the spilled registers
2062 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
2074 /// GetAlignedArgumentStackSize - Make the stack size align e.g 16n + 12 aligned
2075 /// for a 16 byte align requirement.
2076 unsigned X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
2077 SelectionDAG& DAG) {
2078 MachineFunction &MF = DAG.getMachineFunction();
2079 const TargetMachine &TM = MF.getTarget();
2080 const TargetFrameInfo &TFI = *TM.getFrameInfo();
2081 unsigned StackAlignment = TFI.getStackAlignment();
2082 uint64_t AlignMask = StackAlignment - 1;
2083 int64_t Offset = StackSize;
2084 uint64_t SlotSize = TD->getPointerSize();
2085 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
2086 // Number smaller than 12 so just add the difference.
2087 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
2089 // Mask out lower bits, add stackalignment once plus the 12 bytes.
2090 Offset = ((~AlignMask) & Offset) + StackAlignment +
2091 (StackAlignment-SlotSize);
2096 /// IsEligibleForTailCallOptimization - Check whether the call is eligible
2097 /// for tail call optimization. Targets which want to do tail call
2098 /// optimization should implement this function.
2100 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
2101 CallingConv::ID CalleeCC,
2103 const SmallVectorImpl<ISD::InputArg> &Ins,
2104 SelectionDAG& DAG) const {
2105 MachineFunction &MF = DAG.getMachineFunction();
2106 CallingConv::ID CallerCC = MF.getFunction()->getCallingConv();
2107 return CalleeCC == CallingConv::Fast && CallerCC == CalleeCC;
2111 X86TargetLowering::createFastISel(MachineFunction &mf,
2112 MachineModuleInfo *mmo,
2114 DenseMap<const Value *, unsigned> &vm,
2115 DenseMap<const BasicBlock *,
2116 MachineBasicBlock *> &bm,
2117 DenseMap<const AllocaInst *, int> &am
2119 , SmallSet<Instruction*, 8> &cil
2122 return X86::createFastISel(mf, mmo, dw, vm, bm, am
2130 //===----------------------------------------------------------------------===//
2131 // Other Lowering Hooks
2132 //===----------------------------------------------------------------------===//
2135 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) {
2136 MachineFunction &MF = DAG.getMachineFunction();
2137 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2138 int ReturnAddrIndex = FuncInfo->getRAIndex();
2140 if (ReturnAddrIndex == 0) {
2141 // Set up a frame object for the return address.
2142 uint64_t SlotSize = TD->getPointerSize();
2143 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize, -SlotSize);
2144 FuncInfo->setRAIndex(ReturnAddrIndex);
2147 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
2151 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
2152 bool hasSymbolicDisplacement) {
2153 // Offset should fit into 32 bit immediate field.
2154 if (!isInt32(Offset))
2157 // If we don't have a symbolic displacement - we don't have any extra
2159 if (!hasSymbolicDisplacement)
2162 // FIXME: Some tweaks might be needed for medium code model.
2163 if (M != CodeModel::Small && M != CodeModel::Kernel)
2166 // For small code model we assume that latest object is 16MB before end of 31
2167 // bits boundary. We may also accept pretty large negative constants knowing
2168 // that all objects are in the positive half of address space.
2169 if (M == CodeModel::Small && Offset < 16*1024*1024)
2172 // For kernel code model we know that all object resist in the negative half
2173 // of 32bits address space. We may not accept negative offsets, since they may
2174 // be just off and we may accept pretty large positive ones.
2175 if (M == CodeModel::Kernel && Offset > 0)
2181 /// TranslateX86CC - do a one to one translation of a ISD::CondCode to the X86
2182 /// specific condition code, returning the condition code and the LHS/RHS of the
2183 /// comparison to make.
2184 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, bool isFP,
2185 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
2187 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
2188 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
2189 // X > -1 -> X == 0, jump !sign.
2190 RHS = DAG.getConstant(0, RHS.getValueType());
2191 return X86::COND_NS;
2192 } else if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
2193 // X < 0 -> X == 0, jump on sign.
2195 } else if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
2197 RHS = DAG.getConstant(0, RHS.getValueType());
2198 return X86::COND_LE;
2202 switch (SetCCOpcode) {
2203 default: llvm_unreachable("Invalid integer condition!");
2204 case ISD::SETEQ: return X86::COND_E;
2205 case ISD::SETGT: return X86::COND_G;
2206 case ISD::SETGE: return X86::COND_GE;
2207 case ISD::SETLT: return X86::COND_L;
2208 case ISD::SETLE: return X86::COND_LE;
2209 case ISD::SETNE: return X86::COND_NE;
2210 case ISD::SETULT: return X86::COND_B;
2211 case ISD::SETUGT: return X86::COND_A;
2212 case ISD::SETULE: return X86::COND_BE;
2213 case ISD::SETUGE: return X86::COND_AE;
2217 // First determine if it is required or is profitable to flip the operands.
2219 // If LHS is a foldable load, but RHS is not, flip the condition.
2220 if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
2221 !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
2222 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
2223 std::swap(LHS, RHS);
2226 switch (SetCCOpcode) {
2232 std::swap(LHS, RHS);
2236 // On a floating point condition, the flags are set as follows:
2238 // 0 | 0 | 0 | X > Y
2239 // 0 | 0 | 1 | X < Y
2240 // 1 | 0 | 0 | X == Y
2241 // 1 | 1 | 1 | unordered
2242 switch (SetCCOpcode) {
2243 default: llvm_unreachable("Condcode should be pre-legalized away");
2245 case ISD::SETEQ: return X86::COND_E;
2246 case ISD::SETOLT: // flipped
2248 case ISD::SETGT: return X86::COND_A;
2249 case ISD::SETOLE: // flipped
2251 case ISD::SETGE: return X86::COND_AE;
2252 case ISD::SETUGT: // flipped
2254 case ISD::SETLT: return X86::COND_B;
2255 case ISD::SETUGE: // flipped
2257 case ISD::SETLE: return X86::COND_BE;
2259 case ISD::SETNE: return X86::COND_NE;
2260 case ISD::SETUO: return X86::COND_P;
2261 case ISD::SETO: return X86::COND_NP;
2265 /// hasFPCMov - is there a floating point cmov for the specific X86 condition
2266 /// code. Current x86 isa includes the following FP cmov instructions:
2267 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
2268 static bool hasFPCMov(unsigned X86CC) {
2284 /// isUndefOrInRange - Return true if Val is undef or if its value falls within
2285 /// the specified range (L, H].
2286 static bool isUndefOrInRange(int Val, int Low, int Hi) {
2287 return (Val < 0) || (Val >= Low && Val < Hi);
2290 /// isUndefOrEqual - Val is either less than zero (undef) or equal to the
2291 /// specified value.
2292 static bool isUndefOrEqual(int Val, int CmpVal) {
2293 if (Val < 0 || Val == CmpVal)
2298 /// isPSHUFDMask - Return true if the node specifies a shuffle of elements that
2299 /// is suitable for input to PSHUFD or PSHUFW. That is, it doesn't reference
2300 /// the second operand.
2301 static bool isPSHUFDMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2302 if (VT == MVT::v4f32 || VT == MVT::v4i32 || VT == MVT::v4i16)
2303 return (Mask[0] < 4 && Mask[1] < 4 && Mask[2] < 4 && Mask[3] < 4);
2304 if (VT == MVT::v2f64 || VT == MVT::v2i64)
2305 return (Mask[0] < 2 && Mask[1] < 2);
2309 bool X86::isPSHUFDMask(ShuffleVectorSDNode *N) {
2310 SmallVector<int, 8> M;
2312 return ::isPSHUFDMask(M, N->getValueType(0));
2315 /// isPSHUFHWMask - Return true if the node specifies a shuffle of elements that
2316 /// is suitable for input to PSHUFHW.
2317 static bool isPSHUFHWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2318 if (VT != MVT::v8i16)
2321 // Lower quadword copied in order or undef.
2322 for (int i = 0; i != 4; ++i)
2323 if (Mask[i] >= 0 && Mask[i] != i)
2326 // Upper quadword shuffled.
2327 for (int i = 4; i != 8; ++i)
2328 if (Mask[i] >= 0 && (Mask[i] < 4 || Mask[i] > 7))
2334 bool X86::isPSHUFHWMask(ShuffleVectorSDNode *N) {
2335 SmallVector<int, 8> M;
2337 return ::isPSHUFHWMask(M, N->getValueType(0));
2340 /// isPSHUFLWMask - Return true if the node specifies a shuffle of elements that
2341 /// is suitable for input to PSHUFLW.
2342 static bool isPSHUFLWMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2343 if (VT != MVT::v8i16)
2346 // Upper quadword copied in order.
2347 for (int i = 4; i != 8; ++i)
2348 if (Mask[i] >= 0 && Mask[i] != i)
2351 // Lower quadword shuffled.
2352 for (int i = 0; i != 4; ++i)
2359 bool X86::isPSHUFLWMask(ShuffleVectorSDNode *N) {
2360 SmallVector<int, 8> M;
2362 return ::isPSHUFLWMask(M, N->getValueType(0));
2365 /// isSHUFPMask - Return true if the specified VECTOR_SHUFFLE operand
2366 /// specifies a shuffle of elements that is suitable for input to SHUFP*.
2367 static bool isSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2368 int NumElems = VT.getVectorNumElements();
2369 if (NumElems != 2 && NumElems != 4)
2372 int Half = NumElems / 2;
2373 for (int i = 0; i < Half; ++i)
2374 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2376 for (int i = Half; i < NumElems; ++i)
2377 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2383 bool X86::isSHUFPMask(ShuffleVectorSDNode *N) {
2384 SmallVector<int, 8> M;
2386 return ::isSHUFPMask(M, N->getValueType(0));
2389 /// isCommutedSHUFP - Returns true if the shuffle mask is exactly
2390 /// the reverse of what x86 shuffles want. x86 shuffles requires the lower
2391 /// half elements to come from vector 1 (which would equal the dest.) and
2392 /// the upper half to come from vector 2.
2393 static bool isCommutedSHUFPMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2394 int NumElems = VT.getVectorNumElements();
2396 if (NumElems != 2 && NumElems != 4)
2399 int Half = NumElems / 2;
2400 for (int i = 0; i < Half; ++i)
2401 if (!isUndefOrInRange(Mask[i], NumElems, NumElems*2))
2403 for (int i = Half; i < NumElems; ++i)
2404 if (!isUndefOrInRange(Mask[i], 0, NumElems))
2409 static bool isCommutedSHUFP(ShuffleVectorSDNode *N) {
2410 SmallVector<int, 8> M;
2412 return isCommutedSHUFPMask(M, N->getValueType(0));
2415 /// isMOVHLPSMask - Return true if the specified VECTOR_SHUFFLE operand
2416 /// specifies a shuffle of elements that is suitable for input to MOVHLPS.
2417 bool X86::isMOVHLPSMask(ShuffleVectorSDNode *N) {
2418 if (N->getValueType(0).getVectorNumElements() != 4)
2421 // Expect bit0 == 6, bit1 == 7, bit2 == 2, bit3 == 3
2422 return isUndefOrEqual(N->getMaskElt(0), 6) &&
2423 isUndefOrEqual(N->getMaskElt(1), 7) &&
2424 isUndefOrEqual(N->getMaskElt(2), 2) &&
2425 isUndefOrEqual(N->getMaskElt(3), 3);
2428 /// isMOVLPMask - Return true if the specified VECTOR_SHUFFLE operand
2429 /// specifies a shuffle of elements that is suitable for input to MOVLP{S|D}.
2430 bool X86::isMOVLPMask(ShuffleVectorSDNode *N) {
2431 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2433 if (NumElems != 2 && NumElems != 4)
2436 for (unsigned i = 0; i < NumElems/2; ++i)
2437 if (!isUndefOrEqual(N->getMaskElt(i), i + NumElems))
2440 for (unsigned i = NumElems/2; i < NumElems; ++i)
2441 if (!isUndefOrEqual(N->getMaskElt(i), i))
2447 /// isMOVHPMask - Return true if the specified VECTOR_SHUFFLE operand
2448 /// specifies a shuffle of elements that is suitable for input to MOVHP{S|D}
2450 bool X86::isMOVHPMask(ShuffleVectorSDNode *N) {
2451 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2453 if (NumElems != 2 && NumElems != 4)
2456 for (unsigned i = 0; i < NumElems/2; ++i)
2457 if (!isUndefOrEqual(N->getMaskElt(i), i))
2460 for (unsigned i = 0; i < NumElems/2; ++i)
2461 if (!isUndefOrEqual(N->getMaskElt(i + NumElems/2), i + NumElems))
2467 /// isMOVHLPS_v_undef_Mask - Special case of isMOVHLPSMask for canonical form
2468 /// of vector_shuffle v, v, <2, 3, 2, 3>, i.e. vector_shuffle v, undef,
2470 bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
2471 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2476 return isUndefOrEqual(N->getMaskElt(0), 2) &&
2477 isUndefOrEqual(N->getMaskElt(1), 3) &&
2478 isUndefOrEqual(N->getMaskElt(2), 2) &&
2479 isUndefOrEqual(N->getMaskElt(3), 3);
2482 /// isUNPCKLMask - Return true if the specified VECTOR_SHUFFLE operand
2483 /// specifies a shuffle of elements that is suitable for input to UNPCKL.
2484 static bool isUNPCKLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2485 bool V2IsSplat = false) {
2486 int NumElts = VT.getVectorNumElements();
2487 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2490 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2492 int BitI1 = Mask[i+1];
2493 if (!isUndefOrEqual(BitI, j))
2496 if (!isUndefOrEqual(BitI1, NumElts))
2499 if (!isUndefOrEqual(BitI1, j + NumElts))
2506 bool X86::isUNPCKLMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2507 SmallVector<int, 8> M;
2509 return ::isUNPCKLMask(M, N->getValueType(0), V2IsSplat);
2512 /// isUNPCKHMask - Return true if the specified VECTOR_SHUFFLE operand
2513 /// specifies a shuffle of elements that is suitable for input to UNPCKH.
2514 static bool isUNPCKHMask(const SmallVectorImpl<int> &Mask, EVT VT,
2515 bool V2IsSplat = false) {
2516 int NumElts = VT.getVectorNumElements();
2517 if (NumElts != 2 && NumElts != 4 && NumElts != 8 && NumElts != 16)
2520 for (int i = 0, j = 0; i != NumElts; i += 2, ++j) {
2522 int BitI1 = Mask[i+1];
2523 if (!isUndefOrEqual(BitI, j + NumElts/2))
2526 if (isUndefOrEqual(BitI1, NumElts))
2529 if (!isUndefOrEqual(BitI1, j + NumElts/2 + NumElts))
2536 bool X86::isUNPCKHMask(ShuffleVectorSDNode *N, bool V2IsSplat) {
2537 SmallVector<int, 8> M;
2539 return ::isUNPCKHMask(M, N->getValueType(0), V2IsSplat);
2542 /// isUNPCKL_v_undef_Mask - Special case of isUNPCKLMask for canonical form
2543 /// of vector_shuffle v, v, <0, 4, 1, 5>, i.e. vector_shuffle v, undef,
2545 static bool isUNPCKL_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2546 int NumElems = VT.getVectorNumElements();
2547 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2550 for (int i = 0, j = 0; i != NumElems; i += 2, ++j) {
2552 int BitI1 = Mask[i+1];
2553 if (!isUndefOrEqual(BitI, j))
2555 if (!isUndefOrEqual(BitI1, j))
2561 bool X86::isUNPCKL_v_undef_Mask(ShuffleVectorSDNode *N) {
2562 SmallVector<int, 8> M;
2564 return ::isUNPCKL_v_undef_Mask(M, N->getValueType(0));
2567 /// isUNPCKH_v_undef_Mask - Special case of isUNPCKHMask for canonical form
2568 /// of vector_shuffle v, v, <2, 6, 3, 7>, i.e. vector_shuffle v, undef,
2570 static bool isUNPCKH_v_undef_Mask(const SmallVectorImpl<int> &Mask, EVT VT) {
2571 int NumElems = VT.getVectorNumElements();
2572 if (NumElems != 2 && NumElems != 4 && NumElems != 8 && NumElems != 16)
2575 for (int i = 0, j = NumElems / 2; i != NumElems; i += 2, ++j) {
2577 int BitI1 = Mask[i+1];
2578 if (!isUndefOrEqual(BitI, j))
2580 if (!isUndefOrEqual(BitI1, j))
2586 bool X86::isUNPCKH_v_undef_Mask(ShuffleVectorSDNode *N) {
2587 SmallVector<int, 8> M;
2589 return ::isUNPCKH_v_undef_Mask(M, N->getValueType(0));
2592 /// isMOVLMask - Return true if the specified VECTOR_SHUFFLE operand
2593 /// specifies a shuffle of elements that is suitable for input to MOVSS,
2594 /// MOVSD, and MOVD, i.e. setting the lowest element.
2595 static bool isMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT) {
2596 if (VT.getVectorElementType().getSizeInBits() < 32)
2599 int NumElts = VT.getVectorNumElements();
2601 if (!isUndefOrEqual(Mask[0], NumElts))
2604 for (int i = 1; i < NumElts; ++i)
2605 if (!isUndefOrEqual(Mask[i], i))
2611 bool X86::isMOVLMask(ShuffleVectorSDNode *N) {
2612 SmallVector<int, 8> M;
2614 return ::isMOVLMask(M, N->getValueType(0));
2617 /// isCommutedMOVL - Returns true if the shuffle mask is except the reverse
2618 /// of what x86 movss want. X86 movs requires the lowest element to be lowest
2619 /// element of vector 2 and the other elements to come from vector 1 in order.
2620 static bool isCommutedMOVLMask(const SmallVectorImpl<int> &Mask, EVT VT,
2621 bool V2IsSplat = false, bool V2IsUndef = false) {
2622 int NumOps = VT.getVectorNumElements();
2623 if (NumOps != 2 && NumOps != 4 && NumOps != 8 && NumOps != 16)
2626 if (!isUndefOrEqual(Mask[0], 0))
2629 for (int i = 1; i < NumOps; ++i)
2630 if (!(isUndefOrEqual(Mask[i], i+NumOps) ||
2631 (V2IsUndef && isUndefOrInRange(Mask[i], NumOps, NumOps*2)) ||
2632 (V2IsSplat && isUndefOrEqual(Mask[i], NumOps))))
2638 static bool isCommutedMOVL(ShuffleVectorSDNode *N, bool V2IsSplat = false,
2639 bool V2IsUndef = false) {
2640 SmallVector<int, 8> M;
2642 return isCommutedMOVLMask(M, N->getValueType(0), V2IsSplat, V2IsUndef);
2645 /// isMOVSHDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2646 /// specifies a shuffle of elements that is suitable for input to MOVSHDUP.
2647 bool X86::isMOVSHDUPMask(ShuffleVectorSDNode *N) {
2648 if (N->getValueType(0).getVectorNumElements() != 4)
2651 // Expect 1, 1, 3, 3
2652 for (unsigned i = 0; i < 2; ++i) {
2653 int Elt = N->getMaskElt(i);
2654 if (Elt >= 0 && Elt != 1)
2659 for (unsigned i = 2; i < 4; ++i) {
2660 int Elt = N->getMaskElt(i);
2661 if (Elt >= 0 && Elt != 3)
2666 // Don't use movshdup if it can be done with a shufps.
2667 // FIXME: verify that matching u, u, 3, 3 is what we want.
2671 /// isMOVSLDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2672 /// specifies a shuffle of elements that is suitable for input to MOVSLDUP.
2673 bool X86::isMOVSLDUPMask(ShuffleVectorSDNode *N) {
2674 if (N->getValueType(0).getVectorNumElements() != 4)
2677 // Expect 0, 0, 2, 2
2678 for (unsigned i = 0; i < 2; ++i)
2679 if (N->getMaskElt(i) > 0)
2683 for (unsigned i = 2; i < 4; ++i) {
2684 int Elt = N->getMaskElt(i);
2685 if (Elt >= 0 && Elt != 2)
2690 // Don't use movsldup if it can be done with a shufps.
2694 /// isMOVDDUPMask - Return true if the specified VECTOR_SHUFFLE operand
2695 /// specifies a shuffle of elements that is suitable for input to MOVDDUP.
2696 bool X86::isMOVDDUPMask(ShuffleVectorSDNode *N) {
2697 int e = N->getValueType(0).getVectorNumElements() / 2;
2699 for (int i = 0; i < e; ++i)
2700 if (!isUndefOrEqual(N->getMaskElt(i), i))
2702 for (int i = 0; i < e; ++i)
2703 if (!isUndefOrEqual(N->getMaskElt(e+i), i))
2708 /// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
2709 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUF* and SHUFP*
2711 unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
2712 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2713 int NumOperands = SVOp->getValueType(0).getVectorNumElements();
2715 unsigned Shift = (NumOperands == 4) ? 2 : 1;
2717 for (int i = 0; i < NumOperands; ++i) {
2718 int Val = SVOp->getMaskElt(NumOperands-i-1);
2719 if (Val < 0) Val = 0;
2720 if (Val >= NumOperands) Val -= NumOperands;
2722 if (i != NumOperands - 1)
2728 /// getShufflePSHUFHWImmediate - Return the appropriate immediate to shuffle
2729 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFHW
2731 unsigned X86::getShufflePSHUFHWImmediate(SDNode *N) {
2732 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2734 // 8 nodes, but we only care about the last 4.
2735 for (unsigned i = 7; i >= 4; --i) {
2736 int Val = SVOp->getMaskElt(i);
2745 /// getShufflePSHUFLWImmediate - Return the appropriate immediate to shuffle
2746 /// the specified isShuffleMask VECTOR_SHUFFLE mask with PSHUFLW
2748 unsigned X86::getShufflePSHUFLWImmediate(SDNode *N) {
2749 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
2751 // 8 nodes, but we only care about the first 4.
2752 for (int i = 3; i >= 0; --i) {
2753 int Val = SVOp->getMaskElt(i);
2762 /// isZeroNode - Returns true if Elt is a constant zero or a floating point
2764 bool X86::isZeroNode(SDValue Elt) {
2765 return ((isa<ConstantSDNode>(Elt) &&
2766 cast<ConstantSDNode>(Elt)->getZExtValue() == 0) ||
2767 (isa<ConstantFPSDNode>(Elt) &&
2768 cast<ConstantFPSDNode>(Elt)->getValueAPF().isPosZero()));
2771 /// CommuteVectorShuffle - Swap vector_shuffle operands as well as values in
2772 /// their permute mask.
2773 static SDValue CommuteVectorShuffle(ShuffleVectorSDNode *SVOp,
2774 SelectionDAG &DAG) {
2775 EVT VT = SVOp->getValueType(0);
2776 unsigned NumElems = VT.getVectorNumElements();
2777 SmallVector<int, 8> MaskVec;
2779 for (unsigned i = 0; i != NumElems; ++i) {
2780 int idx = SVOp->getMaskElt(i);
2782 MaskVec.push_back(idx);
2783 else if (idx < (int)NumElems)
2784 MaskVec.push_back(idx + NumElems);
2786 MaskVec.push_back(idx - NumElems);
2788 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(1),
2789 SVOp->getOperand(0), &MaskVec[0]);
2792 /// CommuteVectorShuffleMask - Change values in a shuffle permute mask assuming
2793 /// the two vector operands have swapped position.
2794 static void CommuteVectorShuffleMask(SmallVectorImpl<int> &Mask, EVT VT) {
2795 unsigned NumElems = VT.getVectorNumElements();
2796 for (unsigned i = 0; i != NumElems; ++i) {
2800 else if (idx < (int)NumElems)
2801 Mask[i] = idx + NumElems;
2803 Mask[i] = idx - NumElems;
2807 /// ShouldXformToMOVHLPS - Return true if the node should be transformed to
2808 /// match movhlps. The lower half elements should come from upper half of
2809 /// V1 (and in order), and the upper half elements should come from the upper
2810 /// half of V2 (and in order).
2811 static bool ShouldXformToMOVHLPS(ShuffleVectorSDNode *Op) {
2812 if (Op->getValueType(0).getVectorNumElements() != 4)
2814 for (unsigned i = 0, e = 2; i != e; ++i)
2815 if (!isUndefOrEqual(Op->getMaskElt(i), i+2))
2817 for (unsigned i = 2; i != 4; ++i)
2818 if (!isUndefOrEqual(Op->getMaskElt(i), i+4))
2823 /// isScalarLoadToVector - Returns true if the node is a scalar load that
2824 /// is promoted to a vector. It also returns the LoadSDNode by reference if
2826 static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
2827 if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
2829 N = N->getOperand(0).getNode();
2830 if (!ISD::isNON_EXTLoad(N))
2833 *LD = cast<LoadSDNode>(N);
2837 /// ShouldXformToMOVLP{S|D} - Return true if the node should be transformed to
2838 /// match movlp{s|d}. The lower half elements should come from lower half of
2839 /// V1 (and in order), and the upper half elements should come from the upper
2840 /// half of V2 (and in order). And since V1 will become the source of the
2841 /// MOVLP, it must be either a vector load or a scalar load to vector.
2842 static bool ShouldXformToMOVLP(SDNode *V1, SDNode *V2,
2843 ShuffleVectorSDNode *Op) {
2844 if (!ISD::isNON_EXTLoad(V1) && !isScalarLoadToVector(V1))
2846 // Is V2 is a vector load, don't do this transformation. We will try to use
2847 // load folding shufps op.
2848 if (ISD::isNON_EXTLoad(V2))
2851 unsigned NumElems = Op->getValueType(0).getVectorNumElements();
2853 if (NumElems != 2 && NumElems != 4)
2855 for (unsigned i = 0, e = NumElems/2; i != e; ++i)
2856 if (!isUndefOrEqual(Op->getMaskElt(i), i))
2858 for (unsigned i = NumElems/2; i != NumElems; ++i)
2859 if (!isUndefOrEqual(Op->getMaskElt(i), i+NumElems))
2864 /// isSplatVector - Returns true if N is a BUILD_VECTOR node whose elements are
2866 static bool isSplatVector(SDNode *N) {
2867 if (N->getOpcode() != ISD::BUILD_VECTOR)
2870 SDValue SplatValue = N->getOperand(0);
2871 for (unsigned i = 1, e = N->getNumOperands(); i != e; ++i)
2872 if (N->getOperand(i) != SplatValue)
2877 /// isZeroShuffle - Returns true if N is a VECTOR_SHUFFLE that can be resolved
2878 /// to an zero vector.
2879 /// FIXME: move to dag combiner / method on ShuffleVectorSDNode
2880 static bool isZeroShuffle(ShuffleVectorSDNode *N) {
2881 SDValue V1 = N->getOperand(0);
2882 SDValue V2 = N->getOperand(1);
2883 unsigned NumElems = N->getValueType(0).getVectorNumElements();
2884 for (unsigned i = 0; i != NumElems; ++i) {
2885 int Idx = N->getMaskElt(i);
2886 if (Idx >= (int)NumElems) {
2887 unsigned Opc = V2.getOpcode();
2888 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V2.getNode()))
2890 if (Opc != ISD::BUILD_VECTOR ||
2891 !X86::isZeroNode(V2.getOperand(Idx-NumElems)))
2893 } else if (Idx >= 0) {
2894 unsigned Opc = V1.getOpcode();
2895 if (Opc == ISD::UNDEF || ISD::isBuildVectorAllZeros(V1.getNode()))
2897 if (Opc != ISD::BUILD_VECTOR ||
2898 !X86::isZeroNode(V1.getOperand(Idx)))
2905 /// getZeroVector - Returns a vector of specified type with all zero elements.
2907 static SDValue getZeroVector(EVT VT, bool HasSSE2, SelectionDAG &DAG,
2909 assert(VT.isVector() && "Expected a vector type");
2911 // Always build zero vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2912 // type. This ensures they get CSE'd.
2914 if (VT.getSizeInBits() == 64) { // MMX
2915 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2916 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2917 } else if (HasSSE2) { // SSE2
2918 SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
2919 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2921 SDValue Cst = DAG.getTargetConstantFP(+0.0, MVT::f32);
2922 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
2924 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2927 /// getOnesVector - Returns a vector of specified type with all bits set.
2929 static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
2930 assert(VT.isVector() && "Expected a vector type");
2932 // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
2933 // type. This ensures they get CSE'd.
2934 SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
2936 if (VT.getSizeInBits() == 64) // MMX
2937 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
2939 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
2940 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
2944 /// NormalizeMask - V2 is a splat, modify the mask (if needed) so all elements
2945 /// that point to V2 points to its first element.
2946 static SDValue NormalizeMask(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
2947 EVT VT = SVOp->getValueType(0);
2948 unsigned NumElems = VT.getVectorNumElements();
2950 bool Changed = false;
2951 SmallVector<int, 8> MaskVec;
2952 SVOp->getMask(MaskVec);
2954 for (unsigned i = 0; i != NumElems; ++i) {
2955 if (MaskVec[i] > (int)NumElems) {
2956 MaskVec[i] = NumElems;
2961 return DAG.getVectorShuffle(VT, SVOp->getDebugLoc(), SVOp->getOperand(0),
2962 SVOp->getOperand(1), &MaskVec[0]);
2963 return SDValue(SVOp, 0);
2966 /// getMOVLMask - Returns a vector_shuffle mask for an movs{s|d}, movd
2967 /// operation of specified width.
2968 static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
2970 unsigned NumElems = VT.getVectorNumElements();
2971 SmallVector<int, 8> Mask;
2972 Mask.push_back(NumElems);
2973 for (unsigned i = 1; i != NumElems; ++i)
2975 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2978 /// getUnpackl - Returns a vector_shuffle node for an unpackl operation.
2979 static SDValue getUnpackl(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
2981 unsigned NumElems = VT.getVectorNumElements();
2982 SmallVector<int, 8> Mask;
2983 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
2985 Mask.push_back(i + NumElems);
2987 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2990 /// getUnpackhMask - Returns a vector_shuffle node for an unpackh operation.
2991 static SDValue getUnpackh(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
2993 unsigned NumElems = VT.getVectorNumElements();
2994 unsigned Half = NumElems/2;
2995 SmallVector<int, 8> Mask;
2996 for (unsigned i = 0; i != Half; ++i) {
2997 Mask.push_back(i + Half);
2998 Mask.push_back(i + NumElems + Half);
3000 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3003 /// PromoteSplat - Promote a splat of v4f32, v8i16 or v16i8 to v4i32.
3004 static SDValue PromoteSplat(ShuffleVectorSDNode *SV, SelectionDAG &DAG,
3006 if (SV->getValueType(0).getVectorNumElements() <= 4)
3007 return SDValue(SV, 0);
3009 EVT PVT = MVT::v4f32;
3010 EVT VT = SV->getValueType(0);
3011 DebugLoc dl = SV->getDebugLoc();
3012 SDValue V1 = SV->getOperand(0);
3013 int NumElems = VT.getVectorNumElements();
3014 int EltNo = SV->getSplatIndex();
3016 // unpack elements to the correct location
3017 while (NumElems > 4) {
3018 if (EltNo < NumElems/2) {
3019 V1 = getUnpackl(DAG, dl, VT, V1, V1);
3021 V1 = getUnpackh(DAG, dl, VT, V1, V1);
3022 EltNo -= NumElems/2;
3027 // Perform the splat.
3028 int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
3029 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
3030 V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
3031 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
3034 /// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
3035 /// vector of zero or undef vector. This produces a shuffle where the low
3036 /// element of V2 is swizzled into the zero/undef vector, landing at element
3037 /// Idx. This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
3038 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
3039 bool isZero, bool HasSSE2,
3040 SelectionDAG &DAG) {
3041 EVT VT = V2.getValueType();
3043 ? getZeroVector(VT, HasSSE2, DAG, V2.getDebugLoc()) : DAG.getUNDEF(VT);
3044 unsigned NumElems = VT.getVectorNumElements();
3045 SmallVector<int, 16> MaskVec;
3046 for (unsigned i = 0; i != NumElems; ++i)
3047 // If this is the insertion idx, put the low elt of V2 here.
3048 MaskVec.push_back(i == Idx ? NumElems : i);
3049 return DAG.getVectorShuffle(VT, V2.getDebugLoc(), V1, V2, &MaskVec[0]);
3052 /// getNumOfConsecutiveZeros - Return the number of elements in a result of
3053 /// a shuffle that is zero.
3055 unsigned getNumOfConsecutiveZeros(ShuffleVectorSDNode *SVOp, int NumElems,
3056 bool Low, SelectionDAG &DAG) {
3057 unsigned NumZeros = 0;
3058 for (int i = 0; i < NumElems; ++i) {
3059 unsigned Index = Low ? i : NumElems-i-1;
3060 int Idx = SVOp->getMaskElt(Index);
3065 SDValue Elt = DAG.getShuffleScalarElt(SVOp, Index);
3066 if (Elt.getNode() && X86::isZeroNode(Elt))
3074 /// isVectorShift - Returns true if the shuffle can be implemented as a
3075 /// logical left or right shift of a vector.
3076 /// FIXME: split into pslldqi, psrldqi, palignr variants.
3077 static bool isVectorShift(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG,
3078 bool &isLeft, SDValue &ShVal, unsigned &ShAmt) {
3079 int NumElems = SVOp->getValueType(0).getVectorNumElements();
3082 unsigned NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, true, DAG);
3085 NumZeros = getNumOfConsecutiveZeros(SVOp, NumElems, false, DAG);
3089 bool SeenV1 = false;
3090 bool SeenV2 = false;
3091 for (int i = NumZeros; i < NumElems; ++i) {
3092 int Val = isLeft ? (i - NumZeros) : i;
3093 int Idx = SVOp->getMaskElt(isLeft ? i : (i - NumZeros));
3105 if (SeenV1 && SeenV2)
3108 ShVal = SeenV1 ? SVOp->getOperand(0) : SVOp->getOperand(1);
3114 /// LowerBuildVectorv16i8 - Custom lower build_vector of v16i8.
3116 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
3117 unsigned NumNonZero, unsigned NumZero,
3118 SelectionDAG &DAG, TargetLowering &TLI) {
3122 DebugLoc dl = Op.getDebugLoc();
3125 for (unsigned i = 0; i < 16; ++i) {
3126 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
3127 if (ThisIsNonZero && First) {
3129 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3131 V = DAG.getUNDEF(MVT::v8i16);
3136 SDValue ThisElt(0, 0), LastElt(0, 0);
3137 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
3138 if (LastIsNonZero) {
3139 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
3140 MVT::i16, Op.getOperand(i-1));
3142 if (ThisIsNonZero) {
3143 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
3144 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
3145 ThisElt, DAG.getConstant(8, MVT::i8));
3147 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
3151 if (ThisElt.getNode())
3152 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
3153 DAG.getIntPtrConstant(i/2));
3157 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
3160 /// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
3162 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
3163 unsigned NumNonZero, unsigned NumZero,
3164 SelectionDAG &DAG, TargetLowering &TLI) {
3168 DebugLoc dl = Op.getDebugLoc();
3171 for (unsigned i = 0; i < 8; ++i) {
3172 bool isNonZero = (NonZeros & (1 << i)) != 0;
3176 V = getZeroVector(MVT::v8i16, true, DAG, dl);
3178 V = DAG.getUNDEF(MVT::v8i16);
3181 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
3182 MVT::v8i16, V, Op.getOperand(i),
3183 DAG.getIntPtrConstant(i));
3190 /// getVShift - Return a vector logical shift node.
3192 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
3193 unsigned NumBits, SelectionDAG &DAG,
3194 const TargetLowering &TLI, DebugLoc dl) {
3195 bool isMMX = VT.getSizeInBits() == 64;
3196 EVT ShVT = isMMX ? MVT::v1i64 : MVT::v2i64;
3197 unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
3198 SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
3199 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3200 DAG.getNode(Opc, dl, ShVT, SrcOp,
3201 DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
3205 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
3206 DebugLoc dl = Op.getDebugLoc();
3207 // All zero's are handled with pxor, all one's are handled with pcmpeqd.
3208 if (ISD::isBuildVectorAllZeros(Op.getNode())
3209 || ISD::isBuildVectorAllOnes(Op.getNode())) {
3210 // Canonicalize this to either <4 x i32> or <2 x i32> (SSE vs MMX) to
3211 // 1) ensure the zero vectors are CSE'd, and 2) ensure that i64 scalars are
3212 // eliminated on x86-32 hosts.
3213 if (Op.getValueType() == MVT::v4i32 || Op.getValueType() == MVT::v2i32)
3216 if (ISD::isBuildVectorAllOnes(Op.getNode()))
3217 return getOnesVector(Op.getValueType(), DAG, dl);
3218 return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
3221 EVT VT = Op.getValueType();
3222 EVT ExtVT = VT.getVectorElementType();
3223 unsigned EVTBits = ExtVT.getSizeInBits();
3225 unsigned NumElems = Op.getNumOperands();
3226 unsigned NumZero = 0;
3227 unsigned NumNonZero = 0;
3228 unsigned NonZeros = 0;
3229 bool IsAllConstants = true;
3230 SmallSet<SDValue, 8> Values;
3231 for (unsigned i = 0; i < NumElems; ++i) {
3232 SDValue Elt = Op.getOperand(i);
3233 if (Elt.getOpcode() == ISD::UNDEF)
3236 if (Elt.getOpcode() != ISD::Constant &&
3237 Elt.getOpcode() != ISD::ConstantFP)
3238 IsAllConstants = false;
3239 if (X86::isZeroNode(Elt))
3242 NonZeros |= (1 << i);
3247 if (NumNonZero == 0) {
3248 // All undef vector. Return an UNDEF. All zero vectors were handled above.
3249 return DAG.getUNDEF(VT);
3252 // Special case for single non-zero, non-undef, element.
3253 if (NumNonZero == 1) {
3254 unsigned Idx = CountTrailingZeros_32(NonZeros);
3255 SDValue Item = Op.getOperand(Idx);
3257 // If this is an insertion of an i64 value on x86-32, and if the top bits of
3258 // the value are obviously zero, truncate the value to i32 and do the
3259 // insertion that way. Only do this if the value is non-constant or if the
3260 // value is a constant being inserted into element 0. It is cheaper to do
3261 // a constant pool load than it is to do a movd + shuffle.
3262 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
3263 (!IsAllConstants || Idx == 0)) {
3264 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
3265 // Handle MMX and SSE both.
3266 EVT VecVT = VT == MVT::v2i64 ? MVT::v4i32 : MVT::v2i32;
3267 unsigned VecElts = VT == MVT::v2i64 ? 4 : 2;
3269 // Truncate the value (which may itself be a constant) to i32, and
3270 // convert it to a vector with movd (S2V+shuffle to zero extend).
3271 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
3272 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
3273 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3274 Subtarget->hasSSE2(), DAG);
3276 // Now we have our 32-bit value zero extended in the low element of
3277 // a vector. If Idx != 0, swizzle it into place.
3279 SmallVector<int, 4> Mask;
3280 Mask.push_back(Idx);
3281 for (unsigned i = 1; i != VecElts; ++i)
3283 Item = DAG.getVectorShuffle(VecVT, dl, Item,
3284 DAG.getUNDEF(Item.getValueType()),
3287 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
3291 // If we have a constant or non-constant insertion into the low element of
3292 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
3293 // the rest of the elements. This will be matched as movd/movq/movss/movsd
3294 // depending on what the source datatype is.
3297 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3298 } else if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
3299 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
3300 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3301 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
3302 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget->hasSSE2(),
3304 } else if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
3305 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
3306 EVT MiddleVT = VT.getSizeInBits() == 64 ? MVT::v2i32 : MVT::v4i32;
3307 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
3308 Item = getShuffleVectorZeroOrUndef(Item, 0, true,
3309 Subtarget->hasSSE2(), DAG);
3310 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
3314 // Is it a vector logical left shift?
3315 if (NumElems == 2 && Idx == 1 &&
3316 X86::isZeroNode(Op.getOperand(0)) &&
3317 !X86::isZeroNode(Op.getOperand(1))) {
3318 unsigned NumBits = VT.getSizeInBits();
3319 return getVShift(true, VT,
3320 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3321 VT, Op.getOperand(1)),
3322 NumBits/2, DAG, *this, dl);
3325 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
3328 // Otherwise, if this is a vector with i32 or f32 elements, and the element
3329 // is a non-constant being inserted into an element other than the low one,
3330 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
3331 // movd/movss) to move this into the low element, then shuffle it into
3333 if (EVTBits == 32) {
3334 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
3336 // Turn it into a shuffle of zero and zero-extended scalar to vector.
3337 Item = getShuffleVectorZeroOrUndef(Item, 0, NumZero > 0,
3338 Subtarget->hasSSE2(), DAG);
3339 SmallVector<int, 8> MaskVec;
3340 for (unsigned i = 0; i < NumElems; i++)
3341 MaskVec.push_back(i == Idx ? 0 : 1);
3342 return DAG.getVectorShuffle(VT, dl, Item, DAG.getUNDEF(VT), &MaskVec[0]);
3346 // Splat is obviously ok. Let legalizer expand it to a shuffle.
3347 if (Values.size() == 1)
3350 // A vector full of immediates; various special cases are already
3351 // handled, so this is best done with a single constant-pool load.
3355 // Let legalizer expand 2-wide build_vectors.
3356 if (EVTBits == 64) {
3357 if (NumNonZero == 1) {
3358 // One half is zero or undef.
3359 unsigned Idx = CountTrailingZeros_32(NonZeros);
3360 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
3361 Op.getOperand(Idx));
3362 return getShuffleVectorZeroOrUndef(V2, Idx, true,
3363 Subtarget->hasSSE2(), DAG);
3368 // If element VT is < 32 bits, convert it to inserts into a zero vector.
3369 if (EVTBits == 8 && NumElems == 16) {
3370 SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
3372 if (V.getNode()) return V;
3375 if (EVTBits == 16 && NumElems == 8) {
3376 SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
3378 if (V.getNode()) return V;
3381 // If element VT is == 32 bits, turn it into a number of shuffles.
3382 SmallVector<SDValue, 8> V;
3384 if (NumElems == 4 && NumZero > 0) {
3385 for (unsigned i = 0; i < 4; ++i) {
3386 bool isZero = !(NonZeros & (1 << i));
3388 V[i] = getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
3390 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3393 for (unsigned i = 0; i < 2; ++i) {
3394 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
3397 V[i] = V[i*2]; // Must be a zero vector.
3400 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
3403 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
3406 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
3411 SmallVector<int, 8> MaskVec;
3412 bool Reverse = (NonZeros & 0x3) == 2;
3413 for (unsigned i = 0; i < 2; ++i)
3414 MaskVec.push_back(Reverse ? 1-i : i);
3415 Reverse = ((NonZeros & (0x3 << 2)) >> 2) == 2;
3416 for (unsigned i = 0; i < 2; ++i)
3417 MaskVec.push_back(Reverse ? 1-i+NumElems : i+NumElems);
3418 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
3421 if (Values.size() > 2) {
3422 // If we have SSE 4.1, Expand into a number of inserts unless the number of
3423 // values to be inserted is equal to the number of elements, in which case
3424 // use the unpack code below in the hopes of matching the consecutive elts
3425 // load merge pattern for shuffles.
3426 // FIXME: We could probably just check that here directly.
3427 if (Values.size() < NumElems && VT.getSizeInBits() == 128 &&
3428 getSubtarget()->hasSSE41()) {
3429 V[0] = DAG.getUNDEF(VT);
3430 for (unsigned i = 0; i < NumElems; ++i)
3431 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
3432 V[0] = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, V[0],
3433 Op.getOperand(i), DAG.getIntPtrConstant(i));
3436 // Expand into a number of unpckl*.
3438 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
3439 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
3440 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
3441 for (unsigned i = 0; i < NumElems; ++i)
3442 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
3444 while (NumElems != 0) {
3445 for (unsigned i = 0; i < NumElems; ++i)
3446 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + NumElems]);
3455 // v8i16 shuffles - Prefer shuffles in the following order:
3456 // 1. [all] pshuflw, pshufhw, optional move
3457 // 2. [ssse3] 1 x pshufb
3458 // 3. [ssse3] 2 x pshufb + 1 x por
3459 // 4. [all] mov + pshuflw + pshufhw + N x (pextrw + pinsrw)
3461 SDValue LowerVECTOR_SHUFFLEv8i16(ShuffleVectorSDNode *SVOp,
3462 SelectionDAG &DAG, X86TargetLowering &TLI) {
3463 SDValue V1 = SVOp->getOperand(0);
3464 SDValue V2 = SVOp->getOperand(1);
3465 DebugLoc dl = SVOp->getDebugLoc();
3466 SmallVector<int, 8> MaskVals;
3468 // Determine if more than 1 of the words in each of the low and high quadwords
3469 // of the result come from the same quadword of one of the two inputs. Undef
3470 // mask values count as coming from any quadword, for better codegen.
3471 SmallVector<unsigned, 4> LoQuad(4);
3472 SmallVector<unsigned, 4> HiQuad(4);
3473 BitVector InputQuads(4);
3474 for (unsigned i = 0; i < 8; ++i) {
3475 SmallVectorImpl<unsigned> &Quad = i < 4 ? LoQuad : HiQuad;
3476 int EltIdx = SVOp->getMaskElt(i);
3477 MaskVals.push_back(EltIdx);
3486 InputQuads.set(EltIdx / 4);
3489 int BestLoQuad = -1;
3490 unsigned MaxQuad = 1;
3491 for (unsigned i = 0; i < 4; ++i) {
3492 if (LoQuad[i] > MaxQuad) {
3494 MaxQuad = LoQuad[i];
3498 int BestHiQuad = -1;
3500 for (unsigned i = 0; i < 4; ++i) {
3501 if (HiQuad[i] > MaxQuad) {
3503 MaxQuad = HiQuad[i];
3507 // For SSSE3, If all 8 words of the result come from only 1 quadword of each
3508 // of the two input vectors, shuffle them into one input vector so only a
3509 // single pshufb instruction is necessary. If There are more than 2 input
3510 // quads, disable the next transformation since it does not help SSSE3.
3511 bool V1Used = InputQuads[0] || InputQuads[1];
3512 bool V2Used = InputQuads[2] || InputQuads[3];
3513 if (TLI.getSubtarget()->hasSSSE3()) {
3514 if (InputQuads.count() == 2 && V1Used && V2Used) {
3515 BestLoQuad = InputQuads.find_first();
3516 BestHiQuad = InputQuads.find_next(BestLoQuad);
3518 if (InputQuads.count() > 2) {
3524 // If BestLoQuad or BestHiQuad are set, shuffle the quads together and update
3525 // the shuffle mask. If a quad is scored as -1, that means that it contains
3526 // words from all 4 input quadwords.
3528 if (BestLoQuad >= 0 || BestHiQuad >= 0) {
3529 SmallVector<int, 8> MaskV;
3530 MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
3531 MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
3532 NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
3533 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
3534 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
3535 NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
3537 // Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
3538 // source words for the shuffle, to aid later transformations.
3539 bool AllWordsInNewV = true;
3540 bool InOrder[2] = { true, true };
3541 for (unsigned i = 0; i != 8; ++i) {
3542 int idx = MaskVals[i];
3544 InOrder[i/4] = false;
3545 if (idx < 0 || (idx/4) == BestLoQuad || (idx/4) == BestHiQuad)
3547 AllWordsInNewV = false;
3551 bool pshuflw = AllWordsInNewV, pshufhw = AllWordsInNewV;
3552 if (AllWordsInNewV) {
3553 for (int i = 0; i != 8; ++i) {
3554 int idx = MaskVals[i];
3557 idx = MaskVals[i] = (idx / 4) == BestLoQuad ? (idx & 3) : (idx & 3) + 4;
3558 if ((idx != i) && idx < 4)
3560 if ((idx != i) && idx > 3)
3569 // If we've eliminated the use of V2, and the new mask is a pshuflw or
3570 // pshufhw, that's as cheap as it gets. Return the new shuffle.
3571 if ((pshufhw && InOrder[0]) || (pshuflw && InOrder[1])) {
3572 return DAG.getVectorShuffle(MVT::v8i16, dl, NewV,
3573 DAG.getUNDEF(MVT::v8i16), &MaskVals[0]);
3577 // If we have SSSE3, and all words of the result are from 1 input vector,
3578 // case 2 is generated, otherwise case 3 is generated. If no SSSE3
3579 // is present, fall back to case 4.
3580 if (TLI.getSubtarget()->hasSSSE3()) {
3581 SmallVector<SDValue,16> pshufbMask;
3583 // If we have elements from both input vectors, set the high bit of the
3584 // shuffle mask element to zero out elements that come from V2 in the V1
3585 // mask, and elements that come from V1 in the V2 mask, so that the two
3586 // results can be OR'd together.
3587 bool TwoInputs = V1Used && V2Used;
3588 for (unsigned i = 0; i != 8; ++i) {
3589 int EltIdx = MaskVals[i] * 2;
3590 if (TwoInputs && (EltIdx >= 16)) {
3591 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3592 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3595 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3596 pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
3598 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
3599 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3600 DAG.getNode(ISD::BUILD_VECTOR, dl,
3601 MVT::v16i8, &pshufbMask[0], 16));
3603 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3605 // Calculate the shuffle mask for the second input, shuffle it, and
3606 // OR it with the first shuffled input.
3608 for (unsigned i = 0; i != 8; ++i) {
3609 int EltIdx = MaskVals[i] * 2;
3611 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3612 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3615 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3616 pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
3618 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
3619 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3620 DAG.getNode(ISD::BUILD_VECTOR, dl,
3621 MVT::v16i8, &pshufbMask[0], 16));
3622 V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3623 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3626 // If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
3627 // and update MaskVals with new element order.
3628 BitVector InOrder(8);
3629 if (BestLoQuad >= 0) {
3630 SmallVector<int, 8> MaskV;
3631 for (int i = 0; i != 4; ++i) {
3632 int idx = MaskVals[i];
3634 MaskV.push_back(-1);
3636 } else if ((idx / 4) == BestLoQuad) {
3637 MaskV.push_back(idx & 3);
3640 MaskV.push_back(-1);
3643 for (unsigned i = 4; i != 8; ++i)
3645 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3649 // If BestHi >= 0, generate a pshufhw to put the high elements in order,
3650 // and update MaskVals with the new element order.
3651 if (BestHiQuad >= 0) {
3652 SmallVector<int, 8> MaskV;
3653 for (unsigned i = 0; i != 4; ++i)
3655 for (unsigned i = 4; i != 8; ++i) {
3656 int idx = MaskVals[i];
3658 MaskV.push_back(-1);
3660 } else if ((idx / 4) == BestHiQuad) {
3661 MaskV.push_back((idx & 3) + 4);
3664 MaskV.push_back(-1);
3667 NewV = DAG.getVectorShuffle(MVT::v8i16, dl, NewV, DAG.getUNDEF(MVT::v8i16),
3671 // In case BestHi & BestLo were both -1, which means each quadword has a word
3672 // from each of the four input quadwords, calculate the InOrder bitvector now
3673 // before falling through to the insert/extract cleanup.
3674 if (BestLoQuad == -1 && BestHiQuad == -1) {
3676 for (int i = 0; i != 8; ++i)
3677 if (MaskVals[i] < 0 || MaskVals[i] == i)
3681 // The other elements are put in the right place using pextrw and pinsrw.
3682 for (unsigned i = 0; i != 8; ++i) {
3685 int EltIdx = MaskVals[i];
3688 SDValue ExtOp = (EltIdx < 8)
3689 ? DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V1,
3690 DAG.getIntPtrConstant(EltIdx))
3691 : DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, V2,
3692 DAG.getIntPtrConstant(EltIdx - 8));
3693 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, ExtOp,
3694 DAG.getIntPtrConstant(i));
3699 // v16i8 shuffles - Prefer shuffles in the following order:
3700 // 1. [ssse3] 1 x pshufb
3701 // 2. [ssse3] 2 x pshufb + 1 x por
3702 // 3. [all] v8i16 shuffle + N x pextrw + rotate + pinsrw
3704 SDValue LowerVECTOR_SHUFFLEv16i8(ShuffleVectorSDNode *SVOp,
3705 SelectionDAG &DAG, X86TargetLowering &TLI) {
3706 SDValue V1 = SVOp->getOperand(0);
3707 SDValue V2 = SVOp->getOperand(1);
3708 DebugLoc dl = SVOp->getDebugLoc();
3709 SmallVector<int, 16> MaskVals;
3710 SVOp->getMask(MaskVals);
3712 // If we have SSSE3, case 1 is generated when all result bytes come from
3713 // one of the inputs. Otherwise, case 2 is generated. If no SSSE3 is
3714 // present, fall back to case 3.
3715 // FIXME: kill V2Only once shuffles are canonizalized by getNode.
3718 for (unsigned i = 0; i < 16; ++i) {
3719 int EltIdx = MaskVals[i];
3728 // If SSSE3, use 1 pshufb instruction per vector with elements in the result.
3729 if (TLI.getSubtarget()->hasSSSE3()) {
3730 SmallVector<SDValue,16> pshufbMask;
3732 // If all result elements are from one input vector, then only translate
3733 // undef mask values to 0x80 (zero out result) in the pshufb mask.
3735 // Otherwise, we have elements from both input vectors, and must zero out
3736 // elements that come from V2 in the first mask, and V1 in the second mask
3737 // so that we can OR them together.
3738 bool TwoInputs = !(V1Only || V2Only);
3739 for (unsigned i = 0; i != 16; ++i) {
3740 int EltIdx = MaskVals[i];
3741 if (EltIdx < 0 || (TwoInputs && EltIdx >= 16)) {
3742 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3745 pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
3747 // If all the elements are from V2, assign it to V1 and return after
3748 // building the first pshufb.
3751 V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
3752 DAG.getNode(ISD::BUILD_VECTOR, dl,
3753 MVT::v16i8, &pshufbMask[0], 16));
3757 // Calculate the shuffle mask for the second input, shuffle it, and
3758 // OR it with the first shuffled input.
3760 for (unsigned i = 0; i != 16; ++i) {
3761 int EltIdx = MaskVals[i];
3763 pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
3766 pshufbMask.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
3768 V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
3769 DAG.getNode(ISD::BUILD_VECTOR, dl,
3770 MVT::v16i8, &pshufbMask[0], 16));
3771 return DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
3774 // No SSSE3 - Calculate in place words and then fix all out of place words
3775 // With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
3776 // the 16 different words that comprise the two doublequadword input vectors.
3777 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
3778 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
3779 SDValue NewV = V2Only ? V2 : V1;
3780 for (int i = 0; i != 8; ++i) {
3781 int Elt0 = MaskVals[i*2];
3782 int Elt1 = MaskVals[i*2+1];
3784 // This word of the result is all undef, skip it.
3785 if (Elt0 < 0 && Elt1 < 0)
3788 // This word of the result is already in the correct place, skip it.
3789 if (V1Only && (Elt0 == i*2) && (Elt1 == i*2+1))
3791 if (V2Only && (Elt0 == i*2+16) && (Elt1 == i*2+17))
3794 SDValue Elt0Src = Elt0 < 16 ? V1 : V2;
3795 SDValue Elt1Src = Elt1 < 16 ? V1 : V2;
3798 // If Elt0 and Elt1 are defined, are consecutive, and can be load
3799 // using a single extract together, load it and store it.
3800 if ((Elt0 >= 0) && ((Elt0 + 1) == Elt1) && ((Elt0 & 1) == 0)) {
3801 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3802 DAG.getIntPtrConstant(Elt1 / 2));
3803 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3804 DAG.getIntPtrConstant(i));
3808 // If Elt1 is defined, extract it from the appropriate source. If the
3809 // source byte is not also odd, shift the extracted word left 8 bits
3810 // otherwise clear the bottom 8 bits if we need to do an or.
3812 InsElt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16, Elt1Src,
3813 DAG.getIntPtrConstant(Elt1 / 2));
3814 if ((Elt1 & 1) == 0)
3815 InsElt = DAG.getNode(ISD::SHL, dl, MVT::i16, InsElt,
3816 DAG.getConstant(8, TLI.getShiftAmountTy()));
3818 InsElt = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt,
3819 DAG.getConstant(0xFF00, MVT::i16));
3821 // If Elt0 is defined, extract it from the appropriate source. If the
3822 // source byte is not also even, shift the extracted word right 8 bits. If
3823 // Elt1 was also defined, OR the extracted values together before
3824 // inserting them in the result.
3826 SDValue InsElt0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i16,
3827 Elt0Src, DAG.getIntPtrConstant(Elt0 / 2));
3828 if ((Elt0 & 1) != 0)
3829 InsElt0 = DAG.getNode(ISD::SRL, dl, MVT::i16, InsElt0,
3830 DAG.getConstant(8, TLI.getShiftAmountTy()));
3832 InsElt0 = DAG.getNode(ISD::AND, dl, MVT::i16, InsElt0,
3833 DAG.getConstant(0x00FF, MVT::i16));
3834 InsElt = Elt1 >= 0 ? DAG.getNode(ISD::OR, dl, MVT::i16, InsElt, InsElt0)
3837 NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
3838 DAG.getIntPtrConstant(i));
3840 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
3843 /// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
3844 /// ones, or rewriting v4i32 / v2f32 as 2 wide ones if possible. This can be
3845 /// done when every pair / quad of shuffle mask elements point to elements in
3846 /// the right sequence. e.g.
3847 /// vector_shuffle <>, <>, < 3, 4, | 10, 11, | 0, 1, | 14, 15>
3849 SDValue RewriteAsNarrowerShuffle(ShuffleVectorSDNode *SVOp,
3851 TargetLowering &TLI, DebugLoc dl) {
3852 EVT VT = SVOp->getValueType(0);
3853 SDValue V1 = SVOp->getOperand(0);
3854 SDValue V2 = SVOp->getOperand(1);
3855 unsigned NumElems = VT.getVectorNumElements();
3856 unsigned NewWidth = (NumElems == 4) ? 2 : 4;
3857 EVT MaskVT = MVT::getIntVectorWithNumElements(NewWidth);
3858 EVT MaskEltVT = MaskVT.getVectorElementType();
3860 switch (VT.getSimpleVT().SimpleTy) {
3861 default: assert(false && "Unexpected!");
3862 case MVT::v4f32: NewVT = MVT::v2f64; break;
3863 case MVT::v4i32: NewVT = MVT::v2i64; break;
3864 case MVT::v8i16: NewVT = MVT::v4i32; break;
3865 case MVT::v16i8: NewVT = MVT::v4i32; break;
3868 if (NewWidth == 2) {
3874 int Scale = NumElems / NewWidth;
3875 SmallVector<int, 8> MaskVec;
3876 for (unsigned i = 0; i < NumElems; i += Scale) {
3878 for (int j = 0; j < Scale; ++j) {
3879 int EltIdx = SVOp->getMaskElt(i+j);
3883 StartIdx = EltIdx - (EltIdx % Scale);
3884 if (EltIdx != StartIdx + j)
3888 MaskVec.push_back(-1);
3890 MaskVec.push_back(StartIdx / Scale);
3893 V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
3894 V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
3895 return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
3898 /// getVZextMovL - Return a zero-extending vector move low node.
3900 static SDValue getVZextMovL(EVT VT, EVT OpVT,
3901 SDValue SrcOp, SelectionDAG &DAG,
3902 const X86Subtarget *Subtarget, DebugLoc dl) {
3903 if (VT == MVT::v2f64 || VT == MVT::v4f32) {
3904 LoadSDNode *LD = NULL;
3905 if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
3906 LD = dyn_cast<LoadSDNode>(SrcOp);
3908 // movssrr and movsdrr do not clear top bits. Try to use movd, movq
3910 MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
3911 if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
3912 SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
3913 SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
3914 SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
3916 OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
3917 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3918 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3919 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
3927 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
3928 DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
3929 DAG.getNode(ISD::BIT_CONVERT, dl,
3933 /// LowerVECTOR_SHUFFLE_4wide - Handle all 4 wide cases with a number of
3936 LowerVECTOR_SHUFFLE_4wide(ShuffleVectorSDNode *SVOp, SelectionDAG &DAG) {
3937 SDValue V1 = SVOp->getOperand(0);
3938 SDValue V2 = SVOp->getOperand(1);
3939 DebugLoc dl = SVOp->getDebugLoc();
3940 EVT VT = SVOp->getValueType(0);
3942 SmallVector<std::pair<int, int>, 8> Locs;
3944 SmallVector<int, 8> Mask1(4U, -1);
3945 SmallVector<int, 8> PermMask;
3946 SVOp->getMask(PermMask);
3950 for (unsigned i = 0; i != 4; ++i) {
3951 int Idx = PermMask[i];
3953 Locs[i] = std::make_pair(-1, -1);
3955 assert(Idx < 8 && "Invalid VECTOR_SHUFFLE index!");
3957 Locs[i] = std::make_pair(0, NumLo);
3961 Locs[i] = std::make_pair(1, NumHi);
3963 Mask1[2+NumHi] = Idx;
3969 if (NumLo <= 2 && NumHi <= 2) {
3970 // If no more than two elements come from either vector. This can be
3971 // implemented with two shuffles. First shuffle gather the elements.
3972 // The second shuffle, which takes the first shuffle as both of its
3973 // vector operands, put the elements into the right order.
3974 V1 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
3976 SmallVector<int, 8> Mask2(4U, -1);
3978 for (unsigned i = 0; i != 4; ++i) {
3979 if (Locs[i].first == -1)
3982 unsigned Idx = (i < 2) ? 0 : 4;
3983 Idx += Locs[i].first * 2 + Locs[i].second;
3988 return DAG.getVectorShuffle(VT, dl, V1, V1, &Mask2[0]);
3989 } else if (NumLo == 3 || NumHi == 3) {
3990 // Otherwise, we must have three elements from one vector, call it X, and
3991 // one element from the other, call it Y. First, use a shufps to build an
3992 // intermediate vector with the one element from Y and the element from X
3993 // that will be in the same half in the final destination (the indexes don't
3994 // matter). Then, use a shufps to build the final vector, taking the half
3995 // containing the element from Y from the intermediate, and the other half
3998 // Normalize it so the 3 elements come from V1.
3999 CommuteVectorShuffleMask(PermMask, VT);
4003 // Find the element from V2.
4005 for (HiIndex = 0; HiIndex < 3; ++HiIndex) {
4006 int Val = PermMask[HiIndex];
4013 Mask1[0] = PermMask[HiIndex];
4015 Mask1[2] = PermMask[HiIndex^1];
4017 V2 = DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4020 Mask1[0] = PermMask[0];
4021 Mask1[1] = PermMask[1];
4022 Mask1[2] = HiIndex & 1 ? 6 : 4;
4023 Mask1[3] = HiIndex & 1 ? 4 : 6;
4024 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask1[0]);
4026 Mask1[0] = HiIndex & 1 ? 2 : 0;
4027 Mask1[1] = HiIndex & 1 ? 0 : 2;
4028 Mask1[2] = PermMask[2];
4029 Mask1[3] = PermMask[3];
4034 return DAG.getVectorShuffle(VT, dl, V2, V1, &Mask1[0]);
4038 // Break it into (shuffle shuffle_hi, shuffle_lo).
4040 SmallVector<int,8> LoMask(4U, -1);
4041 SmallVector<int,8> HiMask(4U, -1);
4043 SmallVector<int,8> *MaskPtr = &LoMask;
4044 unsigned MaskIdx = 0;
4047 for (unsigned i = 0; i != 4; ++i) {
4054 int Idx = PermMask[i];
4056 Locs[i] = std::make_pair(-1, -1);
4057 } else if (Idx < 4) {
4058 Locs[i] = std::make_pair(MaskIdx, LoIdx);
4059 (*MaskPtr)[LoIdx] = Idx;
4062 Locs[i] = std::make_pair(MaskIdx, HiIdx);
4063 (*MaskPtr)[HiIdx] = Idx;
4068 SDValue LoShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &LoMask[0]);
4069 SDValue HiShuffle = DAG.getVectorShuffle(VT, dl, V1, V2, &HiMask[0]);
4070 SmallVector<int, 8> MaskOps;
4071 for (unsigned i = 0; i != 4; ++i) {
4072 if (Locs[i].first == -1) {
4073 MaskOps.push_back(-1);
4075 unsigned Idx = Locs[i].first * 4 + Locs[i].second;
4076 MaskOps.push_back(Idx);
4079 return DAG.getVectorShuffle(VT, dl, LoShuffle, HiShuffle, &MaskOps[0]);
4083 X86TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
4084 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
4085 SDValue V1 = Op.getOperand(0);
4086 SDValue V2 = Op.getOperand(1);
4087 EVT VT = Op.getValueType();
4088 DebugLoc dl = Op.getDebugLoc();
4089 unsigned NumElems = VT.getVectorNumElements();
4090 bool isMMX = VT.getSizeInBits() == 64;
4091 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
4092 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
4093 bool V1IsSplat = false;
4094 bool V2IsSplat = false;
4096 if (isZeroShuffle(SVOp))
4097 return getZeroVector(VT, Subtarget->hasSSE2(), DAG, dl);
4099 // Promote splats to v4f32.
4100 if (SVOp->isSplat()) {
4101 if (isMMX || NumElems < 4)
4103 return PromoteSplat(SVOp, DAG, Subtarget->hasSSE2());
4106 // If the shuffle can be profitably rewritten as a narrower shuffle, then
4108 if (VT == MVT::v8i16 || VT == MVT::v16i8) {
4109 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4110 if (NewOp.getNode())
4111 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
4112 LowerVECTOR_SHUFFLE(NewOp, DAG));
4113 } else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
4114 // FIXME: Figure out a cleaner way to do this.
4115 // Try to make use of movq to zero out the top part.
4116 if (ISD::isBuildVectorAllZeros(V2.getNode())) {
4117 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4118 if (NewOp.getNode()) {
4119 if (isCommutedMOVL(cast<ShuffleVectorSDNode>(NewOp), true, false))
4120 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(0),
4121 DAG, Subtarget, dl);
4123 } else if (ISD::isBuildVectorAllZeros(V1.getNode())) {
4124 SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, *this, dl);
4125 if (NewOp.getNode() && X86::isMOVLMask(cast<ShuffleVectorSDNode>(NewOp)))
4126 return getVZextMovL(VT, NewOp.getValueType(), NewOp.getOperand(1),
4127 DAG, Subtarget, dl);
4131 if (X86::isPSHUFDMask(SVOp))
4134 // Check if this can be converted into a logical shift.
4135 bool isLeft = false;
4138 bool isShift = getSubtarget()->hasSSE2() &&
4139 isVectorShift(SVOp, DAG, isLeft, ShVal, ShAmt);
4140 if (isShift && ShVal.hasOneUse()) {
4141 // If the shifted value has multiple uses, it may be cheaper to use
4142 // v_set0 + movlhps or movhlps, etc.
4143 EVT EVT = VT.getVectorElementType();
4144 ShAmt *= EVT.getSizeInBits();
4145 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4148 if (X86::isMOVLMask(SVOp)) {
4151 if (ISD::isBuildVectorAllZeros(V1.getNode()))
4152 return getVZextMovL(VT, VT, V2, DAG, Subtarget, dl);
4157 // FIXME: fold these into legal mask.
4158 if (!isMMX && (X86::isMOVSHDUPMask(SVOp) ||
4159 X86::isMOVSLDUPMask(SVOp) ||
4160 X86::isMOVHLPSMask(SVOp) ||
4161 X86::isMOVHPMask(SVOp) ||
4162 X86::isMOVLPMask(SVOp)))
4165 if (ShouldXformToMOVHLPS(SVOp) ||
4166 ShouldXformToMOVLP(V1.getNode(), V2.getNode(), SVOp))
4167 return CommuteVectorShuffle(SVOp, DAG);
4170 // No better options. Use a vshl / vsrl.
4171 EVT EVT = VT.getVectorElementType();
4172 ShAmt *= EVT.getSizeInBits();
4173 return getVShift(isLeft, VT, ShVal, ShAmt, DAG, *this, dl);
4176 bool Commuted = false;
4177 // FIXME: This should also accept a bitcast of a splat? Be careful, not
4178 // 1,1,1,1 -> v8i16 though.
4179 V1IsSplat = isSplatVector(V1.getNode());
4180 V2IsSplat = isSplatVector(V2.getNode());
4182 // Canonicalize the splat or undef, if present, to be on the RHS.
4183 if ((V1IsSplat || V1IsUndef) && !(V2IsSplat || V2IsUndef)) {
4184 Op = CommuteVectorShuffle(SVOp, DAG);
4185 SVOp = cast<ShuffleVectorSDNode>(Op);
4186 V1 = SVOp->getOperand(0);
4187 V2 = SVOp->getOperand(1);
4188 std::swap(V1IsSplat, V2IsSplat);
4189 std::swap(V1IsUndef, V2IsUndef);
4193 if (isCommutedMOVL(SVOp, V2IsSplat, V2IsUndef)) {
4194 // Shuffling low element of v1 into undef, just return v1.
4197 // If V2 is a splat, the mask may be malformed such as <4,3,3,3>, which
4198 // the instruction selector will not match, so get a canonical MOVL with
4199 // swapped operands to undo the commute.
4200 return getMOVL(DAG, dl, VT, V2, V1);
4203 if (X86::isUNPCKL_v_undef_Mask(SVOp) ||
4204 X86::isUNPCKH_v_undef_Mask(SVOp) ||
4205 X86::isUNPCKLMask(SVOp) ||
4206 X86::isUNPCKHMask(SVOp))
4210 // Normalize mask so all entries that point to V2 points to its first
4211 // element then try to match unpck{h|l} again. If match, return a
4212 // new vector_shuffle with the corrected mask.
4213 SDValue NewMask = NormalizeMask(SVOp, DAG);
4214 ShuffleVectorSDNode *NSVOp = cast<ShuffleVectorSDNode>(NewMask);
4215 if (NSVOp != SVOp) {
4216 if (X86::isUNPCKLMask(NSVOp, true)) {
4218 } else if (X86::isUNPCKHMask(NSVOp, true)) {
4225 // Commute is back and try unpck* again.
4226 // FIXME: this seems wrong.
4227 SDValue NewOp = CommuteVectorShuffle(SVOp, DAG);
4228 ShuffleVectorSDNode *NewSVOp = cast<ShuffleVectorSDNode>(NewOp);
4229 if (X86::isUNPCKL_v_undef_Mask(NewSVOp) ||
4230 X86::isUNPCKH_v_undef_Mask(NewSVOp) ||
4231 X86::isUNPCKLMask(NewSVOp) ||
4232 X86::isUNPCKHMask(NewSVOp))
4236 // FIXME: for mmx, bitcast v2i32 to v4i16 for shuffle.
4238 // Normalize the node to match x86 shuffle ops if needed
4239 if (!isMMX && V2.getOpcode() != ISD::UNDEF && isCommutedSHUFP(SVOp))
4240 return CommuteVectorShuffle(SVOp, DAG);
4242 // Check for legal shuffle and return?
4243 SmallVector<int, 16> PermMask;
4244 SVOp->getMask(PermMask);
4245 if (isShuffleMaskLegal(PermMask, VT))
4248 // Handle v8i16 specifically since SSE can do byte extraction and insertion.
4249 if (VT == MVT::v8i16) {
4250 SDValue NewOp = LowerVECTOR_SHUFFLEv8i16(SVOp, DAG, *this);
4251 if (NewOp.getNode())
4255 if (VT == MVT::v16i8) {
4256 SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, DAG, *this);
4257 if (NewOp.getNode())
4261 // Handle all 4 wide cases with a number of shuffles except for MMX.
4262 if (NumElems == 4 && !isMMX)
4263 return LowerVECTOR_SHUFFLE_4wide(SVOp, DAG);
4269 X86TargetLowering::LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op,
4270 SelectionDAG &DAG) {
4271 EVT VT = Op.getValueType();
4272 DebugLoc dl = Op.getDebugLoc();
4273 if (VT.getSizeInBits() == 8) {
4274 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
4275 Op.getOperand(0), Op.getOperand(1));
4276 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4277 DAG.getValueType(VT));
4278 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4279 } else if (VT.getSizeInBits() == 16) {
4280 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4281 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
4283 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4284 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4285 DAG.getNode(ISD::BIT_CONVERT, dl,
4289 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
4290 Op.getOperand(0), Op.getOperand(1));
4291 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
4292 DAG.getValueType(VT));
4293 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4294 } else if (VT == MVT::f32) {
4295 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
4296 // the result back to FR32 register. It's only worth matching if the
4297 // result has a single use which is a store or a bitcast to i32. And in
4298 // the case of a store, it's not worth it if the index is a constant 0,
4299 // because a MOVSSmr can be used instead, which is smaller and faster.
4300 if (!Op.hasOneUse())
4302 SDNode *User = *Op.getNode()->use_begin();
4303 if ((User->getOpcode() != ISD::STORE ||
4304 (isa<ConstantSDNode>(Op.getOperand(1)) &&
4305 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
4306 (User->getOpcode() != ISD::BIT_CONVERT ||
4307 User->getValueType(0) != MVT::i32))
4309 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4310 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
4313 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
4314 } else if (VT == MVT::i32) {
4315 // ExtractPS works with constant index.
4316 if (isa<ConstantSDNode>(Op.getOperand(1)))
4324 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4325 if (!isa<ConstantSDNode>(Op.getOperand(1)))
4328 if (Subtarget->hasSSE41()) {
4329 SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
4334 EVT VT = Op.getValueType();
4335 DebugLoc dl = Op.getDebugLoc();
4336 // TODO: handle v16i8.
4337 if (VT.getSizeInBits() == 16) {
4338 SDValue Vec = Op.getOperand(0);
4339 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4341 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
4342 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
4343 DAG.getNode(ISD::BIT_CONVERT, dl,
4346 // Transform it so it match pextrw which produces a 32-bit result.
4347 EVT EVT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy+1);
4348 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EVT,
4349 Op.getOperand(0), Op.getOperand(1));
4350 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EVT, Extract,
4351 DAG.getValueType(VT));
4352 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
4353 } else if (VT.getSizeInBits() == 32) {
4354 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4358 // SHUFPS the element to the lowest double word, then movss.
4359 int Mask[4] = { Idx, -1, -1, -1 };
4360 EVT VVT = Op.getOperand(0).getValueType();
4361 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4362 DAG.getUNDEF(VVT), Mask);
4363 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4364 DAG.getIntPtrConstant(0));
4365 } else if (VT.getSizeInBits() == 64) {
4366 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
4367 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
4368 // to match extract_elt for f64.
4369 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
4373 // UNPCKHPD the element to the lowest double word, then movsd.
4374 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
4375 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
4376 int Mask[2] = { 1, -1 };
4377 EVT VVT = Op.getOperand(0).getValueType();
4378 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
4379 DAG.getUNDEF(VVT), Mask);
4380 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
4381 DAG.getIntPtrConstant(0));
4388 X86TargetLowering::LowerINSERT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG){
4389 EVT VT = Op.getValueType();
4390 EVT EVT = VT.getVectorElementType();
4391 DebugLoc dl = Op.getDebugLoc();
4393 SDValue N0 = Op.getOperand(0);
4394 SDValue N1 = Op.getOperand(1);
4395 SDValue N2 = Op.getOperand(2);
4397 if ((EVT.getSizeInBits() == 8 || EVT.getSizeInBits() == 16) &&
4398 isa<ConstantSDNode>(N2)) {
4399 unsigned Opc = (EVT.getSizeInBits() == 8) ? X86ISD::PINSRB
4401 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
4403 if (N1.getValueType() != MVT::i32)
4404 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4405 if (N2.getValueType() != MVT::i32)
4406 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4407 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
4408 } else if (EVT == MVT::f32 && isa<ConstantSDNode>(N2)) {
4409 // Bits [7:6] of the constant are the source select. This will always be
4410 // zero here. The DAG Combiner may combine an extract_elt index into these
4411 // bits. For example (insert (extract, 3), 2) could be matched by putting
4412 // the '3' into bits [7:6] of X86ISD::INSERTPS.
4413 // Bits [5:4] of the constant are the destination select. This is the
4414 // value of the incoming immediate.
4415 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
4416 // combine either bitwise AND or insert of float 0.0 to set these bits.
4417 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue() << 4);
4418 // Create this as a scalar to vector..
4419 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
4420 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
4421 } else if (EVT == MVT::i32 && isa<ConstantSDNode>(N2)) {
4422 // PINSR* works with constant index.
4429 X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
4430 EVT VT = Op.getValueType();
4431 EVT EVT = VT.getVectorElementType();
4433 if (Subtarget->hasSSE41())
4434 return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
4439 DebugLoc dl = Op.getDebugLoc();
4440 SDValue N0 = Op.getOperand(0);
4441 SDValue N1 = Op.getOperand(1);
4442 SDValue N2 = Op.getOperand(2);
4444 if (EVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
4445 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
4446 // as its second argument.
4447 if (N1.getValueType() != MVT::i32)
4448 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
4449 if (N2.getValueType() != MVT::i32)
4450 N2 = DAG.getIntPtrConstant(cast<ConstantSDNode>(N2)->getZExtValue());
4451 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
4457 X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
4458 DebugLoc dl = Op.getDebugLoc();
4459 if (Op.getValueType() == MVT::v2f32)
4460 return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f32,
4461 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i32,
4462 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32,
4463 Op.getOperand(0))));
4465 if (Op.getValueType() == MVT::v1i64 && Op.getOperand(0).getValueType() == MVT::i64)
4466 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
4468 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
4469 EVT VT = MVT::v2i32;
4470 switch (Op.getValueType().getSimpleVT().SimpleTy) {
4477 return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
4478 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, AnyExt));
4481 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
4482 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
4483 // one of the above mentioned nodes. It has to be wrapped because otherwise
4484 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
4485 // be used to form addressing mode. These wrapped nodes will be selected
4488 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
4489 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
4491 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4493 unsigned char OpFlag = 0;
4494 unsigned WrapperKind = X86ISD::Wrapper;
4495 CodeModel::Model M = getTargetMachine().getCodeModel();
4497 if (Subtarget->isPICStyleRIPRel() &&
4498 (M == CodeModel::Small || M == CodeModel::Kernel))
4499 WrapperKind = X86ISD::WrapperRIP;
4500 else if (Subtarget->isPICStyleGOT())
4501 OpFlag = X86II::MO_GOTOFF;
4502 else if (Subtarget->isPICStyleStubPIC())
4503 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4505 SDValue Result = DAG.getTargetConstantPool(CP->getConstVal(), getPointerTy(),
4507 CP->getOffset(), OpFlag);
4508 DebugLoc DL = CP->getDebugLoc();
4509 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4510 // With PIC, the address is actually $g + Offset.
4512 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4513 DAG.getNode(X86ISD::GlobalBaseReg,
4514 DebugLoc::getUnknownLoc(), getPointerTy()),
4521 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) {
4522 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
4524 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4526 unsigned char OpFlag = 0;
4527 unsigned WrapperKind = X86ISD::Wrapper;
4528 CodeModel::Model M = getTargetMachine().getCodeModel();
4530 if (Subtarget->isPICStyleRIPRel() &&
4531 (M == CodeModel::Small || M == CodeModel::Kernel))
4532 WrapperKind = X86ISD::WrapperRIP;
4533 else if (Subtarget->isPICStyleGOT())
4534 OpFlag = X86II::MO_GOTOFF;
4535 else if (Subtarget->isPICStyleStubPIC())
4536 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4538 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), getPointerTy(),
4540 DebugLoc DL = JT->getDebugLoc();
4541 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4543 // With PIC, the address is actually $g + Offset.
4545 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4546 DAG.getNode(X86ISD::GlobalBaseReg,
4547 DebugLoc::getUnknownLoc(), getPointerTy()),
4555 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) {
4556 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
4558 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
4560 unsigned char OpFlag = 0;
4561 unsigned WrapperKind = X86ISD::Wrapper;
4562 CodeModel::Model M = getTargetMachine().getCodeModel();
4564 if (Subtarget->isPICStyleRIPRel() &&
4565 (M == CodeModel::Small || M == CodeModel::Kernel))
4566 WrapperKind = X86ISD::WrapperRIP;
4567 else if (Subtarget->isPICStyleGOT())
4568 OpFlag = X86II::MO_GOTOFF;
4569 else if (Subtarget->isPICStyleStubPIC())
4570 OpFlag = X86II::MO_PIC_BASE_OFFSET;
4572 SDValue Result = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlag);
4574 DebugLoc DL = Op.getDebugLoc();
4575 Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
4578 // With PIC, the address is actually $g + Offset.
4579 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
4580 !Subtarget->is64Bit()) {
4581 Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
4582 DAG.getNode(X86ISD::GlobalBaseReg,
4583 DebugLoc::getUnknownLoc(),
4592 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, DebugLoc dl,
4594 SelectionDAG &DAG) const {
4595 // Create the TargetGlobalAddress node, folding in the constant
4596 // offset if it is legal.
4597 unsigned char OpFlags =
4598 Subtarget->ClassifyGlobalReference(GV, getTargetMachine());
4599 CodeModel::Model M = getTargetMachine().getCodeModel();
4601 if (OpFlags == X86II::MO_NO_FLAG &&
4602 X86::isOffsetSuitableForCodeModel(Offset, M)) {
4603 // A direct static reference to a global.
4604 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), Offset);
4607 Result = DAG.getTargetGlobalAddress(GV, getPointerTy(), 0, OpFlags);
4610 if (Subtarget->isPICStyleRIPRel() &&
4611 (M == CodeModel::Small || M == CodeModel::Kernel))
4612 Result = DAG.getNode(X86ISD::WrapperRIP, dl, getPointerTy(), Result);
4614 Result = DAG.getNode(X86ISD::Wrapper, dl, getPointerTy(), Result);
4616 // With PIC, the address is actually $g + Offset.
4617 if (isGlobalRelativeToPICBase(OpFlags)) {
4618 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(),
4619 DAG.getNode(X86ISD::GlobalBaseReg, dl, getPointerTy()),
4623 // For globals that require a load from a stub to get the address, emit the
4625 if (isGlobalStubReference(OpFlags))
4626 Result = DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(), Result,
4627 PseudoSourceValue::getGOT(), 0);
4629 // If there was a non-zero offset that we didn't fold, create an explicit
4632 Result = DAG.getNode(ISD::ADD, dl, getPointerTy(), Result,
4633 DAG.getConstant(Offset, getPointerTy()));
4639 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) {
4640 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
4641 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
4642 return LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
4646 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
4647 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
4648 unsigned char OperandFlags) {
4649 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
4650 DebugLoc dl = GA->getDebugLoc();
4651 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(),
4652 GA->getValueType(0),
4656 SDValue Ops[] = { Chain, TGA, *InFlag };
4657 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 3);
4659 SDValue Ops[] = { Chain, TGA };
4660 Chain = DAG.getNode(X86ISD::TLSADDR, dl, NodeTys, Ops, 2);
4662 SDValue Flag = Chain.getValue(1);
4663 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
4666 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
4668 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4671 DebugLoc dl = GA->getDebugLoc(); // ? function entry point might be better
4672 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
4673 DAG.getNode(X86ISD::GlobalBaseReg,
4674 DebugLoc::getUnknownLoc(),
4676 InFlag = Chain.getValue(1);
4678 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
4681 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
4683 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4685 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
4686 X86::RAX, X86II::MO_TLSGD);
4689 // Lower ISD::GlobalTLSAddress using the "initial exec" (for no-pic) or
4690 // "local exec" model.
4691 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
4692 const EVT PtrVT, TLSModel::Model model,
4694 DebugLoc dl = GA->getDebugLoc();
4695 // Get the Thread Pointer
4696 SDValue Base = DAG.getNode(X86ISD::SegmentBaseAddress,
4697 DebugLoc::getUnknownLoc(), PtrVT,
4698 DAG.getRegister(is64Bit? X86::FS : X86::GS,
4701 SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Base,
4704 unsigned char OperandFlags = 0;
4705 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
4707 unsigned WrapperKind = X86ISD::Wrapper;
4708 if (model == TLSModel::LocalExec) {
4709 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
4710 } else if (is64Bit) {
4711 assert(model == TLSModel::InitialExec);
4712 OperandFlags = X86II::MO_GOTTPOFF;
4713 WrapperKind = X86ISD::WrapperRIP;
4715 assert(model == TLSModel::InitialExec);
4716 OperandFlags = X86II::MO_INDNTPOFF;
4719 // emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
4721 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), GA->getValueType(0),
4722 GA->getOffset(), OperandFlags);
4723 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
4725 if (model == TLSModel::InitialExec)
4726 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
4727 PseudoSourceValue::getGOT(), 0);
4729 // The address of the thread local variable is the add of the thread
4730 // pointer with the offset of the variable.
4731 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
4735 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) {
4736 // TODO: implement the "local dynamic" model
4737 // TODO: implement the "initial exec"model for pic executables
4738 assert(Subtarget->isTargetELF() &&
4739 "TLS not implemented for non-ELF targets");
4740 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
4741 const GlobalValue *GV = GA->getGlobal();
4743 // If GV is an alias then use the aliasee for determining
4744 // thread-localness.
4745 if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
4746 GV = GA->resolveAliasedGlobal(false);
4748 TLSModel::Model model = getTLSModel(GV,
4749 getTargetMachine().getRelocationModel());
4752 case TLSModel::GeneralDynamic:
4753 case TLSModel::LocalDynamic: // not implemented
4754 if (Subtarget->is64Bit())
4755 return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
4756 return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
4758 case TLSModel::InitialExec:
4759 case TLSModel::LocalExec:
4760 return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
4761 Subtarget->is64Bit());
4764 llvm_unreachable("Unreachable");
4769 /// LowerShift - Lower SRA_PARTS and friends, which return two i32 values and
4770 /// take a 2 x i32 value to shift plus a shift amount.
4771 SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) {
4772 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
4773 EVT VT = Op.getValueType();
4774 unsigned VTBits = VT.getSizeInBits();
4775 DebugLoc dl = Op.getDebugLoc();
4776 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
4777 SDValue ShOpLo = Op.getOperand(0);
4778 SDValue ShOpHi = Op.getOperand(1);
4779 SDValue ShAmt = Op.getOperand(2);
4780 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
4781 DAG.getConstant(VTBits - 1, MVT::i8))
4782 : DAG.getConstant(0, VT);
4785 if (Op.getOpcode() == ISD::SHL_PARTS) {
4786 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
4787 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
4789 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
4790 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, ShAmt);
4793 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
4794 DAG.getConstant(VTBits, MVT::i8));
4795 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, VT,
4796 AndNode, DAG.getConstant(0, MVT::i8));
4799 SDValue CC = DAG.getConstant(X86::COND_NE, MVT::i8);
4800 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
4801 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
4803 if (Op.getOpcode() == ISD::SHL_PARTS) {
4804 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4805 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4807 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0, 4);
4808 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1, 4);
4811 SDValue Ops[2] = { Lo, Hi };
4812 return DAG.getMergeValues(Ops, 2, dl);
4815 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
4816 EVT SrcVT = Op.getOperand(0).getValueType();
4818 if (SrcVT.isVector()) {
4819 if (SrcVT == MVT::v2i32 && Op.getValueType() == MVT::v2f64) {
4825 assert(SrcVT.getSimpleVT() <= MVT::i64 && SrcVT.getSimpleVT() >= MVT::i16 &&
4826 "Unknown SINT_TO_FP to lower!");
4828 // These are really Legal; return the operand so the caller accepts it as
4830 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
4832 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
4833 Subtarget->is64Bit()) {
4837 DebugLoc dl = Op.getDebugLoc();
4838 unsigned Size = SrcVT.getSizeInBits()/8;
4839 MachineFunction &MF = DAG.getMachineFunction();
4840 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size);
4841 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4842 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
4844 PseudoSourceValue::getFixedStack(SSFI), 0);
4845 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
4848 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
4850 SelectionDAG &DAG) {
4852 DebugLoc dl = Op.getDebugLoc();
4854 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
4856 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
4858 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
4859 SmallVector<SDValue, 8> Ops;
4860 Ops.push_back(Chain);
4861 Ops.push_back(StackSlot);
4862 Ops.push_back(DAG.getValueType(SrcVT));
4863 SDValue Result = DAG.getNode(useSSE ? X86ISD::FILD_FLAG : X86ISD::FILD, dl,
4864 Tys, &Ops[0], Ops.size());
4867 Chain = Result.getValue(1);
4868 SDValue InFlag = Result.getValue(2);
4870 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
4871 // shouldn't be necessary except that RFP cannot be live across
4872 // multiple blocks. When stackifier is fixed, they can be uncoupled.
4873 MachineFunction &MF = DAG.getMachineFunction();
4874 int SSFI = MF.getFrameInfo()->CreateStackObject(8, 8);
4875 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
4876 Tys = DAG.getVTList(MVT::Other);
4877 SmallVector<SDValue, 8> Ops;
4878 Ops.push_back(Chain);
4879 Ops.push_back(Result);
4880 Ops.push_back(StackSlot);
4881 Ops.push_back(DAG.getValueType(Op.getValueType()));
4882 Ops.push_back(InFlag);
4883 Chain = DAG.getNode(X86ISD::FST, dl, Tys, &Ops[0], Ops.size());
4884 Result = DAG.getLoad(Op.getValueType(), dl, Chain, StackSlot,
4885 PseudoSourceValue::getFixedStack(SSFI), 0);
4891 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
4892 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op, SelectionDAG &DAG) {
4893 // This algorithm is not obvious. Here it is in C code, more or less:
4895 double uint64_to_double( uint32_t hi, uint32_t lo ) {
4896 static const __m128i exp = { 0x4330000045300000ULL, 0 };
4897 static const __m128d bias = { 0x1.0p84, 0x1.0p52 };
4899 // Copy ints to xmm registers.
4900 __m128i xh = _mm_cvtsi32_si128( hi );
4901 __m128i xl = _mm_cvtsi32_si128( lo );
4903 // Combine into low half of a single xmm register.
4904 __m128i x = _mm_unpacklo_epi32( xh, xl );
4908 // Merge in appropriate exponents to give the integer bits the right
4910 x = _mm_unpacklo_epi32( x, exp );
4912 // Subtract away the biases to deal with the IEEE-754 double precision
4914 d = _mm_sub_pd( (__m128d) x, bias );
4916 // All conversions up to here are exact. The correctly rounded result is
4917 // calculated using the current rounding mode using the following
4919 d = _mm_add_sd( d, _mm_unpackhi_pd( d, d ) );
4920 _mm_store_sd( &sd, d ); // Because we are returning doubles in XMM, this
4921 // store doesn't really need to be here (except
4922 // maybe to zero the other double)
4927 DebugLoc dl = Op.getDebugLoc();
4928 LLVMContext *Context = DAG.getContext();
4930 // Build some magic constants.
4931 std::vector<Constant*> CV0;
4932 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x45300000)));
4933 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0x43300000)));
4934 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
4935 CV0.push_back(ConstantInt::get(*Context, APInt(32, 0)));
4936 Constant *C0 = ConstantVector::get(CV0);
4937 SDValue CPIdx0 = DAG.getConstantPool(C0, getPointerTy(), 16);
4939 std::vector<Constant*> CV1;
4941 ConstantFP::get(*Context, APFloat(APInt(64, 0x4530000000000000ULL))));
4943 ConstantFP::get(*Context, APFloat(APInt(64, 0x4330000000000000ULL))));
4944 Constant *C1 = ConstantVector::get(CV1);
4945 SDValue CPIdx1 = DAG.getConstantPool(C1, getPointerTy(), 16);
4947 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4948 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4950 DAG.getIntPtrConstant(1)));
4951 SDValue XR2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4952 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4954 DAG.getIntPtrConstant(0)));
4955 SDValue Unpck1 = getUnpackl(DAG, dl, MVT::v4i32, XR1, XR2);
4956 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
4957 PseudoSourceValue::getConstantPool(), 0,
4959 SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
4960 SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
4961 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
4962 PseudoSourceValue::getConstantPool(), 0,
4964 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
4966 // Add the halves; easiest way is to swap them into another reg first.
4967 int ShufMask[2] = { 1, -1 };
4968 SDValue Shuf = DAG.getVectorShuffle(MVT::v2f64, dl, Sub,
4969 DAG.getUNDEF(MVT::v2f64), ShufMask);
4970 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::v2f64, Shuf, Sub);
4971 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Add,
4972 DAG.getIntPtrConstant(0));
4975 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
4976 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op, SelectionDAG &DAG) {
4977 DebugLoc dl = Op.getDebugLoc();
4978 // FP constant to bias correct the final result.
4979 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL),
4982 // Load the 32-bit value into an XMM register.
4983 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
4984 DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
4986 DAG.getIntPtrConstant(0)));
4988 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
4989 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
4990 DAG.getIntPtrConstant(0));
4992 // Or the load with the bias.
4993 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
4994 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4995 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4997 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
4998 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
4999 MVT::v2f64, Bias)));
5000 Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
5001 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
5002 DAG.getIntPtrConstant(0));
5004 // Subtract the bias.
5005 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
5007 // Handle final rounding.
5008 EVT DestVT = Op.getValueType();
5010 if (DestVT.bitsLT(MVT::f64)) {
5011 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
5012 DAG.getIntPtrConstant(0));
5013 } else if (DestVT.bitsGT(MVT::f64)) {
5014 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
5017 // Handle final rounding.
5021 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
5022 SDValue N0 = Op.getOperand(0);
5023 DebugLoc dl = Op.getDebugLoc();
5025 // Now not UINT_TO_FP is legal (it's marked custom), dag combiner won't
5026 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
5027 // the optimization here.
5028 if (DAG.SignBitIsZero(N0))
5029 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
5031 EVT SrcVT = N0.getValueType();
5032 if (SrcVT == MVT::i64) {
5033 // We only handle SSE2 f64 target here; caller can expand the rest.
5034 if (Op.getValueType() != MVT::f64 || !X86ScalarSSEf64)
5037 return LowerUINT_TO_FP_i64(Op, DAG);
5038 } else if (SrcVT == MVT::i32 && X86ScalarSSEf64) {
5039 return LowerUINT_TO_FP_i32(Op, DAG);
5042 assert(SrcVT == MVT::i32 && "Unknown UINT_TO_FP to lower!");
5044 // Make a 64-bit buffer, and use it to build an FILD.
5045 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
5046 SDValue WordOff = DAG.getConstant(4, getPointerTy());
5047 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl,
5048 getPointerTy(), StackSlot, WordOff);
5049 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
5050 StackSlot, NULL, 0);
5051 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, MVT::i32),
5052 OffsetSlot, NULL, 0);
5053 return BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
5056 std::pair<SDValue,SDValue> X86TargetLowering::
5057 FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG, bool IsSigned) {
5058 DebugLoc dl = Op.getDebugLoc();
5060 EVT DstTy = Op.getValueType();
5063 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
5067 assert(DstTy.getSimpleVT() <= MVT::i64 &&
5068 DstTy.getSimpleVT() >= MVT::i16 &&
5069 "Unknown FP_TO_SINT to lower!");
5071 // These are really Legal.
5072 if (DstTy == MVT::i32 &&
5073 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5074 return std::make_pair(SDValue(), SDValue());
5075 if (Subtarget->is64Bit() &&
5076 DstTy == MVT::i64 &&
5077 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
5078 return std::make_pair(SDValue(), SDValue());
5080 // We lower FP->sint64 into FISTP64, followed by a load, all to a temporary
5082 MachineFunction &MF = DAG.getMachineFunction();
5083 unsigned MemSize = DstTy.getSizeInBits()/8;
5084 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5085 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5088 switch (DstTy.getSimpleVT().SimpleTy) {
5089 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
5090 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
5091 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
5092 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
5095 SDValue Chain = DAG.getEntryNode();
5096 SDValue Value = Op.getOperand(0);
5097 if (isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType())) {
5098 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
5099 Chain = DAG.getStore(Chain, dl, Value, StackSlot,
5100 PseudoSourceValue::getFixedStack(SSFI), 0);
5101 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
5103 Chain, StackSlot, DAG.getValueType(Op.getOperand(0).getValueType())
5105 Value = DAG.getNode(X86ISD::FLD, dl, Tys, Ops, 3);
5106 Chain = Value.getValue(1);
5107 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize);
5108 StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
5111 // Build the FP_TO_INT*_IN_MEM
5112 SDValue Ops[] = { Chain, Value, StackSlot };
5113 SDValue FIST = DAG.getNode(Opc, dl, MVT::Other, Ops, 3);
5115 return std::make_pair(FIST, StackSlot);
5118 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op, SelectionDAG &DAG) {
5119 if (Op.getValueType().isVector()) {
5120 if (Op.getValueType() == MVT::v2i32 &&
5121 Op.getOperand(0).getValueType() == MVT::v2f64) {
5127 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, true);
5128 SDValue FIST = Vals.first, StackSlot = Vals.second;
5129 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
5130 if (FIST.getNode() == 0) return Op;
5133 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5134 FIST, StackSlot, NULL, 0);
5137 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op, SelectionDAG &DAG) {
5138 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG, false);
5139 SDValue FIST = Vals.first, StackSlot = Vals.second;
5140 assert(FIST.getNode() && "Unexpected failure");
5143 return DAG.getLoad(Op.getValueType(), Op.getDebugLoc(),
5144 FIST, StackSlot, NULL, 0);
5147 SDValue X86TargetLowering::LowerFABS(SDValue Op, SelectionDAG &DAG) {
5148 LLVMContext *Context = DAG.getContext();
5149 DebugLoc dl = Op.getDebugLoc();
5150 EVT VT = Op.getValueType();
5153 EltVT = VT.getVectorElementType();
5154 std::vector<Constant*> CV;
5155 if (EltVT == MVT::f64) {
5156 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63))));
5160 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31))));
5166 Constant *C = ConstantVector::get(CV);
5167 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5168 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5169 PseudoSourceValue::getConstantPool(), 0,
5171 return DAG.getNode(X86ISD::FAND, dl, VT, Op.getOperand(0), Mask);
5174 SDValue X86TargetLowering::LowerFNEG(SDValue Op, SelectionDAG &DAG) {
5175 LLVMContext *Context = DAG.getContext();
5176 DebugLoc dl = Op.getDebugLoc();
5177 EVT VT = Op.getValueType();
5179 unsigned EltNum = 1;
5180 if (VT.isVector()) {
5181 EltVT = VT.getVectorElementType();
5182 EltNum = VT.getVectorNumElements();
5184 std::vector<Constant*> CV;
5185 if (EltVT == MVT::f64) {
5186 Constant *C = ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63)));
5190 Constant *C = ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31)));
5196 Constant *C = ConstantVector::get(CV);
5197 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5198 SDValue Mask = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5199 PseudoSourceValue::getConstantPool(), 0,
5201 if (VT.isVector()) {
5202 return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
5203 DAG.getNode(ISD::XOR, dl, MVT::v2i64,
5204 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
5206 DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
5208 return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
5212 SDValue X86TargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
5213 LLVMContext *Context = DAG.getContext();
5214 SDValue Op0 = Op.getOperand(0);
5215 SDValue Op1 = Op.getOperand(1);
5216 DebugLoc dl = Op.getDebugLoc();
5217 EVT VT = Op.getValueType();
5218 EVT SrcVT = Op1.getValueType();
5220 // If second operand is smaller, extend it first.
5221 if (SrcVT.bitsLT(VT)) {
5222 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
5225 // And if it is bigger, shrink it first.
5226 if (SrcVT.bitsGT(VT)) {
5227 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1));
5231 // At this point the operands and the result should have the same
5232 // type, and that won't be f80 since that is not custom lowered.
5234 // First get the sign bit of second operand.
5235 std::vector<Constant*> CV;
5236 if (SrcVT == MVT::f64) {
5237 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 1ULL << 63))));
5238 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5240 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 1U << 31))));
5241 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5242 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5243 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5245 Constant *C = ConstantVector::get(CV);
5246 SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5247 SDValue Mask1 = DAG.getLoad(SrcVT, dl, DAG.getEntryNode(), CPIdx,
5248 PseudoSourceValue::getConstantPool(), 0,
5250 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, SrcVT, Op1, Mask1);
5252 // Shift sign bit right or left if the two operands have different types.
5253 if (SrcVT.bitsGT(VT)) {
5254 // Op0 is MVT::f32, Op1 is MVT::f64.
5255 SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
5256 SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
5257 DAG.getConstant(32, MVT::i32));
5258 SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
5259 SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
5260 DAG.getIntPtrConstant(0));
5263 // Clear first operand sign bit.
5265 if (VT == MVT::f64) {
5266 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, ~(1ULL << 63)))));
5267 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(64, 0))));
5269 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, ~(1U << 31)))));
5270 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5271 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5272 CV.push_back(ConstantFP::get(*Context, APFloat(APInt(32, 0))));
5274 C = ConstantVector::get(CV);
5275 CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
5276 SDValue Mask2 = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
5277 PseudoSourceValue::getConstantPool(), 0,
5279 SDValue Val = DAG.getNode(X86ISD::FAND, dl, VT, Op0, Mask2);
5281 // Or the value with the sign bit.
5282 return DAG.getNode(X86ISD::FOR, dl, VT, Val, SignBit);
5285 /// Emit nodes that will be selected as "test Op0,Op0", or something
5287 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
5288 SelectionDAG &DAG) {
5289 DebugLoc dl = Op.getDebugLoc();
5291 // CF and OF aren't always set the way we want. Determine which
5292 // of these we need.
5293 bool NeedCF = false;
5294 bool NeedOF = false;
5296 case X86::COND_A: case X86::COND_AE:
5297 case X86::COND_B: case X86::COND_BE:
5300 case X86::COND_G: case X86::COND_GE:
5301 case X86::COND_L: case X86::COND_LE:
5302 case X86::COND_O: case X86::COND_NO:
5308 // See if we can use the EFLAGS value from the operand instead of
5309 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
5310 // we prove that the arithmetic won't overflow, we can't use OF or CF.
5311 if (Op.getResNo() == 0 && !NeedOF && !NeedCF) {
5312 unsigned Opcode = 0;
5313 unsigned NumOperands = 0;
5314 switch (Op.getNode()->getOpcode()) {
5316 // Due to an isel shortcoming, be conservative if this add is likely to
5317 // be selected as part of a load-modify-store instruction. When the root
5318 // node in a match is a store, isel doesn't know how to remap non-chain
5319 // non-flag uses of other nodes in the match, such as the ADD in this
5320 // case. This leads to the ADD being left around and reselected, with
5321 // the result being two adds in the output.
5322 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5323 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5324 if (UI->getOpcode() == ISD::STORE)
5326 if (ConstantSDNode *C =
5327 dyn_cast<ConstantSDNode>(Op.getNode()->getOperand(1))) {
5328 // An add of one will be selected as an INC.
5329 if (C->getAPIntValue() == 1) {
5330 Opcode = X86ISD::INC;
5334 // An add of negative one (subtract of one) will be selected as a DEC.
5335 if (C->getAPIntValue().isAllOnesValue()) {
5336 Opcode = X86ISD::DEC;
5341 // Otherwise use a regular EFLAGS-setting add.
5342 Opcode = X86ISD::ADD;
5346 // Due to the ISEL shortcoming noted above, be conservative if this sub is
5347 // likely to be selected as part of a load-modify-store instruction.
5348 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
5349 UE = Op.getNode()->use_end(); UI != UE; ++UI)
5350 if (UI->getOpcode() == ISD::STORE)
5352 // Otherwise use a regular EFLAGS-setting sub.
5353 Opcode = X86ISD::SUB;
5360 return SDValue(Op.getNode(), 1);
5366 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
5367 SmallVector<SDValue, 4> Ops;
5368 for (unsigned i = 0; i != NumOperands; ++i)
5369 Ops.push_back(Op.getOperand(i));
5370 SDValue New = DAG.getNode(Opcode, dl, VTs, &Ops[0], NumOperands);
5371 DAG.ReplaceAllUsesWith(Op, New);
5372 return SDValue(New.getNode(), 1);
5376 // Otherwise just emit a CMP with 0, which is the TEST pattern.
5377 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
5378 DAG.getConstant(0, Op.getValueType()));
5381 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
5383 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
5384 SelectionDAG &DAG) {
5385 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1))
5386 if (C->getAPIntValue() == 0)
5387 return EmitTest(Op0, X86CC, DAG);
5389 DebugLoc dl = Op0.getDebugLoc();
5390 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
5393 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) {
5394 assert(Op.getValueType() == MVT::i8 && "SetCC type must be 8-bit integer");
5395 SDValue Op0 = Op.getOperand(0);
5396 SDValue Op1 = Op.getOperand(1);
5397 DebugLoc dl = Op.getDebugLoc();
5398 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
5400 // Lower (X & (1 << N)) == 0 to BT(X, N).
5401 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
5402 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
5403 if (Op0.getOpcode() == ISD::AND &&
5405 Op1.getOpcode() == ISD::Constant &&
5406 cast<ConstantSDNode>(Op1)->getZExtValue() == 0 &&
5407 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
5409 if (Op0.getOperand(1).getOpcode() == ISD::SHL) {
5410 if (ConstantSDNode *Op010C =
5411 dyn_cast<ConstantSDNode>(Op0.getOperand(1).getOperand(0)))
5412 if (Op010C->getZExtValue() == 1) {
5413 LHS = Op0.getOperand(0);
5414 RHS = Op0.getOperand(1).getOperand(1);
5416 } else if (Op0.getOperand(0).getOpcode() == ISD::SHL) {
5417 if (ConstantSDNode *Op000C =
5418 dyn_cast<ConstantSDNode>(Op0.getOperand(0).getOperand(0)))
5419 if (Op000C->getZExtValue() == 1) {
5420 LHS = Op0.getOperand(1);
5421 RHS = Op0.getOperand(0).getOperand(1);
5423 } else if (Op0.getOperand(1).getOpcode() == ISD::Constant) {
5424 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op0.getOperand(1));
5425 SDValue AndLHS = Op0.getOperand(0);
5426 if (AndRHS->getZExtValue() == 1 && AndLHS.getOpcode() == ISD::SRL) {
5427 LHS = AndLHS.getOperand(0);
5428 RHS = AndLHS.getOperand(1);
5432 if (LHS.getNode()) {
5433 // If LHS is i8, promote it to i16 with any_extend. There is no i8 BT
5434 // instruction. Since the shift amount is in-range-or-undefined, we know
5435 // that doing a bittest on the i16 value is ok. We extend to i32 because
5436 // the encoding for the i16 version is larger than the i32 version.
5437 if (LHS.getValueType() == MVT::i8)
5438 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
5440 // If the operand types disagree, extend the shift amount to match. Since
5441 // BT ignores high bits (like shifts) we can use anyextend.
5442 if (LHS.getValueType() != RHS.getValueType())
5443 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
5445 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
5446 unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
5447 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5448 DAG.getConstant(Cond, MVT::i8), BT);
5452 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5453 unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
5455 SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
5456 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
5457 DAG.getConstant(X86CC, MVT::i8), Cond);
5460 SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
5462 SDValue Op0 = Op.getOperand(0);
5463 SDValue Op1 = Op.getOperand(1);
5464 SDValue CC = Op.getOperand(2);
5465 EVT VT = Op.getValueType();
5466 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
5467 bool isFP = Op.getOperand(1).getValueType().isFloatingPoint();
5468 DebugLoc dl = Op.getDebugLoc();
5472 EVT VT0 = Op0.getValueType();
5473 assert(VT0 == MVT::v4f32 || VT0 == MVT::v2f64);
5474 unsigned Opc = VT0 == MVT::v4f32 ? X86ISD::CMPPS : X86ISD::CMPPD;
5477 switch (SetCCOpcode) {
5480 case ISD::SETEQ: SSECC = 0; break;
5482 case ISD::SETGT: Swap = true; // Fallthrough
5484 case ISD::SETOLT: SSECC = 1; break;
5486 case ISD::SETGE: Swap = true; // Fallthrough
5488 case ISD::SETOLE: SSECC = 2; break;
5489 case ISD::SETUO: SSECC = 3; break;
5491 case ISD::SETNE: SSECC = 4; break;
5492 case ISD::SETULE: Swap = true;
5493 case ISD::SETUGE: SSECC = 5; break;
5494 case ISD::SETULT: Swap = true;
5495 case ISD::SETUGT: SSECC = 6; break;
5496 case ISD::SETO: SSECC = 7; break;
5499 std::swap(Op0, Op1);
5501 // In the two special cases we can't handle, emit two comparisons.
5503 if (SetCCOpcode == ISD::SETUEQ) {
5505 UNORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(3, MVT::i8));
5506 EQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(0, MVT::i8));
5507 return DAG.getNode(ISD::OR, dl, VT, UNORD, EQ);
5509 else if (SetCCOpcode == ISD::SETONE) {
5511 ORD = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(7, MVT::i8));
5512 NEQ = DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(4, MVT::i8));
5513 return DAG.getNode(ISD::AND, dl, VT, ORD, NEQ);
5515 llvm_unreachable("Illegal FP comparison");
5517 // Handle all other FP comparisons here.
5518 return DAG.getNode(Opc, dl, VT, Op0, Op1, DAG.getConstant(SSECC, MVT::i8));
5521 // We are handling one of the integer comparisons here. Since SSE only has
5522 // GT and EQ comparisons for integer, swapping operands and multiple
5523 // operations may be required for some comparisons.
5524 unsigned Opc = 0, EQOpc = 0, GTOpc = 0;
5525 bool Swap = false, Invert = false, FlipSigns = false;
5527 switch (VT.getSimpleVT().SimpleTy) {
5530 case MVT::v16i8: EQOpc = X86ISD::PCMPEQB; GTOpc = X86ISD::PCMPGTB; break;
5532 case MVT::v8i16: EQOpc = X86ISD::PCMPEQW; GTOpc = X86ISD::PCMPGTW; break;
5534 case MVT::v4i32: EQOpc = X86ISD::PCMPEQD; GTOpc = X86ISD::PCMPGTD; break;
5535 case MVT::v2i64: EQOpc = X86ISD::PCMPEQQ; GTOpc = X86ISD::PCMPGTQ; break;
5538 switch (SetCCOpcode) {
5540 case ISD::SETNE: Invert = true;
5541 case ISD::SETEQ: Opc = EQOpc; break;
5542 case ISD::SETLT: Swap = true;
5543 case ISD::SETGT: Opc = GTOpc; break;
5544 case ISD::SETGE: Swap = true;
5545 case ISD::SETLE: Opc = GTOpc; Invert = true; break;
5546 case ISD::SETULT: Swap = true;
5547 case ISD::SETUGT: Opc = GTOpc; FlipSigns = true; break;
5548 case ISD::SETUGE: Swap = true;
5549 case ISD::SETULE: Opc = GTOpc; FlipSigns = true; Invert = true; break;
5552 std::swap(Op0, Op1);
5554 // Since SSE has no unsigned integer comparisons, we need to flip the sign
5555 // bits of the inputs before performing those operations.
5557 EVT EltVT = VT.getVectorElementType();
5558 SDValue SignBit = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()),
5560 std::vector<SDValue> SignBits(VT.getVectorNumElements(), SignBit);
5561 SDValue SignVec = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &SignBits[0],
5563 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SignVec);
5564 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SignVec);
5567 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
5569 // If the logical-not of the result is required, perform that now.
5571 Result = DAG.getNOT(dl, Result, VT);
5576 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
5577 static bool isX86LogicalCmp(SDValue Op) {
5578 unsigned Opc = Op.getNode()->getOpcode();
5579 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI)
5581 if (Op.getResNo() == 1 &&
5582 (Opc == X86ISD::ADD ||
5583 Opc == X86ISD::SUB ||
5584 Opc == X86ISD::SMUL ||
5585 Opc == X86ISD::UMUL ||
5586 Opc == X86ISD::INC ||
5587 Opc == X86ISD::DEC))
5593 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) {
5594 bool addTest = true;
5595 SDValue Cond = Op.getOperand(0);
5596 DebugLoc dl = Op.getDebugLoc();
5599 if (Cond.getOpcode() == ISD::SETCC)
5600 Cond = LowerSETCC(Cond, DAG);
5602 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5603 // setting operand in place of the X86ISD::SETCC.
5604 if (Cond.getOpcode() == X86ISD::SETCC) {
5605 CC = Cond.getOperand(0);
5607 SDValue Cmp = Cond.getOperand(1);
5608 unsigned Opc = Cmp.getOpcode();
5609 EVT VT = Op.getValueType();
5611 bool IllegalFPCMov = false;
5612 if (VT.isFloatingPoint() && !VT.isVector() &&
5613 !isScalarFPTypeInSSEReg(VT)) // FPStack?
5614 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
5616 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
5617 Opc == X86ISD::BT) { // FIXME
5624 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5625 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5628 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
5629 SmallVector<SDValue, 4> Ops;
5630 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
5631 // condition is true.
5632 Ops.push_back(Op.getOperand(2));
5633 Ops.push_back(Op.getOperand(1));
5635 Ops.push_back(Cond);
5636 return DAG.getNode(X86ISD::CMOV, dl, VTs, &Ops[0], Ops.size());
5639 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
5640 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
5641 // from the AND / OR.
5642 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
5643 Opc = Op.getOpcode();
5644 if (Opc != ISD::OR && Opc != ISD::AND)
5646 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5647 Op.getOperand(0).hasOneUse() &&
5648 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
5649 Op.getOperand(1).hasOneUse());
5652 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
5653 // 1 and that the SETCC node has a single use.
5654 static bool isXor1OfSetCC(SDValue Op) {
5655 if (Op.getOpcode() != ISD::XOR)
5657 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
5658 if (N1C && N1C->getAPIntValue() == 1) {
5659 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
5660 Op.getOperand(0).hasOneUse();
5665 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) {
5666 bool addTest = true;
5667 SDValue Chain = Op.getOperand(0);
5668 SDValue Cond = Op.getOperand(1);
5669 SDValue Dest = Op.getOperand(2);
5670 DebugLoc dl = Op.getDebugLoc();
5673 if (Cond.getOpcode() == ISD::SETCC)
5674 Cond = LowerSETCC(Cond, DAG);
5676 // FIXME: LowerXALUO doesn't handle these!!
5677 else if (Cond.getOpcode() == X86ISD::ADD ||
5678 Cond.getOpcode() == X86ISD::SUB ||
5679 Cond.getOpcode() == X86ISD::SMUL ||
5680 Cond.getOpcode() == X86ISD::UMUL)
5681 Cond = LowerXALUO(Cond, DAG);
5684 // If condition flag is set by a X86ISD::CMP, then use it as the condition
5685 // setting operand in place of the X86ISD::SETCC.
5686 if (Cond.getOpcode() == X86ISD::SETCC) {
5687 CC = Cond.getOperand(0);
5689 SDValue Cmp = Cond.getOperand(1);
5690 unsigned Opc = Cmp.getOpcode();
5691 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
5692 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
5696 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
5700 // These can only come from an arithmetic instruction with overflow,
5701 // e.g. SADDO, UADDO.
5702 Cond = Cond.getNode()->getOperand(1);
5709 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
5710 SDValue Cmp = Cond.getOperand(0).getOperand(1);
5711 if (CondOpc == ISD::OR) {
5712 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
5713 // two branches instead of an explicit OR instruction with a
5715 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5716 isX86LogicalCmp(Cmp)) {
5717 CC = Cond.getOperand(0).getOperand(0);
5718 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5719 Chain, Dest, CC, Cmp);
5720 CC = Cond.getOperand(1).getOperand(0);
5724 } else { // ISD::AND
5725 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
5726 // two branches instead of an explicit AND instruction with a
5727 // separate test. However, we only do this if this block doesn't
5728 // have a fall-through edge, because this requires an explicit
5729 // jmp when the condition is false.
5730 if (Cmp == Cond.getOperand(1).getOperand(1) &&
5731 isX86LogicalCmp(Cmp) &&
5732 Op.getNode()->hasOneUse()) {
5733 X86::CondCode CCode =
5734 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5735 CCode = X86::GetOppositeBranchCondition(CCode);
5736 CC = DAG.getConstant(CCode, MVT::i8);
5737 SDValue User = SDValue(*Op.getNode()->use_begin(), 0);
5738 // Look for an unconditional branch following this conditional branch.
5739 // We need this because we need to reverse the successors in order
5740 // to implement FCMP_OEQ.
5741 if (User.getOpcode() == ISD::BR) {
5742 SDValue FalseBB = User.getOperand(1);
5744 DAG.UpdateNodeOperands(User, User.getOperand(0), Dest);
5745 assert(NewBR == User);
5748 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5749 Chain, Dest, CC, Cmp);
5750 X86::CondCode CCode =
5751 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
5752 CCode = X86::GetOppositeBranchCondition(CCode);
5753 CC = DAG.getConstant(CCode, MVT::i8);
5759 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
5760 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
5761 // It should be transformed during dag combiner except when the condition
5762 // is set by a arithmetics with overflow node.
5763 X86::CondCode CCode =
5764 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
5765 CCode = X86::GetOppositeBranchCondition(CCode);
5766 CC = DAG.getConstant(CCode, MVT::i8);
5767 Cond = Cond.getOperand(0).getOperand(1);
5773 CC = DAG.getConstant(X86::COND_NE, MVT::i8);
5774 Cond = EmitTest(Cond, X86::COND_NE, DAG);
5776 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
5777 Chain, Dest, CC, Cond);
5781 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
5782 // Calls to _alloca is needed to probe the stack when allocating more than 4k
5783 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
5784 // that the guard pages used by the OS virtual memory manager are allocated in
5785 // correct sequence.
5787 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
5788 SelectionDAG &DAG) {
5789 assert(Subtarget->isTargetCygMing() &&
5790 "This should be used only on Cygwin/Mingw targets");
5791 DebugLoc dl = Op.getDebugLoc();
5794 SDValue Chain = Op.getOperand(0);
5795 SDValue Size = Op.getOperand(1);
5796 // FIXME: Ensure alignment here
5800 EVT IntPtr = getPointerTy();
5801 EVT SPTy = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
5803 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true));
5805 Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
5806 Flag = Chain.getValue(1);
5808 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
5809 SDValue Ops[] = { Chain,
5810 DAG.getTargetExternalSymbol("_alloca", IntPtr),
5811 DAG.getRegister(X86::EAX, IntPtr),
5812 DAG.getRegister(X86StackPtr, SPTy),
5814 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops, 5);
5815 Flag = Chain.getValue(1);
5817 Chain = DAG.getCALLSEQ_END(Chain,
5818 DAG.getIntPtrConstant(0, true),
5819 DAG.getIntPtrConstant(0, true),
5822 Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
5824 SDValue Ops1[2] = { Chain.getValue(0), Chain };
5825 return DAG.getMergeValues(Ops1, 2, dl);
5829 X86TargetLowering::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
5831 SDValue Dst, SDValue Src,
5832 SDValue Size, unsigned Align,
5834 uint64_t DstSVOff) {
5835 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5837 // If not DWORD aligned or size is more than the threshold, call the library.
5838 // The libc version is likely to be faster for these cases. It can use the
5839 // address value and run time information about the CPU.
5840 if ((Align & 3) != 0 ||
5842 ConstantSize->getZExtValue() >
5843 getSubtarget()->getMaxInlineSizeThreshold()) {
5844 SDValue InFlag(0, 0);
5846 // Check to see if there is a specialized entry-point for memory zeroing.
5847 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
5849 if (const char *bzeroEntry = V &&
5850 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
5851 EVT IntPtr = getPointerTy();
5852 const Type *IntPtrTy = TD->getIntPtrType(*DAG.getContext());
5853 TargetLowering::ArgListTy Args;
5854 TargetLowering::ArgListEntry Entry;
5856 Entry.Ty = IntPtrTy;
5857 Args.push_back(Entry);
5859 Args.push_back(Entry);
5860 std::pair<SDValue,SDValue> CallResult =
5861 LowerCallTo(Chain, Type::getVoidTy(*DAG.getContext()),
5862 false, false, false, false,
5863 0, CallingConv::C, false, /*isReturnValueUsed=*/false,
5864 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args, DAG, dl);
5865 return CallResult.second;
5868 // Otherwise have the target-independent code call memset.
5872 uint64_t SizeVal = ConstantSize->getZExtValue();
5873 SDValue InFlag(0, 0);
5876 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
5877 unsigned BytesLeft = 0;
5878 bool TwoRepStos = false;
5881 uint64_t Val = ValC->getZExtValue() & 255;
5883 // If the value is a constant, then we can potentially use larger sets.
5884 switch (Align & 3) {
5885 case 2: // WORD aligned
5888 Val = (Val << 8) | Val;
5890 case 0: // DWORD aligned
5893 Val = (Val << 8) | Val;
5894 Val = (Val << 16) | Val;
5895 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
5898 Val = (Val << 32) | Val;
5901 default: // Byte aligned
5904 Count = DAG.getIntPtrConstant(SizeVal);
5908 if (AVT.bitsGT(MVT::i8)) {
5909 unsigned UBytes = AVT.getSizeInBits() / 8;
5910 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
5911 BytesLeft = SizeVal % UBytes;
5914 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
5916 InFlag = Chain.getValue(1);
5919 Count = DAG.getIntPtrConstant(SizeVal);
5920 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
5921 InFlag = Chain.getValue(1);
5924 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
5927 InFlag = Chain.getValue(1);
5928 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
5931 InFlag = Chain.getValue(1);
5933 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5934 SmallVector<SDValue, 8> Ops;
5935 Ops.push_back(Chain);
5936 Ops.push_back(DAG.getValueType(AVT));
5937 Ops.push_back(InFlag);
5938 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5941 InFlag = Chain.getValue(1);
5943 EVT CVT = Count.getValueType();
5944 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
5945 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
5946 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
5949 InFlag = Chain.getValue(1);
5950 Tys = DAG.getVTList(MVT::Other, MVT::Flag);
5952 Ops.push_back(Chain);
5953 Ops.push_back(DAG.getValueType(MVT::i8));
5954 Ops.push_back(InFlag);
5955 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, &Ops[0], Ops.size());
5956 } else if (BytesLeft) {
5957 // Handle the last 1 - 7 bytes.
5958 unsigned Offset = SizeVal - BytesLeft;
5959 EVT AddrVT = Dst.getValueType();
5960 EVT SizeVT = Size.getValueType();
5962 Chain = DAG.getMemset(Chain, dl,
5963 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
5964 DAG.getConstant(Offset, AddrVT)),
5966 DAG.getConstant(BytesLeft, SizeVT),
5967 Align, DstSV, DstSVOff + Offset);
5970 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
5975 X86TargetLowering::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
5976 SDValue Chain, SDValue Dst, SDValue Src,
5977 SDValue Size, unsigned Align,
5979 const Value *DstSV, uint64_t DstSVOff,
5980 const Value *SrcSV, uint64_t SrcSVOff) {
5981 // This requires the copy size to be a constant, preferrably
5982 // within a subtarget-specific limit.
5983 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
5986 uint64_t SizeVal = ConstantSize->getZExtValue();
5987 if (!AlwaysInline && SizeVal > getSubtarget()->getMaxInlineSizeThreshold())
5990 /// If not DWORD aligned, call the library.
5991 if ((Align & 3) != 0)
5996 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) // QWORD aligned
5999 unsigned UBytes = AVT.getSizeInBits() / 8;
6000 unsigned CountVal = SizeVal / UBytes;
6001 SDValue Count = DAG.getIntPtrConstant(CountVal);
6002 unsigned BytesLeft = SizeVal % UBytes;
6004 SDValue InFlag(0, 0);
6005 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
6008 InFlag = Chain.getValue(1);
6009 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
6012 InFlag = Chain.getValue(1);
6013 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
6016 InFlag = Chain.getValue(1);
6018 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6019 SmallVector<SDValue, 8> Ops;
6020 Ops.push_back(Chain);
6021 Ops.push_back(DAG.getValueType(AVT));
6022 Ops.push_back(InFlag);
6023 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, &Ops[0], Ops.size());
6025 SmallVector<SDValue, 4> Results;
6026 Results.push_back(RepMovs);
6028 // Handle the last 1 - 7 bytes.
6029 unsigned Offset = SizeVal - BytesLeft;
6030 EVT DstVT = Dst.getValueType();
6031 EVT SrcVT = Src.getValueType();
6032 EVT SizeVT = Size.getValueType();
6033 Results.push_back(DAG.getMemcpy(Chain, dl,
6034 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
6035 DAG.getConstant(Offset, DstVT)),
6036 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
6037 DAG.getConstant(Offset, SrcVT)),
6038 DAG.getConstant(BytesLeft, SizeVT),
6039 Align, AlwaysInline,
6040 DstSV, DstSVOff + Offset,
6041 SrcSV, SrcSVOff + Offset));
6044 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6045 &Results[0], Results.size());
6048 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) {
6049 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
6050 DebugLoc dl = Op.getDebugLoc();
6052 if (!Subtarget->is64Bit()) {
6053 // vastart just stores the address of the VarArgsFrameIndex slot into the
6054 // memory location argument.
6055 SDValue FR = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6056 return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1), SV, 0);
6060 // gp_offset (0 - 6 * 8)
6061 // fp_offset (48 - 48 + 8 * 16)
6062 // overflow_arg_area (point to parameters coming in memory).
6064 SmallVector<SDValue, 8> MemOps;
6065 SDValue FIN = Op.getOperand(1);
6067 SDValue Store = DAG.getStore(Op.getOperand(0), dl,
6068 DAG.getConstant(VarArgsGPOffset, MVT::i32),
6070 MemOps.push_back(Store);
6073 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6074 FIN, DAG.getIntPtrConstant(4));
6075 Store = DAG.getStore(Op.getOperand(0), dl,
6076 DAG.getConstant(VarArgsFPOffset, MVT::i32),
6078 MemOps.push_back(Store);
6080 // Store ptr to overflow_arg_area
6081 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6082 FIN, DAG.getIntPtrConstant(4));
6083 SDValue OVFIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
6084 Store = DAG.getStore(Op.getOperand(0), dl, OVFIN, FIN, SV, 0);
6085 MemOps.push_back(Store);
6087 // Store ptr to reg_save_area.
6088 FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(),
6089 FIN, DAG.getIntPtrConstant(8));
6090 SDValue RSFIN = DAG.getFrameIndex(RegSaveFrameIndex, getPointerTy());
6091 Store = DAG.getStore(Op.getOperand(0), dl, RSFIN, FIN, SV, 0);
6092 MemOps.push_back(Store);
6093 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
6094 &MemOps[0], MemOps.size());
6097 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) {
6098 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6099 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
6100 SDValue Chain = Op.getOperand(0);
6101 SDValue SrcPtr = Op.getOperand(1);
6102 SDValue SrcSV = Op.getOperand(2);
6104 llvm_report_error("VAArgInst is not yet implemented for x86-64!");
6108 SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) {
6109 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
6110 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
6111 SDValue Chain = Op.getOperand(0);
6112 SDValue DstPtr = Op.getOperand(1);
6113 SDValue SrcPtr = Op.getOperand(2);
6114 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
6115 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6116 DebugLoc dl = Op.getDebugLoc();
6118 return DAG.getMemcpy(Chain, dl, DstPtr, SrcPtr,
6119 DAG.getIntPtrConstant(24), 8, false,
6120 DstSV, 0, SrcSV, 0);
6124 X86TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
6125 DebugLoc dl = Op.getDebugLoc();
6126 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6128 default: return SDValue(); // Don't custom lower most intrinsics.
6129 // Comparison intrinsics.
6130 case Intrinsic::x86_sse_comieq_ss:
6131 case Intrinsic::x86_sse_comilt_ss:
6132 case Intrinsic::x86_sse_comile_ss:
6133 case Intrinsic::x86_sse_comigt_ss:
6134 case Intrinsic::x86_sse_comige_ss:
6135 case Intrinsic::x86_sse_comineq_ss:
6136 case Intrinsic::x86_sse_ucomieq_ss:
6137 case Intrinsic::x86_sse_ucomilt_ss:
6138 case Intrinsic::x86_sse_ucomile_ss:
6139 case Intrinsic::x86_sse_ucomigt_ss:
6140 case Intrinsic::x86_sse_ucomige_ss:
6141 case Intrinsic::x86_sse_ucomineq_ss:
6142 case Intrinsic::x86_sse2_comieq_sd:
6143 case Intrinsic::x86_sse2_comilt_sd:
6144 case Intrinsic::x86_sse2_comile_sd:
6145 case Intrinsic::x86_sse2_comigt_sd:
6146 case Intrinsic::x86_sse2_comige_sd:
6147 case Intrinsic::x86_sse2_comineq_sd:
6148 case Intrinsic::x86_sse2_ucomieq_sd:
6149 case Intrinsic::x86_sse2_ucomilt_sd:
6150 case Intrinsic::x86_sse2_ucomile_sd:
6151 case Intrinsic::x86_sse2_ucomigt_sd:
6152 case Intrinsic::x86_sse2_ucomige_sd:
6153 case Intrinsic::x86_sse2_ucomineq_sd: {
6155 ISD::CondCode CC = ISD::SETCC_INVALID;
6158 case Intrinsic::x86_sse_comieq_ss:
6159 case Intrinsic::x86_sse2_comieq_sd:
6163 case Intrinsic::x86_sse_comilt_ss:
6164 case Intrinsic::x86_sse2_comilt_sd:
6168 case Intrinsic::x86_sse_comile_ss:
6169 case Intrinsic::x86_sse2_comile_sd:
6173 case Intrinsic::x86_sse_comigt_ss:
6174 case Intrinsic::x86_sse2_comigt_sd:
6178 case Intrinsic::x86_sse_comige_ss:
6179 case Intrinsic::x86_sse2_comige_sd:
6183 case Intrinsic::x86_sse_comineq_ss:
6184 case Intrinsic::x86_sse2_comineq_sd:
6188 case Intrinsic::x86_sse_ucomieq_ss:
6189 case Intrinsic::x86_sse2_ucomieq_sd:
6190 Opc = X86ISD::UCOMI;
6193 case Intrinsic::x86_sse_ucomilt_ss:
6194 case Intrinsic::x86_sse2_ucomilt_sd:
6195 Opc = X86ISD::UCOMI;
6198 case Intrinsic::x86_sse_ucomile_ss:
6199 case Intrinsic::x86_sse2_ucomile_sd:
6200 Opc = X86ISD::UCOMI;
6203 case Intrinsic::x86_sse_ucomigt_ss:
6204 case Intrinsic::x86_sse2_ucomigt_sd:
6205 Opc = X86ISD::UCOMI;
6208 case Intrinsic::x86_sse_ucomige_ss:
6209 case Intrinsic::x86_sse2_ucomige_sd:
6210 Opc = X86ISD::UCOMI;
6213 case Intrinsic::x86_sse_ucomineq_ss:
6214 case Intrinsic::x86_sse2_ucomineq_sd:
6215 Opc = X86ISD::UCOMI;
6220 SDValue LHS = Op.getOperand(1);
6221 SDValue RHS = Op.getOperand(2);
6222 unsigned X86CC = TranslateX86CC(CC, true, LHS, RHS, DAG);
6223 SDValue Cond = DAG.getNode(Opc, dl, MVT::i32, LHS, RHS);
6224 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
6225 DAG.getConstant(X86CC, MVT::i8), Cond);
6226 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6228 // ptest intrinsics. The intrinsic these come from are designed to return
6229 // an integer value, not just an instruction so lower it to the ptest
6230 // pattern and a setcc for the result.
6231 case Intrinsic::x86_sse41_ptestz:
6232 case Intrinsic::x86_sse41_ptestc:
6233 case Intrinsic::x86_sse41_ptestnzc:{
6236 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
6237 case Intrinsic::x86_sse41_ptestz:
6239 X86CC = X86::COND_E;
6241 case Intrinsic::x86_sse41_ptestc:
6243 X86CC = X86::COND_B;
6245 case Intrinsic::x86_sse41_ptestnzc:
6247 X86CC = X86::COND_A;
6251 SDValue LHS = Op.getOperand(1);
6252 SDValue RHS = Op.getOperand(2);
6253 SDValue Test = DAG.getNode(X86ISD::PTEST, dl, MVT::i32, LHS, RHS);
6254 SDValue CC = DAG.getConstant(X86CC, MVT::i8);
6255 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
6256 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
6259 // Fix vector shift instructions where the last operand is a non-immediate
6261 case Intrinsic::x86_sse2_pslli_w:
6262 case Intrinsic::x86_sse2_pslli_d:
6263 case Intrinsic::x86_sse2_pslli_q:
6264 case Intrinsic::x86_sse2_psrli_w:
6265 case Intrinsic::x86_sse2_psrli_d:
6266 case Intrinsic::x86_sse2_psrli_q:
6267 case Intrinsic::x86_sse2_psrai_w:
6268 case Intrinsic::x86_sse2_psrai_d:
6269 case Intrinsic::x86_mmx_pslli_w:
6270 case Intrinsic::x86_mmx_pslli_d:
6271 case Intrinsic::x86_mmx_pslli_q:
6272 case Intrinsic::x86_mmx_psrli_w:
6273 case Intrinsic::x86_mmx_psrli_d:
6274 case Intrinsic::x86_mmx_psrli_q:
6275 case Intrinsic::x86_mmx_psrai_w:
6276 case Intrinsic::x86_mmx_psrai_d: {
6277 SDValue ShAmt = Op.getOperand(2);
6278 if (isa<ConstantSDNode>(ShAmt))
6281 unsigned NewIntNo = 0;
6282 EVT ShAmtVT = MVT::v4i32;
6284 case Intrinsic::x86_sse2_pslli_w:
6285 NewIntNo = Intrinsic::x86_sse2_psll_w;
6287 case Intrinsic::x86_sse2_pslli_d:
6288 NewIntNo = Intrinsic::x86_sse2_psll_d;
6290 case Intrinsic::x86_sse2_pslli_q:
6291 NewIntNo = Intrinsic::x86_sse2_psll_q;
6293 case Intrinsic::x86_sse2_psrli_w:
6294 NewIntNo = Intrinsic::x86_sse2_psrl_w;
6296 case Intrinsic::x86_sse2_psrli_d:
6297 NewIntNo = Intrinsic::x86_sse2_psrl_d;
6299 case Intrinsic::x86_sse2_psrli_q:
6300 NewIntNo = Intrinsic::x86_sse2_psrl_q;
6302 case Intrinsic::x86_sse2_psrai_w:
6303 NewIntNo = Intrinsic::x86_sse2_psra_w;
6305 case Intrinsic::x86_sse2_psrai_d:
6306 NewIntNo = Intrinsic::x86_sse2_psra_d;
6309 ShAmtVT = MVT::v2i32;
6311 case Intrinsic::x86_mmx_pslli_w:
6312 NewIntNo = Intrinsic::x86_mmx_psll_w;
6314 case Intrinsic::x86_mmx_pslli_d:
6315 NewIntNo = Intrinsic::x86_mmx_psll_d;
6317 case Intrinsic::x86_mmx_pslli_q:
6318 NewIntNo = Intrinsic::x86_mmx_psll_q;
6320 case Intrinsic::x86_mmx_psrli_w:
6321 NewIntNo = Intrinsic::x86_mmx_psrl_w;
6323 case Intrinsic::x86_mmx_psrli_d:
6324 NewIntNo = Intrinsic::x86_mmx_psrl_d;
6326 case Intrinsic::x86_mmx_psrli_q:
6327 NewIntNo = Intrinsic::x86_mmx_psrl_q;
6329 case Intrinsic::x86_mmx_psrai_w:
6330 NewIntNo = Intrinsic::x86_mmx_psra_w;
6332 case Intrinsic::x86_mmx_psrai_d:
6333 NewIntNo = Intrinsic::x86_mmx_psra_d;
6335 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
6340 EVT VT = Op.getValueType();
6341 ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT,
6342 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, ShAmtVT, ShAmt));
6343 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6344 DAG.getConstant(NewIntNo, MVT::i32),
6345 Op.getOperand(1), ShAmt);
6350 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) {
6351 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6352 DebugLoc dl = Op.getDebugLoc();
6355 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
6357 DAG.getConstant(TD->getPointerSize(),
6358 Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
6359 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6360 DAG.getNode(ISD::ADD, dl, getPointerTy(),
6365 // Just load the return address.
6366 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
6367 return DAG.getLoad(getPointerTy(), dl, DAG.getEntryNode(),
6368 RetAddrFI, NULL, 0);
6371 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) {
6372 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
6373 MFI->setFrameAddressIsTaken(true);
6374 EVT VT = Op.getValueType();
6375 DebugLoc dl = Op.getDebugLoc(); // FIXME probably not meaningful
6376 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
6377 unsigned FrameReg = Subtarget->is64Bit() ? X86::RBP : X86::EBP;
6378 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
6380 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr, NULL, 0);
6384 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
6385 SelectionDAG &DAG) {
6386 return DAG.getIntPtrConstant(2*TD->getPointerSize());
6389 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG)
6391 MachineFunction &MF = DAG.getMachineFunction();
6392 SDValue Chain = Op.getOperand(0);
6393 SDValue Offset = Op.getOperand(1);
6394 SDValue Handler = Op.getOperand(2);
6395 DebugLoc dl = Op.getDebugLoc();
6397 SDValue Frame = DAG.getRegister(Subtarget->is64Bit() ? X86::RBP : X86::EBP,
6399 unsigned StoreAddrReg = (Subtarget->is64Bit() ? X86::RCX : X86::ECX);
6401 SDValue StoreAddr = DAG.getNode(ISD::SUB, dl, getPointerTy(), Frame,
6402 DAG.getIntPtrConstant(-TD->getPointerSize()));
6403 StoreAddr = DAG.getNode(ISD::ADD, dl, getPointerTy(), StoreAddr, Offset);
6404 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, NULL, 0);
6405 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
6406 MF.getRegInfo().addLiveOut(StoreAddrReg);
6408 return DAG.getNode(X86ISD::EH_RETURN, dl,
6410 Chain, DAG.getRegister(StoreAddrReg, getPointerTy()));
6413 SDValue X86TargetLowering::LowerTRAMPOLINE(SDValue Op,
6414 SelectionDAG &DAG) {
6415 SDValue Root = Op.getOperand(0);
6416 SDValue Trmp = Op.getOperand(1); // trampoline
6417 SDValue FPtr = Op.getOperand(2); // nested function
6418 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
6419 DebugLoc dl = Op.getDebugLoc();
6421 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
6423 const X86InstrInfo *TII =
6424 ((X86TargetMachine&)getTargetMachine()).getInstrInfo();
6426 if (Subtarget->is64Bit()) {
6427 SDValue OutChains[6];
6429 // Large code-model.
6431 const unsigned char JMP64r = TII->getBaseOpcodeFor(X86::JMP64r);
6432 const unsigned char MOV64ri = TII->getBaseOpcodeFor(X86::MOV64ri);
6434 const unsigned char N86R10 = RegInfo->getX86RegNum(X86::R10);
6435 const unsigned char N86R11 = RegInfo->getX86RegNum(X86::R11);
6437 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
6439 // Load the pointer to the nested function into R11.
6440 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
6441 SDValue Addr = Trmp;
6442 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6445 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6446 DAG.getConstant(2, MVT::i64));
6447 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr, TrmpAddr, 2, false, 2);
6449 // Load the 'nest' parameter value into R10.
6450 // R10 is specified in X86CallingConv.td
6451 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
6452 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6453 DAG.getConstant(10, MVT::i64));
6454 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6455 Addr, TrmpAddr, 10);
6457 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6458 DAG.getConstant(12, MVT::i64));
6459 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 12, false, 2);
6461 // Jump to the nested function.
6462 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
6463 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6464 DAG.getConstant(20, MVT::i64));
6465 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, MVT::i16),
6466 Addr, TrmpAddr, 20);
6468 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
6469 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
6470 DAG.getConstant(22, MVT::i64));
6471 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, MVT::i8), Addr,
6475 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 6) };
6476 return DAG.getMergeValues(Ops, 2, dl);
6478 const Function *Func =
6479 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
6480 CallingConv::ID CC = Func->getCallingConv();
6485 llvm_unreachable("Unsupported calling convention");
6486 case CallingConv::C:
6487 case CallingConv::X86_StdCall: {
6488 // Pass 'nest' parameter in ECX.
6489 // Must be kept in sync with X86CallingConv.td
6492 // Check that ECX wasn't needed by an 'inreg' parameter.
6493 const FunctionType *FTy = Func->getFunctionType();
6494 const AttrListPtr &Attrs = Func->getAttributes();
6496 if (!Attrs.isEmpty() && !Func->isVarArg()) {
6497 unsigned InRegCount = 0;
6500 for (FunctionType::param_iterator I = FTy->param_begin(),
6501 E = FTy->param_end(); I != E; ++I, ++Idx)
6502 if (Attrs.paramHasAttr(Idx, Attribute::InReg))
6503 // FIXME: should only count parameters that are lowered to integers.
6504 InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
6506 if (InRegCount > 2) {
6507 llvm_report_error("Nest register in use - reduce number of inreg parameters!");
6512 case CallingConv::X86_FastCall:
6513 case CallingConv::Fast:
6514 // Pass 'nest' parameter in EAX.
6515 // Must be kept in sync with X86CallingConv.td
6520 SDValue OutChains[4];
6523 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6524 DAG.getConstant(10, MVT::i32));
6525 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
6527 const unsigned char MOV32ri = TII->getBaseOpcodeFor(X86::MOV32ri);
6528 const unsigned char N86Reg = RegInfo->getX86RegNum(NestReg);
6529 OutChains[0] = DAG.getStore(Root, dl,
6530 DAG.getConstant(MOV32ri|N86Reg, MVT::i8),
6533 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6534 DAG.getConstant(1, MVT::i32));
6535 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr, TrmpAddr, 1, false, 1);
6537 const unsigned char JMP = TII->getBaseOpcodeFor(X86::JMP);
6538 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6539 DAG.getConstant(5, MVT::i32));
6540 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, MVT::i8), Addr,
6541 TrmpAddr, 5, false, 1);
6543 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
6544 DAG.getConstant(6, MVT::i32));
6545 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr, TrmpAddr, 6, false, 1);
6548 { Trmp, DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains, 4) };
6549 return DAG.getMergeValues(Ops, 2, dl);
6553 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) {
6555 The rounding mode is in bits 11:10 of FPSR, and has the following
6562 FLT_ROUNDS, on the other hand, expects the following:
6569 To perform the conversion, we do:
6570 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
6573 MachineFunction &MF = DAG.getMachineFunction();
6574 const TargetMachine &TM = MF.getTarget();
6575 const TargetFrameInfo &TFI = *TM.getFrameInfo();
6576 unsigned StackAlignment = TFI.getStackAlignment();
6577 EVT VT = Op.getValueType();
6578 DebugLoc dl = Op.getDebugLoc();
6580 // Save FP Control Word to stack slot
6581 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment);
6582 SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
6584 SDValue Chain = DAG.getNode(X86ISD::FNSTCW16m, dl, MVT::Other,
6585 DAG.getEntryNode(), StackSlot);
6587 // Load FP Control Word from stack slot
6588 SDValue CWD = DAG.getLoad(MVT::i16, dl, Chain, StackSlot, NULL, 0);
6590 // Transform as necessary
6592 DAG.getNode(ISD::SRL, dl, MVT::i16,
6593 DAG.getNode(ISD::AND, dl, MVT::i16,
6594 CWD, DAG.getConstant(0x800, MVT::i16)),
6595 DAG.getConstant(11, MVT::i8));
6597 DAG.getNode(ISD::SRL, dl, MVT::i16,
6598 DAG.getNode(ISD::AND, dl, MVT::i16,
6599 CWD, DAG.getConstant(0x400, MVT::i16)),
6600 DAG.getConstant(9, MVT::i8));
6603 DAG.getNode(ISD::AND, dl, MVT::i16,
6604 DAG.getNode(ISD::ADD, dl, MVT::i16,
6605 DAG.getNode(ISD::OR, dl, MVT::i16, CWD1, CWD2),
6606 DAG.getConstant(1, MVT::i16)),
6607 DAG.getConstant(3, MVT::i16));
6610 return DAG.getNode((VT.getSizeInBits() < 16 ?
6611 ISD::TRUNCATE : ISD::ZERO_EXTEND), dl, VT, RetVal);
6614 SDValue X86TargetLowering::LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
6615 EVT VT = Op.getValueType();
6617 unsigned NumBits = VT.getSizeInBits();
6618 DebugLoc dl = Op.getDebugLoc();
6620 Op = Op.getOperand(0);
6621 if (VT == MVT::i8) {
6622 // Zero extend to i32 since there is not an i8 bsr.
6624 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6627 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
6628 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6629 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
6631 // If src is zero (i.e. bsr sets ZF), returns NumBits.
6632 SmallVector<SDValue, 4> Ops;
6634 Ops.push_back(DAG.getConstant(NumBits+NumBits-1, OpVT));
6635 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6636 Ops.push_back(Op.getValue(1));
6637 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6639 // Finally xor with NumBits-1.
6640 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op, DAG.getConstant(NumBits-1, OpVT));
6643 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6647 SDValue X86TargetLowering::LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
6648 EVT VT = Op.getValueType();
6650 unsigned NumBits = VT.getSizeInBits();
6651 DebugLoc dl = Op.getDebugLoc();
6653 Op = Op.getOperand(0);
6654 if (VT == MVT::i8) {
6656 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
6659 // Issue a bsf (scan bits forward) which also sets EFLAGS.
6660 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
6661 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
6663 // If src is zero (i.e. bsf sets ZF), returns NumBits.
6664 SmallVector<SDValue, 4> Ops;
6666 Ops.push_back(DAG.getConstant(NumBits, OpVT));
6667 Ops.push_back(DAG.getConstant(X86::COND_E, MVT::i8));
6668 Ops.push_back(Op.getValue(1));
6669 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, &Ops[0], 4);
6672 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
6676 SDValue X86TargetLowering::LowerMUL_V2I64(SDValue Op, SelectionDAG &DAG) {
6677 EVT VT = Op.getValueType();
6678 assert(VT == MVT::v2i64 && "Only know how to lower V2I64 multiply");
6679 DebugLoc dl = Op.getDebugLoc();
6681 // ulong2 Ahi = __builtin_ia32_psrlqi128( a, 32);
6682 // ulong2 Bhi = __builtin_ia32_psrlqi128( b, 32);
6683 // ulong2 AloBlo = __builtin_ia32_pmuludq128( a, b );
6684 // ulong2 AloBhi = __builtin_ia32_pmuludq128( a, Bhi );
6685 // ulong2 AhiBlo = __builtin_ia32_pmuludq128( Ahi, b );
6687 // AloBhi = __builtin_ia32_psllqi128( AloBhi, 32 );
6688 // AhiBlo = __builtin_ia32_psllqi128( AhiBlo, 32 );
6689 // return AloBlo + AloBhi + AhiBlo;
6691 SDValue A = Op.getOperand(0);
6692 SDValue B = Op.getOperand(1);
6694 SDValue Ahi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6695 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6696 A, DAG.getConstant(32, MVT::i32));
6697 SDValue Bhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6698 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
6699 B, DAG.getConstant(32, MVT::i32));
6700 SDValue AloBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6701 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6703 SDValue AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6704 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6706 SDValue AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6707 DAG.getConstant(Intrinsic::x86_sse2_pmulu_dq, MVT::i32),
6709 AloBhi = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6710 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6711 AloBhi, DAG.getConstant(32, MVT::i32));
6712 AhiBlo = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
6713 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
6714 AhiBlo, DAG.getConstant(32, MVT::i32));
6715 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
6716 Res = DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
6721 SDValue X86TargetLowering::LowerXALUO(SDValue Op, SelectionDAG &DAG) {
6722 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
6723 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
6724 // looks for this combo and may remove the "setcc" instruction if the "setcc"
6725 // has only one use.
6726 SDNode *N = Op.getNode();
6727 SDValue LHS = N->getOperand(0);
6728 SDValue RHS = N->getOperand(1);
6729 unsigned BaseOp = 0;
6731 DebugLoc dl = Op.getDebugLoc();
6733 switch (Op.getOpcode()) {
6734 default: llvm_unreachable("Unknown ovf instruction!");
6736 // A subtract of one will be selected as a INC. Note that INC doesn't
6737 // set CF, so we can't do this for UADDO.
6738 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6739 if (C->getAPIntValue() == 1) {
6740 BaseOp = X86ISD::INC;
6744 BaseOp = X86ISD::ADD;
6748 BaseOp = X86ISD::ADD;
6752 // A subtract of one will be selected as a DEC. Note that DEC doesn't
6753 // set CF, so we can't do this for USUBO.
6754 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op))
6755 if (C->getAPIntValue() == 1) {
6756 BaseOp = X86ISD::DEC;
6760 BaseOp = X86ISD::SUB;
6764 BaseOp = X86ISD::SUB;
6768 BaseOp = X86ISD::SMUL;
6772 BaseOp = X86ISD::UMUL;
6777 // Also sets EFLAGS.
6778 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
6779 SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
6782 DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
6783 DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
6785 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
6789 SDValue X86TargetLowering::LowerCMP_SWAP(SDValue Op, SelectionDAG &DAG) {
6790 EVT T = Op.getValueType();
6791 DebugLoc dl = Op.getDebugLoc();
6794 switch(T.getSimpleVT().SimpleTy) {
6796 assert(false && "Invalid value type!");
6797 case MVT::i8: Reg = X86::AL; size = 1; break;
6798 case MVT::i16: Reg = X86::AX; size = 2; break;
6799 case MVT::i32: Reg = X86::EAX; size = 4; break;
6801 assert(Subtarget->is64Bit() && "Node not type legal!");
6802 Reg = X86::RAX; size = 8;
6805 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), dl, Reg,
6806 Op.getOperand(2), SDValue());
6807 SDValue Ops[] = { cpIn.getValue(0),
6810 DAG.getTargetConstant(size, MVT::i8),
6812 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6813 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG_DAG, dl, Tys, Ops, 5);
6815 DAG.getCopyFromReg(Result.getValue(0), dl, Reg, T, Result.getValue(1));
6819 SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
6820 SelectionDAG &DAG) {
6821 assert(Subtarget->is64Bit() && "Result not type legalized?");
6822 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6823 SDValue TheChain = Op.getOperand(0);
6824 DebugLoc dl = Op.getDebugLoc();
6825 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6826 SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
6827 SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
6829 SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
6830 DAG.getConstant(32, MVT::i8));
6832 DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
6835 return DAG.getMergeValues(Ops, 2, dl);
6838 SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
6839 SDNode *Node = Op.getNode();
6840 DebugLoc dl = Node->getDebugLoc();
6841 EVT T = Node->getValueType(0);
6842 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
6843 DAG.getConstant(0, T), Node->getOperand(2));
6844 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
6845 cast<AtomicSDNode>(Node)->getMemoryVT(),
6846 Node->getOperand(0),
6847 Node->getOperand(1), negOp,
6848 cast<AtomicSDNode>(Node)->getSrcValue(),
6849 cast<AtomicSDNode>(Node)->getAlignment());
6852 /// LowerOperation - Provide custom lowering hooks for some operations.
6854 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) {
6855 switch (Op.getOpcode()) {
6856 default: llvm_unreachable("Should not custom lower this!");
6857 case ISD::ATOMIC_CMP_SWAP: return LowerCMP_SWAP(Op,DAG);
6858 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
6859 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
6860 case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
6861 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
6862 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
6863 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
6864 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
6865 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
6866 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
6867 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
6868 case ISD::SHL_PARTS:
6869 case ISD::SRA_PARTS:
6870 case ISD::SRL_PARTS: return LowerShift(Op, DAG);
6871 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
6872 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
6873 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
6874 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
6875 case ISD::FABS: return LowerFABS(Op, DAG);
6876 case ISD::FNEG: return LowerFNEG(Op, DAG);
6877 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
6878 case ISD::SETCC: return LowerSETCC(Op, DAG);
6879 case ISD::VSETCC: return LowerVSETCC(Op, DAG);
6880 case ISD::SELECT: return LowerSELECT(Op, DAG);
6881 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
6882 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
6883 case ISD::VASTART: return LowerVASTART(Op, DAG);
6884 case ISD::VAARG: return LowerVAARG(Op, DAG);
6885 case ISD::VACOPY: return LowerVACOPY(Op, DAG);
6886 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
6887 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
6888 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
6889 case ISD::FRAME_TO_ARGS_OFFSET:
6890 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
6891 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
6892 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
6893 case ISD::TRAMPOLINE: return LowerTRAMPOLINE(Op, DAG);
6894 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
6895 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
6896 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
6897 case ISD::MUL: return LowerMUL_V2I64(Op, DAG);
6903 case ISD::UMULO: return LowerXALUO(Op, DAG);
6904 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
6908 void X86TargetLowering::
6909 ReplaceATOMIC_BINARY_64(SDNode *Node, SmallVectorImpl<SDValue>&Results,
6910 SelectionDAG &DAG, unsigned NewOp) {
6911 EVT T = Node->getValueType(0);
6912 DebugLoc dl = Node->getDebugLoc();
6913 assert (T == MVT::i64 && "Only know how to expand i64 atomics");
6915 SDValue Chain = Node->getOperand(0);
6916 SDValue In1 = Node->getOperand(1);
6917 SDValue In2L = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6918 Node->getOperand(2), DAG.getIntPtrConstant(0));
6919 SDValue In2H = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32,
6920 Node->getOperand(2), DAG.getIntPtrConstant(1));
6921 // This is a generalized SDNode, not an AtomicSDNode, so it doesn't
6922 // have a MemOperand. Pass the info through as a normal operand.
6923 SDValue LSI = DAG.getMemOperand(cast<MemSDNode>(Node)->getMemOperand());
6924 SDValue Ops[] = { Chain, In1, In2L, In2H, LSI };
6925 SDVTList Tys = DAG.getVTList(MVT::i32, MVT::i32, MVT::Other);
6926 SDValue Result = DAG.getNode(NewOp, dl, Tys, Ops, 5);
6927 SDValue OpsF[] = { Result.getValue(0), Result.getValue(1)};
6928 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6929 Results.push_back(Result.getValue(2));
6932 /// ReplaceNodeResults - Replace a node with an illegal result type
6933 /// with a new node built out of custom code.
6934 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
6935 SmallVectorImpl<SDValue>&Results,
6936 SelectionDAG &DAG) {
6937 DebugLoc dl = N->getDebugLoc();
6938 switch (N->getOpcode()) {
6940 assert(false && "Do not know how to custom type legalize this operation!");
6942 case ISD::FP_TO_SINT: {
6943 std::pair<SDValue,SDValue> Vals =
6944 FP_TO_INTHelper(SDValue(N, 0), DAG, true);
6945 SDValue FIST = Vals.first, StackSlot = Vals.second;
6946 if (FIST.getNode() != 0) {
6947 EVT VT = N->getValueType(0);
6948 // Return a load from the stack slot.
6949 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot, NULL, 0));
6953 case ISD::READCYCLECOUNTER: {
6954 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6955 SDValue TheChain = N->getOperand(0);
6956 SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
6957 SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
6959 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
6961 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
6962 SDValue Ops[] = { eax, edx };
6963 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops, 2));
6964 Results.push_back(edx.getValue(1));
6967 case ISD::ATOMIC_CMP_SWAP: {
6968 EVT T = N->getValueType(0);
6969 assert (T == MVT::i64 && "Only know how to expand i64 Cmp and Swap");
6970 SDValue cpInL, cpInH;
6971 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6972 DAG.getConstant(0, MVT::i32));
6973 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(2),
6974 DAG.getConstant(1, MVT::i32));
6975 cpInL = DAG.getCopyToReg(N->getOperand(0), dl, X86::EAX, cpInL, SDValue());
6976 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl, X86::EDX, cpInH,
6978 SDValue swapInL, swapInH;
6979 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6980 DAG.getConstant(0, MVT::i32));
6981 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(3),
6982 DAG.getConstant(1, MVT::i32));
6983 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl, X86::EBX, swapInL,
6985 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl, X86::ECX, swapInH,
6986 swapInL.getValue(1));
6987 SDValue Ops[] = { swapInH.getValue(0),
6989 swapInH.getValue(1) };
6990 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
6991 SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
6992 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
6993 MVT::i32, Result.getValue(1));
6994 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
6995 MVT::i32, cpOutL.getValue(2));
6996 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
6997 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, OpsF, 2));
6998 Results.push_back(cpOutH.getValue(1));
7001 case ISD::ATOMIC_LOAD_ADD:
7002 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMADD64_DAG);
7004 case ISD::ATOMIC_LOAD_AND:
7005 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMAND64_DAG);
7007 case ISD::ATOMIC_LOAD_NAND:
7008 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMNAND64_DAG);
7010 case ISD::ATOMIC_LOAD_OR:
7011 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMOR64_DAG);
7013 case ISD::ATOMIC_LOAD_SUB:
7014 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSUB64_DAG);
7016 case ISD::ATOMIC_LOAD_XOR:
7017 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMXOR64_DAG);
7019 case ISD::ATOMIC_SWAP:
7020 ReplaceATOMIC_BINARY_64(N, Results, DAG, X86ISD::ATOMSWAP64_DAG);
7025 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
7027 default: return NULL;
7028 case X86ISD::BSF: return "X86ISD::BSF";
7029 case X86ISD::BSR: return "X86ISD::BSR";
7030 case X86ISD::SHLD: return "X86ISD::SHLD";
7031 case X86ISD::SHRD: return "X86ISD::SHRD";
7032 case X86ISD::FAND: return "X86ISD::FAND";
7033 case X86ISD::FOR: return "X86ISD::FOR";
7034 case X86ISD::FXOR: return "X86ISD::FXOR";
7035 case X86ISD::FSRL: return "X86ISD::FSRL";
7036 case X86ISD::FILD: return "X86ISD::FILD";
7037 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
7038 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
7039 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
7040 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
7041 case X86ISD::FLD: return "X86ISD::FLD";
7042 case X86ISD::FST: return "X86ISD::FST";
7043 case X86ISD::CALL: return "X86ISD::CALL";
7044 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
7045 case X86ISD::BT: return "X86ISD::BT";
7046 case X86ISD::CMP: return "X86ISD::CMP";
7047 case X86ISD::COMI: return "X86ISD::COMI";
7048 case X86ISD::UCOMI: return "X86ISD::UCOMI";
7049 case X86ISD::SETCC: return "X86ISD::SETCC";
7050 case X86ISD::CMOV: return "X86ISD::CMOV";
7051 case X86ISD::BRCOND: return "X86ISD::BRCOND";
7052 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
7053 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
7054 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
7055 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
7056 case X86ISD::Wrapper: return "X86ISD::Wrapper";
7057 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
7058 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
7059 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
7060 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
7061 case X86ISD::PINSRB: return "X86ISD::PINSRB";
7062 case X86ISD::PINSRW: return "X86ISD::PINSRW";
7063 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
7064 case X86ISD::FMAX: return "X86ISD::FMAX";
7065 case X86ISD::FMIN: return "X86ISD::FMIN";
7066 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
7067 case X86ISD::FRCP: return "X86ISD::FRCP";
7068 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
7069 case X86ISD::SegmentBaseAddress: return "X86ISD::SegmentBaseAddress";
7070 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
7071 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
7072 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
7073 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
7074 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
7075 case X86ISD::ATOMADD64_DAG: return "X86ISD::ATOMADD64_DAG";
7076 case X86ISD::ATOMSUB64_DAG: return "X86ISD::ATOMSUB64_DAG";
7077 case X86ISD::ATOMOR64_DAG: return "X86ISD::ATOMOR64_DAG";
7078 case X86ISD::ATOMXOR64_DAG: return "X86ISD::ATOMXOR64_DAG";
7079 case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
7080 case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
7081 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
7082 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
7083 case X86ISD::VSHL: return "X86ISD::VSHL";
7084 case X86ISD::VSRL: return "X86ISD::VSRL";
7085 case X86ISD::CMPPD: return "X86ISD::CMPPD";
7086 case X86ISD::CMPPS: return "X86ISD::CMPPS";
7087 case X86ISD::PCMPEQB: return "X86ISD::PCMPEQB";
7088 case X86ISD::PCMPEQW: return "X86ISD::PCMPEQW";
7089 case X86ISD::PCMPEQD: return "X86ISD::PCMPEQD";
7090 case X86ISD::PCMPEQQ: return "X86ISD::PCMPEQQ";
7091 case X86ISD::PCMPGTB: return "X86ISD::PCMPGTB";
7092 case X86ISD::PCMPGTW: return "X86ISD::PCMPGTW";
7093 case X86ISD::PCMPGTD: return "X86ISD::PCMPGTD";
7094 case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
7095 case X86ISD::ADD: return "X86ISD::ADD";
7096 case X86ISD::SUB: return "X86ISD::SUB";
7097 case X86ISD::SMUL: return "X86ISD::SMUL";
7098 case X86ISD::UMUL: return "X86ISD::UMUL";
7099 case X86ISD::INC: return "X86ISD::INC";
7100 case X86ISD::DEC: return "X86ISD::DEC";
7101 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
7102 case X86ISD::PTEST: return "X86ISD::PTEST";
7103 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
7107 // isLegalAddressingMode - Return true if the addressing mode represented
7108 // by AM is legal for this target, for a load/store of the specified type.
7109 bool X86TargetLowering::isLegalAddressingMode(const AddrMode &AM,
7110 const Type *Ty) const {
7111 // X86 supports extremely general addressing modes.
7112 CodeModel::Model M = getTargetMachine().getCodeModel();
7114 // X86 allows a sign-extended 32-bit immediate field as a displacement.
7115 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
7120 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
7122 // If a reference to this global requires an extra load, we can't fold it.
7123 if (isGlobalStubReference(GVFlags))
7126 // If BaseGV requires a register for the PIC base, we cannot also have a
7127 // BaseReg specified.
7128 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
7131 // If lower 4G is not available, then we must use rip-relative addressing.
7132 if (Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
7142 // These scales always work.
7147 // These scales are formed with basereg+scalereg. Only accept if there is
7152 default: // Other stuff never works.
7160 bool X86TargetLowering::isTruncateFree(const Type *Ty1, const Type *Ty2) const {
7161 if (!Ty1->isInteger() || !Ty2->isInteger())
7163 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
7164 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
7165 if (NumBits1 <= NumBits2)
7167 return Subtarget->is64Bit() || NumBits1 < 64;
7170 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
7171 if (!VT1.isInteger() || !VT2.isInteger())
7173 unsigned NumBits1 = VT1.getSizeInBits();
7174 unsigned NumBits2 = VT2.getSizeInBits();
7175 if (NumBits1 <= NumBits2)
7177 return Subtarget->is64Bit() || NumBits1 < 64;
7180 bool X86TargetLowering::isZExtFree(const Type *Ty1, const Type *Ty2) const {
7181 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7182 return Ty1 == Type::getInt32Ty(Ty1->getContext()) &&
7183 Ty2 == Type::getInt64Ty(Ty1->getContext()) && Subtarget->is64Bit();
7186 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
7187 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
7188 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
7191 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
7192 // i16 instructions are longer (0x66 prefix) and potentially slower.
7193 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
7196 /// isShuffleMaskLegal - Targets can use this to indicate that they only
7197 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
7198 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
7199 /// are assumed to be legal.
7201 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
7203 // Only do shuffles on 128-bit vector types for now.
7204 if (VT.getSizeInBits() == 64)
7207 // FIXME: pshufb, blends, palignr, shifts.
7208 return (VT.getVectorNumElements() == 2 ||
7209 ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
7210 isMOVLMask(M, VT) ||
7211 isSHUFPMask(M, VT) ||
7212 isPSHUFDMask(M, VT) ||
7213 isPSHUFHWMask(M, VT) ||
7214 isPSHUFLWMask(M, VT) ||
7215 isUNPCKLMask(M, VT) ||
7216 isUNPCKHMask(M, VT) ||
7217 isUNPCKL_v_undef_Mask(M, VT) ||
7218 isUNPCKH_v_undef_Mask(M, VT));
7222 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
7224 unsigned NumElts = VT.getVectorNumElements();
7225 // FIXME: This collection of masks seems suspect.
7228 if (NumElts == 4 && VT.getSizeInBits() == 128) {
7229 return (isMOVLMask(Mask, VT) ||
7230 isCommutedMOVLMask(Mask, VT, true) ||
7231 isSHUFPMask(Mask, VT) ||
7232 isCommutedSHUFPMask(Mask, VT));
7237 //===----------------------------------------------------------------------===//
7238 // X86 Scheduler Hooks
7239 //===----------------------------------------------------------------------===//
7241 // private utility function
7243 X86TargetLowering::EmitAtomicBitwiseWithCustomInserter(MachineInstr *bInstr,
7244 MachineBasicBlock *MBB,
7252 TargetRegisterClass *RC,
7253 bool invSrc) const {
7254 // For the atomic bitwise operator, we generate
7257 // ld t1 = [bitinstr.addr]
7258 // op t2 = t1, [bitinstr.val]
7260 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7262 // fallthrough -->nextMBB
7263 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7264 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7265 MachineFunction::iterator MBBIter = MBB;
7268 /// First build the CFG
7269 MachineFunction *F = MBB->getParent();
7270 MachineBasicBlock *thisMBB = MBB;
7271 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7272 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7273 F->insert(MBBIter, newMBB);
7274 F->insert(MBBIter, nextMBB);
7276 // Move all successors to thisMBB to nextMBB
7277 nextMBB->transferSuccessors(thisMBB);
7279 // Update thisMBB to fall through to newMBB
7280 thisMBB->addSuccessor(newMBB);
7282 // newMBB jumps to itself and fall through to nextMBB
7283 newMBB->addSuccessor(nextMBB);
7284 newMBB->addSuccessor(newMBB);
7286 // Insert instructions into newMBB based on incoming instruction
7287 assert(bInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7288 "unexpected number of operands");
7289 DebugLoc dl = bInstr->getDebugLoc();
7290 MachineOperand& destOper = bInstr->getOperand(0);
7291 MachineOperand* argOpers[2 + X86AddrNumOperands];
7292 int numArgs = bInstr->getNumOperands() - 1;
7293 for (int i=0; i < numArgs; ++i)
7294 argOpers[i] = &bInstr->getOperand(i+1);
7296 // x86 address has 4 operands: base, index, scale, and displacement
7297 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7298 int valArgIndx = lastAddrIndx + 1;
7300 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7301 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(LoadOpc), t1);
7302 for (int i=0; i <= lastAddrIndx; ++i)
7303 (*MIB).addOperand(*argOpers[i]);
7305 unsigned tt = F->getRegInfo().createVirtualRegister(RC);
7307 MIB = BuildMI(newMBB, dl, TII->get(notOpc), tt).addReg(t1);
7312 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7313 assert((argOpers[valArgIndx]->isReg() ||
7314 argOpers[valArgIndx]->isImm()) &&
7316 if (argOpers[valArgIndx]->isReg())
7317 MIB = BuildMI(newMBB, dl, TII->get(regOpc), t2);
7319 MIB = BuildMI(newMBB, dl, TII->get(immOpc), t2);
7321 (*MIB).addOperand(*argOpers[valArgIndx]);
7323 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), EAXreg);
7326 MIB = BuildMI(newMBB, dl, TII->get(CXchgOpc));
7327 for (int i=0; i <= lastAddrIndx; ++i)
7328 (*MIB).addOperand(*argOpers[i]);
7330 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7331 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7333 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), destOper.getReg());
7337 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7339 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7343 // private utility function: 64 bit atomics on 32 bit host.
7345 X86TargetLowering::EmitAtomicBit6432WithCustomInserter(MachineInstr *bInstr,
7346 MachineBasicBlock *MBB,
7351 bool invSrc) const {
7352 // For the atomic bitwise operator, we generate
7353 // thisMBB (instructions are in pairs, except cmpxchg8b)
7354 // ld t1,t2 = [bitinstr.addr]
7356 // out1, out2 = phi (thisMBB, t1/t2) (newMBB, t3/t4)
7357 // op t5, t6 <- out1, out2, [bitinstr.val]
7358 // (for SWAP, substitute: mov t5, t6 <- [bitinstr.val])
7359 // mov ECX, EBX <- t5, t6
7360 // mov EAX, EDX <- t1, t2
7361 // cmpxchg8b [bitinstr.addr] [EAX, EDX, EBX, ECX implicit]
7362 // mov t3, t4 <- EAX, EDX
7364 // result in out1, out2
7365 // fallthrough -->nextMBB
7367 const TargetRegisterClass *RC = X86::GR32RegisterClass;
7368 const unsigned LoadOpc = X86::MOV32rm;
7369 const unsigned copyOpc = X86::MOV32rr;
7370 const unsigned NotOpc = X86::NOT32r;
7371 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7372 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7373 MachineFunction::iterator MBBIter = MBB;
7376 /// First build the CFG
7377 MachineFunction *F = MBB->getParent();
7378 MachineBasicBlock *thisMBB = MBB;
7379 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7380 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7381 F->insert(MBBIter, newMBB);
7382 F->insert(MBBIter, nextMBB);
7384 // Move all successors to thisMBB to nextMBB
7385 nextMBB->transferSuccessors(thisMBB);
7387 // Update thisMBB to fall through to newMBB
7388 thisMBB->addSuccessor(newMBB);
7390 // newMBB jumps to itself and fall through to nextMBB
7391 newMBB->addSuccessor(nextMBB);
7392 newMBB->addSuccessor(newMBB);
7394 DebugLoc dl = bInstr->getDebugLoc();
7395 // Insert instructions into newMBB based on incoming instruction
7396 // There are 8 "real" operands plus 9 implicit def/uses, ignored here.
7397 assert(bInstr->getNumOperands() < X86AddrNumOperands + 14 &&
7398 "unexpected number of operands");
7399 MachineOperand& dest1Oper = bInstr->getOperand(0);
7400 MachineOperand& dest2Oper = bInstr->getOperand(1);
7401 MachineOperand* argOpers[2 + X86AddrNumOperands];
7402 for (int i=0; i < 2 + X86AddrNumOperands; ++i)
7403 argOpers[i] = &bInstr->getOperand(i+2);
7405 // x86 address has 4 operands: base, index, scale, and displacement
7406 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7408 unsigned t1 = F->getRegInfo().createVirtualRegister(RC);
7409 MachineInstrBuilder MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t1);
7410 for (int i=0; i <= lastAddrIndx; ++i)
7411 (*MIB).addOperand(*argOpers[i]);
7412 unsigned t2 = F->getRegInfo().createVirtualRegister(RC);
7413 MIB = BuildMI(thisMBB, dl, TII->get(LoadOpc), t2);
7414 // add 4 to displacement.
7415 for (int i=0; i <= lastAddrIndx-2; ++i)
7416 (*MIB).addOperand(*argOpers[i]);
7417 MachineOperand newOp3 = *(argOpers[3]);
7419 newOp3.setImm(newOp3.getImm()+4);
7421 newOp3.setOffset(newOp3.getOffset()+4);
7422 (*MIB).addOperand(newOp3);
7423 (*MIB).addOperand(*argOpers[lastAddrIndx]);
7425 // t3/4 are defined later, at the bottom of the loop
7426 unsigned t3 = F->getRegInfo().createVirtualRegister(RC);
7427 unsigned t4 = F->getRegInfo().createVirtualRegister(RC);
7428 BuildMI(newMBB, dl, TII->get(X86::PHI), dest1Oper.getReg())
7429 .addReg(t1).addMBB(thisMBB).addReg(t3).addMBB(newMBB);
7430 BuildMI(newMBB, dl, TII->get(X86::PHI), dest2Oper.getReg())
7431 .addReg(t2).addMBB(thisMBB).addReg(t4).addMBB(newMBB);
7433 unsigned tt1 = F->getRegInfo().createVirtualRegister(RC);
7434 unsigned tt2 = F->getRegInfo().createVirtualRegister(RC);
7436 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt1).addReg(t1);
7437 MIB = BuildMI(newMBB, dl, TII->get(NotOpc), tt2).addReg(t2);
7443 int valArgIndx = lastAddrIndx + 1;
7444 assert((argOpers[valArgIndx]->isReg() ||
7445 argOpers[valArgIndx]->isImm()) &&
7447 unsigned t5 = F->getRegInfo().createVirtualRegister(RC);
7448 unsigned t6 = F->getRegInfo().createVirtualRegister(RC);
7449 if (argOpers[valArgIndx]->isReg())
7450 MIB = BuildMI(newMBB, dl, TII->get(regOpcL), t5);
7452 MIB = BuildMI(newMBB, dl, TII->get(immOpcL), t5);
7453 if (regOpcL != X86::MOV32rr)
7455 (*MIB).addOperand(*argOpers[valArgIndx]);
7456 assert(argOpers[valArgIndx + 1]->isReg() ==
7457 argOpers[valArgIndx]->isReg());
7458 assert(argOpers[valArgIndx + 1]->isImm() ==
7459 argOpers[valArgIndx]->isImm());
7460 if (argOpers[valArgIndx + 1]->isReg())
7461 MIB = BuildMI(newMBB, dl, TII->get(regOpcH), t6);
7463 MIB = BuildMI(newMBB, dl, TII->get(immOpcH), t6);
7464 if (regOpcH != X86::MOV32rr)
7466 (*MIB).addOperand(*argOpers[valArgIndx + 1]);
7468 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EAX);
7470 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EDX);
7473 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::EBX);
7475 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), X86::ECX);
7478 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG8B));
7479 for (int i=0; i <= lastAddrIndx; ++i)
7480 (*MIB).addOperand(*argOpers[i]);
7482 assert(bInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7483 (*MIB).addMemOperand(*F, *bInstr->memoperands_begin());
7485 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t3);
7486 MIB.addReg(X86::EAX);
7487 MIB = BuildMI(newMBB, dl, TII->get(copyOpc), t4);
7488 MIB.addReg(X86::EDX);
7491 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7493 F->DeleteMachineInstr(bInstr); // The pseudo instruction is gone now.
7497 // private utility function
7499 X86TargetLowering::EmitAtomicMinMaxWithCustomInserter(MachineInstr *mInstr,
7500 MachineBasicBlock *MBB,
7501 unsigned cmovOpc) const {
7502 // For the atomic min/max operator, we generate
7505 // ld t1 = [min/max.addr]
7506 // mov t2 = [min/max.val]
7508 // cmov[cond] t2 = t1
7510 // lcs dest = [bitinstr.addr], t2 [EAX is implicit]
7512 // fallthrough -->nextMBB
7514 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7515 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7516 MachineFunction::iterator MBBIter = MBB;
7519 /// First build the CFG
7520 MachineFunction *F = MBB->getParent();
7521 MachineBasicBlock *thisMBB = MBB;
7522 MachineBasicBlock *newMBB = F->CreateMachineBasicBlock(LLVM_BB);
7523 MachineBasicBlock *nextMBB = F->CreateMachineBasicBlock(LLVM_BB);
7524 F->insert(MBBIter, newMBB);
7525 F->insert(MBBIter, nextMBB);
7527 // Move all successors of thisMBB to nextMBB
7528 nextMBB->transferSuccessors(thisMBB);
7530 // Update thisMBB to fall through to newMBB
7531 thisMBB->addSuccessor(newMBB);
7533 // newMBB jumps to newMBB and fall through to nextMBB
7534 newMBB->addSuccessor(nextMBB);
7535 newMBB->addSuccessor(newMBB);
7537 DebugLoc dl = mInstr->getDebugLoc();
7538 // Insert instructions into newMBB based on incoming instruction
7539 assert(mInstr->getNumOperands() < X86AddrNumOperands + 4 &&
7540 "unexpected number of operands");
7541 MachineOperand& destOper = mInstr->getOperand(0);
7542 MachineOperand* argOpers[2 + X86AddrNumOperands];
7543 int numArgs = mInstr->getNumOperands() - 1;
7544 for (int i=0; i < numArgs; ++i)
7545 argOpers[i] = &mInstr->getOperand(i+1);
7547 // x86 address has 4 operands: base, index, scale, and displacement
7548 int lastAddrIndx = X86AddrNumOperands - 1; // [0,3]
7549 int valArgIndx = lastAddrIndx + 1;
7551 unsigned t1 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7552 MachineInstrBuilder MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rm), t1);
7553 for (int i=0; i <= lastAddrIndx; ++i)
7554 (*MIB).addOperand(*argOpers[i]);
7556 // We only support register and immediate values
7557 assert((argOpers[valArgIndx]->isReg() ||
7558 argOpers[valArgIndx]->isImm()) &&
7561 unsigned t2 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7562 if (argOpers[valArgIndx]->isReg())
7563 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7565 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), t2);
7566 (*MIB).addOperand(*argOpers[valArgIndx]);
7568 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), X86::EAX);
7571 MIB = BuildMI(newMBB, dl, TII->get(X86::CMP32rr));
7576 unsigned t3 = F->getRegInfo().createVirtualRegister(X86::GR32RegisterClass);
7577 MIB = BuildMI(newMBB, dl, TII->get(cmovOpc),t3);
7581 // Cmp and exchange if none has modified the memory location
7582 MIB = BuildMI(newMBB, dl, TII->get(X86::LCMPXCHG32));
7583 for (int i=0; i <= lastAddrIndx; ++i)
7584 (*MIB).addOperand(*argOpers[i]);
7586 assert(mInstr->hasOneMemOperand() && "Unexpected number of memoperand");
7587 (*MIB).addMemOperand(*F, *mInstr->memoperands_begin());
7589 MIB = BuildMI(newMBB, dl, TII->get(X86::MOV32rr), destOper.getReg());
7590 MIB.addReg(X86::EAX);
7593 BuildMI(newMBB, dl, TII->get(X86::JNE)).addMBB(newMBB);
7595 F->DeleteMachineInstr(mInstr); // The pseudo instruction is gone now.
7599 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
7600 // all of this code can be replaced with that in the .td file.
7602 X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
7603 unsigned numArgs, bool memArg) const {
7605 MachineFunction *F = BB->getParent();
7606 DebugLoc dl = MI->getDebugLoc();
7607 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7612 Opc = numArgs == 3 ?
7613 X86::PCMPISTRM128rm :
7614 X86::PCMPESTRM128rm;
7616 Opc = numArgs == 3 ?
7617 X86::PCMPISTRM128rr :
7618 X86::PCMPESTRM128rr;
7621 MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
7623 for (unsigned i = 0; i < numArgs; ++i) {
7624 MachineOperand &Op = MI->getOperand(i+1);
7626 if (!(Op.isReg() && Op.isImplicit()))
7630 BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
7633 F->DeleteMachineInstr(MI);
7639 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
7641 MachineBasicBlock *MBB) const {
7642 // Emit code to save XMM registers to the stack. The ABI says that the
7643 // number of registers to save is given in %al, so it's theoretically
7644 // possible to do an indirect jump trick to avoid saving all of them,
7645 // however this code takes a simpler approach and just executes all
7646 // of the stores if %al is non-zero. It's less code, and it's probably
7647 // easier on the hardware branch predictor, and stores aren't all that
7648 // expensive anyway.
7650 // Create the new basic blocks. One block contains all the XMM stores,
7651 // and one block is the final destination regardless of whether any
7652 // stores were performed.
7653 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
7654 MachineFunction *F = MBB->getParent();
7655 MachineFunction::iterator MBBIter = MBB;
7657 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
7658 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
7659 F->insert(MBBIter, XMMSaveMBB);
7660 F->insert(MBBIter, EndMBB);
7663 // Move any original successors of MBB to the end block.
7664 EndMBB->transferSuccessors(MBB);
7665 // The original block will now fall through to the XMM save block.
7666 MBB->addSuccessor(XMMSaveMBB);
7667 // The XMMSaveMBB will fall through to the end block.
7668 XMMSaveMBB->addSuccessor(EndMBB);
7670 // Now add the instructions.
7671 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7672 DebugLoc DL = MI->getDebugLoc();
7674 unsigned CountReg = MI->getOperand(0).getReg();
7675 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
7676 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
7678 if (!Subtarget->isTargetWin64()) {
7679 // If %al is 0, branch around the XMM save block.
7680 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
7681 BuildMI(MBB, DL, TII->get(X86::JE)).addMBB(EndMBB);
7682 MBB->addSuccessor(EndMBB);
7685 // In the XMM save block, save all the XMM argument registers.
7686 for (int i = 3, e = MI->getNumOperands(); i != e; ++i) {
7687 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
7688 BuildMI(XMMSaveMBB, DL, TII->get(X86::MOVAPSmr))
7689 .addFrameIndex(RegSaveFrameIndex)
7690 .addImm(/*Scale=*/1)
7691 .addReg(/*IndexReg=*/0)
7692 .addImm(/*Disp=*/Offset)
7693 .addReg(/*Segment=*/0)
7694 .addReg(MI->getOperand(i).getReg())
7695 .addMemOperand(MachineMemOperand(
7696 PseudoSourceValue::getFixedStack(RegSaveFrameIndex),
7697 MachineMemOperand::MOStore, Offset,
7698 /*Size=*/16, /*Align=*/16));
7701 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7707 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
7708 MachineBasicBlock *BB) const {
7709 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7710 DebugLoc DL = MI->getDebugLoc();
7712 // To "insert" a SELECT_CC instruction, we actually have to insert the
7713 // diamond control-flow pattern. The incoming instruction knows the
7714 // destination vreg to set, the condition code register to branch on, the
7715 // true/false values to select between, and a branch opcode to use.
7716 const BasicBlock *LLVM_BB = BB->getBasicBlock();
7717 MachineFunction::iterator It = BB;
7723 // cmpTY ccX, r1, r2
7725 // fallthrough --> copy0MBB
7726 MachineBasicBlock *thisMBB = BB;
7727 MachineFunction *F = BB->getParent();
7728 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
7729 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
7731 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
7732 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
7733 F->insert(It, copy0MBB);
7734 F->insert(It, sinkMBB);
7735 // Update machine-CFG edges by transferring all successors of the current
7736 // block to the new block which will contain the Phi node for the select.
7737 sinkMBB->transferSuccessors(BB);
7739 // Add the true and fallthrough blocks as its successors.
7740 BB->addSuccessor(copy0MBB);
7741 BB->addSuccessor(sinkMBB);
7744 // %FalseValue = ...
7745 // # fallthrough to sinkMBB
7748 // Update machine-CFG edges
7749 BB->addSuccessor(sinkMBB);
7752 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
7755 BuildMI(BB, DL, TII->get(X86::PHI), MI->getOperand(0).getReg())
7756 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
7757 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
7759 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7765 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
7766 MachineBasicBlock *BB) const {
7767 switch (MI->getOpcode()) {
7768 default: assert(false && "Unexpected instr type to insert");
7770 case X86::CMOV_V1I64:
7771 case X86::CMOV_FR32:
7772 case X86::CMOV_FR64:
7773 case X86::CMOV_V4F32:
7774 case X86::CMOV_V2F64:
7775 case X86::CMOV_V2I64:
7776 return EmitLoweredSelect(MI, BB);
7778 case X86::FP32_TO_INT16_IN_MEM:
7779 case X86::FP32_TO_INT32_IN_MEM:
7780 case X86::FP32_TO_INT64_IN_MEM:
7781 case X86::FP64_TO_INT16_IN_MEM:
7782 case X86::FP64_TO_INT32_IN_MEM:
7783 case X86::FP64_TO_INT64_IN_MEM:
7784 case X86::FP80_TO_INT16_IN_MEM:
7785 case X86::FP80_TO_INT32_IN_MEM:
7786 case X86::FP80_TO_INT64_IN_MEM: {
7787 const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
7788 DebugLoc DL = MI->getDebugLoc();
7790 // Change the floating point control register to use "round towards zero"
7791 // mode when truncating to an integer value.
7792 MachineFunction *F = BB->getParent();
7793 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
7794 addFrameReference(BuildMI(BB, DL, TII->get(X86::FNSTCW16m)), CWFrameIdx);
7796 // Load the old value of the high byte of the control word...
7798 F->getRegInfo().createVirtualRegister(X86::GR16RegisterClass);
7799 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16rm), OldCW),
7802 // Set the high part to be round to zero...
7803 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
7806 // Reload the modified control word now...
7807 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
7809 // Restore the memory image of control word to original value
7810 addFrameReference(BuildMI(BB, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
7813 // Get the X86 opcode to use.
7815 switch (MI->getOpcode()) {
7816 default: llvm_unreachable("illegal opcode!");
7817 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
7818 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
7819 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
7820 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
7821 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
7822 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
7823 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
7824 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
7825 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
7829 MachineOperand &Op = MI->getOperand(0);
7831 AM.BaseType = X86AddressMode::RegBase;
7832 AM.Base.Reg = Op.getReg();
7834 AM.BaseType = X86AddressMode::FrameIndexBase;
7835 AM.Base.FrameIndex = Op.getIndex();
7837 Op = MI->getOperand(1);
7839 AM.Scale = Op.getImm();
7840 Op = MI->getOperand(2);
7842 AM.IndexReg = Op.getImm();
7843 Op = MI->getOperand(3);
7844 if (Op.isGlobal()) {
7845 AM.GV = Op.getGlobal();
7847 AM.Disp = Op.getImm();
7849 addFullAddress(BuildMI(BB, DL, TII->get(Opc)), AM)
7850 .addReg(MI->getOperand(X86AddrNumOperands).getReg());
7852 // Reload the original control word now.
7853 addFrameReference(BuildMI(BB, DL, TII->get(X86::FLDCW16m)), CWFrameIdx);
7855 F->DeleteMachineInstr(MI); // The pseudo instruction is gone now.
7858 // String/text processing lowering.
7859 case X86::PCMPISTRM128REG:
7860 return EmitPCMP(MI, BB, 3, false /* in-mem */);
7861 case X86::PCMPISTRM128MEM:
7862 return EmitPCMP(MI, BB, 3, true /* in-mem */);
7863 case X86::PCMPESTRM128REG:
7864 return EmitPCMP(MI, BB, 5, false /* in mem */);
7865 case X86::PCMPESTRM128MEM:
7866 return EmitPCMP(MI, BB, 5, true /* in mem */);
7869 case X86::ATOMAND32:
7870 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7871 X86::AND32ri, X86::MOV32rm,
7872 X86::LCMPXCHG32, X86::MOV32rr,
7873 X86::NOT32r, X86::EAX,
7874 X86::GR32RegisterClass);
7876 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR32rr,
7877 X86::OR32ri, X86::MOV32rm,
7878 X86::LCMPXCHG32, X86::MOV32rr,
7879 X86::NOT32r, X86::EAX,
7880 X86::GR32RegisterClass);
7881 case X86::ATOMXOR32:
7882 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR32rr,
7883 X86::XOR32ri, X86::MOV32rm,
7884 X86::LCMPXCHG32, X86::MOV32rr,
7885 X86::NOT32r, X86::EAX,
7886 X86::GR32RegisterClass);
7887 case X86::ATOMNAND32:
7888 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
7889 X86::AND32ri, X86::MOV32rm,
7890 X86::LCMPXCHG32, X86::MOV32rr,
7891 X86::NOT32r, X86::EAX,
7892 X86::GR32RegisterClass, true);
7893 case X86::ATOMMIN32:
7894 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL32rr);
7895 case X86::ATOMMAX32:
7896 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG32rr);
7897 case X86::ATOMUMIN32:
7898 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB32rr);
7899 case X86::ATOMUMAX32:
7900 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA32rr);
7902 case X86::ATOMAND16:
7903 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7904 X86::AND16ri, X86::MOV16rm,
7905 X86::LCMPXCHG16, X86::MOV16rr,
7906 X86::NOT16r, X86::AX,
7907 X86::GR16RegisterClass);
7909 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR16rr,
7910 X86::OR16ri, X86::MOV16rm,
7911 X86::LCMPXCHG16, X86::MOV16rr,
7912 X86::NOT16r, X86::AX,
7913 X86::GR16RegisterClass);
7914 case X86::ATOMXOR16:
7915 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR16rr,
7916 X86::XOR16ri, X86::MOV16rm,
7917 X86::LCMPXCHG16, X86::MOV16rr,
7918 X86::NOT16r, X86::AX,
7919 X86::GR16RegisterClass);
7920 case X86::ATOMNAND16:
7921 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND16rr,
7922 X86::AND16ri, X86::MOV16rm,
7923 X86::LCMPXCHG16, X86::MOV16rr,
7924 X86::NOT16r, X86::AX,
7925 X86::GR16RegisterClass, true);
7926 case X86::ATOMMIN16:
7927 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL16rr);
7928 case X86::ATOMMAX16:
7929 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG16rr);
7930 case X86::ATOMUMIN16:
7931 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB16rr);
7932 case X86::ATOMUMAX16:
7933 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA16rr);
7936 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7937 X86::AND8ri, X86::MOV8rm,
7938 X86::LCMPXCHG8, X86::MOV8rr,
7939 X86::NOT8r, X86::AL,
7940 X86::GR8RegisterClass);
7942 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR8rr,
7943 X86::OR8ri, X86::MOV8rm,
7944 X86::LCMPXCHG8, X86::MOV8rr,
7945 X86::NOT8r, X86::AL,
7946 X86::GR8RegisterClass);
7948 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR8rr,
7949 X86::XOR8ri, X86::MOV8rm,
7950 X86::LCMPXCHG8, X86::MOV8rr,
7951 X86::NOT8r, X86::AL,
7952 X86::GR8RegisterClass);
7953 case X86::ATOMNAND8:
7954 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND8rr,
7955 X86::AND8ri, X86::MOV8rm,
7956 X86::LCMPXCHG8, X86::MOV8rr,
7957 X86::NOT8r, X86::AL,
7958 X86::GR8RegisterClass, true);
7959 // FIXME: There are no CMOV8 instructions; MIN/MAX need some other way.
7960 // This group is for 64-bit host.
7961 case X86::ATOMAND64:
7962 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7963 X86::AND64ri32, X86::MOV64rm,
7964 X86::LCMPXCHG64, X86::MOV64rr,
7965 X86::NOT64r, X86::RAX,
7966 X86::GR64RegisterClass);
7968 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::OR64rr,
7969 X86::OR64ri32, X86::MOV64rm,
7970 X86::LCMPXCHG64, X86::MOV64rr,
7971 X86::NOT64r, X86::RAX,
7972 X86::GR64RegisterClass);
7973 case X86::ATOMXOR64:
7974 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::XOR64rr,
7975 X86::XOR64ri32, X86::MOV64rm,
7976 X86::LCMPXCHG64, X86::MOV64rr,
7977 X86::NOT64r, X86::RAX,
7978 X86::GR64RegisterClass);
7979 case X86::ATOMNAND64:
7980 return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND64rr,
7981 X86::AND64ri32, X86::MOV64rm,
7982 X86::LCMPXCHG64, X86::MOV64rr,
7983 X86::NOT64r, X86::RAX,
7984 X86::GR64RegisterClass, true);
7985 case X86::ATOMMIN64:
7986 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVL64rr);
7987 case X86::ATOMMAX64:
7988 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVG64rr);
7989 case X86::ATOMUMIN64:
7990 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVB64rr);
7991 case X86::ATOMUMAX64:
7992 return EmitAtomicMinMaxWithCustomInserter(MI, BB, X86::CMOVA64rr);
7994 // This group does 64-bit operations on a 32-bit host.
7995 case X86::ATOMAND6432:
7996 return EmitAtomicBit6432WithCustomInserter(MI, BB,
7997 X86::AND32rr, X86::AND32rr,
7998 X86::AND32ri, X86::AND32ri,
8000 case X86::ATOMOR6432:
8001 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8002 X86::OR32rr, X86::OR32rr,
8003 X86::OR32ri, X86::OR32ri,
8005 case X86::ATOMXOR6432:
8006 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8007 X86::XOR32rr, X86::XOR32rr,
8008 X86::XOR32ri, X86::XOR32ri,
8010 case X86::ATOMNAND6432:
8011 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8012 X86::AND32rr, X86::AND32rr,
8013 X86::AND32ri, X86::AND32ri,
8015 case X86::ATOMADD6432:
8016 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8017 X86::ADD32rr, X86::ADC32rr,
8018 X86::ADD32ri, X86::ADC32ri,
8020 case X86::ATOMSUB6432:
8021 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8022 X86::SUB32rr, X86::SBB32rr,
8023 X86::SUB32ri, X86::SBB32ri,
8025 case X86::ATOMSWAP6432:
8026 return EmitAtomicBit6432WithCustomInserter(MI, BB,
8027 X86::MOV32rr, X86::MOV32rr,
8028 X86::MOV32ri, X86::MOV32ri,
8030 case X86::VASTART_SAVE_XMM_REGS:
8031 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
8035 //===----------------------------------------------------------------------===//
8036 // X86 Optimization Hooks
8037 //===----------------------------------------------------------------------===//
8039 void X86TargetLowering::computeMaskedBitsForTargetNode(const SDValue Op,
8043 const SelectionDAG &DAG,
8044 unsigned Depth) const {
8045 unsigned Opc = Op.getOpcode();
8046 assert((Opc >= ISD::BUILTIN_OP_END ||
8047 Opc == ISD::INTRINSIC_WO_CHAIN ||
8048 Opc == ISD::INTRINSIC_W_CHAIN ||
8049 Opc == ISD::INTRINSIC_VOID) &&
8050 "Should use MaskedValueIsZero if you don't know whether Op"
8051 " is a target node!");
8053 KnownZero = KnownOne = APInt(Mask.getBitWidth(), 0); // Don't know anything.
8062 // These nodes' second result is a boolean.
8063 if (Op.getResNo() == 0)
8067 KnownZero |= APInt::getHighBitsSet(Mask.getBitWidth(),
8068 Mask.getBitWidth() - 1);
8073 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
8074 /// node is a GlobalAddress + offset.
8075 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
8076 GlobalValue* &GA, int64_t &Offset) const{
8077 if (N->getOpcode() == X86ISD::Wrapper) {
8078 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
8079 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
8080 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
8084 return TargetLowering::isGAPlusOffset(N, GA, Offset);
8087 static bool isBaseAlignmentOfN(unsigned N, SDNode *Base,
8088 const TargetLowering &TLI) {
8091 if (TLI.isGAPlusOffset(Base, GV, Offset))
8092 return (GV->getAlignment() >= N && (Offset % N) == 0);
8093 // DAG combine handles the stack object case.
8097 static bool EltsFromConsecutiveLoads(ShuffleVectorSDNode *N, unsigned NumElems,
8098 EVT EVT, LoadSDNode *&LDBase,
8099 unsigned &LastLoadedElt,
8100 SelectionDAG &DAG, MachineFrameInfo *MFI,
8101 const TargetLowering &TLI) {
8103 LastLoadedElt = -1U;
8104 for (unsigned i = 0; i < NumElems; ++i) {
8105 if (N->getMaskElt(i) < 0) {
8111 SDValue Elt = DAG.getShuffleScalarElt(N, i);
8112 if (!Elt.getNode() ||
8113 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
8116 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
8118 LDBase = cast<LoadSDNode>(Elt.getNode());
8122 if (Elt.getOpcode() == ISD::UNDEF)
8125 LoadSDNode *LD = cast<LoadSDNode>(Elt);
8126 if (!TLI.isConsecutiveLoad(LD, LDBase, EVT.getSizeInBits()/8, i, MFI))
8133 /// PerformShuffleCombine - Combine a vector_shuffle that is equal to
8134 /// build_vector load1, load2, load3, load4, <0, 1, 2, 3> into a 128-bit load
8135 /// if the load addresses are consecutive, non-overlapping, and in the right
8136 /// order. In the case of v2i64, it will see if it can rewrite the
8137 /// shuffle to be an appropriate build vector so it can take advantage of
8138 // performBuildVectorCombine.
8139 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
8140 const TargetLowering &TLI) {
8141 DebugLoc dl = N->getDebugLoc();
8142 EVT VT = N->getValueType(0);
8143 EVT EVT = VT.getVectorElementType();
8144 ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
8145 unsigned NumElems = VT.getVectorNumElements();
8147 if (VT.getSizeInBits() != 128)
8150 // Try to combine a vector_shuffle into a 128-bit load.
8151 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
8152 LoadSDNode *LD = NULL;
8153 unsigned LastLoadedElt;
8154 if (!EltsFromConsecutiveLoads(SVN, NumElems, EVT, LD, LastLoadedElt, DAG,
8158 if (LastLoadedElt == NumElems - 1) {
8159 if (isBaseAlignmentOfN(16, LD->getBasePtr().getNode(), TLI))
8160 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8161 LD->getSrcValue(), LD->getSrcValueOffset(),
8163 return DAG.getLoad(VT, dl, LD->getChain(), LD->getBasePtr(),
8164 LD->getSrcValue(), LD->getSrcValueOffset(),
8165 LD->isVolatile(), LD->getAlignment());
8166 } else if (NumElems == 4 && LastLoadedElt == 1) {
8167 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
8168 SDValue Ops[] = { LD->getChain(), LD->getBasePtr() };
8169 SDValue ResNode = DAG.getNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops, 2);
8170 return DAG.getNode(ISD::BIT_CONVERT, dl, VT, ResNode);
8175 /// PerformSELECTCombine - Do target-specific dag combines on SELECT nodes.
8176 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
8177 const X86Subtarget *Subtarget) {
8178 DebugLoc DL = N->getDebugLoc();
8179 SDValue Cond = N->getOperand(0);
8180 // Get the LHS/RHS of the select.
8181 SDValue LHS = N->getOperand(1);
8182 SDValue RHS = N->getOperand(2);
8184 // If we have SSE[12] support, try to form min/max nodes.
8185 if (Subtarget->hasSSE2() &&
8186 (LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64) &&
8187 Cond.getOpcode() == ISD::SETCC) {
8188 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
8190 unsigned Opcode = 0;
8191 if (LHS == Cond.getOperand(0) && RHS == Cond.getOperand(1)) {
8194 case ISD::SETOLE: // (X <= Y) ? X : Y -> min
8197 if (!UnsafeFPMath) break;
8199 case ISD::SETOLT: // (X olt/lt Y) ? X : Y -> min
8201 Opcode = X86ISD::FMIN;
8204 case ISD::SETOGT: // (X > Y) ? X : Y -> max
8207 if (!UnsafeFPMath) break;
8209 case ISD::SETUGE: // (X uge/ge Y) ? X : Y -> max
8211 Opcode = X86ISD::FMAX;
8214 } else if (LHS == Cond.getOperand(1) && RHS == Cond.getOperand(0)) {
8217 case ISD::SETOGT: // (X > Y) ? Y : X -> min
8220 if (!UnsafeFPMath) break;
8222 case ISD::SETUGE: // (X uge/ge Y) ? Y : X -> min
8224 Opcode = X86ISD::FMIN;
8227 case ISD::SETOLE: // (X <= Y) ? Y : X -> max
8230 if (!UnsafeFPMath) break;
8232 case ISD::SETOLT: // (X olt/lt Y) ? Y : X -> max
8234 Opcode = X86ISD::FMAX;
8240 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
8243 // If this is a select between two integer constants, try to do some
8245 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
8246 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
8247 // Don't do this for crazy integer types.
8248 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
8249 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
8250 // so that TrueC (the true value) is larger than FalseC.
8251 bool NeedsCondInvert = false;
8253 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
8254 // Efficiently invertible.
8255 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
8256 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
8257 isa<ConstantSDNode>(Cond.getOperand(1))))) {
8258 NeedsCondInvert = true;
8259 std::swap(TrueC, FalseC);
8262 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
8263 if (FalseC->getAPIntValue() == 0 &&
8264 TrueC->getAPIntValue().isPowerOf2()) {
8265 if (NeedsCondInvert) // Invert the condition if needed.
8266 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8267 DAG.getConstant(1, Cond.getValueType()));
8269 // Zero extend the condition if needed.
8270 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
8272 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8273 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
8274 DAG.getConstant(ShAmt, MVT::i8));
8277 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
8278 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8279 if (NeedsCondInvert) // Invert the condition if needed.
8280 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8281 DAG.getConstant(1, Cond.getValueType()));
8283 // Zero extend the condition if needed.
8284 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8285 FalseC->getValueType(0), Cond);
8286 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8287 SDValue(FalseC, 0));
8290 // Optimize cases that will turn into an LEA instruction. This requires
8291 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8292 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8293 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8294 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8296 bool isFastMultiplier = false;
8298 switch ((unsigned char)Diff) {
8300 case 1: // result = add base, cond
8301 case 2: // result = lea base( , cond*2)
8302 case 3: // result = lea base(cond, cond*2)
8303 case 4: // result = lea base( , cond*4)
8304 case 5: // result = lea base(cond, cond*4)
8305 case 8: // result = lea base( , cond*8)
8306 case 9: // result = lea base(cond, cond*8)
8307 isFastMultiplier = true;
8312 if (isFastMultiplier) {
8313 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8314 if (NeedsCondInvert) // Invert the condition if needed.
8315 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
8316 DAG.getConstant(1, Cond.getValueType()));
8318 // Zero extend the condition if needed.
8319 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8321 // Scale the condition by the difference.
8323 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8324 DAG.getConstant(Diff, Cond.getValueType()));
8326 // Add the base if non-zero.
8327 if (FalseC->getAPIntValue() != 0)
8328 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8329 SDValue(FalseC, 0));
8339 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
8340 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
8341 TargetLowering::DAGCombinerInfo &DCI) {
8342 DebugLoc DL = N->getDebugLoc();
8344 // If the flag operand isn't dead, don't touch this CMOV.
8345 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
8348 // If this is a select between two integer constants, try to do some
8349 // optimizations. Note that the operands are ordered the opposite of SELECT
8351 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(N->getOperand(1))) {
8352 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
8353 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
8354 // larger than FalseC (the false value).
8355 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
8357 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
8358 CC = X86::GetOppositeBranchCondition(CC);
8359 std::swap(TrueC, FalseC);
8362 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
8363 // This is efficient for any integer data type (including i8/i16) and
8365 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
8366 SDValue Cond = N->getOperand(3);
8367 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8368 DAG.getConstant(CC, MVT::i8), Cond);
8370 // Zero extend the condition if needed.
8371 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
8373 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
8374 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
8375 DAG.getConstant(ShAmt, MVT::i8));
8376 if (N->getNumValues() == 2) // Dead flag value?
8377 return DCI.CombineTo(N, Cond, SDValue());
8381 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
8382 // for any integer data type, including i8/i16.
8383 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
8384 SDValue Cond = N->getOperand(3);
8385 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8386 DAG.getConstant(CC, MVT::i8), Cond);
8388 // Zero extend the condition if needed.
8389 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
8390 FalseC->getValueType(0), Cond);
8391 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8392 SDValue(FalseC, 0));
8394 if (N->getNumValues() == 2) // Dead flag value?
8395 return DCI.CombineTo(N, Cond, SDValue());
8399 // Optimize cases that will turn into an LEA instruction. This requires
8400 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
8401 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
8402 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
8403 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
8405 bool isFastMultiplier = false;
8407 switch ((unsigned char)Diff) {
8409 case 1: // result = add base, cond
8410 case 2: // result = lea base( , cond*2)
8411 case 3: // result = lea base(cond, cond*2)
8412 case 4: // result = lea base( , cond*4)
8413 case 5: // result = lea base(cond, cond*4)
8414 case 8: // result = lea base( , cond*8)
8415 case 9: // result = lea base(cond, cond*8)
8416 isFastMultiplier = true;
8421 if (isFastMultiplier) {
8422 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
8423 SDValue Cond = N->getOperand(3);
8424 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
8425 DAG.getConstant(CC, MVT::i8), Cond);
8426 // Zero extend the condition if needed.
8427 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
8429 // Scale the condition by the difference.
8431 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
8432 DAG.getConstant(Diff, Cond.getValueType()));
8434 // Add the base if non-zero.
8435 if (FalseC->getAPIntValue() != 0)
8436 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
8437 SDValue(FalseC, 0));
8438 if (N->getNumValues() == 2) // Dead flag value?
8439 return DCI.CombineTo(N, Cond, SDValue());
8449 /// PerformMulCombine - Optimize a single multiply with constant into two
8450 /// in order to implement it with two cheaper instructions, e.g.
8451 /// LEA + SHL, LEA + LEA.
8452 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
8453 TargetLowering::DAGCombinerInfo &DCI) {
8454 if (DAG.getMachineFunction().
8455 getFunction()->hasFnAttr(Attribute::OptimizeForSize))
8458 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
8461 EVT VT = N->getValueType(0);
8465 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
8468 uint64_t MulAmt = C->getZExtValue();
8469 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
8472 uint64_t MulAmt1 = 0;
8473 uint64_t MulAmt2 = 0;
8474 if ((MulAmt % 9) == 0) {
8476 MulAmt2 = MulAmt / 9;
8477 } else if ((MulAmt % 5) == 0) {
8479 MulAmt2 = MulAmt / 5;
8480 } else if ((MulAmt % 3) == 0) {
8482 MulAmt2 = MulAmt / 3;
8485 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
8486 DebugLoc DL = N->getDebugLoc();
8488 if (isPowerOf2_64(MulAmt2) &&
8489 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
8490 // If second multiplifer is pow2, issue it first. We want the multiply by
8491 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
8493 std::swap(MulAmt1, MulAmt2);
8496 if (isPowerOf2_64(MulAmt1))
8497 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
8498 DAG.getConstant(Log2_64(MulAmt1), MVT::i8));
8500 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
8501 DAG.getConstant(MulAmt1, VT));
8503 if (isPowerOf2_64(MulAmt2))
8504 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
8505 DAG.getConstant(Log2_64(MulAmt2), MVT::i8));
8507 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
8508 DAG.getConstant(MulAmt2, VT));
8510 // Do not add new nodes to DAG combiner worklist.
8511 DCI.CombineTo(N, NewMul, false);
8517 /// PerformShiftCombine - Transforms vector shift nodes to use vector shifts
8519 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
8520 const X86Subtarget *Subtarget) {
8521 // On X86 with SSE2 support, we can transform this to a vector shift if
8522 // all elements are shifted by the same amount. We can't do this in legalize
8523 // because the a constant vector is typically transformed to a constant pool
8524 // so we have no knowledge of the shift amount.
8525 if (!Subtarget->hasSSE2())
8528 EVT VT = N->getValueType(0);
8529 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
8532 SDValue ShAmtOp = N->getOperand(1);
8533 EVT EltVT = VT.getVectorElementType();
8534 DebugLoc DL = N->getDebugLoc();
8536 if (ShAmtOp.getOpcode() == ISD::BUILD_VECTOR) {
8537 unsigned NumElts = VT.getVectorNumElements();
8539 for (; i != NumElts; ++i) {
8540 SDValue Arg = ShAmtOp.getOperand(i);
8541 if (Arg.getOpcode() == ISD::UNDEF) continue;
8545 for (; i != NumElts; ++i) {
8546 SDValue Arg = ShAmtOp.getOperand(i);
8547 if (Arg.getOpcode() == ISD::UNDEF) continue;
8548 if (Arg != BaseShAmt) {
8552 } else if (ShAmtOp.getOpcode() == ISD::VECTOR_SHUFFLE &&
8553 cast<ShuffleVectorSDNode>(ShAmtOp)->isSplat()) {
8554 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, EltVT, ShAmtOp,
8555 DAG.getIntPtrConstant(0));
8559 if (EltVT.bitsGT(MVT::i32))
8560 BaseShAmt = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, BaseShAmt);
8561 else if (EltVT.bitsLT(MVT::i32))
8562 BaseShAmt = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, BaseShAmt);
8564 // The shift amount is identical so we can do a vector shift.
8565 SDValue ValOp = N->getOperand(0);
8566 switch (N->getOpcode()) {
8568 llvm_unreachable("Unknown shift opcode!");
8571 if (VT == MVT::v2i64)
8572 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8573 DAG.getConstant(Intrinsic::x86_sse2_pslli_q, MVT::i32),
8575 if (VT == MVT::v4i32)
8576 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8577 DAG.getConstant(Intrinsic::x86_sse2_pslli_d, MVT::i32),
8579 if (VT == MVT::v8i16)
8580 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8581 DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32),
8585 if (VT == MVT::v4i32)
8586 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8587 DAG.getConstant(Intrinsic::x86_sse2_psrai_d, MVT::i32),
8589 if (VT == MVT::v8i16)
8590 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8591 DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
8595 if (VT == MVT::v2i64)
8596 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8597 DAG.getConstant(Intrinsic::x86_sse2_psrli_q, MVT::i32),
8599 if (VT == MVT::v4i32)
8600 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8601 DAG.getConstant(Intrinsic::x86_sse2_psrli_d, MVT::i32),
8603 if (VT == MVT::v8i16)
8604 return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, VT,
8605 DAG.getConstant(Intrinsic::x86_sse2_psrli_w, MVT::i32),
8612 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
8613 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
8614 const X86Subtarget *Subtarget) {
8615 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
8616 // the FP state in cases where an emms may be missing.
8617 // A preferable solution to the general problem is to figure out the right
8618 // places to insert EMMS. This qualifies as a quick hack.
8620 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
8621 StoreSDNode *St = cast<StoreSDNode>(N);
8622 EVT VT = St->getValue().getValueType();
8623 if (VT.getSizeInBits() != 64)
8626 const Function *F = DAG.getMachineFunction().getFunction();
8627 bool NoImplicitFloatOps = F->hasFnAttr(Attribute::NoImplicitFloat);
8628 bool F64IsLegal = !UseSoftFloat && !NoImplicitFloatOps
8629 && Subtarget->hasSSE2();
8630 if ((VT.isVector() ||
8631 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
8632 isa<LoadSDNode>(St->getValue()) &&
8633 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
8634 St->getChain().hasOneUse() && !St->isVolatile()) {
8635 SDNode* LdVal = St->getValue().getNode();
8637 int TokenFactorIndex = -1;
8638 SmallVector<SDValue, 8> Ops;
8639 SDNode* ChainVal = St->getChain().getNode();
8640 // Must be a store of a load. We currently handle two cases: the load
8641 // is a direct child, and it's under an intervening TokenFactor. It is
8642 // possible to dig deeper under nested TokenFactors.
8643 if (ChainVal == LdVal)
8644 Ld = cast<LoadSDNode>(St->getChain());
8645 else if (St->getValue().hasOneUse() &&
8646 ChainVal->getOpcode() == ISD::TokenFactor) {
8647 for (unsigned i=0, e = ChainVal->getNumOperands(); i != e; ++i) {
8648 if (ChainVal->getOperand(i).getNode() == LdVal) {
8649 TokenFactorIndex = i;
8650 Ld = cast<LoadSDNode>(St->getValue());
8652 Ops.push_back(ChainVal->getOperand(i));
8656 if (!Ld || !ISD::isNormalLoad(Ld))
8659 // If this is not the MMX case, i.e. we are just turning i64 load/store
8660 // into f64 load/store, avoid the transformation if there are multiple
8661 // uses of the loaded value.
8662 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
8665 DebugLoc LdDL = Ld->getDebugLoc();
8666 DebugLoc StDL = N->getDebugLoc();
8667 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
8668 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
8670 if (Subtarget->is64Bit() || F64IsLegal) {
8671 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
8672 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(),
8673 Ld->getBasePtr(), Ld->getSrcValue(),
8674 Ld->getSrcValueOffset(), Ld->isVolatile(),
8675 Ld->getAlignment());
8676 SDValue NewChain = NewLd.getValue(1);
8677 if (TokenFactorIndex != -1) {
8678 Ops.push_back(NewChain);
8679 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8682 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
8683 St->getSrcValue(), St->getSrcValueOffset(),
8684 St->isVolatile(), St->getAlignment());
8687 // Otherwise, lower to two pairs of 32-bit loads / stores.
8688 SDValue LoAddr = Ld->getBasePtr();
8689 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
8690 DAG.getConstant(4, MVT::i32));
8692 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
8693 Ld->getSrcValue(), Ld->getSrcValueOffset(),
8694 Ld->isVolatile(), Ld->getAlignment());
8695 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
8696 Ld->getSrcValue(), Ld->getSrcValueOffset()+4,
8698 MinAlign(Ld->getAlignment(), 4));
8700 SDValue NewChain = LoLd.getValue(1);
8701 if (TokenFactorIndex != -1) {
8702 Ops.push_back(LoLd);
8703 Ops.push_back(HiLd);
8704 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, &Ops[0],
8708 LoAddr = St->getBasePtr();
8709 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
8710 DAG.getConstant(4, MVT::i32));
8712 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
8713 St->getSrcValue(), St->getSrcValueOffset(),
8714 St->isVolatile(), St->getAlignment());
8715 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
8717 St->getSrcValueOffset() + 4,
8719 MinAlign(St->getAlignment(), 4));
8720 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
8725 /// PerformFORCombine - Do target-specific dag combines on X86ISD::FOR and
8726 /// X86ISD::FXOR nodes.
8727 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
8728 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
8729 // F[X]OR(0.0, x) -> x
8730 // F[X]OR(x, 0.0) -> x
8731 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8732 if (C->getValueAPF().isPosZero())
8733 return N->getOperand(1);
8734 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8735 if (C->getValueAPF().isPosZero())
8736 return N->getOperand(0);
8740 /// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
8741 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
8742 // FAND(0.0, x) -> 0.0
8743 // FAND(x, 0.0) -> 0.0
8744 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
8745 if (C->getValueAPF().isPosZero())
8746 return N->getOperand(0);
8747 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
8748 if (C->getValueAPF().isPosZero())
8749 return N->getOperand(1);
8753 static SDValue PerformBTCombine(SDNode *N,
8755 TargetLowering::DAGCombinerInfo &DCI) {
8756 // BT ignores high bits in the bit index operand.
8757 SDValue Op1 = N->getOperand(1);
8758 if (Op1.hasOneUse()) {
8759 unsigned BitWidth = Op1.getValueSizeInBits();
8760 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
8761 APInt KnownZero, KnownOne;
8762 TargetLowering::TargetLoweringOpt TLO(DAG);
8763 TargetLowering &TLI = DAG.getTargetLoweringInfo();
8764 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
8765 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
8766 DCI.CommitTargetLoweringOpt(TLO);
8771 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
8772 SDValue Op = N->getOperand(0);
8773 if (Op.getOpcode() == ISD::BIT_CONVERT)
8774 Op = Op.getOperand(0);
8775 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
8776 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
8777 VT.getVectorElementType().getSizeInBits() ==
8778 OpVT.getVectorElementType().getSizeInBits()) {
8779 return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
8784 // On X86 and X86-64, atomic operations are lowered to locked instructions.
8785 // Locked instructions, in turn, have implicit fence semantics (all memory
8786 // operations are flushed before issuing the locked instruction, and the
8787 // are not buffered), so we can fold away the common pattern of
8788 // fence-atomic-fence.
8789 static SDValue PerformMEMBARRIERCombine(SDNode* N, SelectionDAG &DAG) {
8790 SDValue atomic = N->getOperand(0);
8791 switch (atomic.getOpcode()) {
8792 case ISD::ATOMIC_CMP_SWAP:
8793 case ISD::ATOMIC_SWAP:
8794 case ISD::ATOMIC_LOAD_ADD:
8795 case ISD::ATOMIC_LOAD_SUB:
8796 case ISD::ATOMIC_LOAD_AND:
8797 case ISD::ATOMIC_LOAD_OR:
8798 case ISD::ATOMIC_LOAD_XOR:
8799 case ISD::ATOMIC_LOAD_NAND:
8800 case ISD::ATOMIC_LOAD_MIN:
8801 case ISD::ATOMIC_LOAD_MAX:
8802 case ISD::ATOMIC_LOAD_UMIN:
8803 case ISD::ATOMIC_LOAD_UMAX:
8809 SDValue fence = atomic.getOperand(0);
8810 if (fence.getOpcode() != ISD::MEMBARRIER)
8813 switch (atomic.getOpcode()) {
8814 case ISD::ATOMIC_CMP_SWAP:
8815 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8816 atomic.getOperand(1), atomic.getOperand(2),
8817 atomic.getOperand(3));
8818 case ISD::ATOMIC_SWAP:
8819 case ISD::ATOMIC_LOAD_ADD:
8820 case ISD::ATOMIC_LOAD_SUB:
8821 case ISD::ATOMIC_LOAD_AND:
8822 case ISD::ATOMIC_LOAD_OR:
8823 case ISD::ATOMIC_LOAD_XOR:
8824 case ISD::ATOMIC_LOAD_NAND:
8825 case ISD::ATOMIC_LOAD_MIN:
8826 case ISD::ATOMIC_LOAD_MAX:
8827 case ISD::ATOMIC_LOAD_UMIN:
8828 case ISD::ATOMIC_LOAD_UMAX:
8829 return DAG.UpdateNodeOperands(atomic, fence.getOperand(0),
8830 atomic.getOperand(1), atomic.getOperand(2));
8836 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
8837 DAGCombinerInfo &DCI) const {
8838 SelectionDAG &DAG = DCI.DAG;
8839 switch (N->getOpcode()) {
8841 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
8842 case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
8843 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
8844 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
8847 case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
8848 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
8850 case X86ISD::FOR: return PerformFORCombine(N, DAG);
8851 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
8852 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
8853 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
8854 case ISD::MEMBARRIER: return PerformMEMBARRIERCombine(N, DAG);
8860 //===----------------------------------------------------------------------===//
8861 // X86 Inline Assembly Support
8862 //===----------------------------------------------------------------------===//
8864 static bool LowerToBSwap(CallInst *CI) {
8865 // FIXME: this should verify that we are targetting a 486 or better. If not,
8866 // we will turn this bswap into something that will be lowered to logical ops
8867 // instead of emitting the bswap asm. For now, we don't support 486 or lower
8868 // so don't worry about this.
8870 // Verify this is a simple bswap.
8871 if (CI->getNumOperands() != 2 ||
8872 CI->getType() != CI->getOperand(1)->getType() ||
8873 !CI->getType()->isInteger())
8876 const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
8877 if (!Ty || Ty->getBitWidth() % 16 != 0)
8880 // Okay, we can do this xform, do so now.
8881 const Type *Tys[] = { Ty };
8882 Module *M = CI->getParent()->getParent()->getParent();
8883 Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
8885 Value *Op = CI->getOperand(1);
8886 Op = CallInst::Create(Int, Op, CI->getName(), CI);
8888 CI->replaceAllUsesWith(Op);
8889 CI->eraseFromParent();
8893 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
8894 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
8895 std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
8897 std::string AsmStr = IA->getAsmString();
8899 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
8900 std::vector<std::string> AsmPieces;
8901 SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
8903 switch (AsmPieces.size()) {
8904 default: return false;
8906 AsmStr = AsmPieces[0];
8908 SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
8911 if (AsmPieces.size() == 2 &&
8912 (AsmPieces[0] == "bswap" ||
8913 AsmPieces[0] == "bswapq" ||
8914 AsmPieces[0] == "bswapl") &&
8915 (AsmPieces[1] == "$0" ||
8916 AsmPieces[1] == "${0:q}")) {
8917 // No need to check constraints, nothing other than the equivalent of
8918 // "=r,0" would be valid here.
8919 return LowerToBSwap(CI);
8921 // rorw $$8, ${0:w} --> llvm.bswap.i16
8922 if (CI->getType() == Type::getInt16Ty(CI->getContext()) &&
8923 AsmPieces.size() == 3 &&
8924 AsmPieces[0] == "rorw" &&
8925 AsmPieces[1] == "$$8," &&
8926 AsmPieces[2] == "${0:w}" &&
8927 IA->getConstraintString() == "=r,0,~{dirflag},~{fpsr},~{flags},~{cc}") {
8928 return LowerToBSwap(CI);
8932 if (CI->getType() == Type::getInt64Ty(CI->getContext()) &&
8933 Constraints.size() >= 2 &&
8934 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
8935 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
8936 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
8937 std::vector<std::string> Words;
8938 SplitString(AsmPieces[0], Words, " \t");
8939 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
8941 SplitString(AsmPieces[1], Words, " \t");
8942 if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
8944 SplitString(AsmPieces[2], Words, " \t,");
8945 if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
8946 Words[2] == "%edx") {
8947 return LowerToBSwap(CI);
8959 /// getConstraintType - Given a constraint letter, return the type of
8960 /// constraint it is for this target.
8961 X86TargetLowering::ConstraintType
8962 X86TargetLowering::getConstraintType(const std::string &Constraint) const {
8963 if (Constraint.size() == 1) {
8964 switch (Constraint[0]) {
8976 return C_RegisterClass;
8984 return TargetLowering::getConstraintType(Constraint);
8987 /// LowerXConstraint - try to replace an X constraint, which matches anything,
8988 /// with another that has more specific requirements based on the type of the
8989 /// corresponding operand.
8990 const char *X86TargetLowering::
8991 LowerXConstraint(EVT ConstraintVT) const {
8992 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
8993 // 'f' like normal targets.
8994 if (ConstraintVT.isFloatingPoint()) {
8995 if (Subtarget->hasSSE2())
8997 if (Subtarget->hasSSE1())
9001 return TargetLowering::LowerXConstraint(ConstraintVT);
9004 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
9005 /// vector. If it is invalid, don't add anything to Ops.
9006 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
9009 std::vector<SDValue>&Ops,
9010 SelectionDAG &DAG) const {
9011 SDValue Result(0, 0);
9013 switch (Constraint) {
9016 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9017 if (C->getZExtValue() <= 31) {
9018 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9024 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9025 if (C->getZExtValue() <= 63) {
9026 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9032 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9033 if ((int8_t)C->getSExtValue() == C->getSExtValue()) {
9034 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9040 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9041 if (C->getZExtValue() <= 255) {
9042 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9048 // 32-bit signed value
9049 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9050 const ConstantInt *CI = C->getConstantIntValue();
9051 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9052 C->getSExtValue())) {
9053 // Widen to 64 bits here to get it sign extended.
9054 Result = DAG.getTargetConstant(C->getSExtValue(), MVT::i64);
9057 // FIXME gcc accepts some relocatable values here too, but only in certain
9058 // memory models; it's complicated.
9063 // 32-bit unsigned value
9064 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
9065 const ConstantInt *CI = C->getConstantIntValue();
9066 if (CI->isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
9067 C->getZExtValue())) {
9068 Result = DAG.getTargetConstant(C->getZExtValue(), Op.getValueType());
9072 // FIXME gcc accepts some relocatable values here too, but only in certain
9073 // memory models; it's complicated.
9077 // Literal immediates are always ok.
9078 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
9079 // Widen to 64 bits here to get it sign extended.
9080 Result = DAG.getTargetConstant(CST->getSExtValue(), MVT::i64);
9084 // If we are in non-pic codegen mode, we allow the address of a global (with
9085 // an optional displacement) to be used with 'i'.
9086 GlobalAddressSDNode *GA = 0;
9089 // Match either (GA), (GA+C), (GA+C1+C2), etc.
9091 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
9092 Offset += GA->getOffset();
9094 } else if (Op.getOpcode() == ISD::ADD) {
9095 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9096 Offset += C->getZExtValue();
9097 Op = Op.getOperand(0);
9100 } else if (Op.getOpcode() == ISD::SUB) {
9101 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
9102 Offset += -C->getZExtValue();
9103 Op = Op.getOperand(0);
9108 // Otherwise, this isn't something we can handle, reject it.
9112 GlobalValue *GV = GA->getGlobal();
9113 // If we require an extra load to get this address, as in PIC mode, we
9115 if (isGlobalStubReference(Subtarget->ClassifyGlobalReference(GV,
9116 getTargetMachine())))
9120 Op = LowerGlobalAddress(GV, Op.getDebugLoc(), Offset, DAG);
9122 Op = DAG.getTargetGlobalAddress(GV, GA->getValueType(0), Offset);
9128 if (Result.getNode()) {
9129 Ops.push_back(Result);
9132 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
9136 std::vector<unsigned> X86TargetLowering::
9137 getRegClassForInlineAsmConstraint(const std::string &Constraint,
9139 if (Constraint.size() == 1) {
9140 // FIXME: not handling fp-stack yet!
9141 switch (Constraint[0]) { // GCC X86 Constraint Letters
9142 default: break; // Unknown constraint letter
9143 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
9144 if (Subtarget->is64Bit()) {
9146 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX,
9147 X86::ESI, X86::EDI, X86::R8D, X86::R9D,
9148 X86::R10D,X86::R11D,X86::R12D,
9149 X86::R13D,X86::R14D,X86::R15D,
9150 X86::EBP, X86::ESP, 0);
9151 else if (VT == MVT::i16)
9152 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX,
9153 X86::SI, X86::DI, X86::R8W,X86::R9W,
9154 X86::R10W,X86::R11W,X86::R12W,
9155 X86::R13W,X86::R14W,X86::R15W,
9156 X86::BP, X86::SP, 0);
9157 else if (VT == MVT::i8)
9158 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL,
9159 X86::SIL, X86::DIL, X86::R8B,X86::R9B,
9160 X86::R10B,X86::R11B,X86::R12B,
9161 X86::R13B,X86::R14B,X86::R15B,
9162 X86::BPL, X86::SPL, 0);
9164 else if (VT == MVT::i64)
9165 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX,
9166 X86::RSI, X86::RDI, X86::R8, X86::R9,
9167 X86::R10, X86::R11, X86::R12,
9168 X86::R13, X86::R14, X86::R15,
9169 X86::RBP, X86::RSP, 0);
9173 // 32-bit fallthrough
9176 return make_vector<unsigned>(X86::EAX, X86::EDX, X86::ECX, X86::EBX, 0);
9177 else if (VT == MVT::i16)
9178 return make_vector<unsigned>(X86::AX, X86::DX, X86::CX, X86::BX, 0);
9179 else if (VT == MVT::i8)
9180 return make_vector<unsigned>(X86::AL, X86::DL, X86::CL, X86::BL, 0);
9181 else if (VT == MVT::i64)
9182 return make_vector<unsigned>(X86::RAX, X86::RDX, X86::RCX, X86::RBX, 0);
9187 return std::vector<unsigned>();
9190 std::pair<unsigned, const TargetRegisterClass*>
9191 X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
9193 // First, see if this is a constraint that directly corresponds to an LLVM
9195 if (Constraint.size() == 1) {
9196 // GCC Constraint Letters
9197 switch (Constraint[0]) {
9199 case 'r': // GENERAL_REGS
9200 case 'R': // LEGACY_REGS
9201 case 'l': // INDEX_REGS
9203 return std::make_pair(0U, X86::GR8RegisterClass);
9205 return std::make_pair(0U, X86::GR16RegisterClass);
9206 if (VT == MVT::i32 || !Subtarget->is64Bit())
9207 return std::make_pair(0U, X86::GR32RegisterClass);
9208 return std::make_pair(0U, X86::GR64RegisterClass);
9209 case 'f': // FP Stack registers.
9210 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
9211 // value to the correct fpstack register class.
9212 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
9213 return std::make_pair(0U, X86::RFP32RegisterClass);
9214 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
9215 return std::make_pair(0U, X86::RFP64RegisterClass);
9216 return std::make_pair(0U, X86::RFP80RegisterClass);
9217 case 'y': // MMX_REGS if MMX allowed.
9218 if (!Subtarget->hasMMX()) break;
9219 return std::make_pair(0U, X86::VR64RegisterClass);
9220 case 'Y': // SSE_REGS if SSE2 allowed
9221 if (!Subtarget->hasSSE2()) break;
9223 case 'x': // SSE_REGS if SSE1 allowed
9224 if (!Subtarget->hasSSE1()) break;
9226 switch (VT.getSimpleVT().SimpleTy) {
9228 // Scalar SSE types.
9231 return std::make_pair(0U, X86::FR32RegisterClass);
9234 return std::make_pair(0U, X86::FR64RegisterClass);
9242 return std::make_pair(0U, X86::VR128RegisterClass);
9248 // Use the default implementation in TargetLowering to convert the register
9249 // constraint into a member of a register class.
9250 std::pair<unsigned, const TargetRegisterClass*> Res;
9251 Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
9253 // Not found as a standard register?
9254 if (Res.second == 0) {
9255 // GCC calls "st(0)" just plain "st".
9256 if (StringsEqualNoCase("{st}", Constraint)) {
9257 Res.first = X86::ST0;
9258 Res.second = X86::RFP80RegisterClass;
9260 // 'A' means EAX + EDX.
9261 if (Constraint == "A") {
9262 Res.first = X86::EAX;
9263 Res.second = X86::GR32_ADRegisterClass;
9268 // Otherwise, check to see if this is a register class of the wrong value
9269 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
9270 // turn into {ax},{dx}.
9271 if (Res.second->hasType(VT))
9272 return Res; // Correct type already, nothing to do.
9274 // All of the single-register GCC register classes map their values onto
9275 // 16-bit register pieces "ax","dx","cx","bx","si","di","bp","sp". If we
9276 // really want an 8-bit or 32-bit register, map to the appropriate register
9277 // class and return the appropriate register.
9278 if (Res.second == X86::GR16RegisterClass) {
9279 if (VT == MVT::i8) {
9280 unsigned DestReg = 0;
9281 switch (Res.first) {
9283 case X86::AX: DestReg = X86::AL; break;
9284 case X86::DX: DestReg = X86::DL; break;
9285 case X86::CX: DestReg = X86::CL; break;
9286 case X86::BX: DestReg = X86::BL; break;
9289 Res.first = DestReg;
9290 Res.second = X86::GR8RegisterClass;
9292 } else if (VT == MVT::i32) {
9293 unsigned DestReg = 0;
9294 switch (Res.first) {
9296 case X86::AX: DestReg = X86::EAX; break;
9297 case X86::DX: DestReg = X86::EDX; break;
9298 case X86::CX: DestReg = X86::ECX; break;
9299 case X86::BX: DestReg = X86::EBX; break;
9300 case X86::SI: DestReg = X86::ESI; break;
9301 case X86::DI: DestReg = X86::EDI; break;
9302 case X86::BP: DestReg = X86::EBP; break;
9303 case X86::SP: DestReg = X86::ESP; break;
9306 Res.first = DestReg;
9307 Res.second = X86::GR32RegisterClass;
9309 } else if (VT == MVT::i64) {
9310 unsigned DestReg = 0;
9311 switch (Res.first) {
9313 case X86::AX: DestReg = X86::RAX; break;
9314 case X86::DX: DestReg = X86::RDX; break;
9315 case X86::CX: DestReg = X86::RCX; break;
9316 case X86::BX: DestReg = X86::RBX; break;
9317 case X86::SI: DestReg = X86::RSI; break;
9318 case X86::DI: DestReg = X86::RDI; break;
9319 case X86::BP: DestReg = X86::RBP; break;
9320 case X86::SP: DestReg = X86::RSP; break;
9323 Res.first = DestReg;
9324 Res.second = X86::GR64RegisterClass;
9327 } else if (Res.second == X86::FR32RegisterClass ||
9328 Res.second == X86::FR64RegisterClass ||
9329 Res.second == X86::VR128RegisterClass) {
9330 // Handle references to XMM physical registers that got mapped into the
9331 // wrong class. This can happen with constraints like {xmm0} where the
9332 // target independent register mapper will just pick the first match it can
9333 // find, ignoring the required type.
9335 Res.second = X86::FR32RegisterClass;
9336 else if (VT == MVT::f64)
9337 Res.second = X86::FR64RegisterClass;
9338 else if (X86::VR128RegisterClass->hasType(VT))
9339 Res.second = X86::VR128RegisterClass;
9345 //===----------------------------------------------------------------------===//
9346 // X86 Widen vector type
9347 //===----------------------------------------------------------------------===//
9349 /// getWidenVectorType: given a vector type, returns the type to widen
9350 /// to (e.g., v7i8 to v8i8). If the vector type is legal, it returns itself.
9351 /// If there is no vector type that we want to widen to, returns MVT::Other
9352 /// When and where to widen is target dependent based on the cost of
9353 /// scalarizing vs using the wider vector type.
9355 EVT X86TargetLowering::getWidenVectorType(EVT VT) const {
9356 assert(VT.isVector());
9357 if (isTypeLegal(VT))
9360 // TODO: In computeRegisterProperty, we can compute the list of legal vector
9361 // type based on element type. This would speed up our search (though
9362 // it may not be worth it since the size of the list is relatively
9364 EVT EltVT = VT.getVectorElementType();
9365 unsigned NElts = VT.getVectorNumElements();
9367 // On X86, it make sense to widen any vector wider than 1
9371 for (unsigned nVT = MVT::FIRST_VECTOR_VALUETYPE;
9372 nVT <= MVT::LAST_VECTOR_VALUETYPE; ++nVT) {
9373 EVT SVT = (MVT::SimpleValueType)nVT;
9375 if (isTypeLegal(SVT) &&
9376 SVT.getVectorElementType() == EltVT &&
9377 SVT.getVectorNumElements() > NElts)